Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Review of Literature-Based Models for Skin and Eye Irritation and Corrosion Ana Gallegos Saliner, Grace Patlewicz & Andrew P. Worth 2006 EUR 22320 EN EUROPEAN COMMISSION DIRECTORATE GENERAL JOINT RESEARCH CENTRE Institute for Health and Consumer Protection Toxicology and Chemical Substances Unit European Chemicals Bureau I-21020 Ispra (VA) Italy Review of Literature-Based Models for Skin and Eye Irritation and Corrosion Ana Gallegos Saliner, Grace Patlewicz & Andrew P. Worth 2006 EUR 22320 EN The mission of the IHCP is to provide scientific support to the development and implementation of EU policies related to health and consumer protection. European Commission Directorate – General Joint Research Centre Institute for Health and Consumer Protection Contact information Address: E. Fermi, 1, 21020-Ispra (VA) Italy E-mail: [email protected] Tel.: +39 0 332 78 9566 Fax: +39 0 332 78 6717 http:// http://ecb.jrc.it/QSAR/ http:// ihcp.jrc.cec.eu.int http://www.jrc.cec.eu.int LEGAL NOTICE Neither the European Commission nor any person acting on behalf of the Commission is responsible for the use which might be made of this publication. A great deal of additional information on the European Union is available on the Internet. It can be accessed through the Europa server (http://europa.eu.int) EUR 22320 EN © European Communities, 2006 Reproduction is authorised provided the source is acknowledged Printed in Italy ABSTRACT This report reviews the state-of-the-art of in silico and in vitro methods for assessing dermal and ocular irritation and corrosion. Following a general introduction, the current EU legislation for the classification and labelling of chemicals causing irritation and corrosivity is summarised. Then currently available non-animal approaches are reviewed. The main alternative approaches to assess acute local toxic effects are: a) in silico approaches, including SARs, QSARs and expert systems integrating multiple approaches; and b) in vitro test methods. In this review, emphasis is placed on literature-based (Q)SAR models for skin and eye irritation and corrosion as well as computer-based expert systems. TABLE OF CONTENTS 1. Introduction............................................................................................................1 2. Regulatory context in the European Union............................................................3 3. 4. 5. 2.1. REACH ..........................................................................................................3 2.2. Globally Harmonised System ........................................................................4 2.3. Cosmetics Directive .......................................................................................5 2.4. Endpoint – Regulatory definitions .................................................................5 2.4.1. Skin Irritation .........................................................................................7 2.4.2. Skin Corrosion .......................................................................................9 2.4.3. Eye Irritation / Corrosion .....................................................................10 (Q)SAR Models ...................................................................................................14 3.1.1. Skin Irritation .......................................................................................16 3.1.2. Skin Corrosion .....................................................................................22 3.1.3. Eye Irritation / Corrosion .....................................................................29 Alternative test methods ......................................................................................47 4.1. Use of existing data......................................................................................47 4.2. In vitro methods ...........................................................................................47 4.2.1. Skin Irritation .......................................................................................48 4.2.2. Skin Corrosion .....................................................................................49 4.2.3. Eye Irritation / Corrosion .....................................................................49 Expert Systems.....................................................................................................51 5.1. BfR Decision Support System .....................................................................54 5.2. DEREK ........................................................................................................59 5.3. HazardExpert ...............................................................................................61 5.4. MultiCASE ..................................................................................................62 5.5. TOPKAT......................................................................................................63 5.6. Other Expert Systems ..................................................................................65 5.6.1. Cache....................................................................................................65 5.6.2. SuCCSES .............................................................................................65 7 5.6.3. 5.7. OASIS ..................................................................................................66 Integrated Testing Exercise..........................................................................66 6. Conclusions..........................................................................................................68 7. References............................................................................................................69 8 LIST OF ABBREVIATIONS 3D-QSAR Three-Dimensional QSAR 3 Rs Reduction, Refinement, and Replacement (of animal experiments) AD Average Deviation AAD Absolute Average Deviation ALS Adaptive Least Squares BAIA Baseline Activity Identification Algorithm BCOP Bovine Corneal Opacity and Permeability BECAM Bovine Eye Chorio-Allantoic Membrane (in vitro assay) BfR (German) Federal Institute for Risk Assessment (former BgVV) BgVV (German) Federal Institute for Health Protection of Consumers and Veterinary Medicine (now BfR) BLR Binary Logistic Regression BNN Bayesian Neural Network C / NC Corrosive / Non-Corrosive CA Competent Authority (in EU Member State) CAChe Computer-Aided Chemistry CEC Commission of the European Communities CEET Chicken Enucleated Eye Test CHIP Chemicals Hazard Information and Packaging CI Confidence Interval C&L Classification and Labelling CM Classification Model CMC Critical Micelle Concentration COLIPA European Cosmetic, Toiletry and Perfumery Association CT Classification Tree CSA Cluster Significance Analysis DA Discriminant Analysis DEREK Deductive Estimation of Risk from Existing Knowledge 9 DES Draize Eye Score DM Dipole Moment DRD Detailed Review Document DSS Decision Support System ECB European Chemicals Bureau ECETOC European Centre for Ecotoxicology and Toxicology of Chemicals ECM Embedded Cluster Modelling ECVAM European Centre for the Validation of Alternative Methods EEC European Economic Community EHS Environmental Health and Safety EIT Eye Irritation Thresholds ELINCS European List of Notified Chemical Substances EPA Environmental Protection Agency EU European Union E-value Electrotopological state value EWG Endpoint Working Group (within REACH Implementation Project 3.3) F Fischer statistic FDA Food and Drug Administration GFA Genetic Function Approximation GHS Globally Harmonised System HCE Human Corneal Epithelium in vitro reconstituted model HCL Harmonisation of Classification and Labelling HET-CAM Hen's Egg Test on the Chorio-Allantoic Membrane HOMO Highest Occupied Molecular Orbital HSG Human Serum γ-Globulin I / NI Irritant / Non-Irritant IPCS International Programme on Chemical Safety IRAG Interagency Regulatory Alternatives Group ITC Interagency Testing Committee 10 IVSCT In vitro Skin Corrosivity Test LDA Linear Discriminant Analysis LHASA Logic and Heuristics Applied to Synthetic Analysis LMC Laboratory of Mathematical Chemistry LogP Logarithm of the octanol/water partition coefficient LOO Leave-One Out LOO-CV Leave-One-Out Cross-Validation LUMO Lowest Unoccupied Molecular Orbital LVET Low Volume Eye Test MAS Maximum Average Score (eye irritation measure) MCI Molecular Connectivity Index MDES Molar Draize Eye Score MES Molar-adjusted Eye Score MI-QSAR Membrane-Interaction QSAR MLR Multidimensional Linear Regression MMAS Modified Maximum Average Score (eye irritation measure) MP Melting Point MultiCASE Multiple Computer Automated Structure Evaluation MV Mean Value MV Molecular Volume N Absolute Hardness NN Neural Network NOAEL No Observed Adverse Effect Level OECD Organisation for Economic Co-operation and Development OPS Optimum Prediction Space PII Primary (skin/eye) irritation scores PBBK Physiologically Based Bio-Kinetic (model) PC Principal Component PCA Principal Components Analysis 11 PLS Partial Least Squares PM Prediction Model (for converting in vitro to in vivo data) pKa -log10 of the acid dissociation constant (Q)SAR (Quantitative) Structure Activity Relationship QSTR Quantitative Structure-Toxicity Relationship REACH Registration, Evaluation and Authorisation of Chemicals RIP REACH – Implementation Project RIVM Dutch Institute for Public Health and the Environment SAR Structure-Activity Relationship SCCNFP European Commission’s Scientific Committee on Cosmetics and Non Food Products SD Standard Deviation SICRET Skin Irritation Corrosion Rules Estimation Tool SuCCSES Substructure-Based Computerized Chemical Selection Expert System TER (rat skin) Transcutaneous Electrical Resistance test TGD Technical Guidance Document TOPKAT TOxicity Prediction by C(K)omputer Assisted Technology VOC Volatile Organic Compound WDS Weighted Draize Score X Electronegativity 12 1. Introduction This report reviews the state-of-the-art of in silico and in vitro methods for assessing dermal and ocular irritation and corrosion. Following a general introduction (Chapter 1), the current EU legislation for classification and labelling of chemicals causing irritation and corrosivity is summarised (Chapter 2). Then currently available nonanimal approaches are reviewed. The main alternative approaches to assess acute local toxic effects are in silico approaches, particularly SARs, and QSARs (Chapter 3); in vitro test methods (Chapter 4). In this report, emphasis is placed on literaturebased (Q)SAR models for skin and eye irritation and corrosion (Chapter 3), and on expert systems integrating multiple approaches into computer-based tools (Chapter 5). Acute local irritancy and corrosivity are mainly assessed in two contexts: in the hazard classification of industrial chemicals and in the safety assessment of ingredients and mixtures used in industrial, pharmaceutical and consumer products. The intended purpose in each context is different, and therefore the considerations made in the two situations are not equivalent. In the hazard identification of chemicals, the purpose of testing is to assess the irritation/corrosion potential according to classification schemes defined by regulatory authorities. Current regulatory proposals recommend a tiered (stepwise) approach to hazard identification in which chemicals can be classified as irritating/corrosive on the basis of results from non-animal methods. Since testing in animals is only required as a last step to confirm negative results generated by non-animal tests, these testing strategies contribute to the reduction and refinement of in vivo tests. The non-animal methods act as partial replacements of the animal test being used to place chemicals into two or more categories of irritation/corrosion potential, without generating too many false positive results. In such strategies, there is less concern about the generation of false negatives because these can be identified by the animal tests carried out in the last step of the process. In contrast, in the safety assessment of ophthalmologic and cosmetic ingredients, mixtures and products, the purpose is to demonstrate that the products will not cause adverse effects. In this case, the placement of test substances into broad irritation/corrosion categories is often not sufficient, since it is necessary to establish the absence of adverse effects at lower concentrations. Although there is an increasing 1 reliance on non-animal methods, it is more challenging to assess the reliability of predictions for product safety. Since the skin is often exposed to cosmetic products, the potential for a particular product/ingredient to cause skin irritation or corrosion needs to be evaluated as part of the overall safety assessment process. The standard test for assessing skin irritation and corrosion potential has been for many years the Draize rabbit skin test [1]. The eye can be exposed to cosmetic products and their ingredients either through use of products intended to be used around the eyes (e.g. mascaras, eye creams) or through accidental exposure to products that may enter the eye in diluted form during normal use (e.g. shampoos). The evaluation of eye irritation potential for a cosmetic product and its ingredients is essential to ensure that a product is safe for consumers. The conventional test for assessing eye irritation and corrosion potential of chemicals is the Draize rabbit eye test developed by Draize in 1944 [1]. Although corrosivity should not be caused by cosmetics, occasionally it can occur after a manufacturing mistake or misuse. In the formulation process, a cosmetic ingredient that has the intrinsic property to be irritant or corrosive is not necessarily excluded for use in cosmetics. Irritation and corrosivity are dependent on the final concentration in the cosmetic product, on the presence of neutralising substances, the excipients used, the exposure route, and on the conditions of use, among other factors. 2 2. Regulatory context in the European Union 2.1. REACH Currently under European legislation, Directive 67/548/EEC [2] on the classification, packaging and labelling of dangerous substances requires new substances to be tested and assessed for possible risks to human health and the environment before they are marketed in volumes starting at 10 kg. Skin and eye irritation and corrosion (acute dermal toxicity) testing is required for substances produced or imported at levels above 100 kg per year. In contrast, existing substances are assessed under the provisions of Regulation 793/93/EEC [3]-[4]. This Regulation requires the identification of priority substances, which are then subjected to comprehensive risk assessment led by EU member state regulatory authorities. A consequence of this distinction is that testing of low priority existing substances has not been required, which means that data on the health and environmental effects of most existing substances is lacking. In October 2003, to address this knowledge gap, the European Commission published a draft proposal for a new chemicals legislation called REACH, which stands for the Registration, Evaluation and Authorisation of Chemicals in the European Union [5][12]. The proposed REACH legislation aims to obtain the necessary information for all substances imported or manufactured at volumes greater than 1 tonne, with a period of 11 years from the entry into force of the legislation. REACH envisages an extensive system of assessing the physicochemical, (eco)toxicological and environmental properties of chemicals based mainly on their tonnage. Annexes VII to X of REACH describe the information requirements for substances manufactured or imported into the EU in quantities above one tonne per year, with possibilities for adaptations (including derogations) where justified. The exact testing and data requirements depend on the volume, but the base set for substances produced or imported in quantities between 1 and 10 tonnes per year includes physicochemical properties, toxicological and ecotoxicological data. Thus, the risk assessment of New and Existing Chemicals will be driven by the forthcoming REACH legislation, expected to enter into force in 2007. Several studies reporting the estimated economical benefits and animal life savings of using 3 intelligent testing strategies based on (Q)SARs, read-across, and other waiving possibilities have been published by the ECB [13]-[14]. 2.2. Globally Harmonised System In September 1989, the Commission of the European Communities (CEC) raised the need to evaluate the feasibility of rationalising the different classification schemes based on acute toxicity of chemicals existing in different countries, trading blocks, and international organisations [15]-[22]. The schemes for classifying chemicals as dangerous based on toxicological properties, and the national authorities or international organisations responsible for the schemes were identified. In November 1994, the OECD decided to create the Programme on Harmonisation of Classification and Labelling (HCL). The objectives of this programme were to develop a harmonised classification system for chemical substances covering nine hazard endpoints considered in existing classification systems. The objectives also included the development of a harmonised hazard classification system for chemical mixtures, based on the same endpoints as considered for substances. As the work progressed, Detailed Review Documents (DRD) were drafted and subsequently discussed and agreed upon by OECD ad hoc Expert Groups, endorsed by the OECD Task Force on HCL. They were used as the basis for consensus building on the proposals for harmonised classification systems. The various DRD are available as Environmental Health and Safety (EHS) Monographs in the Series on Testing and Assessment. In November 1998, the 28th Joint Meeting recognised the achievements made in the area of classification of substances and approved the concept of an Integrated Document, which brings together the harmonised classification systems for each of the nine recognised hazard categories, including the substances which cause skin irritation/corrosion, and eye irritation/corrosion. The two endpoints are combined into a single classification system in the case of mixtures. The Harmonised System for the Classification of Chemicals which cause Skin Irritation/Corrosion and Eye Irritation/Corrosion are collected in Chapters 2.2 and 2.3, respectively, while the Harmonised System for the Classification of Chemical Mixtures which cause Skin and Eye Corrosion/ Irritation is collected in Chapter 3.3 [20]. 4 2.3. Cosmetics Directive In the EU, the Cosmetics Directive 76/768/EEC [23] imposes the legal obligation that cosmetics sold to consumers should be safe. Thus, the responsibility for product safety lies with the cosmetic manufacturer or the legal person placing a cosmetic product on the Community market, who must be able to demonstrate that the product is safe for the consumer. The Cosmetics Directive does not specify a fixed dataset or methods needed to assess the safety of cosmetic ingredients [24]; however, the safety assessment of finished cosmetic products without animal testing is only possible provided that an adequate toxicological data package on the ingredients is available, according to guidelines developed by the Commission’s Scientific Committee on Cosmetics and Non Food Products (SCCNFP) [25]. Toxicological data requirements for cosmetic ingredients based on the SCCNFP guidelines include eleven endpoints: acute systemic toxicity; skin irritation/corrosion; eye irritation; skin sensitisation; skin absorption; subacute and subchronic toxicity; genotoxicity/mutagenicity; UV-induced toxic effects (phototoxicity, photoallergy, photogenotoxicity); toxicokinetics and metabolism; carcinogenicity; and reproductive and developmental toxicity. The safety assessment of a cosmetic product requires taking into account the general profile of the ingredients, their hazard (which is the intrinsic potential of a substance to cause damage to human health) and the risk (which is the probability of a substance to cause damage under relevant use conditions). For all cosmetic ingredients, appropriate data must be provided to permit an adequate safety assessment, regardless of tonnage. 2.4. Endpoint – Regulatory definitions The classification and labelling of commercial substances within the EU is based on the use of Risk phases (R-phrases). R-phrase definitions are assigned after the toxicological evaluation of a wide range of human health hazard effects, including acute toxicity, repeat dose toxicity, carcinogenicity, mutagenicity, reproductive toxicity, teratogenicity, irritancy and sensitisation. Thus, R-phrases are used as indicators of hazard potential for a large number of new and existing substances. Guidance on the classification and labelling of chemicals is provided under the 5 Chemicals Hazard Information and Packaging for Supply (CHIP) Regulations 1997 [26]-[27]. R-phrase definitions and the grounds upon which chemicals are assigned Rphrases for skin and eye irritation/corrosion are presented in Table 1. Table 1. R-phrases for skin and eye irritation/corrosion Risk Phrase Effect Description R34 C Causes burns Full thickness destruction of the skin occurs as a result of up to 4 hours exposure. R35 C Causes severe burns Full thickness destruction of the skin occurs as a result of up to 3 minutes exposure. R36 Xi Irritating to eyes Defined reversible changes of cornea, iris or conjunctiva R38 Xi Irritating to skin Defined reversible inflammation of the skin as a result of up to four hours of exposure R41 Xi Risk of serious damage to eyes Defined severe reversible changes or irreversible changes of cornea, iris or conjunctiva The protocols recommended for the different required tests are described in the Annex VII of the dangerous substances directive (Directive 67/548/EEC [2]) and in the OECD testing guidelines for chemicals. Test methods in Annex VII for dermal and ocular irritation/corrosion are summarised in Table 2. Table 2. Test methods for dermal and ocular toxicity in Annex V to Directive 67/548/EEC. Method Title Directive B.3 Acute toxicity (dermal) [28] 92/69/EEC B.4 Acute toxicity (skin irritation) [29] 92/69/EEC B.5 Acute toxicity (eye irritation) [30] 92/69/EEC B.40 Skin Corrosion [31] 2000/33/EC The OECD in cooperation with the Coordinating Group on the Harmonisation of Chemical Classification Systems has agreed on a harmonised system of classification to be used internationally [20]. Separate categories have been adopted for skin irritation and for skin corrosion, while eye irritation and corrosion are comprised in a single category. The available OECD test guidelines for these endpoints are summarised in Table 3. 6 Table 3. OECD Guidelines for Testing of Chemicals. No. Title Original Adoption Most Recent Update 404 Acute dermal irritation/corrosion [32] 12 May 1981 24 April 2002 405 Acute eye irritation/corrosion [33] 12 May 1981 24 April 2002 430 In vitro Skin Corrosion: Transcutaneous Electrical Resistance Test (TER) [34] 13 April 2004 - 431 In vitro Skin Corrosion: Human Skin Model Test [35] 13 April 2004 - The TG 404 was first adopted in 1981 and was revised in 1992 [36]; likewise, TG 405 was adopted in 1981 and revised in 1987. Detailed Review Documents on Classification Systems for Eye Irritation/Corrosion and Skin Irritation/Corrosion in OECD Member Countries are gathered in the OECD Series on Testing and Assessment documents No.14 [37] and No.16 [38], respectively. At the OECD workshop in Solna, Sweden, in 1996 [16], it was agreed that both TGs should be updated to include a sequential testing strategy [39]-[40]. The subsequently revised draft documents were submitted for review in May 2001 [41]-[43] and adopted in April 2002. The definitions according to the OECD guidelines and to the GHS are given below in the next sections. 2.4.1. Skin Irritation Definition according to the OECD TG. Dermal irritation is defined in OECD TG 404 as “the production of reversible damage of the skin following the application of a test substance for up to 4 hours” [32]. In the EU this effect is assigned the R-phrases “Xi” and R38. Pathological characteristics and manifestation of skin irritation. Erythema/eschar and edema are manifestations of dermal irritation, but further guidance is lacking. Irritation is initially manifest by redness (erythema), vesicles, serous exudates, serous scabs (eschar) and various degrees of swelling (edema). Over time, other reactions may be manifest, like small areas of alopecia, hyperkeratosis, hyperplasia and scaling. Histopathology is useful in discerning among responses. In most cases inflammation is well developed within the first 72 hours of observation. 7 Grading of dermal responses. For the grading of dermal responses, there is a consensus among all the authorities to use the Draize grading scale for dermal lesions (Table 4). Table 4. Grading of dermal responses (Draize scale) [32]. Erythema and eschar formation Grade No erythema 0 Very slight erythema (barely perceptible) 1 Well defined erythema 2 Moderate to severe erythema 3 Severe erythema (beet redness) to slight eschar formation (injuries in depth) 4 Edema formation Grade No edema 0 Very slight edema (barely perceptible) 1 Slight edema (edges of area well defined by definite raising) 2 Moderate edema (raised approximately 1 mm) 3 Severe edema (raised more than 1 mm and extending beyond the area of exposure) 4 Globally Harmonised System. According to the GHS [20], a single category is adopted for skin irritation. Erythema/eschar and oedema are graded separately; an animal’s mean score from readings over the first three days after exposure must meet a defined level to be positive; and at least 2 of 3 tested animals must be positive for the test to be positive. Positive responses can also be obtained using other less common criteria. The proportion of test substances expected to be positive by the proposed irritant category is within the range of positives among existing classification systems. Some authorities use two hazard categories, i.e. irritants and mild irritants. Criteria for classification. Authorities vary as to the existence of criteria for evaluating irritation; whether mean scores are computed for one time period or across all grading times; whether endpoint grades are used alone (edema or erythema/eschar) or grades are pooled (edema plus erythema/eschar). Authorities also vary as to when animals are observed and scored for dermal lesions and whether animals are followed to evaluate the reversibility of lesions. Most authorities grade lesions at or around 24, 8 48 and 72 hours after patch removal; fewer of them grade lesions within 1 hour of patch removal. None use times beyond 72 hours. EU Criteria. In the in vivo rabbit Draize test, the chemical is applied to the shaved skin of 3-6 rabbits and covered for 4 or 24 hours [29]. The chemical is then wiped off to permit observation for signs of skin irritation, and application sites are examined after 24, 48, and 72 hours after application for assessment of primary skin irritation scores, using an empirical formula. For a three-animal test, mean scores across 3 scoring times (24, 48 and 72 hours after patch removal) are calculated for each animal for edema grades and for erythema/eschar grades, separately. A test is positive when the mean score is 2 or greater. The test is positive for irritation when at least two animals are positive for the same endpoint (erythema/eschar or edema). For a sixanimal test, a mean score is calculated across 3 scoring times (as above) and across all six animals for edema grades and for erythema/eschar grades, separately. A mean score of 2 or more identifies an irritant. EU criteria also recognise the importance of reversibility of lesions over time up to 14 days. 2.4.2. Skin Corrosion Definition according to the OECD TG. Dermal corrosion is defined in OECD TG 404 as “the production of irreversible damage to skin; namely, visible necrosis through the epidermis and into the dermis, following the application of a test substance for up to four hours. Corrosive reactions are typified by ulcers, bleeding, bloody scabs, and, by the end of observation at 14 days, by discolouration due to blanching of the skin, complete areas of alopecia, and scars” [32]. In the EU this effect is assigned “C” and either R34 or R35. Pathological characteristics and manifestation of skin corrosion. In some cases corrosive manifestations are noted soon after application to the skin, however, it can take a period of days for the process to reach maximum intensity. Corrosion involves the destruction of tissue through the epidermis and down into the dermis. When there is bleeding or bloody scabs (eschar) it is obvious that the epidermal basement membrane has been breached and the process has entered the dermis. Over time there is often blanching of the skin due to the loss of pigment cells. Since hair follicles are embedded in the dermis, areas of alopecia also indicate corrosivity. Histopathology 9 can give a definitive diagnosis. Vesicles are an indication of irritation, not corrosion, as they represent disturbances at the epidermal-dermal interface. Global Harmonised System. According to the GHS [20], a single category is adopted for skin corrosion, although some authorities may subdivide the single corrosive category into up to three subcategories. Criteria for classification. The GHS criteria for classification of corrosion are agreed upon (e.g. at least one of three tested animals must show corrosive effects after exposure up to a 4 hour duration before an agent is classified as corrosive) [20]. EU Criteria. EU criteria state that if at least one test animal manifests a corrosive action following exposures of up to 3 minutes or up to 4 hours, then the test material will be judged to cause severe burns (R35) or to cause burns (R34), respectively. 2.4.3. Eye Irritation / Corrosion Definition according to the OECD TG. Eye irritation is defined in OECD TG 405 as “the production of changes in the eye following application of a test substance to the anterior surface of the eye, which are fully reversible within 21 days of application” [33]. In the EU this effect is assigned “Xi” and R36. Eye corrosion is defined in OECD TG 405 as ”the production of tissue damage in the eye, or serious physical decay of vision, following application of a test substance to the anterior surface of the eye, which is not fully reversible within 21 days of application” [33]. In the EU, this effect is assigned “Xi”, and R41. Grading of ocular responses. Essentially four different approaches to the evaluation of ocular hazard have been developed, some of which have been used in the international regulation of chemical products. The classification procedures are similar since they all are based on the rabbit eye model. The effects of the substances on the cornea, iris and conjunctivae are graded on individual scales and given weighted scores, using the Draize scoring system for ranking responses (Table 5). 10 Table 5. Grading of Ocular Responses (Draize scoring system) based on the rabbit eye model [33]. RESPONSE Score CORNEA Opacity: degree of density (area most dense taken for reading) No ulceration or opacity 0 Scattered or diffuse areas of opacity (other than slight dulling of normal lustre), details of iris clearly visible 1 Easily discernible translucent area, details of iris slightly obscured 2 Nacrous area, no details of iris visible, size of pupil barely discernible 3 Opaque cornea, iris not discernible through the opacity 4 IRIS Normal 0 Markedly deepened rugae, congestion, swelling, moderate circumcorneal hyperaemia, or injection, any of these or combination of any thereof, iris still reacting to light (sluggish reaction is positive) 1 No reaction to light, haemorrhage, gross destruction (any or all of these) 2 CONJUNCTIVAE Redness (refers to palpebral and bulbar conjunctivae, cornea and iris); blood vessels normal 0 Some blood vessels definitely hyperaemic (injected) 1 Diffuse, crimson colour, individual vessels not easily discernible 2 Diffuse beefy red 3 Chemosis Swelling (refers to lids and/or nictating membranes) No swelling 0 Any swelling above normal (includes nictating membranes) 1 Obvious swelling with partial eversion of lids 2 Swelling with lids about half closed 3 Swelling with lids more than half closed 4 The scores are processed differently to decide if a test substance is irritant or not, and to grade the severity of the irritation. The different features are of varying importance in each classification scheme [44]: the intensity of the irritant response in different regions of the anterior eye; the weighting of response scores prior to determining an irritation index; the level of response during the first three days after exposure; and the reversibility or irreversibility of a response. 11 Global Harmonised System. A single harmonised hazard group is defined for the classification of eye irritation that reverses within an appropriate observation time [20]. Local effects detected in a Draize rabbit test that reverse within 21 days after instillation of the substance into the eye are classified. Effects on the cornea, effects on the iris and conjunctival erythema and oedema are graded separately. An animal’s mean score from readings over the first three days after instillation must meet a defined level to be positive, and at least 2 of 3 tested animals must be positive for the test to be positive. Some authorities distinguish between mild and moderate eye irritants. Criteria for classification. The criteria for classification are based on mean Draize scores, and on the duration of effects (Table 6). Table 6. EU Criteria for the classification of eye irritation and corrosion. Response R 36 R 41 Corneal opacity > 2, < 3 >3 Iris lesion > 1, < 1.5 > 1.5 Redness of the conjunctivae > 2.5 Oedema of the conjunctivae >2 (Chemosis) The duration of the study should be sufficient to evaluate fully the reversibility or irreversibility of the effects observed. An extended observation period may be necessary if there is persistent corneal involvement or other ocular irritation in any animal in order to determine the progress of the lesions. Persistent colouration of the eye (this effect demonstrating irreversible penetration of the chemical into the cells of the eye) and irreversible corneal opacity are also judged as severe eye damage because these effects diminish crucially the function of the eye as an optical organ. Animals showing severe and enduring signs of distress and pain may need to be humanely killed. In that case the substance has to be labelled R41 without taking into account the possible reversibility of effects. EU Criteria. Generally a Draize eye test is performed by applying the substance to the eye of a living rabbit [30], using one animal (if severe effects are suspected) or three animals (if no severe eye irritation is anticipated). Ocular lesions which occur 12 within 72 hours after exposure and which persist for at least 24 hours after instillation qualify a material to be classified as irritant to the eye, if they also reach certain score levels. 13 3. (Q)SAR Models In vivo testing for skin and eye irritation and corrosion is among the most emotive of all testing procedures. Due to public pressure and legislation, there has been a move to reduce this testing, and replacements have been sought, mainly by exploring the use of in vitro methods, but also non-testing methods (e.g. QSAR and read-across). Under REACH, it is expected that read-across and QSAR data will contribute to the initial assessment of whether testing is scientifically necessary. The OECD principles for (Q)SAR validation [45]-[46] identify the types of information that should be associated with a (Q)SAR model proposed for regulatory use. This section reviews the most representative (Q)SAR approaches developed for skin and eye irritation/corrosion. The underlying mechanisms of skin irritation/corrosion and eye irritation/corrosion are diverse and often not completely understood, and the relevance of the animal tests as indicators of human effects has been questioned. On the other side, it has been argued that the current animal tests provide conservative and protective assessments. For example, there is evidence that animal skin is more permeable to chemicals than human skin; thus, the use of rabbit Draize test as indicator of potential skin irritation in humans has been questioned. Draize test have been pointed as being conservative and protective, i.e. it overestimates the human response. In addition, species-related shifts in classification threshold have been observed in the corrosive / non-corrosive threshold when substituting rat skin for human skin. From the scientific literature, it appears that more emphasis has been placed on the QSAR modelling of ocular irritation compared with dermal irritation. Dermal and ocular responses are usually modelled separately, although some similarities have been established between both endpoints. For instance, the cross-correlation between primary skin irritation scores (PII) and eye irritation measures (i.e. maximum average score (MAS), and molar-adjusted eye score (MES)) showed moderate correlations that suggested both similarities and differences between the skin and eye mechanisms of action. Tiered assessment strategies have been applied jointly and separately to both endpoints. There is strong evidence that chemicals that are corrosive to the skin 14 should also be classified as being corrosive to the eye, especially if the assessment is made from knowledge of acidity and alkalinity [47]-[48]. In particular, in the EU and OECD classification schemes, chemicals that have been found to be corrosive to the skin are automatically considered to be corrosive to the eye as well (and are therefore labelled as such without animal testing). For the prediction of skin toxicity, basic physicochemical properties underlie the irritant response. These properties include the fundamental acidic and basic properties of the chemical, its ability to penetrate the skin (a function of hydrophobicity and molecular size), and cytotoxicity. This last property can be described by in vitro modelling or it can be estimated from physicochemical properties. Thus, skin irritation/corrosivity is postulated to result from the chemical first penetrating the skin; then, if the chemical is sufficiently cytotoxic, death of the cells beneath results in corrosivity. However, irritation and corrosion endpoints have been modelled mostly separately for the dermal response. This is due to the fact that corrosion is considered to be more a physical effect (which can be captured by a simple description of acidity or basicity) than a biologically mediated effect. The description of acidity/basicity can be derived from a measurement of pH of certain strength of solution or acid/alkali reserve, or from the presence of certain structural features known to be associated with acids and bases. The mechanistic pathway is likely to vary with chemical class, and it has generally been assumed that multiple class-specific (Q)SAR models are required to predict the toxic potential of a wide array of chemicals. Thus, (Q)SAR models have generally been developed for small datasets of specific groups of compounds, although in some cases more diverse and larger datasets have been also examined. In general, it has been suggested that basic physicochemical parameters such as acidity, basicity, hydrophobicity, and molecular size as well as electrophilic reactivity, are useful to predict the toxic potential of homologous chemicals. Thus, certain parameters may be used to define cut-offs (e.g. a range of pH values, or molecular weights) so that compounds with a physicochemical parameter below or above a certain value can be classified as toxic (or non-toxic). Most models employ intramolecular physicochemical properties associated with solubility. A few models also consider the intermolecular physicochemical parameters that describe the interactions between the compound of interest and a biological receptor. For example, some models are based 15 on the molecular simulation of molecule-membrane interactions (MI-QSAR). In contrast, models intended to predict the toxic potential of heterogeneous groups of chemicals have emphasised the commonality of structural features. The simplest (Q)SAR approaches are traditional quantitative regression models and classification models (CM). Concerning the statistical technique used to derive the model, linear modelling techniques (linear regression, and linear discriminant analysis for CM) have generally been employed, although a number of non-linear approaches (neural networks, and cluster analysis for CM) have also been explored. Non-linear modelling techniques are thought to be generally useful for the modelling of heterogeneous datasets. A number of selected reviews have been referenced [49]-[60]. Some of the most comprehensive reviews on acute local effects are the PhD thesis of Worth [49]; the series of reviews by Schultz et al. [53]-[54], papers by Cronin et al. [55] and by Patlewicz et al. [56]; and the review of expert systems by Cronin et al. [57]. 3.1.1. Skin Irritation Following in vivo animal testing (Draize rabbit skin test), chemicals are classified as being either irritant or corrosive, depending on specific responses to the tests. In the regulatory assessment of chemicals, relatively few QSAR methods have been applied to the prediction of skin irritation and corrosion. One of the earliest SAR approaches to predict skin irritation is based on the TOPKAT methodology [61]. Other local models have been reported by Berner [62]-[64], Nangia [65], Hayashi [66], Smith [67]-[68], and Kodithala [69], who reported a MI-QSAR model. Enslein et al. [61] reported QSARs to predict rabbit skin irritation severity, based on the Draize assay. The models, based on an extensive dataset of publicly available skin irritation data, were implemented in the TOPKAT software package (Accelrys, Inc.). After harmonising the experimental numerical scores into a consistent severity rating system, the dataset was split into compounds without, and with rings, prior to develop the models. It was noted that compounds without rings (260 compounds) were located at the high range of the severity scale, while the compounds with rings (526 compounds) were concentrated near the low range of the scale. Thus, two sets of independent models were developed for the different categories. For each subset of 16 compounds, two QSAR equations were developed; a first equation to select severe irritant compounds, and a second one to discriminate mild and moderate irritants from non-irritant compounds. The models could not distinguish between mild and moderate irritation. The parameters used for the development of the models were molecular connectivity indices (MCI), MOLSTAC substructural keys (either individual keys describing single chemical features, or combination keys), and molecular length parameters. The QSAR models were developed by stepwise discriminant analysis (DA). The variable selection process was applied separately first to dichotomous parameters and after to continuous parameters. Robustness of the DA was checked, and all the possible regression subsets were calculated to determine the subset of variables that best fit regression criteria. Validation was carried out employing the resubstitution method, in which the discriminant function is used to evaluate the compounds used to develop the equation. The classification accuracy for the QSAR models ranged from 90% to 95% (Table 7). Indeterminate compounds were defined in the region around chance probability (0.5), set between p=0.3 and p=0.7. These QSARs form the basis of the skin irritation module in TOPKAT. Table 7. Classification results of the discriminant QSAR models for skin irritation. Models Non-ring compounds Ring compounds Subequations Overall Accuracy False Positives False Negatives Indeterminate F (p<<0.001) Severes vs Others 94.8% 3.1% (3.8%)* 2.2% (11.1%)* 11.9% 15.56 Negative svs Mild/Moderates 95.0% 1.5% (13.6%)* 3.5% (3.9%)* 2.4% 8.33 Severes vs Others 91.0% 5.6% (6.6%)* 3.4% (22.3%)* 15.8% 13.51 Severes vs Others 89.8% 4.3% (15.8%)* 5.9% (8.1%)* 15.7% 11.93 * Percentage of false predictions in relation to the experimental class Berner et al. published a series of papers [62]-[64] in which they reported that the severity of irritation from reduced sets of organic acids and bases in humans correlated with the acid or basic strength as measured by pKa. For example, one study reports the relationship between pKa and skin irritation in humans for a congeneric series of benzoic acid derivatives [64]. It was found that skin irritation and pKa were 17 strongly correlated; for pKa higher than 4 skin irritation rapidly increased. Related to the same study, Nangia et al. [65] published a one-variable model based on 12 organic bases relating skin irritation with pKa. According to the interpretation of the results, skin irritation would be triggered by local alterations in pH caused by acidic and basic chemicals. Hayashi et al. [66] reported two local QSAR models for predicting the molarweighted skin irritation scores (PII) of a set of 30 phenols. The dataset was split into a training set (24 phenols) and a test set (6 phenols) used to validate the models. The descriptors used to build the models were based on reactivity and permeability. The absolute hardness (N) and electronegativity (X) calculated from the highest occupied molecular orbital (HOMO), and the lowest unoccupied molecular orbital (LUMO) energy levels, accounted for reactivity; the logarithm of the octanol/water partition coefficient (LogP) accounted for permeability. Two different models were obtained according to the sign of the LUMO energy level (positive or negative), i.e. the electronic characteristics of phenols. It was found that phenols with positive LUMO react with amino acids in both the epidermis and the dermis as electron donors, resulting in irritation. Therefore, their irritation potential was based on their skin permeability, LogP (r=0.82). On the other hand, phenols with negative LUMO only reacted with amino acids in the epidermis as electron donors, resulting in relatively low irritation (r=0.72). This model, independent of skin permeability, was correlated with absolute hardness (N). The goodness-of-fit for the different regression equations are summarised in Table 8. The differences between predicted and experimental scores for the six test phenols showed a good correlation (n=30, r=0.83). This study suggested that skin irritation caused by phenols is related to their reactivity with macromolecules and their skin permeability; differences in reactivity among phenols were dependent on the electron-releasing or electron-attracting characteristics of their substituents. 18 Table 8. Goodness-of-fit of the linear regression QSAR models for skin irritant phenols. Model Positive LUMO Negative LUMO N Equation r F s 13 PII = −178 N + 1130 –0.36 1.27 18.00 13 PII = 230 log P + 25.5 0.82 23.12 12.82 11 PII = −474 N + 2200 0.72 3.86 7.85 11 PII = 22.9 log P + 63.4 0.16 0.25 9.21 Smith et al. [67] reported a local QSAR model for 42 esters (13 irritants and 29 nonirritants), for which human skin irritation data were available. The model was based on 19 calculated physicochemical parameters, representing transport, electronic and steric properties. The robustness (stability) of the model was assessed by using multiple random resampling. The 13 irritant esters were paired with ten random sets (with replacement) of 13 non-irritants esters, and statistical analyses were conducted with each test. Best subsets regression was used to select variables for subsequent inclusion in discriminant models, and to identify candidate models for further analysis. Linear discriminant analysis (LDA) was used to distinguish between irritant and non-irritant esters. The sensitivity of the resulting ten models evaluated by leaveone-out (LOO) cross-validation ranged from 0.846 to 0.923; the specificity ranged from 0.615 to 0.923. The averaged performance results for the ten models is summarised in Table 9. Compared with non-irritant esters, the irritant ones had lower density, lower water solubility, lower sum of partial positive charges, higher Hansen hydrogen bonding parameter, and higher Hansen dispersion parameters. The results indicate that transport and electronic parameters related to intermolecular interactions are relevant parameters for modelling skin irritation in humans. Table 9. Averaged performance of the discriminant SAR models for human irritant esters. Squared Distance* Sensitivity Specificity Mean 4.9 0.885 0.738 Upper confidence interval 6.9 0.910 0.799 Lower confidence interval 3.0 0.859 0.678 * Degree of separation between the active and inactive group means 19 In a subsequent study, Smith et al. conducted an experimental validation study [68] to assess the ability of the previously developed QSAR model [67] to predict human skin irritation by esters commonly used as fragrance ingredients. Thirty four esters reported to cause irritation in rabbits, located within the predictive space of the QSAR model, were selected for human testing by using the patch test procedure. Of the 34 rabbit irritants selected, 16 were predicted by the QSAR model to be positive and 18 were predicted to be negative. The comparison of the experimental rabbit Draize test with human patch testing revealed that only two of the 34 esters were human irritants. Training set compounds were then queried by using the previously developed ten QSAR submodels, following a consensus approach [67]. An ester was predicted to be an irritant if at least five of the submodels predicted it to be positive. The QSAR model was refined by incorporating the results from the new experimental patch testing. As before, 19 physicochemical descriptors were calculated; statistical analyses were performed on 10 subsets of the database, each subset consisting of randomly selected irritants and non-irritants. Best subsets regression analysis identified candidate models for subsequent discriminant analysis, and each candidate model was analysed by using linear regression. Sensitivities and specificities were determined by the leave-one-out cross-validation (LOO-CV) method. Although all of the 34 investigated esters had been reported to be dermal irritants on rabbit skin, only 16 were predicted to be potentially irritant according to the QSAR models. Then the dataset of esters was combined with a previously published dataset of 42 esters to create a more robust dataset. Ten random sets of 10 irritant esters and 50 non-irritant esters were formed and statistical analyses were conducted on each set. The most consistent parameters were solubility parameter, water solubility, Hansen dispersion, Hansen hydrogen bonding, and sum of partial positive charges. The sensitivity and specificity values for the submodels ranged from 0.50 to 0.80 and from 0.72 to 0.90, respectively. The averaged results are given in Table 10. 20 Table 10. Averaged predictivity of the discriminant QSAR models for human irritant esters. Squared Distance* Sensitivity Specificity Mean 2.87 0.66 0.81 95% CI ± 0.60 0.06 0.04 * Degree of separation between the active and inactive group means CI: Confidence Interval Kodithala et al. [69] reported a membrane-interaction QSAR (MI-QSAR) study to predict skin irritation of a set of 20 diverse hydroxy alcohol compounds. This study combined the use of an in vitro skin irritation model with QSAR techniques. The main assumption was that skin irritation does not depend on specific interaction receptors but on the absolute uptake into the phospholipid rich regions of cellular membranes, which constitute the effective “receptor”. Different phases of this model comprised modelling of the membrane and molecular dynamics simulations; construction of solute molecules and docking of the molecules in the membrane; calculation of intra and intermolecular QSAR descriptors; and construction of QSAR models by using a genetic function approximation (GFA) as a fitting function in the multidimensional linear regression (MLR) model. The comparison of different intramolecular QSAR models (using descriptors such as LogP and HOMO) with MI-QSAR models suggested that the latter were more robust. After checking the cross-correlation descriptor matrix for each model, the MI-QSAR model with three descriptors (LogP, HOMO and electrostatic plus Van der Waals interaction energy (E(vdw+chg)) with performance of r2=0.76 and r2cv=0.67 was considered the most stable one (see Table 11). The training set was split into two subsets of aliphatic alcohols and phenols; newly developed three-term MI-QSAR models achieved r2=0.95, and r2=0.94, respectively (Table 12). In conclusion, skin irritation was predicted to increase with increasing effective concentration of irritants, with increasing binding of compounds to the phospholipid-rich regions of the cellular membrane, and with increasing chemical reactivity, reflected by HOMO and LUMO. 21 Table 11. Goodness-of-fit and robustness of the QSAR and MI-QSAR models for skin irritant hydroxy alcohols. QSAR models MI-QSAR models N Equation r2 r2cv N Equation r2 r2cv 20 PII = 1.84 + 0.90 log P 0.54 0.37 20 PII = 1.96 + 0.93 log P 0.54 0.37 20 PII = 56.3 + 3.63HOMO + +0.41FOCT 0.61 0.43 20 PII = 34.3 − 0.06 E (vdw + chg ) +2.42 HOMO 0.62 0.48 0.64 0.40 20 PII = 3.51 − 0.26 E (vdw + chg ) −0.82 log P + 1.14 HOMO 0.76 0.67 0.67 0.37 20 0.86 0.72 20 20 PII = 42.4 + 2.57 HOMO + 0.16MR +0.93FOCT PII = −19.84 − 0.57Vm + 0.61FOCT +3.71LUMO + 2.36MR PII = 68.07 − 0.06MW + 0.66ETO −0.18E (vdw + chg ) + 4.72 HOMO FOCT ,Vm, E (vdw + chg ), ETO : intermolecular solute-membrane interaction descriptors Table 12. Goodness-of-fit and robustness of the MI-QSAR models for skin irritant aliphatic alcohols and phenols. Submodel N MI-QSAR Equation r2 r2cv LSE Aliphatic Alcohols 13 PII = 49.60 − 0.09E (vdw + chg ) + 3.46HOMO + 0.22EHB 0.95 0.84 0.35 Phenols 9 PII = −19.04 − 0.61E (vdw + chg ) + 3.11LUMO + 2.44 log P − 0.04SA 0.94 0.81 0.42 E (vdw + chg ), EHB, SA : intermolecular solute-membrane interaction descriptors 3.1.2. Skin Corrosion Various models for skin corrosion have been reported by Barratt and collaborators [70]-[75]. Most of these studies do not provide explicit classification models but principal component (PC) analyses based on physicochemical properties, which allow visual discrimination between corrosive (C) and non-corrosive (NC) chemicals. The models are based on homogeneous datasets, i.e. acids, bases, electrophiles or neutral organics. In addition to PC analyses, discriminant analysis (DA), stepwise regression, and neural network (NN) analyses are also reported. In these studies, Barratt also points out the potential of using combined in vitro cytotoxicity data together with more conventional physicochemical parameters. Tiered testing strategies for skin corrosion have been developed and assessed by Worth and Cronin [76]-[78]. The evaluation of a two-step strategy, based on the 22 sequential use of pH measurements and in vitro data, indicated that the combined use of these data improved the ability to predict corrosion potential [76]. Subsequently, a three-step strategy was reported, based on the sequential use of QSARs, pH measurements and in vitro data [79]. A separate study by Worth confirmed the usefulness of pH as a predictor of skin corrosion potential, and provided a new prediction model (PM) for identifying corrosive chemicals by a pH-dependent mechanism [78]. Barratt et al. [70]-[71] reported a study to predict the presence or absence of skin corrosivity potential for four datasets containing 30 mixed organic chemicals (acids, bases, and neutral), 20 organic acids, 21 organic bases, and 15 phenols, respectively. The descriptors were selected taking into account postulated mechanisms for skin corrosion. It was hypothesised that the chemical must first penetrate into the skin, and then it can kill the cells beneath the stratum corneum if it is sufficiently cytotoxic. Thus, LogP values, molecular volume (MV), and melting point (MP) was used to model percutaneous absorption, while the intrinsic acidity or basicity of chemicals expressed by the pKa value was used to account for cytotoxicity. Classification data and structure-corrosivity relationships were investigated separately for each subset of chemicals, and an exploratory similarity analysis was carried out by using principal components analysis (PCA). For each group of chemicals, plots of the first two principal components (PC) for the four descriptors were presented. In all cases, the chemicals classified as corrosive were clearly separated from those classified as noncorrosive. The counterbalancing effects of the different variables on corrosivity were also illustrated by the plots. Chemicals with low LogP values, large molecular volumes or high melting points (features associated with low skin permeability and low solubility) were likely to be found in non-corrosive areas of the plots, unless they were particularly acidic or basic (very low pKa values). On the other hand, the mechanism of cytotoxicity of phenols was associated with hydrophobicity and the proton-donating ability; this mechanism was characterised by the pKa value. Barratt also reported models relating skin irritation and corrosivity data of neutral and electrophilic organic chemicals [72]. As before [70], the parameters used modelled skin permeability (LogP, and molecular volume), reactivity (dipole moment (DM)) and compensated for the fact that skin irritation/corrosivity testing is carried out with 23 a fixed mass or volume of chemical. Datasets were analysed by using stepwise regression, discriminant analysis, and principal components analysis. Cross validation was carried out by using a fixed deletion pattern of three subsets of data; three trials were carried out with each subset omitted once. Although the models were reported in the same paper, skin irritation and skin corrosion were modelled separately. Finally, a stepwise approach to the skin corrosivity/irritation classification of neutral and electrophilic organic chemicals was outlined. Skin irritation. For the set of 52 neutral and electrophilic organic chemicals, discriminant analysis between irritant and non-irritant compounds based on the above parameters was able to discriminate well for only 73.1% of the dataset, and 67.3% when cross-validated (Table 13). The poor discrimination at the irritant/non-irritant classification boundary was attributed to biological variability. Stepwise regression analysis of the Primary Irritation Index (PII) for the same dataset showed a poor correlation (r2= 0.422; r2cv= 0.201) with a positive dependence on LogP and dipole moment, and a negative dependence on molecular volume, which confirmed the putative model for skin irritation (Equation 1). When the QSAR equation was used to predict the PII values, the 17 misclassified chemicals formed a band with predicted PII values lying between 2.07 and 3.48. This range of predicted PII values represented the range of uncertainty for the discriminant analysis, for which no firm predictions could be made. Table 13. Classification results for each step of the discriminant model for skin irritation. Step Entering Variable Total fraction well classified Unclassified Irritant well classified 1 MV 0.615 0.645 0.571 2 DM 0.615 0.613 0.619 3 LogP 0.712 0.742 0.667 4 1/MW 0.731 0.742 0.714 0.673 0.677 0.667 Cross-Validation Equation 1. Goodness-of-fit and robustness of the QSAR model for skin irritation. PII = 1.047 log P − 0.244MV + 0.888 DM + 0.353 n = 52; r 2 = 0.422; rcv2 = 0.201; s = 1.376; F = 11.70 24 Skin corrosion. For the set of 40 electrophiles and two pro-electrophiles, discriminant analysis based on LogP, MV and DM, was able to discriminate reasonably well between corrosive and non-corrosive chemicals (92.9% correctly classified). The addition of the third parameter, DM, did not increase the overall discrimination (Table 14). A plot of the first two principal components of the same parameters, accounting for the 89.9% of the total variance, showed a clear demarcation between corrosive and non-corrosive electrophiles. The predominant parameter in determining the skin corrosivity of electrophilic organic chemicals appeared to be the molar dose at which they are tested, because skin corrosivity testing is conducted using a fixed mass or volume of chemical. Table 14. Classification results for each step of the discriminant model for skin corrosion. Step Entering Variable Total fraction well classified Non-corrosive well classified Corrosive well classified 1 LogP 0.833 0.867 0.750 2 MV 0.929 0.933 0.917 3 DM 0.929 0.900 1.0 0.929 0.933 0.917 Cross-Validation In another study [73], the corrosive potential of a series of six fatty acids was investigated in the in vitro skin corrosivity test (IVSCT), using both rat and human skin. The variables used were the same as those used in the previously reported studies, i.e. LogP, MV, and MP for percutaneous absorption, and pKa for cytotoxicity. The dataset previously reported in [70], was extended with the six additional fatty acids, and consisted of 27 organic acids. A plot of the first two principal components, containing 84.3% of the variance of the original data for organic acids, was able to discriminate between corrosive and non-corrosive chemicals on the basis of effects on rabbit skin. However, for human skin there was a shift in the corrosive/non-corrosive threshold. This was explained in terms of the effect of LogP: although skin permeability increases almost linearly with 1ogP, at a certain point it begins to level off. As LogP increases further, there is evidence that skin permeability of fatty acids starts to decrease in human skin [47]. The consequence of this latter effect, together with the increase in melting point with increasing alkyl chain length, is that fatty acids with alkyl chains longer than 9 carbons are not corrosive to human skin. Due to 25 differences in barrier properties (stratum corneum thickness and lipid composition) between rabbit and human skin, the shift in the corrosive/non-corrosive threshold between animal and human skin was rationalised in terms of changes in the skin permeability parameters. Chemicals on the borderline could be expected to be those in which the classification was most sensitive to changes in skin type. Thus, the classification of chemicals located in the boundary could be equivocal as a consequence of variability of biological data. In contrast, the classification of chemicals that were away from boundary regions were least likely to be affected by such changes. The datasets of organic acids, bases and phenols published in earlier works were extended in a subsequent study by Barratt [74]. As in the previous studies, PCA on LogP, MV, MP, and pKa confirmed that the analysis was able to discriminate between corrosive and non-corrosive chemicals. In addition to PCA, the datasets were also analysed employing neural networks (NN), using the same parameters as inputs. NN were trained in the range 0.0 to 1.0, with non-corrosive chemicals being assigned the value 0 and corrosive chemicals the value 1. The accuracies of the NN were 97.02%, 93.86%, and 95.85% for the acids, bases and phenols, respectively. As well as yielding classification predictions in agreement with those in the training sets, predicted outputs in the 0 to 1 range gave an indication of the confidence of the predicted classification, especially in regions near to the classification boundaries, where there can be uncertainty in the prediction. The regions of biological uncertainty at the classification threshold between corrosive and non-corrosive chemicals were also examined. In a subsequent study, Barratt et al. [75] derived structure-corrosivity relationships relating the severity of skin corrosivity of 27 acids (from [74]) designated by the EU risk phrases R34 and R35, to descriptors for skin permeability and cytotoxicity. Skin permeability was modelled by the usual parameters (LogP, MV and MP) whereas the cytotoxicity was accounted for by pKa values and an in vitro cytotoxicity parameter for the effects of the sodium salts of the acids on mouse embryo cells (pEC50). The dataset was analysed by using PCA and back-propagation neural network (NN) analysis. Plots of the first two PCs using only the conventional physicochemical descriptors failed to distinguish the acids classified as R34 from those as classified as R35. A plot of PCs 2 and 3 gave better resolution, but with some embedded clustered 26 data, possibly due to the presence of electrophilic groups conferring additional cytotoxicity to that expressed by the pKa parameter. When in vitro cytotoxicity parameters, expressed as pEC50 values, were included in the PCA plot, PCs 2 and 3 showed the best discrimination between acids classified as R34 and R35. Backpropagation networks contained a single hidden layer with four nodes. The values of the inputs were transformed so that the lower limit was 0.0 and the upper limit 1.0. The accuracy for predicted classifications ranged from 84 to 87%. The classification predictions were in agreement for 26 of the 27 acids; when an outlier was omitted, the accuracy achieved for classification increased to 96.26%. PCA and NN analysis were both able to discriminate well between chemicals with the EU classifications R34, and R35. Tiered assessment schemes for the prediction of skin irritation and corrosion have been designed and evaluated by Worth [49]. The first step of the process consisted of classifying corrosivity based on melting point (MP) and molecular weight (MW). This was followed by the use of a classification model (CM) based on pH (Step 2). Subsequently, in vitro data from the EPISKIN assay were used (Step 3). If the compound was ultimately classified as non-corrosive in the third step, a similar iterative process was performed to identify skin irritants. Ultimately, if the compound was predicted non-irritant, the in vivo Draize skin test was applied. Thus, the purpose of such a scheme was to classify chemicals and to confirm negative classifications with the use of animal tests. In a subsequent study, Worth et al. [76] evaluated the uncertainty associated with the predictive abilities of two-group classification models (CM), expressed in terms of their Cooper statistics. Standard and percentile bootstrap resampling techniques were used to judge whether predicted classifications were significantly better than the predictions made by a different CM, or whether the performance of a CM exceeded predefined performance criteria in a statistically significant way. This method was illustrated by constructing 95% confidence intervals (CI) for the Cooper statistics of four alternative skin corrosivity tests (TER, EPISKIN, Skin2, and CORROSITEX), as well as two-step sequences in which each in vitro test was used in combination with a physicochemical test for skin corrosion based on pH measurements. The PMs were applied to a dataset of 60 chemicals already published [79]. Cooper statistics were used to determine whether the four two-step sequences, with sensitivities greater than 27 or equal to 70%, were significantly more predictive than the four stand-alone in vitro tests. This study showed that the performances of the TER, EPISKIN, Skin2 and CORROSITEX tests in combination with the pH test were better than the individual performances of the in vitro tests (Table 15). Table 15. Classification results of the individual in vitro tests and the in vitro tests in sequence with the pH test. Bootstrap mean estimates In vitro test Concordance Sensitivity Specificity TER 78 83* 87 96* 71 73* EPISKIN 81 88* 83 96* 80 82* Skin2 75 83* 45 70* 100 94* CORROSITEX 73 74* 71 72* 76 78* * Use of the in vitro test in sequence with the pH test The bootstrap resampling technique was subsequently applied to assess the variability of Draize tissue scores to estimate acute dermal and ocular effects [77]. This technique was used to estimate biological variability arising from the use of different animals, and temporal variability arising from different time-points. The estimates of variability where then used to determine the extent to which Draize skin and eye test tissue scores could be predicted. A dataset of 143 ECETOC chemicals was used for the variability of Draize skin scores, while for eye scores the dataset consisted of 92 ECETOC chemicals. The results indicated that the variability in Draize skin scores was such that no model for predicting PII could be expected to have r2>0.57; in contrast, the variability in Draize eye scores was such that no model for predicting MMAS could be expected to have r2> 0.81. Finally, a paper by Walker et al. reports the so-called Skin Irritation Corrosion Rules Estimation Tool (SICRET) that was developed to estimate whether chemicals are likely to cause skin irritation or skin corrosion [80]. SICRET is a tiered approach that uses physicochemical property limits, structural alerts and in vitro tests to classify chemicals. 28 3.1.3. Eye Irritation / Corrosion In comparison to skin irritation and corrosion, a more significant number of (Q)SAR analyses have been reported to model local eye toxicity. The first reported models were incorporated into or derived from expert systems. The model published by Enslein et al. [81], based on structural fragments, was implemented in TOPKAT. In contrast, the models published by Klopman et al. [82]-[84] used chemical substructures implemented in MultiCASE (MultiCASE Inc.). Enslein et al. [81] reported a QSAR study for the estimation of rabbit eye irritation scores, using the same basis as the study performed on skin irritation [61]. For statistical and modelling purposes, the original dataset (consisting of 592 compounds) was divided into compounds without rings and compounds with at least one ring. The descriptors used were Kier and Hall molecular connectivity indices and computergenerated MOLSTAC substructural keys, which account for the contribution of structural fragments to eye irritation. Separate models were developed for the two subsets by using discriminant analysis (DA); furthermore, two sub-equations were developed for each model in order to separate the various classes of irritancy. After collinearity analyses and detection of outliers, the subset of variables which explained most of the variance was determined by an all-possible regressions algorithm, which generates all the possible combinations for a predefined number of variables. DA produces probabilities ranging from 0 to 1; the region from 0.3 to 0.7 was considered as indeterminate, and therefore a region in which predictions should not be used. For non-ring models, it was not possible to separate the mild irritants from the moderate ones. The classification accuracy of the equation which separated negative compounds from all others was 89.8%, with an indeterminate rate of 15%, while the model separating severe irritant compounds from all others yielded an accuracy of 90.2%, with an indeterminate rate of 30.9%. In contrast, for ring-models, there was 92.7% accuracy for the discrimination between negative/mild irritants and moderate/severe ones, and a 87.8% accuracy for the separation of severe irritants from all others. The classification accuracies of the QSAR models is summarised in Table 16. In general, the false negative rate turned out to be substantially higher than the false positive one. 29 Table 16. Classification results of the discriminant QSAR models for eye irritation. Models Non-ring compounds Ring compounds Subequations Overall Accuracy False Positives False Negatives Indeterminate F (p<0.001) Negatives vs Others 89.8% 3.2% (22.5%)* 7.0% (8.2%)* 15.0% 25.3 Severes vs Others 90.2% 2.7% (4.3%)* 7.1% (19.1%)* 30.9% 15.2 Negative/Milds vs Moderate/Severes 92.7% 2.9% (8.1%)* 4.45 (6.9%)* 29.3% 8.63 Severes vs Others 87.8% 6.6% (9.9%)* 5.6% (16.7%)* 24.3% 10.0 * Percentage of false predictions in relation to the experimental class The Multi-CASE SAR method was used to develop a predictive model of ocular irritation based on the evaluation of the significance of chemical structure in the induction of eye irritation [82]-[83]. The original Draize dataset was augmented and a new QSAR model was reported [84]. Multi-CASE was trained with a non-congeneric database of 186 organic chemicals previously evaluated in a modified Draize test for their ability to cause eye irritation in vivo (by Klopman et al. [82]). The eye toxicity of the compounds was ranked by giving an index of 15 to non-toxic molecules, 25 to those listed as marginally toxic, 35 to toxic molecules, 45 to very toxic molecules, and 55 to those found to be extremely toxic. A total of 37 functionalities and substructural attributes were found to be relevant for eye irritation potential, including carboxylate and sulphonate groups (in anionic surfactants), phenolic moieties (in non-ionic surfactants), amino groups, halogenated and unsaturated structures, anhydrides and epoxides. Ester groups were recognised as having a low irritation potential. The Multi-CASE program was then used to predict the eye irritation potential of 21 test chemicals, chemical mixtures, and polymers. Five of the compounds contained substructural fragments not encountered in the training set, but the remaining 18 compounds were successfully predicted. The eye irritation database of the previous study consisting of 186 compounds, extended with 21 compounds, was studied by using a Baseline Activity Identification Algorithm (BAIA) [83] that uncovers the underlying relations between the biological activity of a diverse set of molecules and global parameters. Chemicals were ranked by a qualitative score, according to the irritation potential. Following elimination of 40 molecules, the correlation between LogP and the remaining 167 molecules yielded 30 r=0.60, and F=93.2 (Equation 2). Correlations between irritancy and LogP revealed that hydrophilic molecules tend to be more toxic. However, the variation in activity with changes in LogP was rather small. The program identified 20 substructures accounting for the toxic activity of the molecules (biophores). Equation 2. Goodness-of-fit and robustness of the QSAR model for eye irritation. eye scores = 34.5 − 3.155 log P n = 167; r = 0.60; F = 93.2; SD = 13.637 A different SAR study based on the results of 297 chemicals tested in the Draize eye irritation assay was developed with Multi-CASE [84]. Prior to the SAR analyses, the major differences between the physicochemical properties of ocular irritants (I) and non-irritants (NI) were determined: in comparison with NI chemicals, I chemicals had significantly lower molecular weights, higher aqueous solubilities, and lower lipophilicities. The SAR analysis indicated that chemical reactivity was not associated with irritancy; in fact, I were significantly less electrophilic (lower LUMO), less nucleophilic (lower HOMO), less electronegative, and less reactive (greater electron gap) than NI. The authors used the Multi-CASE expert system to identify biophores (substructures which occur with a significantly greater frequency in irritants than in non-irritants) and biophobes (substructures which occur significantly more frequently in non-irritants than in irritants). The major structural determinants included primary, secondary and tertiary amine groups (i.e. basic groups), as well as carboxylate, organosulphate and sulphonate groups (i.e. acidic groups). Multi-CASE analyses resulted in the development of SAR models which explained the 98.8% and 99.3 % of the activity, respectively. However, the activities of many compounds (n=22) were explained by unique biophores that occurred only once. This finding suggested that chemicals in the training set had a low informational content (41.5%), and a restricted chemical coverage. The predictivity of the SAR submodels derived from the dataset was determined by a ten-fold cross-validation procedure. The overall predictivity determined for the individual SAR submodels ranged from 66.3% to 69.7%, yielding 69.7% for the consensus model (Table 17). The predictive performance determined after removing chemicals with unique biophores (out of the applicability domain) improved to 74.2% for the consensus model. Thus, the major structural determinants included hydrophilicity, alkalinity, acidity, and specific receptor-binding interactions. 31 Table 17. Predictivity of the SAR submodels and consensus model for eye irritation. SAR submodel Cut-off Sensitivity Specificity Concordance MultiCASE potency 18 units 0.47 0.56 * 0.85 0.86 * 66.7% 72.4% * MultiCASE prob. 49% 0.49 0.58 * 0.83 0.83 * 66.3% 72.0% * CASE potency 18 units 0.56 0.59 * 0.83 0.82 * 69.7% 71.3% * CASE prob. 66% 0.46 0.51 * 0.87 0.87 * 67.0% 70.6% * Consensus Model 50% 0.60 0.69 * 0.80 0.79 * 69.7% 74.2% * * without unique fragments MultiCASE is the updated version of CASE, which includes CASE algorithm as an option; the new program MultiCASE includes the ability to deal with subtle geometrical differences, with synergies, and with logical analysis and hierarchical decision making. Later, Sugai and coworkers published two papers: a QSAR model developed by applying the adaptive least squares (ALS) method to a heterogeneous dataset [85], and a QSAR for a reduced group of salicylates [86]. The first paper [85] reported a study of the relationship between the structural features of a set of 131 heterogeneous chemicals and the primary eye irritation in rabbits. The primary eye irritation ratings were classified into three classes on the basis of the recovery time of corneal and conjunctival damages: class I (recovering of damages within 24 hours); class II (damages persisting for more than 24 hours but recovering within 21 days); and class III (damages not recovering within 21 days). Thirty six substructural features were used to describe the molecules, i.e. a physicochemical descriptor, 34 descriptors accounting for the presence/absence/number of occurrences of structural fragments, and an indicator variable accounting for structural fragments and moieties. A QSAR model was developed by using the nonparametric pattern classifier adaptive least squares (ALS) method. After a preliminary selection of descriptors by backward stepwise elimination, several subsets of descriptors were selected by using an everypossible-combination technique that minimised the number of misclassified chemicals. The derived three-class discriminant function included 18 descriptors. The descriptors with high positive contributions in the discrimination function (rings with one or two N atoms, –SO2–, –NO2, C=O bonds, and –COOH and –OH on benzene ring, and halogens) indicated the substructural features which contribute to the enhancement of the affinity of chemicals for proteins or lipids in ocular tissue due to electronic effects, resulting in irritation. In contrast, the descriptors with negative 32 contributions (R–O–R’, and –NH2) were associated with a reduced likelihood of irritation. The accuracy in classifying the chemicals into three classes was 86.3% in the recognition step, and 74.0% in the LOO prediction step. A prediction of irritation potential of an external set of 28 diverse chemicals supported the reliability of the discriminant function, with an accuracy of 72.4% and all the misclassified chemicals assigned to the nearest neighbouring classes (Table 18). Table 18. Classification results of the discriminant models for eye irritation. N. misclassifications N. misclassifications (by 2 grades) Spearman rank correlation coeff. Accuracy Recognition 18 0 0.869 86.3% LOO prediction 34 0 0.736 74.0% External prediction 18 0 - 72.4% In the second part of the previously reported study [85], the primary eye irritation potential of ten salicylates was assessed by in vivo and in vitro studies [86]. The effects of salicylates on human serum γ-globulin (HSG) and lipid matrix were investigated with an in vitro model for the prediction of eye irritation. A nondestructive structural analysis of the corneal surface treated with salicylates was performed with IR spectroscopy. QSARs were derived for the irritation potential and in vitro effects of salicylates. A good correlation was obtained between the ability of salicylates to produce HSG aggregation and protein denaturation; in addition, statistical correlations were obtained between HSG aggregation of and the corneal and conjunctival scores. However, poor correlations were obtained between the Draize score and the lipid-substance interaction parameter. The physicochemical descriptors (acid dissociation constants and partition coefficients) were correlated with in vivo and in vitro data (Table 19). For in vivo data, pKa values were correlated with the cornea scores (r=–0.73) but not with the conjunctiva scores (r=–0.37). Salicylates with low pKa values led to an increase in the cornea score. LogP values did not correlate with either the cornea scores or the scores of conjunctiva. For in vitro data, a significant correlation was found between the aggregation of HSG and pKa values (r=–0.82), and salicylates with low pKa values led to an increase in the production of HSG aggregates. A poor correlation was obtained between the aggregation of HSG and the LogP values. Thus, for salicylates, increasing dissociation potential and electron withdrawing properties were important factors in inducing corneal 33 opacification and HSG aggregation. The corneal irritation and protein aggregation potential of salicylates were found to correlate with the acid dissociation constant more closely than the octanol/water partiion coefficient. Thus, eye irritation caused by salicylates was mainly attributed to the denaturation of proteins in ocular tissue, while effects on proteins depended on the dissociation potential of molecules. Table 19. Goodness-of-fit of the linear regression QSAR models for eye irritant salicylates. Correlation N. chemicals Equation r (P<0.05) pKa vs S (cornea) 10 S = −31.0 pKa + 144.1 –0.725 pKa vs S (conjunctiva) 10 S = −1.53 pKa + 11.7 –0.365 LogP vs S (cornea) 10 S = −3.72 log P + 60.0 –0.195 LogP vs S (conjunctiva) 10 S = −0.184 log P + 6.80 –0.100 Log(1/C) vs pKa 10 log(1 C ) = −0.277 pKa + 2.84 –0.820 Log(1/C) vs LogP 10 log(1 C ) = 0.023 log P + 1.98 0.155 S: in vivo Draize scores; log(1/C): in vitro aggregation of HSG Lipnick et al. [87] analysed a dataset of 42 aliphatic alcohols and ketones, consisting of eye scores on an integer scale of l-10. Parabolic and bilinear models using LogP were fitted to model the eye score; the bilinear model yielded a better fit. Cronin started a series of papers analysing different statistical methods applied to an eye irritation dataset. He attempted to find linear regression relationships for a physically heterogeneous set of compounds [88]. However, the impossibility of deriving meaningful (Q)SAR models suggested the need to restrict the analysis to pure bulk liquids, developing submodels for limited datasets of alcohols and acetates. Discriminant analysis revealed an embedded cluster of irritant chemicals within the non-irritant chemicals, which was investigated further by Cronin [89]. Subsequently, the method of cluster significance analysis (CSA), which can be used to determine the significance of embedded clustering, was applied to the eye irritation dataset, and the technique of Embedded Cluster Modelling (ECM) was developed to generate elliptic prediction models for the embedded clusters [90]. The series of papers is closed with a comparative analysis [91] that illustrated the use of several linear statistical techniques to develop classification models (CM), i.e. linear discriminant analysis (LDA), binary logistic regression (BLR) and classification tree (CT) analysis. 34 Cronin et al. [88] reported a number of QSAR investigations based on an ECETOC dataset of 53 chemical diverse compounds with measured Draize eye irritation scores (DES). These numerical pseudo-quantitative parameters based on the summation of subjective scores were adjusted with the molarity of pure chemicals into molar eye scores (MES). The dataset was first analysed by using a set of 23 physicochemical descriptors accounting for hydrophobicity, size and shape, and electronic characteristics (Table 20). Three QSAR models obtained by stepwise regression analysis to all compounds in the dataset (n=38) and to the 23 physicochemical descriptors yielded poor statistical results: from r2=0.16 (using the square of LogP) to r2=0.35 (adding LUMO and the zero-order Kappa-alpha index). These results suggested that different chemical groups were driven by varying mechanisms of toxic action. QSAR analyses were attempted for defined chemical classes (alcohols and acetates) expected to follow a similar mechanism of action. When the dataset was split into subclasses, the QSAR model for 9 alcohols correlating the molar eye scores with LogP and HOMO achieved r2=0.80. For the subgroup of 7 acetates a correlation with LogP yielded r2=0.74. Stepwise linear discriminant analysis on the entire dataset with the 23 descriptors found no significant variables that could discriminate between irritants and non-irritants. This suggested that there was no linear distribution of the irritants with any of the variables considered. Principal components analysis of the 32 parameters revealed that four components were sufficient to account for 92.6% of the original information. As a pattern recognition technique, the pairwise score plots of the four PCs for the 23 variables revealed that active compounds formed an embedded cluster surrounded by the diffuse cluster of inactive chemicals. Table 20. Goodness-of-fit of the linear regression QSAR models for eye irritants. Compounds All compounds Alcohols Acetates r2 S F MES = −0.118(log P ) + 3.33 0.16 2.56 8.2 38 MES = −0.125(log P ) + 0.713LUMO + 2.23 0.27 2.4 7.7 38 MES = −0.156(log P ) + 1.13LUMO + 0.397κα 0 − 0.44 0.35 2.26 7.6 9 MES = 2.06 log P + 4.13 0.72 1.89 21.3 9 MES = 1.72 log P + 9.04 HOMO + 102 0.80 1.58 17.3 7 MES = −1.37 log P + 2.97 0.74 0.52 18.2 n Equation 38 2 2 2 MES: Molar Eye Score 35 The presence of an embedded cluster was corroborated by an further paper [89], which identified the parameters that formed an embedded cluster of eye irritants among non irritants. Out of the total of 23 physicochemical descriptors, the five most significant were LogP, LogP2, heat of formation, dipole moment, and zero-order valence corrected molecular connectivity. A representation of heat of formation against LogP did not achieve full separation. However, PCA on the significant parameters allowed the embedded cluster to be visualised. Worth and Cronin reported another example of an embedded cluster prediction model derived from the ECETOC dataset [90]. The statistical significance of the embedded cluster was assessed by cluster significance analysis (CSA). The embedded cluster modelling (ECM) method allows the generation of elliptic models of two or more dimensions that can be used to identify embedded data. This paper reported a twodimensional ECM model for eye irritation potential based on 73 diverse organic ECETOC chemicals, using LogP (which models chemical diffusion) and a sizeindependent molecular connectivity index (dV1, which accounts for the degree of branching and cyclicity). The sensitivity and specificity were 73% and 78%, respectively. The irritant region of parameter space was defined by a LogP range between 0.7 and 1.43, and an intermediate degree of branching or cyclicity. An further investigation by Worth and Cronin [91] illustrates the use of several linear statistical techniques to develop classification models (CM), i.e. linear discriminant analysis (LDA), binary logistic regression (BLR) and classification tree (CT) analysis, for predicting the presence or absence of eye irritation. The different methods were applied to a dataset of 119 organic liquids classified as I or NI according to EU classification criteria. The CMs were developed by applying LDA, BLR, and CT analyses, using a single predictor variable (molecular weight), and assigning equal probabilities for the two classes (I/NI). The cut-off values below which a chemical should be predicted to be irritating to the eye were 121, 77, and 137 g/mol, in the LDA, BLR, and CT classification models, respectively (Table 21). Although all the statistical techniques yielded comparable models, with concordances between 61 and 76%, the most appropriate model was developed by the non-parametric CT analysis, since the assumptions of LDA and BLR had not been met. Some differences outlined for the three methods were that: a) LDA and BLR make a number of assumptions about the underlying data, whereas CT analysis is a non-parametric technique; b) 36 LDA and BLR can be used to generate a distinct probability of group membership for individual chemicals, whereas CT analysis only produces average probabilities for the different groups; c) LDA and BLR use all (statistically significant) predictor variables simultaneously in the model, whilst CT analysis uses the predictor variables in a hierarchical and recursive manner; d) LDA and BLR models assign all objects to a pre-defined group in a single step, whereas CT have the flexibility of assigning the objects in one or more steps. Table 21. Classification results of the different models of eye irritancy. CM (p<0.01) Cut-off value Sensitivity Specificity Accuracy Linear Discriminant Analysis (LDA) if MW ≤ 121 g/mol, then predict I; otherwise, predict NI 73 62 65 Binary Logistic Regression (BLR) if MW ≤ 77 g/mol, then predict I; otherwise, predict NI 27 93 76 Classification Tree (CT) if MW ≤ 137 g/mol, then predict I; otherwise, predict NI 97 49 61 Following the same PCA approach used to assess skin irritation and skin corrosion [70], Barratt developed a series of “models” for eye irritation. He reported principal component analysis (PCA) able to discriminate between I and NI, among reduced series of neutral organic chemicals [92]-[93]. In the first paper [92], Barratt used a variation of the extended ECETOC dataset of neutral organic chemicals previously used by Cronin [88], composed of 46 compounds. The parameters used to model eye irritation potential, i.e. LogP, the minor principal inertial axes (Ry and Rz) and dipole moment (DM), were selected according to mechanistic considerations. The choice of DM as a reactivity parameter was based on the premise that low molecular-weight neutral molecules can alter electrical resistance and thus the ion permeability of cell membranes. It was hypothesised that changes in electrical resistance in the membranes of eye might lead to cytotoxicity and result in irritation. The choice of LogP and parameters accounting for molecular dimensions was based on the expectation that chemicals should be able to partition into the membrane and hence should possess appropriate hydrophobic/hydrophilic properties. All three principal inertial axes were used in the initial studies, but the size and shape parameters which gave the best discrimination were Ry and Rz, which model the cross-sectional area of 37 the long axis of molecules. The four descriptors were analysed by PCA. The plot of the first two PCs of the four parameters was able to discriminate between irritant and non-irritant chemicals in the dataset. Thus an eye irritant would be a chemical hydrophobic enough to allow partitioning into a biological membrane whilst also possessing a moderate dipole moment. To support the validity of the original four-parameter eye irritation model for neutral organic chemicals, the previous study of Barratt [92] was followed up by additional analyses [93]. In the latter study, the same descriptors used to predict the eye irritation potential of neutral organic chemicals (LogP, the minor principal inertial axes (Ry and Rz) and DM) were applied to a dataset of 34 aliphatic alcohols previously reported by Lipnick et al. [87], with measured eye irritation scores on an integer scale of 1-10. In addition, the previous dataset of neutral organic chemicals was extended to 57 chemicals, including some aliphatic alcohols to increase the chemical parameter space covered by the prediction model. As in other studies [74]-[75], for the first dataset of aliphatic alcohols, back propagation NN analysis was carried out by using the four selected parameters as input, and using an integer scale from 1 to 10 for the eye irritation scores output. The most important variables turned out to be Rz and LogP, with the accuracy of the network being 97.27%. A plot of predicted values of the eye score against actual eye scores showed a strong correlation (r=0.9). However, despite this good fit, there was a notable uncertainty in the prediction of eye score values, possibly due to biological variability. For the extended dataset of 57 neutral organic chemicals and aliphatic alcohols used in the previous paper [92], PCA based on the same four physicochemical parameters concentrated 89.7% of the variance into the first two components, which showed a clear discrimination between irritant and nonirritant chemicals. The dataset was also investigated by NN analysis, using the same inputs and an integer scale from 0 (for NI) to 1 (for I). The accuracy was 97.37%, with the dipole moment being the most important variable. Abraham et al. [94] used parameters derived from chromatography experiments to model a dataset comprising 38 of the 53 organic liquids previously analysed by Cronin et al. [88]. Subsequently, Abraham et al. [95]-[97] extended the dataset and studied the relationship between the mechanism of action in the Draize test and in humans. In Cronin et al. [88], it was concluded that the lack of success of the QSAR analysis was possibly due to varying mechanisms of toxic action of the different 38 chemical groups, and to the choice of descriptors. In the study by Abraham et al. [94], a transport-driven mechanism was hypothesised for the process of transfer of an irritant from the state of a pure organic liquid to a dilute solution in an organic solvent biophase (model for the biological structure of the eye). In order to obtain a response related to the transfer process, DES was related to the activity coefficient (γo), which is the equilibrium constant for the transport mechanism, by means of a vapour solubility in the solvent phase (L) and the saturated vapour pressure of the pure liquid (Po). The descriptors used for the QSAR model were compound (or solute) parameters: excess molar refraction (R2), polarizability/ dipolarity ( π 2H ), effective hydrogen bond acidity and basicity ( ∑ α 2H and ∑β H 2 ), and the logarithm of the vapour solubility in hexadecane (logL16). Two QSARs were modelled (Table 22). When using the DES, a very poor equation was obtained (r2=0.27). However, a reasonable correlation was achieved (r2=0.89) when using the modified log (DES/Po); the results improved with the removal of an outlier and the removal of a nonsignificant descriptor (r2=0.95). On transforming the calculated log (DES/Po) values back to calculated DES, the correlation between log(DES) observed and calculated yielded a good agreement r2=0.77. Table 22. Goodness-of-fit of QSAR models for eye irritant pure organic liquids. N Descriptors r2 SD F 38 log(DES/Po) vs R2, π 2H , ∑ α 2H , ∑ β 2H and logL16 0.894 0.46 54.0 0.953 0.32 125.5 37 37 log(DES/Po) vs π 2H , ∑ α 2H , ∑ β 2H and logL16 0.951 0.32 155.9 37 log(DES) observed vs log(DES) calculated 0.771 0.30 117.6 In a subsequent study, Abraham et al. [95] combined Draize eye scores (DES) and human eye irritation thresholds (EIT) into a single QSAR, assuming that the main feature of both in vivo responses was the transport of the compound into the biophase. They extended the original dataset of measured DES for 38 pure organic liquids with a dataset of 17 volatile organic compounds (VOC) for which EIT had been measured. The two scales of irritation, DES and EIT, were compared, establishing 0.66 logarithmic units of difference. The models found in the previous study were redeveloped for the new extended dataset of 54 compounds (Table 23), obtaining r2=0.928 and r2=0.924 for the models with and without the R2 variable. Finally, EIT 39 scores were predicted from DES values, and they were compared with experimental EIT. Table 23. Goodness-of-fit of QSAR models for the eye irritancy of pure organic liquids. r2 r2cv SD F log(DES/P ) vs R2, π 2H , ∑ α 2H , ∑ β 2H and logL 0.928 0.913 0.363 123.8 log(DES/Po) vs π 2H , ∑ α 2H , ∑ β 2H and logL16 0.924 0.910 0.368 149.6 N. Descriptors 54 54 o 16 In a more recent study, Abraham et al. [97] extended the connection between the eye irritation effects on rabbits (MMAS) modified by vapour pressure, and the effects on humans (EIT) [95]. They combined two different datasets into an extended dataset of 91 compounds. The total dataset for 91 compounds was analysed by the same general solvation equation (Table 24). To establish the relationship between MMAS and EIT and to determine whether the two datasets were compatible, an indicator variable (I) to account for the dataset was set. The same model for eye irritation used in the first paper of the series [94] was re-developed, and a single QSAR covering a wide range of compound types fit the two sets with r2=0.94. When the equation was applied to the subset with calculated MMAS values, comparable statistics were obtained (r2=0.95). To test the predictive power of the model, the dataset was split into a training set of 45 compounds; the corresponding equation for the training set compounds (r2=0.921) was used to predict the remaining 46 test set compounds. The mechanistic assumption underlying this combined model was that rabbit DES and human EIT result from the passive transfer of the compound from the bulk liquid or the vapour phase to a biophase. This passive transfer is assumed to be non-specific, unless there is some particular interaction between the substituents for which there is a transfer from the biophase to a receptor. In such a case, the shape of a compound could be crucial, so to investigate whether QSARs might be improved by the inclusion of other types of descriptors, three shape descriptors were included by using an energy minimised conformation: three-dimensional volume (VG), the longest length of the minimum energy conformation (LG), and a shape descriptor that takes into account branching and the distance of atoms from a centre of gravity of a molecule (DPO1). The additional shape descriptors had almost no effect (r2=0.94). 40 Table 24. Goodness-of-fit of QSAR models for eye irritants. N. Descriptors r2 SD AD AAD F 91 log(DES/Po) vs R2, π 2H , ∑ α 2H , ∑ β 2H , 0.936 0.433 0.000 0.340 204.5 68 log(DES/Po) vs R2, π 2H , ∑ α 2H , ∑ β 2H 0.954 0.398 0.000 0.318 255.8 45 log(DES/Po) vs R2, π 2H , ∑ α 2H , ∑ β 2H 0.921 0.469 0.000 0.354 74.0 log(DES/Po) vs R2, π 2H , ∑ α 2H , ∑ β 2H , 0.941 0.422 0.000 0.334 144.0 91 logL16, I and logL16 and logL16 – internal prediction logL16, I, DPO1, VG, LG AD: Average Deviation; AAD: Absolute Average Deviation DPO1, VG, LG: shape descriptors Kulkarni and Hopfinger developed a new approach, called Membrane-Interaction QSAR (MI-QSAR) analysis [98], which combines structure-based design techniques with the traditional QSAR paradigm. Molecular dynamic simulations (MDS) are used to generate membrane–solute interaction properties for the interaction of organic chemicals with model phospholipid membranes. This methodology allows investigating structurally diverse training sets using a genetic algorithm approach. MIQSAR was applied to the neutral organic chemicals within the ECETOC eye irritation dataset [99]. The seminal study of Kulkarni et al. [98] reports a MI-QSAR model based on a hypothesised mechanism of eye irritation and developed by using a diverse Draize eye irritation dataset of 38 compounds. Molecular dynamic simulation was used to generate intermolecular membrane-solute interaction properties. These intermolecular properties were combined with intramolecular physicochemical properties and features of the solute (irritant) to construct QSAR models by using multi-dimensional linear regression and the genetic function approximation (GFA) algorithm. The principal correlation descriptors, solute aqueous solvation free energy and solutemembrane interaction energies, were selected from a trial set of 95 descriptors for 16 structurally diverse compounds fully representative of the whole dataset (aliphatic and aromatic hydrocarbons, and aliphatic ketones, alcohols and acetates). A QSAR based 41 on three parameters (two energy terms and a topological index) showed a good fit (r2= 0.92). Kulkarni et al. [99] reported a second MI-QSAR study to predict molar-adjusted Draize scores for the ECETOC dataset of 38 structurally diverse compounds. The mechanistic hypothesis for eye irritation relied on the uptake and diffusion of an irritant into the keratocytes of the corneal epithelium. Thus, uptake and interaction of each solute molecule with a model phospholipid membrane were simulated. The membrane-solute and solute-solvent intermolecular interaction properties estimated from molecular simulations were added to intramolecular physicochemical property descriptors to provide an extended set of descriptors. Thus, the descriptors used were physicochemical parameters associated with the solute molecule and with the proposed receptor (intramolecular properties), as well as parameters describing the interaction between both (intermolecular solute-membrane properties). Linear MIQSAR models for the 38 compounds were built including the outliers from the previous study [98] (Table 25); the MES was correlated with the aqueous solvation free energy, and additively with the electrostatic solute-membrane interaction energy, obtaining r2=0.15 for both cases. In contrast, a GFA parabolic non-linear model based on the previous descriptors, Van der Waals solute-membrane interaction energy, and octanol solvation free energy, yielded r2=0.78. However, when removing the only two irritants from the model, a good-linear MI-QSAR model (r2=0.71) could be achieved. Eye irritancy was predicted to increase with the aqueous solubility of the molecule, with the binding energy of the solute to the phospholipid regions of the membrane, and with LUMO (as reactivity increases). 42 Table 25. Goodness-of-fit and robustness of the linear and non-linear MI-QSAR models for eye irritants. Model N. Model r2 r2cv LSE Linear 38 MES = 1.77 − 0.18F (H 2 O ) 0.15 –0.18 6.40 Linear 38 MES = 1.74 − 0.004 E (chg ) − 0.18F (H 2O ) 0.15 –0.20 6.30 Nonlinear GFA 38 0.78 0.73 1.52 (F=32.4) Linear 36 0.71 0.65 1.15 (F=38.9) { MES = 1.76 − 0.06[F (H 2O ) − 6.5] + 0.05[F (OCT ) − 4.5] 2 2 } − 0.01[E (vdw) − 3.9]2 −0.13E (chg ) + 0.05LUMO 2 MES = −0.66 − 0.01E (chg + vdw) − 0.48F (H 2O ) + 0.39 LUMO F (H 2O ), E (chg ), F (OCT ), E (vdw), E (chg + vdw) : intermolecular solute-membrane interaction descriptors A more recent MI-QSAR study has been published by Li et al. [100]. QSAR models were developed to predict drug permeability coefficients across cornea and its component layers (epithelium, stroma, and endothelium). From a training set of 25 structurally diverse drugs, QSAR models were constructed and compared for the permeability of the component layers. Cornea permeability was found to be dependent on the measured distribution coefficient, the cohesive energy of the drug, the total potential energy of the drug-membrane "complex" and three other energy refinement descriptor terms. The endothelium might be a more important barrier in cornea permeation than the stroma. Thirteen structurally diverse drugs, whose molar-adjusted eye irritation scores (MES) were measured by using the Draize test, were chosen as an eye irritation comparison set. A poor correlation (r2= 0.02) between the measured MES and the predicted cornea permeability coefficients for the drugs in the eye irritation set suggested there was no significant relationship between eye irritation potency and the corneal permeability. Patel [101] reported a method that considers mixtures of chemicals, called Quantitative Component Analysis of Mixtures (QCAM). The QCAM method was used to analyze an existing set of product formulation data to determine if the irritating ingredients in the mixtures could be identified [101]. Four eye irritation scores, based on a rat model, were analysed by QCAM for 18 mixtures having a composite total of 37 components. QCAM permits the construction of quantitative models for extracting the individual toxicity profile of each of the components of a mixture from knowledge of their relative concentrations and the net toxicity of the entire mixture. Thus, QCAM relates a net toxicity measure of a mixture to the 43 toxicities of the individual components through linear, quadratic, and pairwise crosscomponent concentration-dependent interactions. A correlation model was established by using a genetic function approximation (GFA) algorithm employing either multidimensional linear regression (MLR) or partial least-squares (PLS) regression. A correlation analysis of the eye irritation scores was carried out (Table 26). Each of the standard deviations (SD) was correlated with its respective mean value (MV): cornea and average SD correlated significantly with MV, and the larger the MV (greater eye irritation score), the larger the SD. This indicated that corneal and average eye irritation scores define NI compounds. Linear correlation equations between eye irritation scores and selected components (Table 27) showed that corneal eye irritation and average eye irritation are well-explained in terms of a linear model containing, at most, three components over the set of mixtures (r2= 0.83). It was also found that extensive cornea and average eye irritations were due to only one of these three components of the mixtures. In addition, one of the three significant components was predicted to decrease the extent of eye irritation, and subsequently identified as an anti-irritant in contact lens solutions. A reasonable linear correlation model could also be developed for conjunctival irritation, but no significant iris irritation model could be constructed. The addition of quadratic and/or cross-component concentration terms to a linear correlation model did not statistically improve the overall resultant model. The QCAM models permitted estimation of the intrinsic (self) toxicity of each of the components of a mixture. 44 Table 26. Linear Correlations between Standard Deviations (SD) and Mean Values (MV) for four eye irritation scores. Eye irritation scores Linear Equations r2 Cornea SD = 0.81MV – 1.2 0.91 Iris SD = 0.17MV – 1.1 0.08 Conjunctiva SD = – 0.03MV + 0.3 Average SD = 0.77MV – 3.8 0.26 0.90 SD: Standard Deviation; MV: Mean Value Table 27. Goodness-of-fit and robustness of QCAM models for eye irritation. Cornea Irritation Scores Iris Irritation Scores Conjunctiva Irritation Scores Average Eye Irritation Scores r2 r2cv r2 r2cv r2 r2cv r2 r2cv 1 0.71 0.60 0.23 0.09 0.52 0.41 0.74 0.65 2 0.83 0.74 0.42 0.29 0.63 0.36 0.84 0.77 3 0.88 0.83 0.62 0.10 0.71 0.38 0.88 0.81 N. components Patlewicz et al. [102] used a non-linear neural network (NN) analysis applied to cationic surfactants (Hayashi et al. also reported a QSAR for surfactants [103]). They reported a non-linear QSAR model based on 29 in vivo rabbit eye irritation tests for 19 different cationic surfactants, and derived by back propagation NN analysis [102]. The parameters used were LogP and molecular volume to model the partitioning of the surfactants into the membranes of the eye; the logarithm of the critical micelle concentration (logCMC), which is a measure of the reactivity of the surfactants with the eye; and surfactant concentration. MAS scores showed strongly positive, nonlinear correlations with surfactant concentration, and logCMC showed a strongly negative, non-linear correlation with LogP. The accuracy of the NN was 91.18%, while the Pearson correlation between the actual and predicted values of MAS was 0.838, showing that around 70% (r2=0.702) of the variance in the dataset was explained by the model. Patlewicz et al. [104] also proposed a refined QSAR model for prediction of MAS, using the same cationic surfactant dataset. In this case, the non-linear approach used was a Bayesian neural network (BNN) that provided a robust and accurate model (r2=0.89). The model was used to explore the relationship between MAS score and 45 molecular properties. The simulated results were used to identify the range of molecular properties of potential molecules with a probability greater than 75% of causing more than mild irritancy. Within the parameter space of the model, the probability of causing severe irritation was highest for molecules with MV < 320 Å3, and contained within the following ranges: - 2 < log P < 14 , and - 6 < log CMC < 3 . 46 4. Alternative test methods An alternative method is any method used to replace, reduce or refine (3 Rs) the use of animal testing methods for hazard and risk assessment. In relation to alternative methods, Council Directive 86/609/EEC states that “An experiment shall not be performed if another scientifically satisfactory method of obtaining the result sought, not entailing the use of an animal, is reasonably and practicably available” [105]. The objectives of this directive are to diminish the number of materials needing any animal testing, to decrease the number of animals per test, to minimize animal distress, to decrease ancillary testing and to ensure development of information suitable for hazard classification. Annex XI of REACH describes possibilities for waiving testing requirements, on the basis of exposure considerations, the use of existing data, (Q)SARs and chemical grouping, and in vitro tests in order to minimise the need for animal testing [10]. 4.1. Use of existing data The use of existing data includes the use of data on physicochemical properties, human and animal data, and non-guideline studies. Concerning the use of existing human data, an assessment of available human data is requested for all toxicological endpoints, including skin irritation/corrosion, and eye irritation. In addition, Annex XI states that historical human data, such as epidemiological studies on exposed populations, accidental or occupational exposure data and clinical studies, shall be considered. In addition, articles from peer-reviewed literature as well as study reports from companies can constitute a resource for the safety evaluation of chemicals and substances. 4.2. In vitro methods The number of stand-alone in vitro tests that have been validated and incorporated into an official test guideline, replacing an in vivo procedure, is limited. Recently, OECD test guidelines based on the use of in vitro tests on skin corrosion were adopted. In these cases in vitro methods may be used to replace the animal testing 47 currently required for hazard and risk assessment. For ocular and dermal irritation several well-studied in vitro tests are available [106]- [108] but none of them has yet been formally validated or adopted as an OECD guideline. Although the in vitro tests have not been formally included in OECD test guidelines, positive results obtained in several non-adopted in vitro tests for eye irritation are often considered sufficient for classification and labelling. In other cases, in vitro tests provide supplementary information, which may be used in interpretation of the relevance for humans of data from animal studies or for a better understanding of the toxicological mechanism of action of the substance. In addition, they are used in tiered testing strategies and weight of evidence approaches, in order to reduce the need to test further or to reduce the number of animals needed at a final confirmatory stage. 4.2.1. Skin Irritation In vitro alternatives vary from simple models such as keratinocyte cultures to more complex organotypic cultures and reconstituted human skin models. Although a number of in vitro alternatives have been proposed in the literature, to date there are no validated in vitro tests for predicting acute skin irritation and for replacing the classical Draize test [1]. Several prevalidation efforts on in vitro methods have been published [109]-[112]. In these studies, some of the more promising in vitro methods have been evaluated: EpiDerm™, EPISKIN™, PREDISKIN™, the non-perfused pig ear model, and the in vitro mouse skin integrity function test (SIFT). As a follow-up to these prevalidation efforts, an ECVAM validation study should be completed during 2006. Two tests (EpiDerm and EPISKIN) are being validated for the purpose of classification and labelling on the basis of skin irritation. At present there is no OECD TG for in vitro skin irritation. Until validated in vitro tests for skin irritation are available or guidance is provided on the use of non-adopted tests, in vivo skin irritation tests should be performed to meet the information requirements in the REACH proposal. However, it is foreseen that suitable in vitro tests could be used for identifying positives (skin irritants). 48 4.2.2. Skin Corrosion The traditional animal test used in the past to assess skin corrosion has been replaced by alternative in vitro validated methods [71],[112]-[114], which have been included in Annex V of the Dangerous Substances Directive [2]. These are the in vitro skin corrosion rat skin transcutaneous electrical resistance (TER) test, which uses excised rat skin as a test system and its electrical resistance as an endpoint; and the human skin models such as EpiDermTM and EPISKINTM, which are reconstructed human epidermal equivalents and use the cell viability as an endpoint. OECD TGs for in vitro skin corrosion testing have also been adopted [31],[34]-[35]. 4.2.3. Eye Irritation / Corrosion The Draize eye irritation tests has been criticised for animal welfare reasons and scientific reasons, including the subjectivity in allocation of the response and differences in sensitivity to tested substances between the rabbit and man [115]-[122]. Some modifications have been proposed to reduce the pain endured by the animal and achieve higher rates of recovery, such as the low volume eye test (LVET), where only a tenth of the sample volume used in the standard Draize test is applied to the rabbit's cornea and the eyes are not held shut after instillation of the test material. In vitro methods for eye irritation comprise organotypic models such as isolated eyes or components thereof, tissue and cell culture systems and physicochemical tests. Several in vitro validation studies to identify alternative tests capable of replacing the Draize rabbit eye irritation test have taken place, e.g. the EC/HO study [115], and the COLIPA study [116]. Although no test was found capable of replacing the Draize rabbit eye test, some of the assays showed considerable promise as screens for ocular irritancy [120]-[121]. At present there is no OECD TG based on the use of in vitro eye irritation tests. Until validated in vitro tests are available, it was agreed at the 64th Competent Authority (CA) meeting in November 2002 that a substance can be considered a severe eye irritant (R41) if it gives a positive result in one of several non validated in vitro tests, whilst negative results require confirmation by further testing in vivo. The same approach is expressed by Annex XI of the proposed REACH legislation. Promising non validated in vitro eye irritation tests include: two isolated organ-based methods (Bovine Corneal Opacity and Permeability (BCOP) test, Chicken Enucleated Eye Test 49 (CEET)); an organotypic method (Hen's Egg Test on the Chorio-Allantoic Membrane (HET-CAM)); and two tridimensional methods (EpiOcular™ and in vitro reconstituted Human Corneal Epithelium model (HCE)). 50 5. Expert Systems Expert systems are computer programs that guide hazard assessment and that are able to predict toxicity endpoints of certain chemical structures based on the available information. These systems can be used to provide an indication of the potential toxic activity of chemicals, for either screening or assessment purposes. Expert systems for the prediction of chemical toxicity are built upon experimental data representing one or more toxic manifestations of chemicals in biological systems (database), and/or rules derived from such data (rulebase). Ideally, an expert system should be heuristic (able to reason both with theories and with expert knowledge); transparent (able to motivate the line of reasoning); and flexible (able to integrate new knowledge into the existing base). Expert systems are used to make predictions on the basis of prior information, and in some cases mimic the ways in which groups of human experts solve problems. They are either knowledge-based systems, which use specific rules to relate existing physicochemical properties to specific toxicity endpoints, or automated rule induction systems, which aim to discover patterns within data. Some of the more commonly used expert systems to predict skin and eye irritation and corrosion are included in the OECD's Database on Chemical Risk Assessment Models [123]. Several reviews make comparative analyses of different expert systems [124]-[132]. The paper by Hulzebos et al. [132] assesses the results of three expert systems by applying the OECD Principles for (Q)SAR Validation. The most widely used expert systems for assessing irritation and corrosion are the BfR-DSS, DEREK for Windows (DEREKfW), MULTICASE, and TOPKAT, and will be described in the following sections. Within a rulebase, rules can be based on different approaches: explicit rules derived from knowledge and human expert judgement (expert rules and knowledge-based rules); and rules based on mathematical induction (induced rules), i.e. rules derived by statistical correlation. Induced rules have the advantage of extending existing knowledge without being biased toward particular mechanisms of toxic action; however, they can be empirical relationships devoid of biological meaning. In contrast, expert rules or knowledge-based rules are likely to have a mechanistic basis, but they are expressions of existing knowledge rather than of new knowledge. 51 Typically, (Q)SARs are induced rules, whereas expert rules are often based on reactive chemistry. An expert system based on entirely expert rules is referred to as a knowledge-based system (e.g. BfR-DSS, and DEREK), whereas a system based on induced rules is called an automated rule-induction system or a correlative system (e.g. MultiCASE and TOPKAT) [127]. Table 28 summarises the main characteristics of these expert systems [123]. 52 Table 28. Characteristics of DSS that model skin and eye irritation. Features BfR – DSS DEREK HazardExpert MultiCASE TOPKAT Model Name Decision Support System (DSS) for Local Lesions DEREK HazardExpert MCASE, MC4PC, TOXPC, META TOPKAT Developer / Supplier (German) Federal Institute for Risk Assessment LHASA Ltd. CompuDrug Ltd. MultiCASE Inc. Accelrys Inc. Country Germany United Kingdom Hungary United States United States Qualitative Type of Hazard information Provided Qualitative Qualitative Qualitative Categorical Quantitative Range Point estimate Categorical Qualitative Quantitative Point estimate QSARs Model Approach SARs SARs SARs / QSARs Deterministic or probabilistic Artificial intelligence expert system Model Applicability (Inside) Organic compounds (with purity >95%) Organic salts Industrial chemicals Can not evaluate Organic compounds No physicochemical properties; Model Applicability (Outside) Inorganic substances; Organic chelates; Diluted substances; Mixtures Model Format (Computerised or not ) Computerised algorithms No searchable database No searchable database Defaults and assumptions are provided for the user Computerised model Computerised model Input Structural and empirical formulas, melting point, molecular weight, aqueous solubility, octanol/water partition coefficient. Inorganic compounds Evaluation / Validation of Model Computerised algorithms Computerised algorithms Searchable database Conditions are defined by the user Searchable database Computerised model Model Availability Defaults and assumptions selected / by default Chemical structures (SMILES codes, MOL files, SDF files) Graphical interface Further requirements of in vivo testing Toxicophore (i.e. substituent with potential toxic effect), pertinent endpoint and literature references Toxicophores and substructures causing modulatory effects; Physicochemical parameters Predicted activity, rational for conclusions and documentation Evaluated with measured data; Evaluated by several companies using inhouse experimental data Evaluated in the literaure Validated using multiple 10%-off and recreation of model Proprietary Proprietary Validated using 400 datasets Non-proprietary, but not yet publicly available Peer-reviewed evaluation and validation Proprietary Organometallics; Inorganics Mixtures. Structural information (MOL,SDF, ISIS files). EU risk phrases for skin corrosion or serious damage to eyes; Organic molecules Polymers; Lipid solubility, surface activity and pH, vapour pressure Output Salts; Metallorganics; Pharmaceuticals; Peptides and Protein; Molecules >200 atoms Computerised algorithms Organic compounds Availability or low chemical diversity of datasets No rank chemicals by activity; No 3D information Small molecules; QSPR/QSAR (Optimum Prediction Space (OPS)) 53 Computerised algorithms Defaults and assumptions provided (inflexible) Searchable database Computerised model Molecular structure (SMILES strings or SD files) In formation on the 16 Physical, Chemical, Fate, and Hazard endpoints; Rabbit Skin Irritancy; Eye Irritancy Evaluated with measured data; Validated using "leaveone out" crossvalidation Proprietary 5.1. BfR Decision Support System The German Federal Institute for Risk Assessment (BfR), which is the former Federal Institute for Health Protection of Consumers and Veterinary Medicine (BgVV), has developed a Decision Support System (DSS) for predicting local lesions. It is used to assess chemicals for skin/eye irritation/corrosion potential in support of regulatory decisions (e.g. on classification and labelling). The new substances data within the database were derived by using standardised test protocols and were evaluated by regulators. The system also recommends when a specific alternative test should be used and/or when to conduct an animal test. The DSS provides categorical predictions of hazard. The predictions are based on the physicochemical data of the query chemical, by applying physicochemical exclusion rules, and on the structural formula of the chemical, by applying structural inclusion rules (i.e. SARs).. The system evaluates the irritation/corrosion caused by single contact of a chemical with skin and/or eyes. The DSS predict whether a chemical: a) is likely to cause serious damage or corrosive effects (in which case no testing should be necessary); b) might have corrosive effects (in which case in vitro testing is recommended); or c) is unlikely to produce any effect or only marginal effects (in which case, animal testing may be necessary to confirm the negative prediction). The DSS applies the appropriate EU risk phrase according to EU regulations, providing reliable data for legal classification and labelling. The system can be applied to evaluate organic compounds with structural formulae including only atoms of carbon, hydrogen, oxygen, nitrogen, sulphur, phosphorus, silicon and halogens. Structural formulae of organic salts can also be evaluated. However, inorganic substances, organic chelates, diluted substances, and mixtures cannot be dealt with. A number of papers on the BfR-DSS have been reported by the same group [80],[133]-[139], focusing on the different aspects of the expert system, including the development of a relational database [134], the one-dimensional representation of molecular structures [135], the creation of the rulebase [137] (physicochemical property limit rules [138], and structural alerts [139]), the application of self-learning classifiers [136], and the skin irritation corrosion rules estimation tool (SICRET) [80]. 54 Database. A computerised database was developed from datasets based on toxicological test protocols relating to substance properties responsible for skin and eye irritation/corrosion [134]. The electronic relational database contains physicochemical and toxicological data for more than 1,000 proprietary chemical substances with purity > 95% submitted according to the EU Notification of New Substances procedure and the German Chemical Substances act [136]. Eleven of the physicochemical parameters of this database were used to construct the DSS: specific particular substructures within the structural formula; pH value of a saturated aqueous solution at 20º; molecular weight (MW); aqueous solubility at 20º; hydrolytic properties; lipid solubility at 37º; n-octanol water partition coefficient at 20º; melting point; boiling point at atmospheric pressure; surface tension of a saturated aqueous solution at 20º; and vapour pressure at 20º. The overall toxicological assessments of local corrosive or irritant properties of a chemical substance are represented within the database by risk EU phrases for skin and eye irritancy/corrosivity. The exclusion rules were developed by using only measured values of the physicochemical data, and the structural rules were based on the neighbourhood relations of atoms within the structural formula without any mechanistic hypothesis. Representation of molecular structures. The computer-based method used to analyse and interpret structural formulae characterising chemical molecules from onedimensional representations applies recursive principles and identifies partial isomorphic graphs [135]. Rulebase rationale. In order to develop the DSS that uses information extracted from a database, chemical groups were categorised on the basis of their empirical formulae. Reactive chemical substructures relevant to the formation of local lesions and rules for the prediction of the absence of skin irritation potential were then identified. In the evaluation of physical data, relationships were established between local irritation/corrosion and physicochemical properties. Two types of rules can be clearly differentiated: Physicochemical exclusion rules. The data were examined by applying statistical software and correlations were established between the measured physicochemical properties and the toxicological endpoints of skin/eye irritation/corrosion. The rules were checked against all data in the database. The rules were implemented as exclusion rules of the kind: 55 IF (physicochemical property) A THEN not (toxic) Effect B (Example: IF MW> 1200 g/mol THEN no effect or only marginal effects to the skin are expected) The examination of all the physicochemical parameters led to the definition of physicochemical exclusion rules, derived from the combination of a single physicochemical parameter and a particular substructure that serves to subdivide the total of all chemical substances into different chemical classes. The process of prediction was made reliable by covering most chemicals by more than one exclusion rule; i.e. only the rules which covered at least four chemicals were accepted. More recently, limit values for specific physicochemical properties that are appropriate for identifying chemical substances that have no skin irritation or corrosion potential have been described within clearly defined chemical groups. For skin irritation/corrosion, the physicochemical rules have been updated by evaluating a regulatory (confidential) database compiled with data submitted within German as well as within EU chemicals regulations made of 1833 chemicals [138]. The physicochemical properties include melting point, molecular weight, octanol-water partition coefficient, surface tension, vapour pressure, aqueous solubility and lipid solubility. Physicochemical properties limit values were defined to determine that when the physicochemical properties for a chemical were either greater or smaller than these limits the chemical would have no skin irritation or corrosion potential. The application domains of these limits were constructed to correspond with the EU Rphrases for chemicals classified for skin irritation/corrosion, in order to facilitate classification and labelling. Structural inclusion rules. Substructures that might be responsible for severe damaging effects caused by chemicals to skin and eyes were detected. Structural inclusion rules were derived from correlations between specific local effects and the submitted structural formulae in the form: IF (substructure) A THEN Effect B (Example: IF (phenol) THEN “corrosive”) These rules are statistically and chemically motivated. In order to make the prediction process more reliable, nearly a quarter of the labelled chemicals were covered by more than one exception rule. To ensure the validity of structural rules, they were cross- checked against the whole database. 56 More recently, structural alerts for acute skin lesions were categorised as irritation or corrosion alerts or as a combination of corrosion/irritation alerts according to the mechanisms of action [139]. The rules, which were evaluated for causing skin irritation/corrosion by using 1654 substances, were also connected with physicochemical property limits and their domain of applicability was characterised. Combination of rules. The physicochemical exclusion rules and the structural rules should be used in combination to form a final prediction of the irritant and corrosive properties of a chemical to skin and eyes. To achieve this, the rules were divided into two parts, predicting: the irritant/corrosive properties to skin (R34/R35, R38, and no statement); and the irritant/corrosive properties to eyes (R36, R41, and no statement). The system predicts not only irritant or corrosive properties, but also whether a chemical exhibits no chemical irritation, or whether no prediction can be given if an insufficient result is obtained. The system is designed to avoid conflicting rules. To avoid the conflict of structural rules, the more-severe effect outweighs minor effects. Similarly, the exclusion rules themselves cannot conflict because they are defined to produce a unique result. Classifiers. Procedures for the extension of the DSS for the prediction of the local irritation/corrosion potential of chemicals by using self-learning classifiers were investigated [136]. Different approaches (decision trees, distance examinations in the multidimensional space, and k-nearest neighbour methods) have been implemented, tested and evaluated independently. A combination of all of the established extension approaches was also developed and tested. In contrast to knowledge-based expert systems, self-learning classifiers are automatically computed. Using these classifiers increased significantly the predictive power of the extended DSS; however, automatically calculated results of self-learning classifiers were not able to motivate the toxicological relevance of the results obtained, which was not considered adequate for regulatory hazard assessment purposes. Evaluation. To evaluate the DSS, it was trained with more than 1,000 substances and the resulting DSS prototype was evaluated within the current EU notification procedure by predicting local irritant/corrosive effects of new chemicals not included in the training set [133]. The results of this initial evaluation showed that the majority of the tested substances that cause skin corrosion (66.7%) were predicted by the SARs; a high percentage (20.8%) should be confirmed with a suitable in vitro method, 57 and thus a total of 87.5 % of all skin corrosive substances could be classified without animal testing. The “no statement” rate was 8.3%. For the prediction of serious damage to eyes the prediction rates were similar; 73.7% of substances were correctly predicted to be causing serious eye damage. The no statement rate was 21.1% but the false prediction rate was low (1.8%). Recently, evaluations of the DSS physicochemical exclusion rules according to the OECD principles for (Q)SAR validation were reported. Skin irritation and corrosion physicochemical property limits have been evaluated by RIVM [140], whilst skin irritation and corrosion has been assessed by ECB [141]. Assessment of the structural inclusion rules has also been carried out by ECB [142]-[143]. SICRET. The Skin Irritation Corrosion Rules Estimation Tool (SICRET) [80] is a tiered approach that uses physicochemical property limits, structural alerts and in vitro tests to identify and classify chemicals that cause skin irritation or skin corrosion without animal testing. SICRET uses physicochemical property limits to identify chemicals with no skin corrosion or skin irritation potential; if the exclusion physicochemical rules do not identify the chemicals with no skin corrosion or skin irritation potential, then the structural alerts are used to identify chemicals with skin corrosion or skin irritation potential. If a chemical does not contain structural alerts that indicate it has skin corrosion or skin irritation potential, then in vitro skin corrosion or skin irritation testing is evaluated. If the in vitro test is positive, then the data are included in feedback loops for development of new structural alerts to identify chemicals with skin corrosion or skin irritation potential. If the in vitro test is negative then the data are included in feedback loops for development of new physicochemical property limits to identify chemicals with no skin corrosion or skin irritation potential. SICRET is a tiered approach that it has been proposed to complement the current OECD skin corrosion and skin irritation testing strategy as described in TG 404 for acute dermal irritation/corrosion [39]. A significant difference between SICRET and the strategy described in the OECD TG 404 is that the former only uses information from in vitro tests, whereas the latter also uses in vivo tests information. Although the in vitro corrosion testing proposed by SICRET has been adopted by the OECD member states, in-vitro skin irritation tests have not been yet validated. After the external validation of the physicochemical property limits and the structural alerts, 58 SICRET software will be coded to allow users determining whether a chemical is likely to cause either skin irritation or skin corrosion by providing a probability. 5.2. DEREK The DEREK system, originally referring to the “Deductive Estimation of Risk from Existing Knowledge”, [144]-[148] is a computer and knowledge-based expert system developed by LHASA (Logic and Heuristics Applied to Synthetic Analysis) Ltd and its members at the School of Chemistry, University of Leeds, UK. The software running under Windows is called DEREK for Windows and abbreviated as DEREKfw. DEREK is used for the qualitative prediction of possible toxic action of compounds on the basis of their chemical structure. The system contains a number of structural rules that allow the identification of the structural features of a chemical that may result in the manifestation of toxicity. These structural rules aim to encode structuretoxicity information with an emphasis on mechanisms. It includes 40 human health endpoints, mostly related to mammalian toxicity, i.e. mutagenicity, neurotoxicity, developmental toxicity, oncogenicity, skin sensitisation, and skin and eye irritancy and corrosivity. DEREK has several generic rulebases, based on sets of related chemicals rather than on specific chemicals, and derived from mechanistic organic chemistry. They consist of descriptions of molecular substructures (structural alerts), which have been associated with toxic endpoints on the basis of existing knowledge. To predict toxicity, the program checks whether any alerts within the query structure match previously characterised toxicophores (substructure with potential toxic effect) in the knowledge base. The reasoning engine then assesses the likelihood of a structure being toxic, and a message indicating the nature of the toxicological hazard is provided together with relevant literature references. There are nine levels of confidence: certain, probable, plausible, equivocal, doubted, improbable, impossible, open, contradicted. The reasoning model considers the following information: the toxicological endpoint, the alerts that match toxicophores in the query structure, the physicochemical property values calculated for the query structure, and the presence of an exact match between the query structure and a supporting example within the knowledge base. The presence of several toxicophores in the molecule means that 59 there is a greater risk, but the additivity of the risks is decided by the user. Structures can be drawn directly on a computer graphics terminal or automatically retrieved from an in-house database. The output of the model provides the identification of the toxicophore together with the endpoint prediction and literature references. Some limitations of the model are that physicochemical properties cannot be included, compounds cannot be ranked according to their activity because quantitative information is not available, metabolic aspects are not uniformly realised in the models, and three-dimensional information is not considered. The model cannot evaluate peptides or proteins. The internal performance of the system is biased towards positive structural alerts. This rule-based system is used for research purposes, and to estimate human health hazard. Since DEREK is rule based, it relies on experimental data, and it has been tested and evaluated by several companies using in-house experimental data. Validation exercises have been carried out by using as input a class of specified test structures with a known toxicophore. The criteria for the validation were whether the program was able to detect them. However, external validation exercises have been limited because, as the data are proprietary, the test set can contain chemicals of the training set. Only a few external validations for skin and eye irritation and corrosion potential in DEREK have been published [127],[132],[147]. The DEREK rulebase has nine rules for the prediction of irritancy; two rules are specific to eye irritancy, but none is specific to skin irritancy or corrosivity. To assess the performance of the DEREK rulebase, 300 chemical structures were assessed by using the irritancy information published in the EU classifications for new and existing chemicals, the ECETOC reference data banks for eye and skin irritation, and in vivo data [127]. Of the chemicals labelled as skin and eye irritants, which constituted distinct but overlapping subsets, 23% and 30%, respectively, were predicted to be irritants. Of the new chemicals not labelled as skin or eye irritants, 19 out of 45 were predicted to be irritants. DEREK predicted 20-40% of the corrosive chemicals to be irritants. Thus, probably due to the fact that physicochemical properties are not taken into account, the sensitivity of DEREK for irritant and corrosive effects is low. The external performance of DEREK was assessed by predicting the irritancy of a number of chemicals not included in the DEREK training set [147]. DEREKfW 5.01 60 contains 248 structural fragments (toxicophores) and rules that identify adverse effects; 27 of the 248 toxicophores could be used to assess 54 randomly selected and new notified chemicals for which experimental data were available. DEREK predicted skin irritation for one out of 54 chemicals, but the rationale for selection of the toxicophore was not referenced. The 53 remaining experimental results were compared with structural fragments for corrosion/irritation derived from literature: 17 fragments from the literature predicted corrosion/irritation, while 24 positive test results were found. The Cooper statistics yielded 14% false negatives, 45% specificity, 33% sensitivity; and 8% false positives. The authors, Hulzebos and Posthumus, concluded that DEREK identified irritation poorly, and when structural irritation fragments were included, the number of false negatives decreased. 5.3. HazardExpert HazardExpert is a rule-based software tool developed by CompuDrug (CompuDrug Chemistry Ltd., Hungary) [149] for predicting the toxicity of organic compounds in humans and in animals [150]. HazardExpert predictions are based on known toxic fragments collected from in vivo experiments and reported by the US EPA. The knowledge base was developed based on the list of toxic fragments reported by more than 20 experts, but it can also be used as an open architecture knowledge base, where the user can add new fragments. The input of the rules can be performed by means of a graphical interface, and the toxicity estimation results are also displayed by a graph. It contains multiple toxicity endpoints combined with a capability for estimating toxicokinetic effects, such as bioaccumulation and bioavailability, on the basis of predicted physicochemical values. Human health hazard effects include oncogenicity, mutagenicity, teratogenicity, membrane-irritation, sensitisation, immunotoxicity, and neurotoxicity. The user defines the species, route of administration, dose level, and duration of exposure, and each endpoint is predicted as one of four levels of toxicity, taking into account the effects of bioavailability and bioaccumulation. Substructures which may exert a positive or negative modulatory effect are identified. In addition, several physicochemical parameters are calculated. HazardExpert works by searching the query structure for known toxicophores held in the literature-based knowledge base of toxic fragments. The identification of a toxicophore leads to estimates of the toxicity endpoints by triggering rules in the 61 knowledge-base. The rules describe toxic segments and their effects on various biological systems, and are based on the combined use of toxicological knowledge, expert judgement, QSAR models, and fuzzy logics (which simulates the effects of different exposure conditions). The system is also able to examine potential metabolites of a compound. An evaluation of HazardExpert has been published by Lewis et al. [151]. 5.4. MultiCASE The Computer Automated Structure Evaluation (CASE) program, upgraded to Multiple CASE (MultiCASE) [152]-[157], is an automated rule induction tool, based on an artificial intelligence expert system that is capable of automatically identifying molecular fragments likely to be relevant to the activity of molecules. It is also able to assess the importance of these fragments in relation to the potency of the molecules that contain them, without previous a priori assumptions or knowledge about the mechanism involved in the toxic phenomenon. No pre-classification of the chemicals is required with respect to chemical classes or nature of use. The program is been designed to find structural entities that discriminate between active molecules from inactive ones in non-congeneric datasets. It works by generating all possible substrates of a defined dimension in each compound and finds substructure characteristics of the toxic activity on the basis of the recurrence of the residue within the series of toxic compounds; similarly, inactivating structures are also identified. MultiCASE descriptors are automatically generated in an unbiased way. The molecule of interest is divided into chemical fragments of 2–10 heavy atoms, and a statistical distribution is performed to determine which fragments are biophores or biophobes, according to whether they are associated with the biological activity of interest, or no activity, respectively. This knowledge-based expert system has been used for both scientific and regulatory purposes by the Danish EPA, the US FDA, and Health Canada authorities. The knowledge base consists of a collection of empirically derived heuristics (rules of thumb) relating chemical structural features with the likely areas of toxic action, and an inference engine which is able to identify the structural features present in the target molecule and relate these to the information stored in the knowledge base. The 62 structural commonality between molecules is based on a global as well as local rational evaluation of their substructures rather than on a common substructure. MultiCASE is able to learn or update itself automatically from user data. It selects its own descriptors from the practically infinite number of possible structural assemblies and creates automatically an ad hoc library of fragments. It is able to deal with subtle geometrical differences, with the potency of fragments to cause toxicity (synergies), with logical analysis and hierarchical decision making. MultiCASE can be used to generate SARs and QSARs, in a deterministic or probabilistic way, providing qualitative and quantitative information, ranges and point estimates. The Bayesian MultiCASE has been used in the prediction of physicochemical and environmental fate properties, and in environmental and human health hazard effects, including mutagenicity, oncogenicity, reproductive, developmental and systemic toxicity, and skin and eye irritation [82],[84]. MultiCASE requires a large number of high quality, heterogeneous biological data for the endpoint in question, covering a wide range of activity. The data are then converted by the CASE system into scalar units in the range of 10-99. For the creation and detection of structural alerts, the structure of each molecule is divided up into all possible fragments, starting from two heavy atoms. Statistical methods are then used to classify the fragments as biophores or biophobes. The fragments are then combined to give a QSAR equation. The output of the program provides the predicted activity, and a rationale for conclusions, together with documentation. The model is able to deal mainly with small organic molecules, but the application to salts, organometallics, polymers, and mixtures is limited. The model has been evaluated with measured data, and validated by using a leavemany-out cross validation process for several endpoints [82], [84]. 5.5. TOPKAT TOxicity Prediction by C(K)omputer Assisted Technology (TOPKAT, Accelrys Inc.) [158]-[164] is a PC-based system for the in silico estimation of toxicological properties of compounds from their molecular structure. TOPKAT allows the prediction of physicochemical and environmental fate properties, and environmental and human health hazard effects, including mutagenicity, developmental and systemic 63 toxicity, oncogenicity, and rabbit skin and eye irritation. In particular, the TOPKAT models for skin [61] and eye irritation/corrosion [81] were developed by Enslein and have been updated successively. TOPKAT provides qualitative and quantitative information, and point estimates. TOPKAT QSPR/QSAR models are based on descriptors for electronic structure, connectivity, shape, and substructures. Substructure descriptors are selected from a library of about 3,000 molecular fragments. QSARs are derived using multiple regression equations for continuous endpoints, and two-group linear discriminant regression functions for dichotomous endpoints. It is a modular integrated package, each module consisting of a specific database and several chemical subclass QSAR models, which perform automatic rule induction. The system first scans the database for molecules with substructures overlapping with the query compound, and then it reports the fragments found. The accuracy of the toxicity assessment is estimated by a patented Optimum Prediction Space (OPS) technology that assesses whether or not the query compound is within the space defined by the training set of compounds, and the distance of the query structure from other compounds in the data base (degree of structural coverage). The predictions are based solely on physical chemistry, not biochemical mechanisms. TOPKAT generates a quantitative structure-toxicity relationship model (QSTR) for each toxicological endpoint, using two-dimensional descriptors of electrotopological state values (E-values), symmetry and shape descriptors. These descriptors describe the electronic, shape and symmetry attributes of the molecule. The output of the program is a point estimate of toxicity. The reliability of the estimated toxicity can be assessed by ascertaining whether all of the fragments of the query structure are represented in the database compounds that were used to develop the model, whether the submitted structure fits within, or near the periphery of the Optimum Prediction Space (OPS) of the equation, and whether the model accurately computes the experimental values of the data base compounds most similar to the query structure. In addition to the toxicity point estimate, the output also provides additional information that allows evaluation of the reliability of the prediction, including: fragment coverage of the dataset used to calculate the model with respect to the query compound presented for evaluation; query structure relationship to the Optimum Prediction Space as defined by the training dataset; probability scores which 64 contribute to identification of a degree of accuracy for the prediction; and literature references. Even if a considerable variety of structures have been included in each model in order to maximise the chemical space covered by the models, the limitations of the model are inevitably defined by the contents of the dataset used for the model development. For example, the program has some limitations in dealing with organometallics and inorganics. 5.6. Other Expert Systems 5.6.1. Cache The Computer-Aided Chemistry (CAChe) program [165]-[167] is a software tool developed by Fujitsu for creating new models to predict a number of physical, chemical, and biological properties, and for generating a wide range of descriptors for QSAR including quantum chemistry descriptors. CAChe includes the quantitative prediction of physicochemical and environmental fate properties, and environmental and human health hazard effects. CAChe is able to deal with either organic or inorganic molecules, and predicts both structure and properties. The input structures can be provided from X-ray data or in the form of standard database formats. It is able to dock molecules in reaction sites, and to perform conformational searches to identify structure-energy relationships. CAChe can provide a visualisation of reactive sites on molecules, and UV-Visible and IR spectra, which allow locating transition states and reaction intermediates. CAChe also includes tools for regression analysis, and the output information for the QSAR models includes correlation coefficients and cross validated statistics. One limitation of the model is that it is based on calibration data and statistical analysis results, and it is unable to deal with any organic or inorganic chemical containing elements up to Lawrencium. 5.6.2. SuCCSES The Substructure-Based Computerized Chemical Selection Expert System (SuCCSES) [168] was developed to facilitate the US EPA review of large groups of chemicals with similar substructures and modes of action. SuCCSES has been 65 described previously [57],[169]-[173]. SuCCSES was developed based on historical information and expert opinions. Historical information was obtained from the US EPA’s InterAgency Testing Committee (ITC) scoring exercises performed from 1978 to 1983. For health effects, numerous international experts were sent a questionnaire listing more than 100 different chemical substructures and were asked to predict (based on their toxicological expertise and knowledge of modes or mechanisms of action) the potential for chemicals containing any of the substructures to cause acute, chronic, mutagenic, carcinogenic, developmental, reproductive, or neurotoxic effects or membrane irritation. Opinions from these health effects experts were converted to codes that identified chemical substructures and indicated the potential of chemicals containing one or more substructures to cause specific health effects. The substructures in SuCCSES that were associated with membrane irritation were included in [58]. SuCCSES is applied in accordance with the ITC's statutory mandate to use SARs before the US EPA recommends chemicals for testing. SuCCSES is not publicly available because it contains confidential business information. 5.6.3. OASIS OASIS is a database platform developed by the Laboratory of Mathematical Chemistry of University of Bourgas for developing three-dimensional QSAR (3DQSAR) models [174]-[176]. The system provides selection of active conformers as 3D structure representatives, including options for conformation multiplication, and quantum-chemical optimisation. The system also assesses chemical similarity, taking account of the flexibility of chemical structures. Calculated steric, electronic and physicochemical descriptors of selected conformers can be calculated. The database stores a number of (Q)SAR models on environmental fate properties and environmental and human health effects, including the MultiCASE model for severe skin irritation. The output of the program assesses if the (Q)SAR predictions are inside or outside the applicability domain of the model. 5.7. Integrated Testing Exercise An example of application of expert systems to the prediction of skin and eye irritation for the safety evaluation of new and existing chemicals has been reported in an integrated testing exercise [131]. The integrated approach consisted of the use of 66 existing non-animal data and the generation of alternative data by using in silico and in vitro methods. These methods were applied to ten substances listed on ELINCS and associated with a complete base-set. The hazard assessment for these substances was performed on the basis of available non-animal data, QSAR, PBBK-modelling and additional in vitro testing was applied to the following endpoints: acute fish toxicity, skin and eye irritation, acute oral, dermal and inhalatory toxicity, sensitisation, and systemic toxicity after repeated dosing. The predictions were compared with the outcome of the in vivo tests. The methods used for this endpoint were (Q)SARs using DEREKfw (v7.0) for the qualitative estimation, and TOPKAT (v6.1) for the quantitative prediction; and the in vitro BECAM assay. Skin and eye irritation predictions were based on (Q)SARs and the outcome of the BECAM test in a weight of evidence approach, assuming that any substance not irritating to the eyes was not irritating to the skin. After the predictions were made by the integrated method, they were made available to four expert groups, to check the bias of predictions. Experts associated a reliability score to their own predictions (from 0 to 5 in order of increasing reliability), and the scores were combined to calculate and overall prediction value. However, no record on the method of assessment of the endpoints by the experts was made. After the results of the in vivo tests for the 10 substances were released, two comparisons of the predicted values with the in vivo outcome were made: a direct comparison with the in vitro and in vivo predictions and a comparison with the overall prediction value from the expert predictions. When a predicted value (or predicted range) would lead to the same classification according to the in vivo data, a prediction was considered to be correct. Skin and eye irritation were correctly predicted in 70% of the cases, with one false negative for both endpoints. However, experts reached 9 correct predictions in skin irritation and 8 correct predictions in eye irritation. For this endpoint, a test in one animal would be sufficient to confirm the results of the QSAR and in vitro assay (only if negative) and to prevent false negatives. This approach is similar to that promoted by the current OECD guidelines for skin and eye irritation. 67 6. Conclusions In this report, we have provided a detailed review of alternative approaches for assessing skin and eye irritation/corrosion based on the use of in silico methods (SARs, QSARs and expert systems). Literature-based models tend to be applicable to specific groups of chemicals, whereas expert systems tend to have broader applicability domains. Some of these systems are already well characterised (e.g. the BfR rulebases and the DEREK system) and may therefore be suitable for regulatory use. In general, we conclude that the further development, validation and documentation of in silico systems for local toxicity to the skin and eye are necessary. A considerable effort to promote the availability of valid (Q)SARs is already being carried out by the ECB [45]-[46]. This includes the need to present information on the characteristics of the models in a transparent way. At present, so-called “QSAR Model Reporting Formats” (QMRFs) are being developed for this purpose, and are being made available from the ECB website. The most effective approach is to integrate all appropriate information to make a weight-of-evidence-based assessment of the chemical hazard and risk. Integrated Testing Strategies combine all possible sources of information from (Q)SARs, expert systems, read-across and other grouping approaches, and test methods (especially in vitro tests). Current work within the context of one of the REACH-Implementation Projects (RIP 3.3-2) is developing tiered approaches for the assessment of skin and eye irritation/corrosion potential As increasing experience is gained in the regulatory use of (Q)SARs and expert systems, it is expected that specific research needs will be identified concerning the development of new, tailor-made models, specifically designed for inclusion within testing strategies. 68 7. References [1] Draize, J.H., Woodard, G., Calvery, H.O. (1944) Methods for the study of irritation and toxicity of substances applied topically to the skin and mucous membranes. Journal of Pharmacology and experimental Therapeutics 82, 377-390. [2] European Economic Community (1967) Council Directive 67/548/EEC of 27 June 1967 on the approximation of laws, regulations and administrative provisions relating to the classification, packaging and labelling of dangerous substances. Official Journal of European Communities 196, 16/08/1967. Latest updates: 9th Amendment by Directive 1999/33/EC; Adapting to Technical Progress for the 29th time by Directive 2004/73/EC. [3] European Economic Community (1993) Council Regulation (EEC) No 793/93 of 23 March 1993 on the evaluation and control of the risks of existing substances. Official Journal of European Communities L 84, 5/4/1993. [4] Commission of the European Communities (2005) Commission Regulation (EC) No 642/2005 of 27 April 2005 imposing testing and information requirements on the importers or manufacturers of certain priority substances in accordance with Council Regulation (EEC) No 793/93 on the evaluation and control of the risks of existing substancesText with EEA relevance. Official Journal of European Communities L 107, 28/04/2005. [5] The Complete REACH documentation. http://europa.eu.int/comm/enterprise/reach/overview.htm [6] Commission of the European Communities (2002) White Paper on the Strategy for a future Chemicals Policy, COM(2001)88. European Commission, Brussels, Belgium. Available at: http://europa.eu.int/comm/environment/chemicals/whitepaper.htm [7] Commission of the European Communities (2003) Proposal for a Regulation of the European Parliament and of the Council concerning the Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH), establishing a European Chemicals Agency and amending Directive 1999/45/EC and Regulation (EC) {on Persistent Organic Pollutants}. Available at: http://europa.eu.int/comm/enterprise/chemicals/chempol/whitepaper/reach.htm [8] European Economic Community (1986) Council Directive of 24 November 1986 on the approximation of laws, regulations and administrative provisions of the Member 69 States regarding the protection of animals used for experimental and other scientific purposes (86/609/EEC). Official Journal of European Communities L 358, 18.12.1986, pp. 1-32. [9] TAPIR (2005). Three point three – A Project for the Information Requirements of REACH. Lot 3: Scoping study on the Development of a Technical Guidance Document on Information requirements on Intrinsic Properties of Substances (RIP 3.3-1). Final TAPIR Report RIP 3.3-1. Submitted to European Chemical Bureau, Joint Research Centre, European Commission under Contract No. – 22553-2004-12 F1SC ISP BE. [10] TAPIR (2005) Final report appendix 8, Report of Information Working Group 2 “Testing” (IWG-2). REACH Implementation Project 3.3-1. [11] TAPIR (2005) Final report appendix 9, Report of Endpoint Working Group 3 “No Testing Considerations”. REACH Implementation Project 3.3-1. [12] TAPIR Final report appendix 12 (2005) Report of Endpoint Working Group 4 “Irritation / Corrosion”. REACH Implementation Project 3.3-1. [13] Pedersen, F., de Bruijn, J., Munn, S., van Leeuwen, K. (2003) Report “Assessment of additional testing needs under REACH. Effects of (Q)SARS, risk based testing and voluntary industry initiatives”. Ispra, Italy: European Commission, Joint Research Centre. [14] van der Jagt, K., Munn, S., Tørsløv, J., de Bruijn, J. (2004). Addendum to the report “Assessment of additional testing needs under REACH. Effects of (Q)SARS, risk based testing and voluntary industry initiatives”. EUR 21405 EN. Ispra, Italy: European Commission, Joint Research Centre. [15] Walker A.P. (1990): A proposal for the harmonisation of classification schemes for chemicals. Contract No. B6680/90/6065 for the Commision of the European Communities, DG XI-A-2. [16] Organisation for Economic Co-operation and Development (1996). Final Report of the OECD Workshop on Harmonisation of Validation and Acceptance Criteria for Alternative Toxicological Test Methods. OECD, Paris, France. Held in Solna, Sweden, 22-24 January 1996, 62pp. Available at: http://www.oecd.org/ehs/test/background.htm [17] Organisation for Economic Co-operation and Development (1998) Harmonized Integrated Hazard Classification System for Human Health and Environmental 70 Effects of Chemical Substances, as endorsed by the 28th Joint Meeting of the Chemicals Committee and the Working Party on Chemicals, November 1998. Available at: http://www.oecd.org/ehs/Class/HCL6.htm [18] Organisation for Economic Co-operation and Development (1998). Revised Proposal for the Harmonisation of Hazard Classification Based on Eye Irritation/Corrosion. ENV/MC/CHEM/HCL(98)5, 10pp. OECD, Paris, France. [19] Organisation for Economic Co-operation and Development (1998). Harmonised Integrated Classification System for Human Health and Environmental Hazards of Chemical Substances and Mixtures, ENV/JM/HCL(2001)1, OECD Paris, France. Available at: http://www.oecd.org/ehs/Class/hcl/finaw.pdf [20] Organisation for Economic Co-operation and Development (2001). Series on Testing and Assessment No.33: Harmonised Integrated Classification System for Human Health and Environmental Hazards of Chemical Substances and Mixtures ENV/JM/MONO(2001)6. Available at: http://www.olis.oecd.org/olis/2001doc.nsf/LinkTo/env-jm-mono(2001)6 [21] United Nations (2001). Report of the first session of the Sub-Committee of Experts on the Globally Harmonized System of Classification and Labelling of Chemicals ST/SG/AC.10/C.4/2001/4 [22] United Nations (2003). Globally Harmonized System of Classification and Labelling of Chemicals (GHS). Part 3. Health and Environmental Hazards, 122pp. New York, NY, USA, and Geneva, Switzerland: United Nations. Available at: http:// www.unece.org/trans/danger/publi/ghs [23] European Economic Community (1976) Council Directive 76/768/EEC of 27 July 1976 on the approximation of the laws of the Member states relating to cosmetic products. Official Journal of European Communities L 262, 27/9/1976, p. 169. Update: 7th amendment by Directive 2003/15/EC of the European Parliament and of the Council of 27 February 2003; Commission Directive 2005/80/EC of 21 November 2005 amending Council Directive 76/768/EEC, concerning cosmetic products, for the purposes of adapting Annexes II and III thereto to technical progress. Official Journal of European Communities L 303, 22/11/2005 pp. 0032 – 0037. [24] ECVAM (2004) Report for establishing the timetable for phasing out animal testing for the purpose of the cosmetics directive. 71 [25] SCCNFP/0690/03, Final: The Scientific Committee on Cosmetic Products and NonFood Products Intended for Consumers Notes of Guidance For Testing of Cosmetic Ingredients For Their Safety Evaluation. 5th Revision, adopted by the SCCNFP during the plenary meeting of 20 October 2003. [26] http://www.hse.gov.uk/chip/phrases.htm [27] HSC (1997) Approved Guide to the Classification and Labelling of Substances and Preparations Dangerous for Supply (Third Edition) Chemicals (Hazard Information and Packaging for Supply) (Amendment) Regulations 1997 (L100), Sudbury, UK, HSE Books. [28] European Economic Community (1992) Annex V-Methods for the determination of toxicity. B.3. Acute toxicity (Dermal). Official Journal of European Communities L383 A 35. [29] European Economic Community (1992) Annex V-Methods for the determination of toxicity. B.4. Acute toxicity (Skin Irritation). Official Journal of European Communities L383 A 35. [30] European Economic Community (1992) Annex V-Methods for the determination of toxicity. B.5. Acute toxicity (Eye Irritation). Official Journal of European Communities L383 A 35. [31] Commission of the European Communities (2000) Directive 2000/33/EC of 25 April 2000 Adapting to Technical Progress for the 27th time the Council Directive 67/548/EEC on the Approximation of Laws, Regulations and Administrative Provisions Relating to the Classification, Packaging and Labelling of Dangerous Substances. Annex 1. B.40. Skin corrosion. Official Journal of European Communities L 136. [32] Organisation for Economic Co-operation and Development (1981). Guideline for Testing of Chemicals No. 404: Acute dermal irritation/corrosion. OECD, Paris, France, adopted on 12 May 1981, revised in 1992. Updated Guideline, adopted on 24 April 2002. [33] Organisation for Economic Co-operation and Development (1981). Guideline for Testing of Chemicals No. 405: Acute Eye Irritation/Corrosion. OECD, Paris, France, adopted on 12 May 1981, revised on 24 February 1987. Updated Guideline, adopted on 24 April 2002. 72 [34] Organisation for Economic Co-operation and Development (2004) Guideline for Testing of Chemicals 430 In Vitro Skin Corrosion: Transcutaneous Electrical Resistance Test (TER). OECD, Paris, France, adopted on 13 April 2004. [35] Organisation for Economic Co-operation and Development (2004) Guideline for Testing of Chemicals 431 In Vitro Skin Corrosion: Human Skin Model Test. OECD, Paris, France, adopted on 13 April 2004. [36] Organisation for Economic Co-operation and Development (2001) Brief Chronology of the Draft Test Guidelines 404: Acute Dermal Irritation/Corrosion, 405: Acute Eye Irritation/Corrosion, and 429: Skin Sensitisation: Local Lymph Node Assay. [37] Organisation for Economic Co-operation and Development (1999). Series on Testing and Assessment No.14. Detailed Review Document on Classification Systems for Eye Irritation/Corrosion in OECD Member Countries. ENV/JM/MONO(99)4. Available at: http://www.olis.oecd.org/olis/1999doc.nsf/LinkTo/env-jm-mono(99)4 [38] Organisation for Economic Co-operation and Development (1999). Series on Testing and Assessment No.16: Detailed Review Document on Classification Systems for Skin Irritation/Corrosion in OECD Member Countries ENV/JM/MONO(99)6. Available at: http://www.olis.oecd.org/olis/1999doc.nsf/LinkTo/env-jm-mono(99)6 [39] Organisation for Economic Co-operation and Development (2001) OECD Attachment to the Test Guideline 404: A Sequential Testing Strategy for Skin Irritation and Corrosion. OECD, Paris, France. [40] Organisation for Economic Co-operation and Development (2001) OECD Attachment to the Test Guideline 405: A Sequential Testing Strategy for Eye Irritation and Corrosion. OECD, Paris, France. [41] Organisation for Economic Co-operation and Development (2001) Environment Directorate, Joint meeting of the Chemicals Committee and the working party on chemicals, pesticides and biotechnology. Test Guidelines Programme. Revised proposals for updated test guidelines 404 and 405: dermal and eye corrosion/ irritation studies. ENV/JM/TG (2001) 2, 112 pp, Paris, France, OECD. [42] Organisation for Economic Co-operation and Development (2001) OECD Guideline for Testing of Chemicals – Draft Revised Guideline 404: Acute Dermal Irritation/Corrosion. OECD, Paris, France. 73 [43] Organisation for Economic Co-operation and Development (2001) OECD Guideline for Testing of Chemicals—Draft Revised Guideline 405: Acute Eye Irritation/Corrosion. OECD, Paris, France. [44] European Centre for Ecotoxicology and Toxicology of Chemicals (1988) Eye Irritation Testing. Monograph No. 11. ECETOC, Brussels, Belgium. [45] Organisation for Economic Co-operation and Development (2004) “The Report From The Expert Group On (Quantitative) Structure-Activity Relationships [(Q)Sars] On The Principles For The Validation Of (Q)SARs” 2nd Meeting of the ad hoc Expert Group on QSARs OECD Headquarters, 20-21 September, 2004. [46] Worth, A.P., Bassan, A., Gallegos, A., Netzeva, T.I., Patlewicz, G., Pavan, M., Tsakovska, I., Vracko, M. (2005) The Characterisation of (Quantitative) StructureActivity Relationships: Preliminary Guidance. JRC Technical EUR Report. EUR 21866 EN, 95pp. Ispra, Italy: European Commission, Joint Research Centre. [47] Hinz R.S., Lorence. C.R., Hodson C.D., Hansch C., Hall L.L. and Guy R.H. (1991) Percutaneous penetration of para-substituted phenols in vitro. Fundamental and Applied Toxicology 17, 575-583. [48] Young, J.R., How, M.J., Walker, A.P., Worth W.M.H. (1988) Classification as Corrosive or Irritant to Skin of Preparations Containing Acidic or Alkaline Substance Without Testing on Animals. Toxicol. In vitro, 2, 19 - 26. [49] Worth, A.P. (2000) The integrated use of physicochemical and in vitro data for predicting chemical toxicity. PhD thesis. Liverpool John Moores University, Liverpool, UK. [50] Barratt, M.D. (2000) Prediction of toxicity from chemical structure. Cell Biology and Toxicology 16, 1-13. [51] Barratt, M.D., Rodford, R. (2001) The computational prediction of toxicity. Current Opinion in Chemical Biology 5, 383-388. [52] European Chemical Industry Council (2002) Report of the CEFIC Workshop on Regulatory Acceptance of QSARs for Human Health and the Environment. Held in Setubal, Portugal, 4-6 March, 2002, pp. 60. [53] Schultz, T.W., Cronin, M.T.D., Walker, J.D., Aptula, A.O. (2003) Quantitative structure–activity relationships (QSARs) in toxicology: a historical perspective. Journal of Molecular Structure (Theochem) 622, 1-22. 74 [54] Schultz, T.W., Cronin, M.T.D., Netzeva, T.I. (2003) The present status of QSAR in toxicology. Journal of Molecular Structure (Theochem) 622, 23-38. [55] Cronin, M.T.D., Dearden, J.C., Walker, J.D., Worth, A.P. (2003) Quantitative Structure-Activity Relationships for Human Health Effects: Commonalities with other Endpoints. Environmental Toxicology and Chemistry 22, 1829-1843. [56] Patlewicz, G., Rodford, R., Walker, J.D. (2003) Quantitative Structure-Activity Relationships for Predicting Skin and Eye Irritation. Environmental Toxicology and Chemistry 22, 1862-1869. [57] Cronin, M.T.D., Jaworska, J.S., Walker, J.D., Comber, M.H.I., Watts, C.D., Worth, A.P. (2003) Use of QSARs in International Decision-Making Frameworks to Predict Health Effects of Chemical Substances. Environmental Health Perspectives 111(10), 1391-1401. [58] Hulzebos, E.M., Maslankiewicz, L., Walker, J.D. (2003) Verification of literaturederived SARs for skin irritation and corrosion. QSAR & Combinatorial Science 22, 351-363. [59] Gerner, I., Spielmann, H., Hoefer, T., Liebsch, M., Herzler, M. (2004). Regulatory use of (Q)SARs in Toxicological Hazard Assessment Strategies. SAR and QSAR in Environmental Research 15(5-6), 359-366. [60] Walker, J.D., Gerner, I., Hulzebos, E., Schlegel, K. (2004) (Q)SARs for Predicting Skin Irritation and Corrosion: Mechanisms, Transparency and Applicability of Predictions. QSAR & Combinatorial Science 23, 721-725. [61] Enslein, K., Borgstedt, H.H., Blake, B.W., Hart, J.B. (1987) Prediction of rabbit skin irritation severity by structure activity relationships. In vitro Toxicology 1, 129-147. [62] Berner, B., Wilson, D.R., Guy, R.H., Mazzenga, G.C., Clarke, F.H., Maibach, H.I. (1988) The relationship of pKa and acute skin irritation in man. Pharmaceutical Research 5, 660-663. [63] Berner, B., Wilson, D.R., Steffens, R.J., Mazzenga, G.C., Hinz, R., Guy, R.H., Maibach, H.I. (1990) The relationship between pKa and acute skin irritation for a series of basic penetrants in man. Fundamental and Applied Pharmacology 15, 760– 766. [64] Berner, B., Wilson, D.R., Guy, R.H., Mazzenga, G.C., Clarke, F.H., Maibach, H.I. (1989/1990) Relationship of pKa and acute skin irritation in humans. Journal of Toxicology-Cutaneous and Ocular Toxicology 9, 481–492. 75 [65] Nangia, A., Anderson, P.H., Berner, B., Maibach, H.I. (1996). High dissociation constants (pKa) of baasic permeants are associated with in vivo skin irritation in man. Contact Dermatitis i34, 237-242. [66] Hayashi, M., Nakamura, Y., Higashi, K., Kato, H., Kishida, F., Kaneko, H. (1999) A Quantitative Structure-Activity Relationship Study of the Skin Irritation Potential of Phenols. Toxicology in vitro 13, 915-922. [67] Smith, J.S., Macina, O.T., Sussman, N.B., Luster, M.I., Karol, M.H. (2000) A Robust Structure-Activity Relationship (SAR) Model for Esters that Cause Skin Irritation in Humans. Toxicological Sciences 55, 215-222. [68] Smith, J.S., Macina, O.T., Sussman, N.B., Karol, M.H. Maibach, H.I. (2000) Experimental validation of a structure-activity relationship model of skin irritation by esters. Quantitative Structure-Activity Relationships 19, 467–474. [69] Kodithala, K., Hopfinger, A.J., Thompson, E.D., Robinson, M.K. (2002) Prediction of Skin Irritation from Organic Chemicals Using Membrane-Interaction QSAR Analysis. Toxicological Sciences 66, 336-346. [70] Barratt, M.D. (1995) Quantitative structure activity relationships for skin corrosivity of organic acids, bases and phenols. Toxicology Letters 75, 169-176. [71] Botham, P., Chamberlain, M., Barratt, M. D., Curren, R.D., Esdaile, D.J., Gardner, J.R., Gordon, V.C., Hildebrand, B., Lewis, R.D., Liebsch, M., Logemann, P., Osborne, R., Ponec, M., Régnier, J.-F., Steiling, W., Walker, A.P., Balls, M. (1995) A Prevalidation Study on In vitro Skin Corrosivity Testing. The Report and Recommendations of ECVAM Workshop 6. ATLA – Alternatives to Laboratory Animals 23, 219-255. [72] Barratt, M.D. (1996) Quantitative Structure-Activity Relationships for Skin Irritation and Corrosivity of Neutral and Electrophilic Organic Chemicals. Toxicology in vitro 10, 247-256. [73] Whittle, E., Barratt, M.D., Carter, J.A., Basketter, D.A., Chamberlain, M. (1996) Skin Corrosivity Potential of Fatty Acids: In vitro Rat and Human Skin Testing and QSAR Studies. Toxicology in vitro 10, 95-100. [74] Barratt, M.D. (1996) Quantitative Structure-Activity Relationships for Skin Corrosivity of Organic Acids, Bases and Phenols: Principal Components and Neural Network Analysis of Extended Sets. Toxicology in vitro 10, 85-94. 76 [75] Barratt, M.D., Dixit, M.B., Jones, P.A. (1996) The Use of In vitro Cytotoxicity Measurements in QSAR Methods for the Prediction of the Skin Corrosivity Potential of Acids. Toxicology in vitro 10, 283-290. [76] Worth, A.P., Cronin, M.T.D. (2001) The Use of Bootstrap Resampling to Assess the Uncertainty of Cooper Statistics. ATLA – Alternatives to Laboratory Animals 29, 447-459. [77] Worth, A.P., Cronin, M.T.D. (2001) The use of bootstrap resampling to assess the variability of Draize tissue scores. ATLA – Alternatives to Laboratory Animals 29, 557-573. [78] Worth, A.P., Cronin, M.T.D. (2001). The use of pH measurements to predict the potential of chemicals to cause acute dermal and ocular toxicity. Toxicology 169, 119-131. [79] Worth, A.P., Fentem, J., Balls, M., Botham, P.A., Curren, R.D., Earl, L.K., Esdaile, D.J., Liebsch, M. (1998) An Evaluation of the Proposed OECD Testing Strategy for Skin Corrosion. ATLA – Alternatives to Laboratory Animals 26, 709-720. [80] Walker, J.D., Gerner, I., Hulzebos, E., Schlegel, K. (2005) The Skin Irritation Corrosion Rules Estimation Tool (SICRET). QSAR & Combinatorial Science 24, 378-384. [81] Enslein, K., Blake, B.W., Tuzzeo, T.M., Borgstedt, H.H., Hart, J.B., Salem, H. (1988) Estimation of rabbit eye irritation scores by structure activity equations. In vitro Toxicology 2, 1-14. [82] Klopman, G., Ptchelintsev, D., Frierson, M., Pennisi, S., Renskers, K., Dickens, M. (1993) Multiple computer automated structure evaluation methodology as an alternative to in vitro eye irritation testing. ATLA – Alternatives to Laboratory Animals 21, 14-27. [83] Klopman, G. (1998) The MultiCASE Program II. Baseline Activity Identification Algorithm (BAIA). Journal of Chemical Information and Computer Sciences 38, 7881. [84] Rosenkranz, H.S., Zhang, Y.P., Klopman, G. (1998) The Development and Characterisation of a Structure-activity Relationship Model of the Draize Eye Irritation Test. ATLA – Alternatives to Laboratory Animals 26, 779-809. [85] Sugai, S, Murata, K, Kitagaki, T, Tomita, I. (1990) Studies on eye irritation caused by chemicals in rabbits—1. A quantitative structure-activity relationships approach to 77 primary eye irritation of chemicals in rabbits. The Journal of Toxicological Sciences 15, 245-262. [86] Sugai, S., Murata, K., Kitagaki, T., Tomita, I. (1991) Studies on eye irritation caused by chemicals in rabbits—II. Structure-activity relationships and in vitro approach to primary eye irritation of salicylates in rabbits. The Journal of Toxicological Sciences 16, 111-130. [87] Lipnick, R.L., Bentley, J.L., Bentley, D.L. (1992) QSAR analysis of ocular toxicity. Abstracts, The 5th International Workshop on QSAR in Environmental Toxicology, Duluth, MN, USA. [88] Cronin, M.T.D., Basketter, D.A., York, M. (1994) A quantitative structure- activity relationship (QSAR) investigation of a Draize eye irritation database. Toxicology in vitro 8, 21-28. [89] Cronin, M.T.D. (1996) The use of cluster significance analysis to identify asymmetric QSAR datasets in toxicology. An example with eye irritation data. SAR and QSAR in Environmental Research 5, 167-175. [90] Worth, A.P., Cronin, M.T.D. (2000) Embedded cluster modelling: a novel quantitative structure-activity relationhip for generating elliptic models of biological activity. In, Progress in the Reduction, Refinement and Replacement of Animal Experimentation. Balls, M., van Zeller, A.-M., Halder, M.E. (Eds.) Elsevier Science, Amsterdam, The Netherlands, pp. 479-491. [91] Worth, A.P., Cronin, M.T.D. (2003) The use of discriminant analysis, logistic regression and classification tree analysis in the development of classification models for human health effects. Journal of Molecular Structure (Theochem) 622, 97-111. [92] Barratt, M.D. (1995) A quantitative structure-activity relationship for the eye irritation potential neutral organic chemicals. Toxicology Letters 80, 69-74. [93] Barratt, M.D. (1997) QSARS for the Eye Irritation Potential of Neutral Organic Chemicals. Toxicology in vitro 11, 1-8. [94] Abraham, M.H., Kumarsingh, R., Cometto-Muñiz, J.E., Cain, W.S. (1998) A Quantitative Structure-Activity Relationship (QSAR) for a Draize Eye Irritation Database. Toxicology in vitro 12, 201-207. [95] Abraham, M.H., Kumarsingh, R., Cometto-Muñiz, J.E., Cain, W.S. (1998) Draize Eye Scores and Eye Irritation Thresholds in Man Combined into one Quantitative Structure-Activity Relationship. Toxicology in vitro 12, 403-408. 78 [96] Abraham, M.H., Kumarsingh, R., Cometto-Muñiz, J.E., Cain, W.S. (1998) Draize eye scores and eye irritation thresholds in man can be combined into one QSAR. Olfaction and Taste XII 855, 652-656. [97] Abraham, M.H, Hassanisadi, M., Jalali-Heravi, M., Ghafourian, T., Cain, W.S., Cometto-Muñiz, J.E. (2003). Draize Rabbit Eye Test Compatibility with Eye Irritation Thresholds in Humans: A Quantitative Structure-Activity Relationship Analysis. Toxicological Sciences 76, 384-391. [98] Kulkarni, A.S., Hopfinger, A.J. (1999). Membrane-interaction QSAR analysis: application to the estimation of eye irritation by organic compounds. Pharmaceutical Research 16, 1245-1253. [99] Kulkarni, A., Hopfinger, A.J., Osborne, R., Bruner, L.H., Thompson, E.D. (2001) Prediction of Eye Irritation from Organic Chemicals Using Membrane-Interaction QSAR Analysis. Toxicological Sciences 59, 335-345. [100] Li, Y., Liu, J., Pan, D., Hopfinger, A.J., (2005) A study of the relationship between cornea permeability and eye irritation using membrane-interaction QSAR analysis. Toxicological Sciences 88, 434-446. [101] Patel, H.C., Duca, J.S., Hopfinger, A.J., Glendening, C.D., Thompson, E.D. (1999) Quantitative component analysis of mixtures for risk assessment: Application to eye irritation. Chemical Research in Toxicology 12, 1050–1056. [102] Patlewicz, G.Y., Rodford, R.A., Ellis, G., BArratt, M.D. (2000) A QSAR Model for the Eye Irritation of Cationic Surfactants. Toxicology in vitro 14, 79-84. [103] Hayashi, T., Itagaki, H., Fukuda, T., Tamura, U., Kato, S. (1993) Quantitative structure-activity relationship of surfactants on eye irritation predicted by hemoglobin denaturation. Alternatives to Animal Test Experiments 2, 49-55. [104] Patlewicz GY, El-Deredy W. 1999. Identification of molecular properties of cationic surfactants causing eye irritation: A QSAR Bayesian neural network model. Proceedings, Predictive Toxicology of Chemicals: Experiences and Impact of AI Tools AAAI Spring Symposium, Menlo Park, CA, USA, March 22–24, pp 123-127. [105] European Parliament resolution on Council Directive 86/609/EEC on the protection of animals used for experimental and other scientific purposes (2001/2259(INI)). Official Journal of European Communities C 27 E, 30/01/2004, pp. 0151-0155. [106] Balls, M., De Klerck, W., Baker, F., van Beek, M., Bouillon, C., Bruner, L., Carstensen, J., Chamberlain, M., Cottin, M., Curren, R., Dupuis, J., Fairweather, F., 79 Faure, U., Fentem, J., Fisher, C., Galli, C., Kemper, F., Knaap, A., Langley, G., Loprieno, G., Loprieno, N., Pape, W., Pechovitch, G., Spielmann, H., Ungar, K., White, I., Zuang, V. (1995) Development and Validation of Non-Animal Tests and Testing Strategies: the Identification of a Coordinated Response to the Challenge and the Opportunity Presented by the Sixth Amendment to the Cosmetics Directive (76/768/EEC). ATLA – Alternatives To Laboratory Animals 23, 398-409. [107] Barratt, M.D., Castell, J.V., Chamberlain, M., Combes, R.D., Dearden, J.C., Fentem, J.H., Gerner, I., Giuliani, A., Gray, T.J.B., Livingstone, D.J., Provan, W.M., Rutten, F.A.J.J.L., Verhaar, H.J.M., Zbinden, P. (1995) The Integrated Use of Alternative Approaches for Predicting Toxic Hazard. Alternatives To Laboratory Animals 23, 410-429. [108] Liebsch, M., Spielmann, H. (2002) Currently available in vitro methods used in regulatory toxicology. Toxicology Letters 127, 127-134. [109] van de Sandt, J., Roguet, R., Cohen, C., Esdaile, D., Ponec, M., Corsini, E., Barker, C., Fusenig, N., Liebsch, M., Benford, D., de Brugerolle de Fraissinette, A., Fartasch, M. (1999) The Use of Human Keratinocytes and Human Skin Models for Predicting Skin Irritation. The Report and Recommendations of ECVAM Workshop 38. ATLA – Alternatives to Laboratory Animals 27, 723-743. [110] Harbell, J.W., Curren, R.D. (2001) In vitro methods for the prediction of ocular and dermal toxicity. In, Handbook of Toxicology, 2nd Ed. M.J. Derelanko and M.A. Hollinger ED., CRC Press, Boca Raton, pp. 835-866. [111] Fentem, J., Botham, P.A. (2004) Update on the Validation and Regulatory Acceptance of Alternative Tests for Skin Corrosion and Irritation. ATLA – Alternatives to Laboratory Animals 32 (Supplement 1), 29-40. [112] Zuang, V., Alonso, M.A., Botham, P.A., Eskes, C., Fentem, J., Liebsch, M., van de Sandt, J.J.M. (2005) ECVAM Report: Subgroup 2. Skin Irritation/Corrosion. ATLA – Alternatives to Laboratory Animals 33 (Supplement 1), 35-46. [113] Barratt M.D., Brantom P.G., Fentem J.H., Gerner I., Walker, J.D., Worth A.P. (1998) The ECVAM International validation study on In vitro Tests for Skin Corrosivity. I. Selection and distribution of the test chemicals. Toxicology in vitro 12, 471-482. [114] Fentem J.H., Archer G.E.B., Balls M., Botham P.A., Curren R.D., Earl L.K., Esdaile D.J., Holzhütter H.G., Liebsch M. (1998) The ECVAM international validation study on in vitro tests for skin corrosivity. II. Results and evaluation by the Management Team. Toxicology in vitro 12, 483-524. 80 [115] Balls, M., Botham,PA., Bruner, LH., Spielmann,H. (1995) The EC/HO international validation study on alternatives to the Draize eye irritation test. Toxicology in vitro 9, 871-929. [116] Brantom, P.G., Bruner, L.H., Chamberlain, M., Desilva, O., Dupuis, J., Earl, L.K., Lovell, D.P., Pape, W.J.W., Uttley, M., Bagley, D.M., Baker, F.W., Brachter, M., Courtellemont, P., Declercq, L., Freeman, S., Steiling, W., Walker, A.P., Carr, G.J., Dami, N., Thomas, G., Harbell, J., Jones, P.A., Pfannenbecker, U., Southee, J.A., Tcheng, M., Argembeaux, H., Castelli, D., Clothier, R., Esdaile, D.J., Itigaki, H., Jung, K., Kasai, Y., Kojima, H., Kristen, U., Larnicol, M., Lewis, R.W., Marenus, K., Moreno, O., Peterson, A., Rasmussen, E.S., Robles, C., Stern, M. (1997) A summary report of the COLIPA international validation study on alternatives to the Draize rabbit eye irritation test. Toxicology in vitro 11(N1-2), 141-179. [117] Bagley, D.M., Bruner, L.H., de Silva, O, Cottin, M., O'Brien, K.A.F., Uttley, M., Walker, A.P. (1992) An evaluation of five potential alternatives in vitro to the rabbit eye irritation test in vivo. Toxicology in vitro 6, 136- 149. [118] Sina, J.F., Galer, D.M., Sussman, R.G., Gautheron, P.D., Sargent, E.V., Leong, B., Shah, P.V., Curren, R.D., Miller K. (1995) A collaborative evaluation of seven alternatives to the Draize eye irritation test using pharmaceutical intermediates. Fundamental and applied toxicology 26(1), 20-31. [119] Bradlaw, J., Gupta, K., Green, S., Hill, R., Wilcox, N. (1997) Practical application of non –whole animal alternatives: summary IRAG workshop on eye irritation. Food and chemical toxicology 35, 175-178. [120] Balls, M., Berg, N., Bruner, L.H., Curren, R., de Silva, O., Earl, L.K., Esdaile, D.J., Fentem, J.H., Liebsch, M., Ohno, Y., Prinsen, M.K., Spielmann, H., Worth, A.P. (1999) Eye irritation testing: the way forward. The report and recommendations of ECVAM workshop 34. Alternatives to laboratory animals. ATLA – Alternatives to Laboratory Animals 27, 53-77. [121] Balls, M., Brantom, P.G., Cassidy, S., Esdaile, D., Fentem, J., Liebsch, M., McPherson, J., Pfannenbecker, U., Prinsen, M. (1999) Preliminary evaluation of the application of reference standards in the prevalidation and validation of in vitro test for eye irritation. In: D.G. Clark, S.G. Lisansky, R. Macmillan: Alternatives to animal testing II, proceedings, Brussels. CPL Press, Berkshire, 201-204. [122] Eskes, C., Bessou, S., Bruner, L., Curren, R., Harbell, J., Jones, P., Kreiling, R., Liebsch, M., McNamee, P., Pape, W., Prinsen, M.K., Seidle, T., Vanparys, P., Worth, 81 A., Zuang, V. (2005) ECVAM Report: Subgroup 3. Eye Irritation. ATLA – Alternatives to Laboratory Animals 33 (Supplement 1), 47-81. [123] Organisation for Economic Co-operation and Development (2002) OECD’s Database on Chemical Risk Assessment Models. Paris: Organisation for Economic Cooperation and Development. OECD, Paris, France. Available at: http://webdomino1.oecd.org/comnet/env/models.nsf [124] Cronin, M.T.D., Dearden, J.C. (1995) QSAR in toxicology. 4. Prediction of nonlethal mammalian toxicological endpoints and expert systems for toxicity prediction. Quantitative Structure-Activity Relationships 14, 518-523. [125] Combes, R.D.; Judson, P. (1995) The use of artificial intelligent expert systems for the prediction of toxicity. Pesticide Science 45, 179-194. [126] Benfenati, E., Gini, G. (1997) Computational predictive programs (expert systems) in toxicology. Toxicology 119, 213-225. [127] Dearden, J.C., Barratt, M.D., Benigni, R., Bristol, D.W., Combes, R.D., Cronin, M.T.D., Judson, P.N., Payne, M.P., Richard, A.M., Tichy, M., Worth, A.P., Yourick, J.J. (1997) The Development and Validation of Expert Systems for Predicting Toxicity. The Report and Recommendations of an ECVAM/ECB Workshop (ECVAM Workshop 24). ATLA – Alternatives to Laboratory Animals 25, 223-252. [128] Darvas, F.; Papp, Á.; Allardyce, A.; Benfenati, E.; Gini, G.; Tichy, M.; Sobb, N.; Citti, A. (1999) Overview of Different Artificial Intelligence Approaches Combined with a Deductive Logic-based Expert System for Predicting Chemical Toxicity. Oral presentation at the "Predictive Toxicology of Chemicals: Experiences and Impact of AI Tools Symposium" within the American Association for Artificial Intelligence Spring Symposium Series, Stanford University, CA, pp. 94-99. [129] Cronin, M.T.D. (2000) Computational methods for the prediction of drug toxicity. Current Opinion in Drug Discovery & Development 3, 292-297. [130] Bhogal, N.; Grindon, C.; Combes, R.; Balls, M. (2005) Toxicity testing: creating a revolution based on new technologies. Trends in Biotechnology 23, 299-307. [131] Gubbels-van Hal, W.M., Blaauboer, B.J., Barentsen, H.M., Hoitink, M.A., Meerts, I.A., van der Hoeven, J.C. (2005) An alternative approach for the safety evaluation of new and existing chemicals, an exercise in integrated testing. Regulatory Toxicology and Pharmacology 42, 284-295. 82 [132] Hulzebos, E., Sijm, D., Traas, T., Posthumus, R., Maslankiewicz, L. (2005) Validity and validation of expert (Q)SAR systems. SAR and QSAR in Environmental Research 16, 385-401. [133] Zinke, S., Gerner, I., Graetschel, G., Schlede, E. (1999) Local Irritation/Corrosion Testing Strategies: Development of a Decision Support System for the Introduction of Alternative Methods. ATLA – Alternatives to Laboratory Animals 28, 29-40. [134] Gerner, I., Graetschel, G., Kahl, J., Schlede, E. (2000) Development of a Decision Support System for the Introduction of Alternative Methods into Local Irritation/Corrosivity Testing Strategies. Development of a Relational Database. ATLA – Alternatives to Laboratory Animals 28, 11-28. [135] Zinke, S., Gerner, I. (2000) A computer-based structure activity relationship method for predicting the toxic effects of organic chemicals from one-dimensional representations of their molecular structures. ATLA – Alternatives to Laboratory Animals 28, 609-620. [136] Zinke, S., Gerner, I. (2000) Local irritation/corrosion testing strategies: Extending a decision support system by applying self-learning classifiers. ATLA – Alternatives to Laboratory Animals 28, 651-663. [137] Gerner, I., Zinke, S., Graetschel, G., Schlede, E. (2000) Development of a Decision Support System for the Introduction of Alternative Methods into Local Irritation/Corrosivity Testing Strategies. Creation of Fundamental Rules for a Decision Support System. ATLA – Alternatives to Laboratory Animals 28, 665-698. [138] Gerner, I., Schlegel, K., Walker, J.D., Hulzebos, E. (2004) Use of Physicochemical Property Limits to Develop Rules for Identifying Chemical Substances with no Skin Irritation or Corrosion Potential. QSAR & Combinatorial Science 23, 726-733. [139] Hulzebos, E., Walker, J.D., Gerner, I., Schlegel, K. (2005) Use of structural alerts to develop rules for identifying chemical substances with skin irritation or skin corrosion potential. QSAR & Combinatorial Science 24, 332-342. [140] Rorije, E., Hulzebos, E. (2005) Evaluation of (Q)SARs for the prediction of Skin Irritation/Corrosion Potential. Physico-chemical exclusion rules. Report submitted by the National Institute of Public Health and Environment (RIVM) to the European Chemicals Bureau (ECB). September 2005. 83 [141] Tsakovska, I., Netzeva, T., Worth, A. (2006) Evaluation of (Q)SARs for the prediction of Eye Irritation/Corrosion Potential. Structural Inclusion rules. EUR technical report. Ispra, Italy: European Commission, Joint Research Centre. [142] Tsakovska, I., Gallegos, A., Netzeva, T., Pavan, M., Worth, A. (2006) Evaluation of structure-activity relationships for the prediction of eye irritation and corrosion. SAR and QSAR in Environmental Research. Accepted. [143] Gallegos Saliner, A., Tsakovska, I., Pavan, M., Patlewicz, G., Worth, A. (2006) Evaluation of SARs for the prediction of skin irritation/corrosion potential - structural inclusion rules in the BfR decision support system. SAR and QSAR in Environmental Research. Accepted. [144] LHASA Ltd. (2005). DEREK for Windows. Leeds, U.K. Available at: http://www.chem.leeds.ac.uk/luk/derek [145] Sanderson, D.M., Earnshaw, C.G. (1991) Computer Prediction of Possible Toxic Action from Chemical Structure; The DEREK System. Human & Experimental Toxicology 10, 261-273. [146] Greene, N., Judson, P.N., Langowski, J.J., Marchant, C.A. (1999) Knowledge-based expert systems for toxicity and metabolism prediction: DEREK, StAR, METEOR. SAR and QSAR in Environmental Research 10, 299-314. [147] Hulzebos, E.M., Posthumus, R. (2003) (Q)SARs: Gatekeepers against risk on chemicals? SAR and QSAR in Environmental Research 14, 285-316. [148] Combes, R.D.; Rodford, R. (2004) The Use of Expert Systems for Toxicity Prediction - Illustrated With Reference to the DEREK Program. In, Predicting Toxicity and Fate. Cronin, M.T.D.; Livingstone, D.J. (Eds.) CRC Press, pp. 193-204. [149] CompuDrug Chemistry Ltd. Budapest, Hungary. Available at: http://www.compudrug.com [150] Smithing, M.P., Darvas, F. (1992) HazardExpert: An Expert System for Predicting Chemical Toxicity. In, Food Safety Assessment. Finlay, S.F., Robinson, S.F., Armstrong, D.J. (Eds.). American Chemical Society, Washington, DC, pp. 191-200. [151] Lewis, D.F., Bird, M.G., Jacobs, M.N. (2002) Human carcinogens: an evaluation study via the COMPACT and HazardExpert procedures. Human & Experimental Toxicology 21, 115-122. 84 [152] MultiCASE Inc. MultiCASE. Cleveland, OH, USA. Available at: http://www.multicase.com [153] Klopman, G. (1984) Artificial intelligence approach to structure-activity studies. Computer automated structure evaluation of biological activity of organic molecules. Journal of the American Chemical Society 106, 7315-7321. [154] Klopman, G. (1992) MULTICASE.1. A Hierarchical Computer Automated Structure Evaluation Program. Quantitative Structure-Activity Relationships 11, 176-184. [155] Klopman, G., Ptchelinstev, D., Frierson, M., Pennisi, S., Renskers, K., Dickens, M. (1993) Multiple computer automated structure evaluation methodology as an alternative to in vivo eye irritation testing. ATLA – Alternatives to Laboratory Animals 21, 14-27. [156] Rosenkranz, H.S., Cunningham, A.R., Zhang, Y.P., Klopman, G. (1999) Applications of the case/multicase SAR method to environmental and public health situations. SAR and QSAR in Environmental Research 10, 263-276. [157] Rosenkranz, H.S., Cunningham, A.R., Zhang, Y.P., Claycamp, H.G., Macina, O.T., Sussman, N.B., Grant, S.G., Klopman, G. (1999) Development, characterization and application of predictive-toxicology models. SAR and QSAR in Environmental Research 10, 277-298. [158] Accelrys Inc. (2005) TOPKAT. Cambridge, UK. (Developed by: Health Design Inc. & Oxford Molecular Group, Inc.) Available at: http://www.accelrys.com/products/topkat [159] Enslein, K. (1988) An overview of structure-activity relationships as an alternative to testing in animals for carcinogenicity, mutagenicity, dermal and eye irritation, and acute oral toxicity. Toxicology and Industrial Health 4, 479-498. [160] Gombar, V.K., and Enslein, K. (1990) Quantitative structure-activity relationship (QSAR) studies using electronic descriptors calculated from topological and molecular orbital (MO) methods. Quantitative Structure-Activity Relationships 8, 321-325. [161] Enslein, K. (1993) The future of toxicity prediction with QSAR. In vitro Toxicology 6, 163-169. [162] Gombar, V.K. (1997) Quantitative structure-activity relationships in toxicology: from fundamentals to applications. In, Advances In Molecular Toxicology, pp. 125-139. 85 [163] Enslein, K. (1997) QSTR applications in acute, chronic, and developmental toxicity, and carcinogenicity. In, Adv. In Molecular Toxicology, 141-164. [164] Gombar, V.K., Enslein, K., Blake. B.W., Reid, D.A. (1997) Quantitative StructureActivity Relationships In Predicting Chemical Toxicity, When is the Prediction Reliable? In The proceedings of the Workshop on QSARs in Environmental Sciences 27, 399-411. [165] http://www.CACheSoftware.com/cache/ [166] Check, C.E.; Gilbert, T.M. (2000) CAChe: Computer-Aided Chemistry for Macintosh and Windows. Version 3.2 for Windows; Version 4.1 for Macintosh. Journal of the American Chemical Society (Computer Software Review) 122, 9566-9567. [167] Rajamani, R. (2004) CAChe 6.1 WorkSystem Pro plus Active Site Add-On. CAChe Group, Fujitsu America, Inc., 15244 Greenbriar Parkway, Beaverton, OR 97006. http://www.cachesoftware.com/. Contact CAChe group for pricing information. Journal of the American Chemical Society (Computer Software Review) 126, 88568857. [168] Walker, J.D., Gray, D.A. (2001) The Substructure-based Computerized Chemical Selection Expert System (SuCCSES): Providing Chemical Right-to-Know (CRTK) Information on Potential Actions of Structurally-Related Chemical Classes on the Environment and Human Health. In, Quantitative Structure Activity Relationships (QSARs) for Predicting Effects of Chemicals on Environmental-Human Health Interactions. Walker, J. D. (Ed). SETAC Press, Pensacola, FL. [169] Walker, J.D., Brink, R.H. (1989) New cost-effective, computerized approaches to selecting chemicals for priority testing consideration. In, Aquatic Toxicology and Environmental Fate, Vol 11, ASTM STP 1007. Suter, G.W.II, Lewis, M.A. (Eds). Philadelphia American Society for Testing and Materials, pp. 507-536. [170] Walker, J.D. (1991) Chemical selection by the TSCA Interagency Testing Committee: use of computerized substructure searching to identify chemical groups for health effects, chemical fate and ecological effects testing. Science of the Total Environment 109/110, 691-700. [171] Walker, J.D. (1993) The TSCA Interagency Testing Committee, 1977 to 1992: creation, structure, functions and contributions. In, Environmental Toxicology and Risk Assessment, Vol 2. Gorsuch, J.W., Dwyer, F.J., Ingersoll, C.G., La Pointe, T.W., (Eds.). ASTM STP 1216. American Society for Testing and Materials, Philadelphia, PA, pp. 451-509. 86 [172] Walker, J.D. (1993) The TSCA Interagency Testing Committee’s approaches to screening and scoring chemicals and chemical groups: 1977-1983. In, Access and Use of Information Resources in Assessing Health Risks from Chemical Exposure. Lu, P.Y. (Ed). Oak Ridge National Laboratories, Oak Ridge, TN, pp. 77-93. [173] Walker, J.D. (1995) Estimation methods used by the TSCA Interagency Testing Committee to prioritize chemicals for testing: exposure and biological effects scoring and structure activity relationships. Toxicological Modelling 1, 123-141. [174] Laboratory of Mathematical Chemistry. (2004) University of Bourgas, Bourgas, Bulgaria. Available at: http://www.oasis-lmc.org [175] Mekenyan, O.G., Karabunarliev, S., Bonchev, D. (1990) The microcomputer OASIS system for predicting of the biological activity of chemical compounds. Computers & Chemistry 14, 193-200. [176] Mekenyan, O.G., Karabunarliev, S., Ivanov, J., Dimitrov, D. (1994) A new development of the OASIS computer system for modelling of molecular properties. Computers & Chemistry 18, 173-187. 87 European Commission EUR 22320 EN – DG Joint Research Centre, Institute for Health and Consumer Protection Luxembourg: Office for Official Publications of the European Communities 2006 – 87 pp. – 21x 29.7 cm Scientific and Technical Research series Abstract This report reviews the state-of-the-art of in silico and in vitro methods for assessing dermal and ocular irritation and corrosion. Following a general introduction, the current EU legislation for the classification and labelling of chemicals causing irritation and corrosivity is summarised. Then currently available non-animal approaches are reviewed. The main alternative approaches to assess acute local toxic effects are: a) in silico approaches, including SARs, QSARs and expert systems integrating multiple approaches; and b) in vitro test methods. In this review, emphasis is placed on literature-based (Q)SAR models for skin and eye irritation and corrosion as well as computer-based expert systems. Mission of the JRC The mission of the JRC is to provide customer-driven scientific and technical support for the conception, development, implementation and monitoring of EU policies. As a service of the European Commission, the JRC functions as a reference centre of science and technology for the Union. Close to the policy-making process, it serves the common interest of the Member States, while being independent of special interests, whether private or national.