Prediction of Skin Sensitizers using Alternative Methods to Animal Experimentation

The 3Ranker: An AI-based Algorithm for Finding Non-animal Alternative Methods

The search for existing non-animal alternative methods for use in experiments is currently challenging because of the lack of both comprehensive structured databases and balanced keyword-based search strategies to mine unstructured textual databases. In this paper we describe 3Ranker, which is a fast, keyword-independent algorithm for finding non-animal alternative methods for use in biomedical research. The 3Ranker algorithm was created by using a machine learning approach, consisting of a Random Forest model built on a dataset of 35 million abstracts and constructed with weak supervision, followed by iterative model improvement with expert curated data. We found a satisfactory trade-off between sensitivity and specificity, with Area Under the Curve (AUC) values ranging from 0.85-0.95. Trials showed that the AI-based classifier was able to identify articles that describe potential alternatives to animal use, among the thousands of articles returned by generic PubMed queries on dermatitis and Parkinson's disease. Application of the classification models on time series data showed the earlier implementation and acceptance of Three Rs principles in the area of cosmetics and skin research, as compared to the area of neurodegenerative disease research. The 3Ranker algorithm is freely available at www.open3r.org; the future goal is to expand this framework to cover multiple research domains and to enable its broad use by researchers, policymakers, funders and ethical review boards, in order to promote the replacement of animal use in research wherever possible.

Predicting in vivo effect levels for repeat-dose systemic toxicity using chemical, biological, kinetic and study covariates

In an effort to address a major challenge in chemical safety assessment, alternative approaches for characterizing systemic effect levels, a predictive model was developed. Systemic effect levels were curated from ToxRefDB, HESS-DB and COSMOS-DB from numerous study types totaling 4379 in vivo studies for 1247 chemicals. Observed systemic effects in mammalian models are a complex function of chemical dynamics, kinetics, and inter- and intra-individual variability. To address this complex problem, systemic effect levels were modeled at the study-level by leveraging study covariates (e.g., study type, strain, administration route) in addition to multiple descriptor sets, including chemical (ToxPrint, PaDEL, and Physchem), biological (ToxCast), and kinetic descriptors. Using random forest modeling with cross-validation and external validation procedures, study-level covariates alone accounted for approximately 15% of the variance reducing the root mean squared error (RMSE) from 0.96 log10 to 0.85 log10 mg/kg/day, providing a baseline performance metric (lower expectation of model performance). A consensus model developed using a combination of study-level covariates, chemical, biological, and kinetic descriptors explained a total of 43% of the variance with an RMSE of 0.69 log10 mg/kg/day. A benchmark model (upper expectation of model performance) was also developed with an RMSE of 0.5 log10 mg/kg/day by incorporating study-level covariates and the mean effect level per chemical. To achieve a representative chemical-level prediction, the minimum study-level predicted and observed effect level per chemical were compared reducing the RMSE from 1.0 to 0.73 log10 mg/kg/day, equivalent to 87% of predictions falling within an order-of-magnitude of the observed value. Although biological descriptors did not improve model performance, the final model was enriched for biological descriptors that indicated xenobiotic metabolism gene expression, oxidative stress, and cytotoxicity, demonstrating the importance of accounting for kinetics and non-specific bioactivity in predicting systemic effect levels. Herein, we generated an externally predictive model of systemic effect levels for use as a safety assessment tool and have generated forward predictions for over 30,000 chemicals.

Multivariate models for prediction of human skin sensitization hazard

One of the Interagency Coordinating Committee on the Validation of Alternative Method's (ICCVAM) top priorities is the development and evaluation of non-animal approaches to identify potential skin sensitizers. The complexity of biological events necessary to produce skin sensitization suggests that no single alternative method will replace the currently accepted animal tests. ICCVAM is evaluating an integrated approach to testing and assessment based on the adverse outcome pathway for skin sensitization that uses machine learning approaches to predict human skin sensitization hazard. We combined data from three in chemico or in vitro assays - the direct peptide reactivity assay (DPRA), human cell line activation test (h-CLAT) and KeratinoSens™ assay - six physicochemical properties and an in silico read-across prediction of skin sensitization hazard into 12 variable groups. The variable groups were evaluated using two machine learning approaches, logistic regression and support vector machine, to predict human skin sensitization hazard. Models were trained on 72 substances and tested on an external set of 24 substances. The six models (three logistic regression and three support vector machine) with the highest accuracy (92%) used: (1) DPRA, h-CLAT and read-across; (2) DPRA, h-CLAT, read-across and KeratinoSens; or (3) DPRA, h-CLAT, read-across, KeratinoSens and log P. The models performed better at predicting human skin sensitization hazard than the murine local lymph node assay (accuracy 88%), any of the alternative methods alone (accuracy 63-79%) or test batteries combining data from the individual methods (accuracy 75%). These results suggest that computational methods are promising tools to identify effectively the potential human skin sensitizers without animal testing. Published 2016. This article has been contributed to by US Government employees and their work is in the public domain in the USA.

Analysis of the Local Lymph Node Assay (LLNA) variability for assessing the prediction of skin sensitisation potential and potency of chemicals with non-animal approaches

The knowledge of the biological mechanisms leading to the induction of skin sensitisation has favoured in recent years the development of alternative non-animal methods. During the formal validation process, results from the Local Lymph Node Assay (LLNA) are generally used as reference data to assess the predictive capacity of the non-animal tests. This study reports an analysis of the variability of the LLNA for a set of chemicals for which multiple studies are available and considers three hazard classification schemes: POS/NEG, GHS/CLP and ECETOC. As the type of vehicle used in a LLNA study is known to influence to some extent the results, two analyses were performed: considering the solvent used to test the chemicals and without considering the solvent. The results show that the number of discordant classifications increases when a chemical is tested in more than one solvent. Moreover, it can be concluded that study results leading to classification in the strongest classes (1A and EXT) seem to be more reliable than those in the weakest classes. This study highlights the importance of considering the variability of the reference data when evaluating non-animal tests.

Predicting chemically-induced skin reactions. Part I: QSAR models of skin sensitization and their application to identify potentially hazardous compounds

Repetitive exposure to a chemical agent can induce an immune reaction in inherently susceptible individuals that leads to skin sensitization. Although many chemicals have been reported as skin sensitizers, there have been very few rigorously validated QSAR models with defined applicability domains (AD) that were developed using a large group of chemically diverse compounds. In this study, we have aimed to compile, curate, and integrate the largest publicly available dataset related to chemically-induced skin sensitization, use this data to generate rigorously validated and QSAR models for skin sensitization, and employ these models as a virtual screening tool for identifying putative sensitizers among environmental chemicals. We followed best practices for model building and validation implemented with our predictive QSAR workflow using Random Forest modeling technique in combination with SiRMS and Dragon descriptors. The Correct Classification Rate (CCR) for QSAR models discriminating sensitizers from non-sensitizers was 71-88% when evaluated on several external validation sets, within a broad AD, with positive (for sensitizers) and negative (for non-sensitizers) predicted rates of 85% and 79% respectively. When compared to the skin sensitization module included in the OECD QSAR Toolbox as well as to the skin sensitization model in publicly available VEGA software, our models showed a significantly higher prediction accuracy for the same sets of external compounds as evaluated by Positive Predicted Rate, Negative Predicted Rate, and CCR. These models were applied to identify putative chemical hazards in the Scorecard database of possible skin or sense organ toxicants as primary candidates for experimental validation.