ChemTunes/ToxGPS (Database and Knowledgebase for Safety Evaluation and Risk Assessment, Altamira LLC and Molecular Networks GmbH) is a knowledge base of in vitro and in vivo toxicity information and comprises multiple components/workflows to support the safety and risk assessment of chemical compounds, including an expert-QC'ed database and the MoA-based ToxGPS prediction system for a series of human health and regulatory-relevant toxicity endpoints, as well as the Liver BioPath, a tool for human metabolism prediction.
ToxGPS prediction system comprises a set of albums for a series of human health toxicity endpoints, including genetic toxicity, carcinogenicity, developmental and reproductive toxicity, skin sensitization. Each knowledgebase album consists of the following components:
- alerting chemotypes (structural alerts)
- mechanistically-informed (mode-of-action driven) QSAR models, i.e. an approach used at US FDA CERES (Chemical Evaluation and Risk Estimation System)
- nearest neighbors (“structural analogues”) analysis and optional access to training sets (with expert-aggregated study calls)
- optional access to ToxGPS toxicity database (QC’ed by experts) for provided endpoints.
All ToxGPS QSAR models consist of chemical mode-of-action (MoA) category models as well as a general global model. The computational modelling approach is a hybrid of partial least squares (PLS)/ordinal logistic regression methods. For model building, global molecular and shape descriptors (from CORINA Symphony) and quantum-mechanic parameters are used. The models return probabilistic predictions (positive and negative probabilities plus a quantitative estimate of the associated uncertainty) and an overall prediction (positive/negative/equivocal).
Unlike QSAR models, chemotype alerts generate only positive predictions. The reliability of each alert is determined by exploring the ability of the alert to hit positive compounds in a large training set. Different training sets were used for the QSAR models and the alerts, so that predictions from these are indeed independent. A mathematically rigorous and quantitative weight-of-evidence (WoE) decision theory approach (i.e., Dempster-Shafer theory (DST)) is used to obtain the final overall assessment (by combining the predictions from QSAR and alerts), and to provide a quantitative estimation of the uncertainty associated with the prediction.
Applicability domain analysis is performed based on the QSAR global models and reports whether the target compound is out-of-domain.