SbD4Skin by EosCloud: Integrating multi-view molecular representation for predicting skin sensitization, irritation,...

Area of Science:

• Computational toxicology
• In silico methods
• Chemical risk assessment

Background:

• Ethical, financial, and scientific challenges necessitate non-animal testing methods for chemical toxicity assessment.
• Predicting skin sensitization, irritation/corrosion, and acute dermal toxicity is crucial for human health risk evaluation.
• Existing methods often rely on animal testing, driving the need for alternative in silico approaches.

Purpose of the Study:

• To develop and validate a computational framework for predicting key skin toxicity endpoints.
• To evaluate the efficacy of diverse molecular representations (MACCS keys, Morgan fingerprints, Mordred descriptors) in toxicity prediction.
• To enhance model transparency and provide mechanistic insights into chemical toxicity.

Main Methods:

• Utilized machine learning algorithms (Random Forest, SVM, k-NN) to evaluate individual molecular representations.
• Developed a multi-view fully connected neural network (FCNN) integrating various molecular descriptors.
• Employed Shapley Additive exPlanations (SHAP) for feature importance analysis and model interpretability.

Main Results:

• The multi-view FCNN model demonstrated superior or comparable predictive performance to single-representation models.
• Achieved high AUC values: 0.91 for irritation/corrosion, 0.88 for sensitization, and 0.82 for acute dermal toxicity.
• Successfully identified 86% of skin sensitizers, 89% of irritants, and 86% of dermally toxic compounds on test sets.

Conclusions:

• The computational framework effectively predicts skin toxicity endpoints, offering a viable alternative to animal testing.
• The SbD4Skin web platform provides free access to validated models, supporting regulatory decision-making.
• FAIRified datasets and models promote transparency, reusability, and acceptance in chemical risk assessment.