Accurate clinical toxicity prediction using multi-task deep neural nets and contrastive molecular explanations

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Molecular toxicity prediction is crucial for efficient drug development and chemical safety.
• Current methods often involve costly and time-consuming experimental testing, raising ethical concerns.
• Explainable machine learning (ML) offers a promising alternative to reduce experimental burden.

Purpose of the Study:

• To develop a deep learning framework for simultaneous prediction of in vitro, in vivo, and clinical molecular toxicity.
• To evaluate the efficacy of different molecular representations (Morgan fingerprints, pre-trained SMILES embeddings) for toxicity prediction.
• To provide explainable predictions for molecular toxicity using a post-hoc contrastive explanation method.

Main Methods:

• A multi-task deep learning model was employed to integrate diverse toxicity data.
• Two molecular input representations were utilized: Morgan fingerprints and pre-trained SMILES embeddings.
• A post-hoc contrastive explanation method was adapted to identify key predictive features (toxicophores).

Main Results:

• The multi-task deep learning model accurately predicted toxicity across all endpoints, including clinical.
• Pre-trained SMILES embeddings improved clinical toxicity predictions compared to existing benchmarks.
• The model demonstrated comparable performance to state-of-the-art methods for specific endpoints.
• Transfer learning indicated a reduced need for in vivo data in clinical toxicity prediction.
• Contrastive explanations identified known toxicophores, with higher recovery for in vitro and in vivo data.

Conclusions:

• Explainable ML provides an efficient and ethical approach to molecular toxicity prediction.
• The developed multi-task deep learning model with pre-trained SMILES embeddings shows significant promise for drug development.
• Contrastive explanations enhance model interpretability and aid in understanding toxicity mechanisms.