MARS: Multi-anchor reasoning for reliable toxicity prediction under distribution shift.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning for toxicology

Background:

• Toxicity and safety prediction models are crucial in drug discovery.
• Conventional benchmarks may overestimate model reliability under distribution shift (e.g., scaffold or out-of-distribution shifts).
• High confidence predictions from models may be unreliable when applied to novel chemical spaces.

Purpose of the Study:

• To develop a reliability-centered framework (MARS) for toxicity prediction that accounts for distribution shift.
• To separate toxicity prediction from confidence estimation.
• To enhance the trustworthiness of toxicity predictions in drug discovery.

Main Methods:

• MARS (Multi-Anchor Reasoning framework) combines a molecular encoder with evidence from multiple training anchors.
• A dedicated shift-aware reliability head is trained to distinguish prediction scores from trust signals.
• The framework utilizes target representations, anchor similarities, labels, local disagreement, and novelty for reasoning.

Main Results:

• MARS evaluated on seven public toxicity tasks (24,256 molecules) across various distribution shifts.
• Ours-Full achieved 0.776 AUROC and 0.131 ECE on scaffold summaries.
• Ours-Anchor reduced ECE by 0.049 compared to RF-Morgan on transfer tasks and improved error detection AUROC from 0.672 to 0.719.

Conclusions:

• Toxicity models require evaluation beyond discrimination, focusing on reliability under chemical shifts.
• MARS provides a robust approach to assess and maintain confidence in toxicity predictions across different chemical spaces.
• The proposed framework enhances the usability of confidence signals in drug discovery pipelines facing distribution shifts.