Reliable Aquatic Toxicity Prediction via Embedding-Aware Applicability Domains.

Area of Science:

• Computational toxicology
• cheminformatics
• machine learning

Background:

• Accurate and auditable toxicity predictions are crucial for regulatory assessment.
• Current methods often lack reliability and scalability.
• Structure-activity relationships (SAR) are key to predicting chemical toxicity.

Purpose of the Study:

• To propose a novel framework, the structure-activity landscape with an embedding-compatible applicability domain (SAL-AD), for enhanced toxicity prediction.
• To link prediction reliability to representation-space geometry using transformer embeddings.
• To establish interpretable metrics for identifying reliable prediction regions.

Main Methods:

• Utilized transformer-driven embeddings (768-dimensional) for chemical representations.
• Developed SAL-AD with cosine top-k neighborhoods to define similarity density and activity inconsistency metrics.
• Employed gradient-boosted learners and merged-endpoint training across 11 EPA ECOTOX endpoints.
• Validated the approach on 1499 substances from the Chinese Hazardous Chemicals Inventory.

Main Results:

• Transformer embeddings with SAL-AD consistently outperformed descriptor-based methods.
• Merged-endpoint training improved prediction accuracy through generalization and interpolation.
• SAL-AD established a practical operating window and measurable boundaries for prediction reliability.
• Fine-tuning sharpened toxicity clustering, aligning representation geometry with SAR.

Conclusions:

• SAL-AD provides a robust and interpretable method for assessing toxicity prediction reliability.
• The framework enables high-throughput screening with transparent uncertainty control for regulatory applications.
• This approach advances computational toxicology by transforming applicability domain assessment from a heuristic to a measurable boundary.