Reliable Aquatic Toxicity Prediction via Embedding-Aware Applicability Domains

Regulatory assessment requires toxicity predictions that are both accurate and auditable at scale. We propose a transformer-driven structure-activity landscape with an embedding-compatible applicability domain (SAL-AD), which links prediction reliability to representation-space geometry. Using cosine top-k neighborhoods, SAL-AD defines two interpretable metrics, similarity density and activity inconsistency, to identify reliable in-domain regions. Across 11 end points curated from EPA ECOTOX (algae, fish, aquatic invertebrates), 768-dimensional transformer embeddings paired with gradient-boosted learners consistently outperformed descriptor-based baselines, and merged-end point training further improved prediction through compound-level generalization and cross-end point interpolation. SAL-AD raised in-domain accuracy while revealing a practical operating window (top-k ≈ 0.05-0.2) and clear degradation beyond the calibrated domain, turning applicability domain from a heuristic into a measurable boundary. Fine-tuning sharpened toxicity clustering, indicating alignment between representation geometry and structure-activity organization. Deployed to 1499 substances from the Chinese Hazardous Chemicals Inventory under harmonized acute (EC₅₀/LC₅₀) and chronic (EC₁₀) settings, the framework enabled high-throughput screening with transparent uncertainty control.