AmesNet: A Task-Conditioned Deep Learning Model with Enhanced Sensitivity and Generalization in Ames Mutagenicity Prediction

Abstract

Regulatory agencies require comprehensive genotoxicity assessments for novel small-molecule therapeutics prior to human trials. Developers often delay these studies until a candidate is nearing regulatory submission because they are expensive and secondary to bioactivity. This timing creates a bottleneck where late-stage failures can jeopardize >\$10 million in capital and multiple years of developmental progress per candidate. The Ames assay detects a molecule’s mutagenic potential. Regulators now explicitly support the use of in silico Ames mutagenicity models through enabling legislation, dedicated FDA AI toxicology programs, internationally harmonized guidelines, and benchmark challenges. However, current Ames models suffer from a sensitivity drop-off when they evaluate molecules outside their training domain. The out-of-domain (OOD) test set from the Second Ames/QSAR International Challenge Project yielded a participant average sensitivity of 0.46. Sensitivity is the most important metric in Ames prediction because false negatives allow mutagenic compounds to advance undetected and trigger the most costly late-stage failures. Attempts to fix this sensitivity drop-off often reduce overall model performance, which can be represented by balanced accuracy (BA). For example, DeepAmes reports high levels of sensitivity (0.87) by sacrificing its BA (0.52). We introduce AmesNet, a task-conditioned Ames model that achieves class-leading performance in novel chemical spaces. AmesNet utilizes a dual branch architecture containing a molecular encoder and a dedicated channel to condition Ames assay context such as metabolic activation and bacterial strain type. In comparative benchmarks, AmesNet reached a sensitivity of 0.72 (95% CI: 0.68–0.76) and a simultaneous BA of 0.81 (95% CI: 0.78–0.83) on the OOD test data. This performance improves discrimination without the sensitivity trade-off limiting existing approaches. Structural analysis demonstrates that AmesNet recovers difficult-to-detect mutagenic compounds missed by existing models. This framework provides a high-confidence filtering mechanism that enables drug developers to turn a costly late-stage safety bottleneck into an earlier-stage mutagenicity triage.