Scalable and Interpretable Chemical Neurotoxicity Prediction via Multi-Modal Gated Fusion and ChemBERTa

This is the authors' abstract. We don't add key points for this paper.

Chemical exposure causes adverse neurological outcomes, highlighting the need for high-throughput screening of chemical neurotoxicity. However, existing computational methods rely on single-modality inputs and lack complementary molecular information to sufficiently enhance the accuracy and efficiency of neurotoxicity risk assessment. Here, we present NeuroToxPredictor, a scalable and interpretable multimodal framework built on a curated chemical neurotoxicity dataset for accurate and efficient chemical neurotoxicity prediction. It integrates graph attention networks and molecular fingerprints with ChemBERTa, a pretrained chemical language model that provides contextual encoding of SMILES. A gated fusion module adaptively weights these complementary representations according to their predictive relevance to enhance modality-specific weight allocation through a gating fusion mechanism. NeuroToxPredictor achieved competitive performance across five independent training seeds (AUC = 0.919 ± 0.004; accuracy = 0.845 ± 0.014; F1-score = 0.861 ± 0.011; recall = 0.884 ± 0.013; precision = 0.839 ± 0.024; MCC = 0.688 ± 0.028), outperforming both single-modality and reduced-modality baselines. It demonstrates strong generalizability with an AUC of 0.861, as externally validated on an independent set of 279 compounds. NeuroToxPredictor is deployed on a publicly platform (https://www.ai4environ.cn/NeuroToxPredictor), enabling an end-to-end intelligent workflow from chemical retrieval, neurotoxicity prediction, applicability-domain assessment, and large language model-assisted result interpretation.