Integrating Multiview Information for Enhanced Deep Learning-Based Acute Dermal Toxicity Prediction

Accurate prediction of acute dermal toxicity is vital for the safe and effective development of contact drugs. While numerous deep learning models have been created to replace costly and ethically challenging animal toxicity tests, most approaches overlook the multiview information on molecules. To overcome this limitation, we introduce a novel model named MVIToxNet, which integrates multiview features from both molecular fingerprints and SMILES sequences. To capture the multiview information on SMILES, MVIToxNet incorporates character-level and atom-level features. In addition, byte-pair encoding tokenization is utilized to capture substructural details within molecules, allowing the model to differentiate similar SMILES by assigning distinct tokens to different substructures. Since the data sets in this study are small and imbalanced, we argue that selecting a single model based solely on the best validation performance may not reliably reflect the best generalization for test sets. Therefore, we propose a weighted model averaging approach that combines multiple trained models according to their top-K validation scores into one model, yielding an improved model for inference. Extensive experimental results demonstrate that MVIToxNet significantly outperforms existing baselines in acute dermal toxicity prediction, validating the effectiveness of utilizing multiview features and the weighted model averaging strategy. Furthermore, our proposed methods demonstrate the potential for data-driven model design.