iDILI-MT: identifying drug-induced liver injury compounds with a multi-head Transformer

Drug-induced liver injury (DILI) is a leading cause of late-stage drug attrition and post-approval withdrawals, making early <i>in silico</i> risk assessment essential for drug safety. We present iDILI-MT (identifying drug-induced liver injury compounds with a multi-head Transformer), a self-contained computational framework that integrates a feed-forward network for sequential feature extraction, a multi-head Transformer encoder for contextual representation learning, and a squeeze-and-excitation attention module for channel-wise feature recalibration. Evaluated on a curated set of 1,919 small-molecule compounds, iDILI-MT outperformed traditional machine-learning classifiers and state-of-the-art graph neural networks, achieving a mean area under the receiver-operating-characteristic curve (AUC-ROC) of 0.8499, area under the precision-recall curve (AUC-PR) of 0.8905, and F1 score of 0.8173 across ten trials. Attention-weight analysis reveals that the combined multi-head and squeeze-and-excitation attention mechanisms effectively highlight key substructural and chemical motifs associated with hepatotoxicity. These findings indicate that iDILI-MT provides an useful method for detecting compounds at risk of DILI, potentially accelerating safety assessments in drug development.