Controlled evaluation of architectural, classifier, and training refinements in MolTrans-based drug-target interaction prediction

Abstract Drug–target interaction (DTI) prediction is a central task in computational drug discovery, but performance improvements can be difficult to attribute when architectural modifications and training changes are introduced simultaneously. This study evaluates a Bidirectional Cross-Attention and Global Aggregation DTI model (BCAG-DTI) through eight controlled configurations that separate bidirectional cross-attention, global average–max pooling, classifier design, and optimisation strategy. All principal experiments use fixed training, validation, and test partitions and five matched random seeds on BindingDB, BIOSNAP, and DAVIS. Relative to the MolTrans baseline, the complete BCAG-DTI configuration improves mean area under the receiver operating characteristic curve (AUROC) from 0.8815 to 0.9063 on BindingDB, from 0.8631 to 0.8891 on BIOSNAP, and from 0.8808 to 0.8955 on DAVIS. The corresponding gains in area under the precision–recall curve (AUPRC) are 0.0831, 0.0346, and 0.0619, respectively. Controlled ablation shows that the enhanced classifier achieves the highest mean AUROC on BindingDB and BIOSNAP and the highest mean AUPRC and F1-score on all three datasets, whereas the cross-attention-plus-pooling configuration with enhanced training achieves the highest mean AUROC on DAVIS. The enhanced training strategy also provides substantial improvements, while cross-attention alone reduces performance under the baseline training configuration. In BIOSNAP robustness experiments, BCAG-DTI improves MolTrans for unseen drugs, unseen proteins, and 70–90% missing-interaction settings, whereas its AUROC and F1-score are slightly lower at 95% missing data. As an external same-split reference, CPI-GGS evaluated on the fixed BIOSNAP partitions achieves 0.8619 ± 0.0023 AUROC, 0.8645 ± 0.0042 AUPRC, and 0.7938 ± 0.0028 F1-score, compared with 0.8891 ± 0.0078, 0.8992 ± 0.0067, and 0.8178 ± 0.0094 for BCAG-DTI. This comparison is interpreted in the context of different input preprocessing pipelines and substantial differences in model capacity. Attention case analysis further indicates that cross-attention weights should be treated as model-internal allocation patterns rather than validated binding contacts. Overall, the results show that classifier design and optimisation account for a substantial portion of the improvement over MolTrans, while the contribution of cross-modal architectural components is optimisation-sensitive and dataset-dependent.