Optimizing drug discovery through long short-term memory recurrent neural networks (LSTM RNNs): A hybrid multi-modal...

Drug discovery is a resource intensive and time consuming process, signifying a need for machine learning frameworks that can scale large biochemical datasets on modest hardware. In this study, we introduce a hybrid CNN-LSTM-GNN model (HMLCG) for drug-target interaction (DTI) prediction in 2 sets of workflows: the full pipeline with BindingDB, and the reduced regime with 20% of BindingDB. The proposed model (∼1.0 M parameters) is trained for 32,749 ligand-target pairs in less than 9 min via CPU with mixed-precision, label smoothing and OneCycleLR scheduling with ROC-AUC 0.7917, accuracy 75.11% and F1-score 0.4885. The ablation results based on a randomly selected subset of 10,000 samples prove that the prediction results are very important and that if we remove the LSTM pathway from the architecture, the ROC-AUC score deteriorates to 0.4955, while the F1-score falls to 0. CNN-only baselines with no GNNs achieve a yes/no ROC - AUC of 0.5396 and GNN-only baselines with no CNN achieves a yes/no ROC - AUC of 0.4741, suggesting that either CNN or GNN components can learn informative but complementary structural features, albeit with modest contribution over and above sequence modeling in current design. Interestingly, multimodal integration does not always assure the superior performance: the best AUC values are obtained in LSTM-only configurations. A paired comparison with Adam and quantum-inspired optimizer (QIO) for 20-seed results demonstrates statistically insignificant differences in mean AUC (Adam: 0.6279, QIO: 0.6236) or convergence epoch (16.2 ± 0.4, 16.2 ± 0.4 respectively; Bonferroni-corrected p-values > 0.01), indicating limited impact of optimizer choice relative to architectural design and data representation. Evaluation of imbalanced data from PubChem BioAssay shows that directed enrichment factors are better than random at EF@1-10 and that a cost-benefit analysis suggests that the optimal screening set will be the top 5% of compounds ranked by the enrichment factor. Still, there are also moderate performances in high-throughput screening conditions (unbalanced AUC 0.52; balanced AUC 0.63), so that HMLCG can be used as a prioritization tool, but not as a standalone decision system. The overall results show that achieving computational efficiency is possible without significant compromises in predictive power, and that sequence modeling is the primary contributor to DTI prediction in this framework, with the incremental benefits of additional modalities coming thereafter in a context-dependent way.