Sequence-Driven Drug–Target Interaction Prediction Modelling Using Deep Learning Models with Cancer Specific Enhancements.

Recent breakthroughs in computational drug research have mostly focused on deep learning-based models for anticipating drug-target binding affinity (DTA), an important problem in determining effective therapeutics. This is especially essential in cancer research, where molecular heterogeneity, dynamic protein configurations, and the desire for personalised therapy options provide significant challenges. Traditional experimental tests, while precise, are time consuming, expensive, and difficult to scale across large chemical libraries. Structure-based computational techniques usually depend on regulated elements and high-resolution 3D data, limiting their utility in real-world biological applications. In addition, many existing DTA models lack disease-specific adaptability and interpretability, rendering them unsuitable for translational applications in oncology. To address these constraints, the current study introduces a modular, sequence-driven deep learning system that predicts binding affinity directly from raw Simplified Molecular Input Line Entry System (SMILES) strings and amino acid sequences, without the need for structure input. The full DAVIS dataset was used to benchmark several architectures, such as Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), and RNN and LSTM hybrid systems. Performance of the model was evaluated using robust metrics such as Mean Squared Error (MSE), Concordance Index (CI), Area Under the Precision-Recall Curve (AUPR), modified R² (R²ₘ), Area Under the ROC Curve (AUROC), sensitivity and specificity. The results show that the RNN and LSTM architectures outperform CNN in capturing dependencies over long periods and enhancing sensitivity. To improve therapeutic relevance, the model was fine-tuned on a curated cancer-specific subset of Drug Affinity Validation of Interaction via Structure (DAVIS), and then extended into a hybrid CNN, RNN, LSTM, and Multi-Head Attention architecture. This final model showed increased ranking fidelity, interpretability, and predictive robustness in cancer-specific, molecularly diverse environments. The current study fills the major gaps in DTA prediction research and presents a biologically informed, interpretable, and scalable deep learning approach with significant potential for personalised treatment screening and precision oncology.