CS-DTA: a language model-driven framework for robust drug-target affinity prediction under strict cold-start scenarios.

Introduction: Accurate prediction of drug-target affinity is important for computational drug discovery, yet many deep learning models show limited robustness when applied to unseen compounds or previously uncharacterized proteins. Methods: We developed CS-DTA, a modular DTA prediction framework that integrates large language models (LLMs) for compound and protein representation learning with a cross-modal interaction module. By leveraging the strong transferability of LLM-based encoders, the model can capture rich semantic and structural patterns from both molecular sequences and protein sequences. The cross-modal attention mechanism further enhances prediction by explicitly modeling fine-grained interactions between compound and protein representations, enabling more effective fusion of complementary biochemical information. Results: CS-DTA achieved state-of-the-art predictive performance across warm-start and strict cold-start scenarios. Ablation analyses showed that the dual-encoder representation backbone was the primary source of predictive robustness. Interpretability analyses highlighted compact ligand substructures and localized protein regions with plausible relevance to binding. In downstream screening, CS-DTA prioritized known target-specific interactions, and selected predictions were supported by molecular docking. The non-kinase validation provided preliminary evidence of partial transferability beyond the kinase-focused benchmark setting. Discussion: The above results demonstrate that CS-DTA offers a robust and interpretable approach for DTA prediction. By integrating LLM-based encoding with a modular interaction architecture, the framework enhances generalization to previously unseen compounds and proteins while maintaining biologically meaningful interpretability. This combination makes CS-DTA a promising framework for virtual screening and early-stage drug discovery.