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

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning in drug discovery

Background:

• Accurate drug-target affinity (DTA) prediction is crucial for computational drug discovery.
• Existing deep learning models often lack robustness when applied to novel compounds or proteins.

Purpose of the Study:

• To develop a robust and interpretable framework for DTA prediction.
• To enhance generalization capabilities for unseen chemical and biological entities.

Main Methods:

• Developed CS-DTA, a modular framework integrating large language models (LLMs) for representation learning.
• Utilized a cross-modal interaction module with attention mechanisms for fine-grained compound-protein interactions.
• Employed LLM-based encoders for capturing semantic and structural patterns from molecular and protein sequences.

Main Results:

• CS-DTA achieved state-of-the-art performance in both warm-start and strict cold-start scenarios.
• Ablation studies identified the dual-encoder backbone as key to predictive robustness.
• Interpretability analyses revealed biologically relevant ligand substructures and protein regions.
• Downstream screening demonstrated CS-DTA's ability to prioritize known target-specific interactions.

Conclusions:

• CS-DTA provides a robust and interpretable approach for DTA prediction, enhancing generalization.
• The integration of LLM-based encoding and modular interaction architecture makes CS-DTA promising for virtual screening and early-stage drug discovery.
• The framework shows potential for advancing computational drug discovery pipelines.