LLM-Enhanced Knowledge Distillation for Sequence-Based Protein-Ligand Interaction Prediction.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Accurate protein-ligand interaction prediction is crucial for drug discovery.
• Existing structure-based methods face scalability and robustness limitations.
• Structure-limited settings hinder the application of many computational approaches.

Purpose of the Study:

• To introduce a sequence-based framework, Multi-Combinatorial Knowledge Distillation (MCKD), for protein-ligand interaction prediction.
• To overcome the limitations of structure-based methods by not requiring 3D structures at inference time.
• To enable scalable and robust prediction for structure-limited and data-limited drug discovery scenarios.

Main Methods:

• MCKD represents proteins and ligands as 2D molecular graphs from sequences and physicochemical properties.
• A hybrid distillation strategy combines cross-modal and self-distillation to integrate structural knowledge.
• A bilinear attention network models residue-atom level interactions for affinity regression and classification.

Main Results:

• MCKD consistently outperforms existing sequence-based methods on benchmark datasets.
• Performance is comparable to structure-based approaches, demonstrating effectiveness.
• The model shows good generalization to unseen proteins and novel ligand scaffolds.

Conclusions:

• MCKD provides a scalable and effective solution for protein-ligand interaction prediction.
• The framework is particularly valuable for structure-free and data-limited drug discovery.
• MCKD offers interpretable insights into critical molecular interaction regions.