CompBind: Complex Guided Pretraining-Based Structure-Free Protein-Ligand Affinity Prediction.

Area of Science:

• Computational biology
• Drug discovery
• Machine learning

Background:

• Accurate protein-ligand binding affinity prediction is crucial for drug discovery.
• Current structure-based deep learning methods are limited by the scarcity and cost of experimental complex structures.

Purpose of the Study:

• To develop a novel framework, CompBind, for binding affinity prediction that does not require 3D structural inputs.
• To enable accurate predictions using only protein and ligand sequences.

Main Methods:

• CompBind integrates bidirectional cross-attention with a dual-objective pretraining strategy.
• Contrastive learning enforces feature consistency between monomers and complexes.
• Generative learning reconstructs interaction features for sequence-based inference.

Main Results:

• CompBind outperforms non-complex-based methods and rivals complex-based approaches in affinity prediction.
• The model shows strong performance in cold-start and sparse-label conditions.
• CompBind successfully identified known inhibitors in a drug repurposing case study targeting GPX4.

Conclusions:

• CompBind offers a scalable and generalizable solution for binding affinity prediction, decoupling accuracy from experimental structure availability.
• The framework accelerates drug discovery pipelines and enhances model interpretability through its attention mechanism.