From Proteins to Ligands: Decoding Deep Learning Methods for Binding Affinity Prediction.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Accurate in silico prediction of protein-ligand binding affinity is crucial for early-stage drug discovery.
• Current deep learning methods lack generalizability compared to conventional techniques like giga-docking.
• Understanding model learning from protein and ligand data is key to improving generalizability.

Purpose of the Study:

• To systematically investigate a sequence-based deep learning framework for predicting binding affinities.
• To assess the impact of various protein and ligand encodings on prediction accuracy.
• To evaluate the influence of different protein contact generation methods and ligand perturbations.

Main Methods:

• Utilized convolutional neural network (CNN) and graph neural network (GNN) based encodings for proteins and ligands.
• Employed protein contact maps generated by AlphaFold2, Pconsc4, and ESM-1b, alongside a random control.
• Tested ligand encodings by randomizing node and edge properties to assess data reliance.

Main Results:

• Protein encodings showed no substantial impact on binding affinity predictions across multiple metrics (concordance index, Pearson's R, Spearman's Rank, RMSE) for the KIBA dataset.
• Significant differences were observed with ligand encodings, particularly when using random ligands or random ligand node properties.
• Combining protein and ligand encodings in different ways did not yield significant performance improvements.

Conclusions:

• Deep learning models for binding affinity prediction demonstrate a stronger reliance on ligand data features than on protein encodings.
• The choice of protein contact generation method has a limited effect on prediction performance.
• Further research should focus on optimizing ligand representation and feature engineering to enhance model generalizability in drug discovery.