BindPred: a framework for predicting protein-protein binding affinity from language model embeddings.

Area of Science:

• Computational Biology
• Bioinformatics
• Drug Discovery

Background:

• Accurate prediction of protein-protein binding affinities is crucial for molecular biology and drug development.
• Existing computational methods often require 3D structural models, which are not always available.
• There is a need for structure-agnostic approaches to predict binding affinities.

Purpose of the Study:

• To develop a novel framework, BindPred, for predicting protein-protein binding affinities.
• To enable predictions directly from amino acid sequences, bypassing the need for 3D structures.
• To facilitate large-scale computational screening of protein interactions.

Main Methods:

• BindPred utilizes a structure-agnostic input framework.
• It combines embeddings from large protein language models with gradient boosting trees.
• The model is trained and validated on the protein-protein binding (PPB)-Affinity benchmark.

Main Results:

• BindPred achieved a Pearson correlation coefficient of 0.86 on the PPB-Affinity benchmark using cross-validation.
• Evolutionary embeddings were found to capture most predictive signals, with minimal improvement from physics-based energy terms.
• The model demonstrated robust generalization to novel protein interactions, even with stringent protein-level splits.
• Rapid inference allows for proteome-scale screening, with approximately 3 million complexes predicted per GPU hour.

Conclusions:

• BindPred offers a computationally efficient and accurate method for predicting protein-protein binding affinities from sequences alone.
• The structure-agnostic approach significantly expands the applicability of binding affinity prediction.
• This method holds promise for accelerating drug discovery and understanding molecular biology.