Deep Local Analysis deconstructs protein-protein interfaces and accurately estimates binding affinity changes upon...

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in structural biology

Background:

• Advances in protein structure prediction enable large-scale interactome reconstruction.
• Accurate modeling of sequence variation effects on protein-protein interactions is crucial.

Purpose of the Study:

• To introduce Deep Local Analysis (DLA), a novel deep learning framework for predicting the impact of mutations on protein binding affinity.
• To evaluate DLA's performance and generalization capabilities on unseen protein complexes.

Main Methods:

• DLA deconstructs protein interfaces into residue-centered 3D cubes.
• It employs 3D convolutions to recognize patterns within these cubes.
• Predictions are made using wild-type and mutant residue cubes to estimate binding affinity changes.

Main Results:

• DLA accurately estimates binding affinity changes, achieving a Pearson correlation coefficient of 0.735 on approximately 400 mutations.
• The framework demonstrates superior generalization on blind datasets compared to state-of-the-art methods.
• Incorporating evolutionary constraints enhances prediction accuracy.

Conclusions:

• DLA provides an efficient and accurate method for predicting mutation effects on protein binding affinity.
• The DLA framework is versatile and applicable to various structural bioinformatics tasks, including residue identification and complex function prediction.