Training bias and sequence alignments shape protein-peptide docking by AlphaFold and related methods

Area of Science:

• Structural biology
• Computational biology
• Bioinformatics

Background:

• Protein-peptide interactions are crucial for biological processes.
• Accurate structural models are vital for understanding function and designing inhibitors.
• Computational models like AlphaFold3 show promise for predicting these structures.

Purpose of the Study:

• To analyze the performance of four protein-peptide structure prediction models: AlphaFold2-Multimer, AlphaFold3, Boltz-1, and Chai-1.
• To investigate how these models utilize multiple sequence alignments (MSAs) for predictions.
• To identify limitations and areas for improvement in deep learning-based peptide docking.

Main Methods:

• Evaluation of four protein-peptide structure prediction models using a dataset of experimentally resolved structures.
• Analysis of model performance regarding generalization to novel protein-peptide complexes.
• Investigation of the contribution of protein and peptide unpaired and paired MSAs to prediction accuracy.

Main Results:

• Models demonstrated high accuracy but exhibited bias towards previously observed structures, limiting generalization to novel complexes.
• Shallow or poor-quality MSAs for peptides were noted.
• Weak evidence for the use of coevolutionary information from paired MSAs was found.
• Both protein and peptide unpaired MSAs significantly contributed to prediction accuracy.

Conclusions:

• Deep learning models show significant promise for protein-peptide docking.
• Model performance is influenced by the quality and diversity of training data, particularly interface geometries.
• Future improvements require addressing model biases and enhancing the representation of novel interactions in training datasets.