Challenging AlphaFold in predicting proteins with large-scale allosteric transitions.

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Proteins often exist in multiple conformations to perform functions.
• Existing protein structure prediction models are typically trained on static structures, limiting their ability to capture conformational dynamics.
• Autoinhibited proteins represent a class of proteins with inherent conformational flexibility, existing in equilibrium between active and inactive states.

Purpose of the Study:

• To benchmark the performance of AlphaFold2, AlphaFold3, and related variants in predicting the structures of autoinhibited proteins.
• To assess the capability of these models in capturing protein conformational diversity.
• To identify potential improvements for protein structure prediction of dynamic protein systems.

Main Methods:

• Benchmarking AlphaFold2, AlphaFold3, and variants on experimentally determined structures of autoinhibited proteins.
• Evaluating prediction accuracy and confidence scores against experimental data.
• Analyzing the impact of different subsampling strategies (uniform vs. local) on AlphaFold2's performance in capturing conformational diversity.

Main Results:

• AlphaFold2 demonstrated significant limitations in accurately predicting the experimental structures of many autoinhibited proteins, often resulting in lower confidence scores.
• In contrast, AlphaFold2 showed high accuracy and confidence for non-autoinhibited multi-domain proteins.
• While AlphaFold3 and BioEmu showed improvements over AlphaFold2, they still faced challenges in precisely reproducing experimental structural details.
• Uniform subsampling improved AlphaFold2's ability to capture conformational diversity compared to local subsampling.

Conclusions:

• Predicting the structures of autoinhibited proteins remains a significant challenge for current protein structure prediction tools.
• The complex energy landscapes governing the conformational dynamics of these proteins pose persistent difficulties for computational modeling.
• Further advancements are needed to enhance the accuracy and reliability of protein structure prediction for proteins exhibiting conformational heterogeneity.