Challenge for Deep Learning: Protein Structure Prediction of Ligand-Induced Conformational Changes at Allosteric and...

Area of Science:

• Computational biology
• Structural biology
• Drug discovery

Background:

• Protein structure dictates function and drug interactions.
• Traditional structure determination methods are resource-intensive.
• Deep learning shows promise in predicting static protein structures.

Purpose of the Study:

• To evaluate deep learning models' ability to predict dynamic protein conformational changes upon ligand binding, particularly at allosteric sites.
• To assess the accuracy of predicting allosteric induced-fit conformations compared to orthosteric ones.
• To provide a curated dataset and evaluation framework for advancing protein structure prediction.

Main Methods:

• Curated a dataset of 578 X-ray structures with orthosteric and allosteric binding.
• Evaluated deep learning methods including AlphaFold2, NeuralPLexer, and RoseTTAFold All-Atom.
• Assessed prediction accuracy for static structures and dynamic conformational changes.

Main Results:

• Deep learning models accurately predict orthosteric bound conformations.
• Predicting allosteric induced-fit conformations remains a challenge.
• Modifying MSA depth in AlphaFold2 increased conformational diversity but not allosteric fit accuracy.

Conclusions:

• Deep learning shows potential for predicting dynamic protein changes but has limitations.
• Allosteric site binding prediction requires further algorithmic development.
• The study provides valuable resources for the protein structure prediction community.