ConforFold recovers alternative protein conformations beyond MSA subsampling

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Protein conformational changes are crucial for function, but current prediction tools struggle to model diverse structural ensembles.
• Systematic sampling of protein structural ensembles remains a significant challenge in computational biology.

Purpose of the Study:

• To develop a novel computational framework for predicting multiple protein conformational states.
• To improve the recovery of alternative protein conformations using deep learning and secondary structure information.

Main Methods:

• ConforPSSP generates diverse 8-state protein secondary structure predictions (PSSPs) using a transformer model.
• ConforFold, a retrained OpenFold model, uses PSSPs to condition structure prediction for conformer recovery.
• The framework integrates secondary-structure sampling into deep learning for enhanced ensemble prediction.

Main Results:

• ConforFold achieved state-of-the-art performance in recovering protein conformers, identifying both states in 84% of test cases (TM-scores ≥0.8).
• Outperformed diffusion-based (AlphaFlow) and MSA clustering (Cfold) methods in conformer recovery.
• Demonstrated superior performance over BioEmu in cases with significant secondary structure differences between conformers (83% vs. 76%).

Conclusions:

• ConforFold provides a broadly applicable framework for modeling protein structural ensembles.
• Explicitly integrating secondary structure allows recovery of conformations missed by other methods.
• Offers a new approach for studying protein conformational heterogeneity and function.