Alternative Conformation Prediction Using Deep Learning With Multi-MSA Strategy and Structural Clustering in CASP16.

Area of Science:

• Structural Biology
• Computational Biology
• Bioinformatics

Background:

• Accurate prediction of protein and nucleic acid structures is crucial for understanding biological function.
• Ensemble prediction, which captures dynamic conformational changes, is essential for a complete structural picture.
• Existing methods often struggle to represent the full conformational landscape of biomolecules.

Purpose of the Study:

• To evaluate the performance of the EnsembleFold pipeline for structure ensemble predictions in CASP16.
• To assess the efficacy of deep learning methods and clustering for capturing diverse conformational states.
• To compare EnsembleFold's performance against established tools like AlphaFold3.

Main Methods:

• Utilized DeepMSA2 and rMSA for generating multiple sequence alignments (MSAs).
• Employed deep learning models (D-I-TASSER2, DMFold2, ExFold, DeepProtNA) for initial structure decoy generation.
• Applied MolClust for structural clustering and replica-exchange Monte Carlo (REMC) simulations for refinement.

Main Results:

• EnsembleFold achieved an average TM-score of 0.657 for 19 CASP16 ensemble targets, a 10.2% improvement over AlphaFold3.
• Demonstrated strong performance on hybrid protein/nucleic-acid targets.
• Identified that distinct MSAs, REMC simulations, and structural clustering contribute to accurate ensemble predictions.

Conclusions:

• The EnsembleFold pipeline significantly enhances structure ensemble prediction accuracy and conformational diversity.
• Deep learning integration with advanced refinement and clustering offers a powerful approach for biomolecular modeling.
• Future improvements in Quality Assessment (QA) scoring can further boost the reliability of ensemble predictions.