DeepFold

DeepFold improves protein structure prediction accuracy by refining backbone and side-chain modeling through optimized loss functions, conditional random field-based template alignment, and conformational space annealing with molecular mechanics energy functions.

Key Features:

• Optimized loss functions: Modifies loss functions to better capture side-chain torsion angles and frame-aligned point errors and adds losses for side-chain confidence and secondary structure prediction.
• Conditional random field template alignment: Replaces traditional template feature generation with an alignment method based on conditional random fields to improve template quality for modeling.
• Re-optimized energy function: Employs conformational space annealing integrating molecular mechanics energy functions that use potential energies derived from distogram and side-chain prediction models.
• AlphaFold2 basis: Extends the AlphaFold2 foundation by introducing the above enhancements to backbone and side-chain modeling.

Scientific Applications:

• CASP15 benchmarking: In the CASP15 blind test for single protein and domain modeling (109 domains), DeepFold ranked fourth among 132 participating groups with a median GDT-TS of 88.64.
• Backbone accuracy comparison: DeepFold achieved a median GDT-TS of 88.64, exceeding AlphaFold2's reported 85.88 for the same benchmark in backbone accuracy.
• TBM-easy and TBM-hard performance: Demonstrated strong performance across TBM-easy and TBM-hard targets based on Z-scores for GDT-TS.
• Side-chain and stereochemistry evaluation: Analysis of 55 domains from 39 publicly available structures showed improved side-chain accuracy and better MolProbity scores compared with other top groups.

Methodology:

Modifies loss functions to capture side-chain torsion angles and frame-aligned point errors and adds side-chain confidence and secondary structure losses; replaces template feature generation with a conditional random field alignment method; and re-optimizes the energy function using conformational space annealing integrating molecular mechanics energy functions with potential energies derived from distogram and side-chain prediction models.