OPUS-Rota4

OPUS-Rota4 performs gradient-based refinement combined with deep-learning predictions to model protein side-chain dihedral angles and side-chain contact maps, improving side-chain conformations for studies of protein folding and protein design.

Key Features:

• Modular architecture: Composed of three modules—OPUS-RotaNN2, OPUS-RotaCM, and OPUS-Fold2—that together predict, constrain, and refine side-chain conformations.
• OPUS-RotaNN2: Predicts protein side-chain dihedral angles as initial estimates for modeling.
• OPUS-RotaCM: Measures distance and orientation information between side chains of residue pairs to generate spatial constraints.
• OPUS-Fold2: Refines side-chain conformations using gradient-based optimization implemented in Python with TensorFlow 2.4 and supports integration of differentiable energy terms.
• Gradient-based refinement: Converts side-chain modeling into a side-chain contact map prediction task and applies gradient-based optimization augmented by deep learning predictors instead of discrete rotamer-library sampling.
• Constraint-driven modeling: Uses distance and orientation constraints derived from OPUS-RotaCM together with predicted dihedral angles from OPUS-RotaNN2 to guide refinement in OPUS-Fold2.

Scientific Applications:

• Improving side-chain accuracy: Demonstrated closer-to-native side-chain conformations than Alphafold2 in a comparison of 15 CASP14 free-modeling predictions.
• Protein structure prediction: Enhances the accuracy of predicted side-chain conformations within computational protein structure models.
• Protein folding and design studies: Provides refined side-chain conformations and contact maps useful for investigating protein folding mechanisms and informing protein design.

Methodology:

OPUS-Rota4 converts side-chain modeling into a side-chain contact map prediction task; OPUS-RotaNN2 predicts dihedral angles; OPUS-RotaCM derives inter-side-chain distance and orientation constraints; OPUS-Fold2 performs gradient-based refinement in Python using TensorFlow 2.4 with support for differentiable energy terms.