SASA-Net

SASA-Net predicts protein three-dimensional (3D) structures directly from estimated inter-residue distances using a spatial-aware self-attention deep learning model.

Key Features:

• Direct Structure Learning: Learns 3D protein structure directly from estimated inter-residue distances, bypassing handcrafted potential functions.
• Residue Pose Representation: Represents protein structure via a per-residue coordinate frame that fixes backbone atom positions to define residue poses.
• Spatial-Aware Self-Attention: Uses a spatial-aware self-attention mechanism that adjusts residue poses based on features from all other residues and their inter-residue distances.
• Iterative Refinement: Applies the spatial-aware self-attention mechanism iteratively to progressively refine the predicted 3D structure.
• End-to-End Neural Network Integration: Integrates with a neural network that predicts inter-residue distances to form an end-to-end prediction model.

Scientific Applications:

• CATH35 Benchmarking: Demonstrated accurate and efficient reconstruction of representative proteins from the CATH35 dataset.
• Structural Biology: Provides predicted 3D structures useful for interpreting protein function and fold analysis.
• Drug Discovery: Supplies structural models that can inform structure-based drug design and ligand modeling.
• Molecular Mechanism Studies: Enables modeling of structural changes relevant to understanding molecular mechanisms.

Methodology:

Inputs are estimated inter-residue distances; each residue is represented by a pose with fixed backbone atom coordinates; a spatial-aware self-attention mechanism updates residue poses using features from all residues and inter-residue distances with iterative refinement; the system is integrated with a neural network that predicts inter-residue distances for end-to-end structure prediction.