PointSite

PointSite identifies ligand binding atoms on protein structures by segmenting atom-level point clouds to enable precise ligand binding site characterization.

Key Features:

• Atom-Level Identification: Identifies ligand binding atoms directly on protein structures, avoiding reliance on pseudo surface points (PSPs) outside the protein.
• Point Cloud Transformation: Converts 3D protein structures into point clouds to represent atomic positions and local connectivity for analysis.
• Submanifold Sparse Convolution U-Net: Uses a Submanifold Sparse Convolution (SSC)-based U-Net architecture for fine-grained atom-level segmentation capturing local and global 3D geometry.
• Performance Metrics: Optimizes atom Intersection over Union (IoU) as a primary segmentation performance measure.
• False Positive Reduction: Acts as a filter on pocket-centric predictions to reduce false positives among ligand binding site candidates.
• Cascaded Filtering and Re-ranking: Employs cascaded filtering and re-ranking using segmented atoms, reporting improved performance on benchmarks and challenging CAMEO targets using Distance Contact Area (DCA) criteria.

Scientific Applications:

• Protein function analysis: Enables atom-level mapping of ligand interactions to support interpretation of protein function and molecular mechanisms.
• Structure-based drug design: Provides precise atom-level binding site identification to prioritize and refine ligand docking and lead optimization.

Methodology:

Convert 3D protein structures to point clouds, segment atoms using a Submanifold Sparse Convolution (SSC)-based U-Net, apply cascaded filtering and re-ranking of pocket-centric candidates using segmented atoms, and evaluate performance via atom Intersection over Union (IoU) and Distance Contact Area (DCA) on benchmarks including CAMEO targets.