DISTEMA

DISTEMA predicts the quality of single protein structural models by applying an attentive 2D convolutional neural network to residue-residue distance map differences (inter-residue distance maps) to estimate model accuracy.

Key Features:

• Deep Learning Framework: Employs an attentive 2D convolutional neural network with channel-wise attention that processes raw input without relying on expert-curated features.
• Input Utilization: Uses raw difference maps between inter-residue distance maps calculated from a protein model and distance maps predicted from the protein sequence.
• Network Architecture: Combines multiple convolutional layers, batch normalization layers, dense layers, and Squeeze-and-Excitation blocks with attention mechanisms to automatically extract features relevant to model quality.
• Performance Evaluation: Evaluated on CASP13 targets with a reported ranking loss of 0.079 in terms of GDT-TS, demonstrating performance superior to several state-of-the-art single-model quality assessment methods.

Scientific Applications:

• Structural Biology: Assesses protein model quality to support selection of accurate models for structural biology studies, interpretation of biological function, and design of therapeutic interventions.
• Bioinformatics: Supports selection of high-quality single protein structural models for downstream bioinformatics analyses and applications.

Methodology:

Uses raw difference maps derived from inter-residue distance comparisons; applies an attentive CNN architecture including channel-wise attention, convolutional layers, batch normalization, dense layers, and Squeeze-and-Excitation blocks for automatic feature extraction; predicts protein model quality and was evaluated on CASP13 using ranking loss in terms of GDT-TS.