Protein Domain Segmentor

Protein Domain Segmentor segments multi-domain protein structures at the residue level by classifying residues into 38 CATH architecture classes using convolutional neural networks to enable domain segmentation and fold-quality assessment.

Key Features:

• Semantic Segmentation: Residues are classified into 38 CATH architecture classes for residue-level domain segmentation.
• Accuracy Metrics: Achieves 90.8% per-residue accuracy, 95.0% average per-class accuracy, and 87.8% average per-structure accuracy.
• Fold Conformity Assessment: Class probabilities quantify structural similarity to known folds and identify non-conformative regions.
• Structural Sampling Guidance: Highlights non-conforming regions to inform design and prediction workflows.

Scientific Applications:

• Structural Prediction: Identifies candidate structures that conform to native folds or indicate non-native regions for prediction assessment.
• De Novo Design: Enables exploration of novel protein designs via analysis of structural motifs and fold conformity at residue resolution.

Methodology:

A convolutional neural network is trained on multi-domain protein data to classify residues into 38 CATH architecture classes for fold identification and structural-quality assessment.