Multi-scale structural analysis of proteins by deep semantic segmentation.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in structural biology

Background:

• Protein structure prediction and de novo design are advancing rapidly due to computational methods.
• Identifying candidate structures that lead to native folds or are designable remains challenging.
• Existing metrics often fail to capture high-level structural features like architectures, folds, and conserved motifs.

Purpose of the Study:

• To apply semantic segmentation using Convolutional Neural Networks (CNNs) to protein structures.
• To develop a novel strategy for protein fold identification and structure quality assessment.
• To create a tool that aids in guiding structural sampling for protein prediction and design.

Main Methods:

• Trained a CNN to assign each residue in multi-domain proteins to one of 38 CATH architecture classes.
• Utilized semantic segmentation, a technique from image classification, for protein structure analysis.
• Evaluated model performance on a test set, achieving high per-residue accuracy.

Main Results:

• The CNN model achieved 90.8% per-residue accuracy on the test set.
• Demonstrated that class probabilities from the CNN can serve as a metric for fold conformity.
• Showed that the model can identify non-conformative regions within known protein classes.

Conclusions:

• The developed CNN-based semantic segmentation approach offers a novel strategy for fold identification and quality assessment.
• The model's ability to quantify fold conformity and identify deviations provides a powerful tool for structural biology.
• This method facilitates improved structural sampling for both protein structure prediction and de novo protein design.