U-Net-based reactive center loop-identifier for serpins.

Area of Science:

• Proteomics
• Bioinformatics
• Structural Biology

Background:

• Serine protease inhibitors (serpins) are crucial proteins with diverse functions in animals.
• Despite a conserved folding scheme, identifying functional specificity, particularly the reactive center loop (RCL), is challenging due to high variability.
• Current annotation methods fail to identify the RCL in most of the over 48,000 known serpins.

Purpose of the Study:

• To develop an automated method for accurate annotation of the serpin reactive center loop (RCL).
• To overcome the limitations of traditional motif-matrix-based annotation strategies for RCL identification.

Main Methods:

• Tested various neural network architectures and encoding strategies for automatic RCL annotation.
• Utilized an expert-annotated training/validation dataset for model development.
• Evaluated model performance on an independent test dataset.

Main Results:

• U-Net-based neural network models demonstrated superior performance for RCL annotation.
• Achieved approximately 98% accuracy in identifying the RCL at the per-sequence level on an independent test set.
• The developed models significantly enhance the ability to annotate serpin RCLs.

Conclusions:

• Neural network-based approaches, specifically U-Net models, are highly effective for automatic serpin RCL annotation.
• This advancement addresses a critical gap in serpin research, enabling better functional characterization.
• The developed tools and datasets are publicly available to facilitate further research.