Enhancing peptide identification in metaproteomics through curriculum learning in deep learning

Area of Science:

• Microbiology
• Computational Biology
• Biochemistry

Background:

• Metaproteomics enables the study of microbial community functions, but challenges exist in accurately identifying peptides.
• Large, incomplete protein databases from metagenomes create computational bottlenecks in peptide-spectrum match (PSM) filtering.

Purpose of the Study:

• To develop and evaluate WinnowNet, a deep learning method for enhanced PSM filtering in metaproteomics.
• To address the computational challenges posed by large metagenomic databases.

Main Methods:

• WinnowNet utilizes deep learning, with variants based on transformers and convolutional neural networks.
• A curriculum learning strategy trains models from simple to complex examples, handling unordered PSM data.

Main Results:

• WinnowNet achieves higher true identifications at equivalent false discovery rates compared to existing tools like Percolator, MS2Rescore, and DeepFilter.
• The method outperforms filters integrated into standard metaproteomic analysis pipelines.
• WinnowNet identified more gut microbiome biomarkers associated with diet and health.

Conclusions:

• WinnowNet offers a significant advancement in PSM filtering for metaproteomics.
• The tool enhances the ability to analyze complex microbial communities and identify biomarkers.
• WinnowNet holds potential for advancing personalized medicine through improved microbiome analysis.