Enhancing Peptide Identification in Metaproteomics through Curriculum Learning in Deep Learning

Metaproteomics offers a powerful window into the active functions of microbial communities, but accurately identifying peptides remains challenging due to the size and incompleteness of protein databases derived from metagenomes. These databases often contain vastly more sequences than those from single organisms, creating a computational bottleneck in peptide-spectrum match (PSM) filtering. Here we present WinnowNet, a deep learning–based method for PSM filtering, available in two versions: one using transformers and the other convolutional neural networks. Both variants are designed to handle the unordered nature of PSM data and are trained using a curriculum learning strategy that moves from simple to complex examples. WinnowNet consistently achieves more true identifications at equivalent false discovery rates compared to leading tools, including Percolator, MS$^2$Rescore, and DeepFilter, and outperforms filters integrated into popular analysis pipelines. It also uncovers more gut microbiome biomarkers related to diet and health, highlighting its potential to support advances in personalized medicine