Proteomics of natural bacterial isolates powered by deep learning-based de novo identification.
Metaproteomics has been increasingly utilized for high-throughput molecular characterization in complex environments and has been demonstrated to provide insights into microbial composition and functional roles in soil systems. Despite its potential for the study of microbiomes, significant challenges remain in data analysis, including the creation of a sample-specific protein sequence database as the taxonomic composition of soil is often unknown. Almost all metaproteome analysis tools require this database and their accuracy and sensitivity suffer when the database is incomplete or contains extraneous sequences from organisms which are not present. Here, we leverage a de novo peptide sequencing approach to identify sample composition directly from metaproteomic data. First, we created a deep learning model, Kaiko, to predict the peptide sequences from mass spectrometry data, and trained it on 5 million peptide-spectrum matches from 55 phylogenetically diverse bacteria. After training, Kaiko successfully identified unsequenced soil isolates directly from proteomics data. Finally, we created a pipeline for metaproteome database generation using Kaiko. We tested the pipeline on native soils collected in Kansas, showing that the de novo sequencing model can be employed to construct the sample-specific protein database instead of relying on (un)matched metagenomes. Our pipeline identified all highly abundant taxa from 16S ribosomal RNA sequencing of the soil samples and also uncovered several additional species which were strongly represented only in proteomic data. Our pipeline offers an alternative and complementary method for metaproteomic data analysis by creating a protein database directly from proteomic data, thus removing the need for metagenomic sequencing. Significance StatementProteomic characterization of environmental samples, or metaproteomics, reveals microbial activity critical to our understanding of climate, nutrient cycling and human health. Metaproteomic samples originate from diverse environs, such as soil and oceans. One option for data analysis is a de novo interpretation of the mass spectra. Unfortunately, the current generation of de novo algorithms were primarily trained on data originating from human proteins. Therefore, these algorithms struggle with data from environmental samples, limiting our ability to analyze metaproteomics data. To address this challenge, we trained a new algorithm with data from dozens of diverse environmental bacteria and achieved significant improvements in accuracy across a broad range of organisms. This generality opens proteomics to the world of natural isolates and microbiomes.