ProteoStorm: An Ultrafast Metaproteomics Database Search Framework

Abstract

Shotgun metaproteomics has the potential to reveal the functional landscape of microbial communities but lacks appropriate methods for complex samples with unknown compositions. In the absence of prior taxonomic information, tandem mass spectra would be searched against large pan-microbial databases, which requires heavy computational workload and reduces sensitivity. We present ProteoStorm, an efficient database search framework for large-scale metaproteomics studies, which identifies high-confidence peptide-spectrum matches (PSMs) while achieving a two-to-three orders-of-magnitude speedup over popular tools. A reanalysis of a urinary tract infection (UTI) dataset of 110 individuals revealed a complex pattern of polymicrobial expression, including sub-types of UTIs, cases of bacterial vaginosis, and evidence of no underlying disease. Importantly, compared to the initial UTI study that restricted the search database to a manually curated list of 20 genera, ProteoStorm identified additional genera that were previously unreported, including a case of infection with the rare pathogen Propionimicrobium.

Keywords: LC-MS/MS; metaproteomics; microbial communities; microbiology; proteome informatics; proteomics.

Figure 1.. ProteoStorm search framework: performance and scalability

(a) Each stage of ProteoStorm is composed of three modules: i) data partitioning, ii) peptide filtering, and iii) p-value computation. Identifications from a fully-tryptic search are used to construct a refined protein database for a semi-tryptic search. PSMs with p-values are reported. (b) In the first module, database and spectra partitioning dramatically reduces search space from all peptides (black dotted lines) to peptides within spectra parent mass ranges (red vertical lines). (c) In the second module, spectrum-peptide pairs are filtered based on shared counts of prominent spectral peaks (red) to b-/y-ions of peptides (numbers within blocks) using an ion-mass indexing data structure. Each colored block represents a unique peptide within the parent mass tolerance of a given spectrum. (d) 946,845 spectra were searched against the UniProtKB database using ProteoStorm, MSGF+, Comet, and MSFragger. ProteoStorm required 9.7 CPU-hours, while other tools required CPU-weeks to complete. (e) Breakdown of ProteoStorm runtime by module. S1 and S2 represent the two different stages of ProteoStorm. See also Figure S3.

Figure 2.. ProteoStorm identifies bi-clusters of individuals with similar microbial compositions

Searching the full UTI dataset against the RefUP++ database (2,259 genera) using a genera-restriction approach, ProteoStorm identified 64 genera. Out of 73,092 peptides, 28.5% (20,833) mapped uniquely to a single genus. Four bi-clusters (white boxes) were inferred from clusters with an approximately unbiased (au) p-value greater than 0.90 (magenta boxes), indicating a complex pattern of polymicrobial expression, including sub-types of urinary tract infections, cases of bacterial vaginosis, and evidence of no underlying disease. Pathology groups: Healthy, ERY (erythrocyte/vascular injury), EXF (exfoliation of squamous epithelial and urothelial cells), and UTI (urinary tract infection). See also Table S1 and Table S2.