Predicting strain-specific metabolic capabilities in the Genus Pseudomonas using a Flux-to-AI approach unravels hidden cell envelope properties.

Abstract

Bacteria from the Pseudomonas genus are omnipresent in air, soil, and water. They have been widely studied for their broad metabolic versatility and presence in the epidemiological chain, bioproduction, bioremediation, and disease processes. Each year, more genomic sequences are reported in databases and repositories. However, the relationship between the genomic variability of Pseudomonas strains and the diversity in their metabolic capabilities under environmentally relevant phenotypes remains unknown. Additionally, predictive tools for the analysis of different strains in a systematic framework are limited. Here, we reconstructed genome-scale metabolic models (GEMs) of 44 Pseudomonas strains from various environments and investigated their capabilities to metabolize different carbon sources and metabolic intermediaries. By systematically testing the substrate utilization of the models, we demonstrate how GEM-predicted capabilities can differentiate between strains and that high metabolic versatility is associated with the ability of the strains to remove toxic compounds while maintaining core functionalities. Hundreds of model simulations were used as input for a classification schema that uses machine learning algorithms, resulting in the identification of metabolic capabilities that better differentiate between species. Interestingly, transcription and expression models validated these findings by showing how pathways in which those metabolites are members change their proteome allocation across strains (e.g. phenylalanine metabolism as well as in the carbohydrate metabolism). Author summaryA central challenge in understanding how Pseudomonas strains can play symbiotic, competitive, and pathogenic roles depending on their environment can be potentially addressed with advanced computational tools that combine systems biology and machine learning approaches. Remarkable progress in understanding relationships between metabolism and phenotypes of Pseudomonas has been achieved through the collection of multi-omics datasets (e.g. genomic, transcriptomic, proteomics, fluxomics, etc.) from different settings such as air, clinical, soil, and water. However, maximum utilization of those tools has been limited by the lack of computational tools that can capture the metabolism at genome-scale. In this work, we are making available source genome-scale metabolic models for the most studied Pseudomonas strains and generated flux predictions across 44 strains under 425 different growth conditions by changing the carbon source in turn. Predicted growth phenotypes were used as input for machine learning to identify critical metabolic activities that change among strains. Computational resources leveraged here will enable deeper understanding of the commonalities and differences triggering the metabolic capabilities of Pseudomonas strains.