High-throughput genomic feature extraction reveals environmental adaptations of prokaryotes

Abstract

Understanding the adaptations of microorganisms to their environment is key to predicting the stability and dynamics of microbial communities. To uncover molecular mechanisms of environmental response, we extracted genomic features from 13,554 prokaryotic isolates, and trained machine learning models to identify which ones are most strongly associated with the microbial salinity, temperature, oxygen, and pH preferences. To extract these features in high throughput, including gene families, non-coding RNAs (ncRNAs), oligonucleotides, and amino acid usage, we built FxTractor, a scalable and adjustable pipeline available at: https://github.com/MGXlab/FxTractor. We validated the performance of our models with experimental data from a newly isolated deep-sea extremophile belonging to the genus Limnochorda that is not well-represented among the ML training sets, showing strong agreement between predictions and the conditions used to isolate this strain. Our analysis revealed specific gene and ncRNA families associated with each of the four environmental parameters, uncovering both established and potentially new molecular mechanisms. Examples include the bacterial large Signaling Recognition Particle in isolates that are able to grow at high temperatures (≥55°C), suggesting a role in translational pausing and structural stability under thermal stress. We also found the anti-hemB ncRNA to be associated with low-salinity (<0.7% NaCl), indicating a conserved antisense mechanism regulating the energetic costs of heme biosynthesis. Together, these findings provide new insights into microbe-environment interactions, and show how FxTractor enables high throughput discovery of genomic associations. Importance Microbes have different growth preferences and understanding how this is reflected in the genome is central to microbial ecology and biotechnology. Genomic signatures of environmental adaptation, such as salinity, temperature, oxygen, and pH, may vary from specific encoded proteins and non-coding RNAs to nucleotide and amino acid usage. To analyse these, we developed FxTractor, a flexible open-source bioinformatics pipeline to extract genomic features, that may be readily extended with additional feature mining tools in the future. After applying it to >13,000 prokaryotic genomes, we used machine learning to identify which features were associated with growth preferences, recovering both known and previously uncharacterized mechanisms, including little-researched non-coding RNAs. This work provides both a reusable, extendable tool for genomic feature mining, and new biological insights into the molecular basis of environmental adaptation.