TolRad: A model for predicting radiation tolerance using Pfam annotations identifies novel radiosensitive bacterial species from reference genomes and MAGs

Abstract

The trait of ionizing radiation (IR) tolerance is variable between bacterial species, with radiosensitive bacteria succumbing to acute doses around 100Gy and extremophiles able to survive doses exceeding 10,000Gy. While survival screens have identified multiple highly radioresistant bacteria, such systemic searches have not been conducted for radiosensitive bacteria. The taxonomy-level diversity of IR of intolerance across bacteria is poorly understood, as are genetic elements that influence IR sensitivity. Using the protein domain frequencies from 61 bacterial species with experimentally determined IR D10 values (the dose at which only 10% of the population survives) we trained TolRad, a random forest binary classifier, to distinguish between radiosensitive bacteria (D10 < 200Gy) and radiation tolerant bacteria (D10 > 200Gy). On the hidden species, TolRad had an accuracy of 0.900. We applied TolRad to 152 UniProt-hosted bacterial proteomes, including 37 strains from the ATCC Human Microbiome Collection, and classified 34 species as radiosensitive. Whereas IR intolerance (D10 < 200Gy) in the training dataset had been confined to the phylum Proteobacterium , this initial TolRad screen identified radiosensitive bacteria in 2 additional phyla. We experimentally validated the predicted radiosensitivity of a key species of the human microbiome from the Bacteroidota phyla. To demonstrate that TolRad can be applied to Metagenome-Assembled Genome (MAGs), we tested the accuracy of TolRad on Egg-NOG assembled proteomes (0.965) and partial proteomes. Finally, three collections of MAGs were screened using TolRad, identifying further phylum with radiosensitive species and suggesting that environmental conditions influence the abundance of radiosensitive bacteria. Importance: Bacterial species have vast genetic diversity, allowing for life in extreme environments and the conduction of complex chemistry. The ability to harness the full potential of bacterial diversity is hampered by the lack of high-throughput experimental or bioinformatic methods for characterizing bacterial traits. Here, we present a computational model that uses de novo generated genome annotations to classify a bacterium as tolerant of ionizing radiation (IR) or as radiosensitive. This model allows for rapid screening of bacterial communities for low-tolerance species that are of interest for both mechanistic studies into bacterial sensitivity to IR and biomarkers of IR exposure.