Machine-learning classification suggests that many alphaproteobacterial prophages may instead be gene transfer agents

Abstract

21Many of the sequenced bacterial and archaeal genomes encode regions of viral provenance. Yet, not all of 22 these regions encode bona fide viruses. Gene transfer agents (GTAs) are thought to be former viruses that 23 are now maintained in genomes of some bacteria and archaea and are hypothesized to enable exchange of 24 DNA within bacterial populations. In Alphaproteobacteria, genes homologous to the 'head-tail' gene 25 cluster that encodes structural components of the Rhodobacter capsulatus GTA (RcGTA) are found in 26 many taxa, even if they are only distantly related to Rhodobacter capsulatus. Yet, in most genomes 27 available in GenBank RcGTA-like genes have annotations of typical viral proteins, and therefore are not 28 easily distinguished from their viral homologs without additional analyses. Here, we report a 'support 29 vector machine' classifier that quickly and accurately distinguishes RcGTA-like genes from their viral 30 homologs by capturing the differences in the amino acid composition of the encoded proteins. Our open-31 source classifier is implemented in Python and can be used to scan homologs of the RcGTA genes in 32 newly sequenced genomes. The classifier can also be trained to identify other types of GTAs, or even to 33 detect other elements of viral ancestry. Using the classifier trained on a manually curated set of 34 homologous viruses and GTAs, we detected RcGTA-like 'head-tail' gene clusters in 57.5% of the 1,423 35 examined alphaproteobacterial genomes. We also demonstrated that more than half of the in silico 36prophage predictions are instead likely to be GTAs, suggesting that in many alphaproteobacterial 37 genomes the RcGTA-like elements remain unrecognized. 38 39 Keywords 40 Virus exaptation, GTA, Rhodobacter capsulatus, support vector machine, binary classification, carbon 41 depletion 42 43 7 value < 0.001; query and subject overlap by at least 60% of their length) and PSI-BLASTP searches (E-132 value < 0.001; query and subject overlap by at least 40% of their length; maximum of six iterations) of 133 the viral RefSeq database release 90 (last accessed in November 2018; accession numbers of the viral 134 entries are provided in Supplementary Table S2). BLASTP and PSI-BLAST executables were from the 135 BLAST v. 2.6.0+ package (Altschul et al. 1997) . The obtained homologs are listed in Supplementary 136

Phylogenetic evidence of headful packaging strategy in gene transfer agents

“…In the phylogenetic tree of TerLs (Figure 1), a group of 616 TerLs was labeled as the "RcGTA-like containing group." This group includes 507 TerLs of the 526 RcGTA-like TerLs that were identified and curated by Kogay et al (2019). Nineteen RcGTA-like TerLs from the dataset of Kogay et al (2019) are absent in our dataset because they were not in GenBank at the time of our data collection (December 2018).…”

“…This group includes 507 TerLs of the 526 RcGTA-like TerLs that were identified and curated by Kogay et al (2019). Nineteen RcGTA-like TerLs from the dataset of Kogay et al (2019) are absent in our dataset because they were not in GenBank at the time of our data collection (December 2018). The 616 TerLs were classified as either "RcGTAlike" or "virus-like" using a machine learning approach implemented in the GTA-Hunter program (Kogay et al, 2019).…”

“…For the 604 TerLs within the "RcGTA-like containing" group ( Figure 1) that originated from bacterial and archaeal genomes, the presence of 11 other RcGTA-like genes near the terL gene was examined. To this end, the RcGTA-like genes from the training set of Kogay et al (2019) were used as queries in a BlastP search against the assemblies of the bacterial and archaeal genomes (E-value < 0.001; query and subject overlap by at least 60% of their length) (Altschul et al, 1997). The detected RcGTA gene homologs were classified as "RcGTA-like" or "virus-like" using GTA-Hunter (Kogay et al 2019).…”

“…To this end, the RcGTA-like genes from the training set of Kogay et al (2019) were used as queries in a BlastP search against the assemblies of the bacterial and archaeal genomes (E-value < 0.001; query and subject overlap by at least 60% of their length) (Altschul et al, 1997). The detected RcGTA gene homologs were classified as "RcGTA-like" or "virus-like" using GTA-Hunter (Kogay et al 2019). The detected RcGTA gene homologs were also clustered into regions using DBSCAN, with a maximum distance cutoff of 8,000 bp between adjacent genes (Ester et al, 1996;Kogay et al, 2019).…”

“…The detected RcGTA gene homologs were classified as "RcGTA-like" or "virus-like" using GTA-Hunter (Kogay et al 2019). The detected RcGTA gene homologs were also clustered into regions using DBSCAN, with a maximum distance cutoff of 8,000 bp between adjacent genes (Ester et al, 1996;Kogay et al, 2019). If a terL gene was embedded in a region containing at least 6 of the 11 RcGTA-like genes, it was classified as being in a "large RcGTA-like element."…”