Batch-Learning Self-Organizing Map Identifies Horizontal Gene Transfer Candidates and Their Origins in Entire Genomes

Abstract

Horizontal gene transfer (HGT) has been widely suggested to play a critical role in the environmental adaptation of microbes; however, the number and origin of the genes in microbial genomes obtained through HGT remain unknown as the frequency of detected HGT events is generally underestimated, particularly in the absence of information on donor sequences. As an alternative to phylogeny-based methods that rely on sequence alignments, we have developed an alignment-free clustering method on the basis of an unsupervised neural network "Batch-Learning Self-Organizing Map (BLSOM)" in which sequence fragments are clustered based solely on oligonucleotide similarity without taxonomical information, to detect HGT candidates and their origin in entire genomes. By mapping the microbial genomic sequences on large-scale BLSOMs constructed with nearly all prokaryotic genomes, HGT candidates can be identified, and their origin assigned comprehensively, even for microbial genomes that exhibit high novelty. By focusing on two types of Alphaproteobacteria, specifically psychrotolerant Sphingomonas strains from an Antarctic lake, we detected HGT candidates using BLSOM and found higher proportions of HGT candidates from organisms belonging to Betaproteobacteria in the genomes of these two Antarctic strains compared with those of continental strains. Further, an origin difference was noted in the HGT candidates found in the two Antarctic strains. Although their origins were highly diversified, gene functions related to the cell wall or membrane biogenesis were shared among the HGT candidates. Moreover, analyses of amino acid frequency suggested that housekeeping genes and some HGT candidates of the Antarctic strains exhibited different characteristics to other continental strains. Lys, Ser, Thr, and Val were the amino acids found to be increased in the Antarctic strains, whereas Ala, Arg, Glu, and Leu were decreased. Our findings strongly suggest a low-temperature adaptation process for microbes that may have arisen convergently as an independent evolutionary strategy in each Antarctic strain. Hence, BLSOM analysis could serve as a powerful tool in not only detecting HGT candidates and their origins in entire genomes, but also in providing novel perspectives into the environmental adaptations of microbes.

Unsupervised explainable AI for molecular evolutionary study of forty thousand SARS-CoV-2 genomes

“…The phylogenetic method based on sequence alignment is a well-established and irreplaceable method for molecular evolutionary studies. The presently developed sequence alignment-free method is suitable for analyzing a massive amount of sequence data and can analyze over ve million sequences simultaneously 8 ; notably, this method is highly robust against sequencing errors, and therefore, no special pretreatment is required. Furthermore, the BLSOM is unsupervised AI that can be used without special models or presumptions, and has powerful visualization capabilities that enable e cient knowledge discovery from big data.…”

“…When we constructed a BLSOM for oligonucleotide compositions in fragment sequences (e.g., 10 kb) from a wide variety of species, the sequences were clustered (self-organized) primarily according to species, despite no species information being used during machine learning 5,7 . Importantly, the BLSOM is suitable for large-scale analysis and has been used to analyze ve million genomic fragments from over one thousand genera 8 . In addition, the BLSOM is explainable AI and can reveal the drivers of species-speci c clustering (self-organization).…”

Unsupervised explainable AI for the collective analysis of a massive number of genome sequences: various examples from the small genome of pandemic SARS-CoV-2 to the human genome

“…In Fig. 4A, the tetranucleotide composition of over 5 million 5-kb fragments of almost all microorganisms (including viral sequences) available in INSDC (DDBJ, ENA/EBI and NCBI) was analyzed by a BLSOM (Abe et al, 2020). In INSDC, only one strand of complementary sequences is registered, and the strand is chosen rather arbitrarily in the registration of fragment sequences.…”

“…By analyzing the compositions of short oligonucleotides (e.g., 4-and 5-mers) in a large number of genomic fragments (e.g., 10 kb) derived from a wide variety of species, the BLSOM enables the separation (self-organization) of the genomic sequences by species and phylogeny and identifies oligonucleotides that significantly contribute to the separation (Abe et al, 2003(Abe et al, , 2006cIwasaki et al, 2013b;Bai et al, 2014;Kikuchi et al, 2015). In an analysis of the genomic fragments of a wide range of microbial genomes, over 5 million sequences were successfully separated by phylogenetic groups with high accuracy (Abe et al, 2020). A BLSOM program suitable for PC cluster systems is available on the following website: http://bioinfo.ie.niigata-u.ac.jp/?BLSOM.…”

Unsupervised explainable AI for simultaneous molecular evolutionary study of forty thousand SARS-CoV-2 genomes

“…When we constructed a BLSOM for oligonucleotide compositions in fragment sequences (e.g., 10 kb) from a wide variety of species, the sequences were clustered (self-organized) primarily according to species, despite no species information being used during machine learning 5,7 . Importantly, the BLSOM is suitable for large-scale analysis and has been used to analyze five million genomic fragments from over one thousand genera 8 . In addition, the BLSOM is explainable AI and can reveal the drivers of species-specific clustering (self-organization).…”