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

Abstract

In genetics and related fields, huge amounts of data, such as genome sequences, are accumulating, and the use of artificial intelligence (AI) suitable for big data analysis has become increasingly important. Unsupervised AI that can reveal novel knowledge from big data without prior knowledge or particular models is highly desirable for analyses of genome sequences, particularly for obtaining unexpected insights. We have developed a batch-learning self-organizing map (BLSOM) for oligonucleotide compositions that can reveal various novel genome characteristics. Here, we explain the data mining by the BLSOM: an unsupervised AI. As a specific target, we first selected SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2) because a large number of viral genome sequences have been accumulated via worldwide efforts. We analyzed more than 0.6 million sequences collected primarily in the first year of the pandemic. BLSOMs for short oligonucleotides (e.g., 4-6-mers) allowed separation into known clades, but longer oligonucleotides further increased the separation ability and revealed subgrouping within known clades. In the case of 15-mers, there is mostly one copy in the genome; thus, 15-mers that appeared after the epidemic started could be connected to mutations, and the BLSOM for 15-mers revealed the mutations that contributed to separation into known clades and their subgroups. After introducing the detailed methodological strategies, we explain BLSOMs for various topics, such as the tetranucleotide BLSOM for over 5 million 5-kb fragment sequences derived from almost all microorganisms currently available and its use in metagenome studies. We also explain BLSOMs for various eukaryotes, including fishes, frogs and Drosophila species, and found a high separation ability among closely related species. When analyzing the human genome, we found enrichments in transcription factor-binding sequences in centromeric and pericentromeric heterochromatin regions. The tDNAs (tRNA genes) could be separated according to their corresponding amino acid.

AI-based search for convergently expanding, advantageous mutations in SARS-CoV-2 by focusing on oligonucleotide frequencies

“…It should be noted that when nodes that have sequences with distinctly different frequencies are located nearby, empty nodes, to which no sequences are attributed after the learning, often appear between them, and these nodes are left blank (no color) on the BLSOM [8,9]. As in our previous BLSOM analysis of SARS-CoV-2 strains [9,23,24], which were mainly isolated within the first year of the pandemic, good clustering by lineage was obtained, and Omicron strains formed their own territory (red) on the right side of the map (Fig 2A).…”

“…After the machine learning, to determine whether the clustering (self-organization) achieved by the BLSOM corresponded to the known lineages, nodes containing genomes of a single lineage were colored to indicate each lineage, and those containing genomes of multiple lineages were displayed in black. It should be noted that when nodes that have sequences with distinctly different frequencies are located nearby, empty nodes, to which no sequences are attributed after the learning, often appear between them, and these nodes are left blank (no color) on the BLSOM [8,9]. As in our previous BLSOM analysis of SARS-CoV-2 strains [9,23,24], which were mainly isolated within the first year of the pandemic, good clustering by lineage was obtained, and Omicron strains formed their own territory (red) on the right side of the map (Fig 2A).…”

“…The BLSOM in Fig 2A included all Omicron strains isolated from Nov. 2021 to Jan. 2022, so we can examine territories of these three sublineages on the BLSOM by mapping the strains in the three sublineages separately onto the BLSOM (Fig 2B). In more detail, using the vectorial data of the occurrence frequency in each genome, we searched for the node with the closest Euclidean distance, attributed each genome to that node and colored the node for each sublineage, as described previously [7,9]. Separation between BA.1 (reddish brown) and BA.2 (pink) was clear, and BA.3 (violet), which remained in a small population, was in the middle of the former two sublineages (Fig 2B).…”

“…Unsupervised machine learning can provide new information without relying on particular models or assumptions. We previously established an unsupervised AI, batch-learning self-organizing map (BLSOM), for oligonucleotide composition, which can reveal various new characteristics of a wide variety of genomes by analyzing a large number of genome sequences [5][6][7][8][9].…”

“…We previously analyzed time-series changes in composition of short and long oligonucleotides from SARS-CoV-2 strains isolated mainly in the first year of the pandemic and found many long oligonucleotides (e.g., 20-mers) that had expanded remarkably in the viral population, which allowed us to predict candidates of advantageous mutations that had spread among humans [9,23,24]. In the present study, we analyzed 20-mer frequencies to search for mutations that expanded markedly in the Omicron population.…”