Affiliations: [a] Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, Universidad Nacional Autónoma de México, México | [b] Postgraduation Program in Computer Science, Universidad Nacional Autónoma de México, México
Correspondence:
[*]
Corresponding author. Antonio Neme, Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, Universidad Nacional Autónoma de México, México. E-mail: [email protected]; ORCID: 0000-0002-9952-1943.
Note: [1] ORCID: 0000-0003-2281-0891. All authors contributed equally.
Abstract: Horizontal gene transference is a biological process that involves the donation of DNA or RNA from an organism to a second, unrelated organism. This process is different from the more common one, vertical transference, which is present whenever an organism or pair of organisms reproduce and transmit their genetic material to the descendants. The identification of segments of genetic material that are the result of horizontal transference is relevant to construct accurate phylogenetic trees, on one hand, and to detect possible drug-resistance mechanisms, on the other, since this movement of genetic material is the main cause behind antibiotic resistance in bacteria. Here, we describe a novel algorithm able to detect sequences of foreign origin, and thus, possible acquired via horizontal transference. The general idea of our method is that within the genome of an organism, there might be sequences that are different from the vast majority of the remaining sequences from the same organism. The former are candidate anomalies, and thus, their origin may be explained by horizontal transference. This approach is equivalent to a particular instance of the authorship attribution problem, that in which from a set of texts or paragraphs, almost all of them were written by the same author, whereas a minority has a different authorship. The constraint is that the author of each text is not known, so the algorithm has to attribute the authorship of each one of the texts. The texts detected to be written by a different author are the equivalent of the sequences of foreign origin for the case of genetic material. We describe here a novel method to detect anomalous sequences, based on interpretable embeddings derived from a common attention mechanism in humans, that of identifying novel tokens within a given sequence. Our proposal achieves novel and consistent results over the genome of a well known organism.
