Unsupervised explainable AI reveals similar oligonucleotide-usage zones matching the highest-resolution human chromosome bands

Abstract

Unsupervised and explainable AI can uncover genomic features that extend beyond human expectations. Oligonucleotides such as penta- and hexanucleotides often function as core binding motifs for regulatory proteins, and their usage provides a powerful tool in functional genomics. We applied an unsupervised, explainable AI approach to odds ratio (observed/expected) profiles of all 1-Mb euchromatic fragments in the human genome. This odds-ratio analysis identified oligonucleotide features independent of mononucleotide composition, thereby highlighting functional roles of oligonucleotide motifs. AI-based clustering of all 1-Mb euchromatic fragments, using either penta- or hexanucleotide odds ratios (1,024 or 4,096 variables), unexpectedly revealed nearly 2,000 distinct zones, despite the large differences in dimensionality. If these [~]2,000 zones represent biologically meaningful segmentations, comparable structures would be expected to emerge when other oligonucleotide types are analyzed. Consistent with this expectation, CG-containing penta- and hexanucleotides (244 and 1,185 variables) produced comparable [~]2,000 zones, indicating that the underlying segmental structures reflect fundamental functional divisions within the genome. Human chromosomes exhibit well-established Giemsa-banding patterns comprising 850 bands at prometaphase and 2,000 bands at prophase. Since single-nucleotide coordinates are available for the 850 bands, we identified a diagnostic oligonucleotide set that distinguishes Giemsa-negative and -positive regions. Computational pseudo-band reconstruction based on this set generated genome segmentations that more closely paralleled the AI-derived [~]2,000 clusters than the 850 bands. These unexpected findings indicate that AI captures the characteristic features of chromosomal bands and can predict high-resolution banding from genome sequences alone, thereby bridging classical cytogenetics and modern AI-based genomics.