Unduliner reaches 0.897 accuracy in cancer-genome preprint

Halimat Chisom Atanda and Adam D Ewing posted a bioRxiv preprint on Oct. 11 describing unduliner, an AI framework that reached 0.897 accuracy on held-out test data.

Evidence grade: Early (small, animal or preprint). The bioRxiv preprint reported 0.803 validation accuracy, below the held-out test result, and has not been certified by peer review.

Unduliner is designed to connect somatic variants with methylation changes in cancer genomes. Somatic variants are DNA changes acquired by cells rather than inherited, while methylation is a chemical mark on DNA that can affect gene regulation.

The model uses a convolutional neural network, a type of AI system built to detect patterns in structured data. The authors said the framework predicts single-nucleotide variants, or one-letter DNA changes, and structural variants, larger genomic rearrangements, associated with altered methylation patterns.

The work relies on nanopore sequencing, a long-read technology that can capture genetic variation and DNA base modifications from the same sequencing reads. That matters because standard variant annotation often focuses on DNA sequence changes, while methylation adds another layer of biological context.

The authors trained the model using phased sequence reads, single-nucleotide germline variants, per-base methylation probability, and variant proximity to or presence within haplotype-specific differentially methylated regions. A haplotype is a block of DNA inherited together from one parent; a differentially methylated region is a genomic stretch where methylation differs between comparison groups.

As a proof of concept, Atanda and Ewing applied unduliner to subline-specific variants from MCF7 breast cancer cell lines. They said the model identified methylation differences stratified by variant independently of methylation differences linked to haplotypes.

The authors also applied the framework to a pancreatic ductal adenocarcinoma sample. In that analysis, they reported coding and non-coding single-nucleotide variants and structural variants associated with tumour-normal differentially methylated regions, allele-specific methylation and intratumour differentially methylated regions.

For researchers, the output is intended to be more than a variant list. According to the preprint, unduliner returns variant information, the size and direction of methylation change relative to the variant, and the genomic location of the methylation change.

The potential use case is interpretation, not diagnosis. The preprint frames unduliner as a way to add functional context to coding and non-coding cancer variants, particularly where conventional variant-impact tools may not explain whether a variant is linked to an epigenetic change.

The limits are substantial for ordinary readers and clinicians. The supplied abstract does not report a clinical validation cohort size, does not show that unduliner improves cancer diagnosis or treatment selection, and does not establish that the reported methylation associations cause tumour behaviour.

The preprint status is also important. bioRxiv states that the article has not been peer reviewed, meaning independent reviewers have not yet evaluated the methods, results or interpretation through a journal process.

The authors declared funding from the Australian Department of Health Medical Research Future Fund and Mater Foundation. Adam D Ewing reported reimbursement for travel, accommodation and conference fees to speak at events organised by Oxford Nanopore Technologies; the remaining authors declared no competing interests.

Unduliner is publicly available on GitHub, according to the preprint. The article footnotes also list links to Zenodo, Figshare and National Center for Biotechnology Information resources connected to the work.

Source: academic preprint, bioRxiv, Oct. 11