Somatic copy-number alteration calling from tumor whole-genome sequencing, using Mamba and Hyena blocks over genome-scale inputs of about 1M loci.
No providers recorded yet. Browse all providers
Somatic copy-number alterations — gains, losses, and whole-genome doublings of chromosomal segments — are among the most pervasive features of cancer genomes. The established tools that call them from whole-genome sequencing (ASCAT, Battenberg, CNVkit, HMMCopy) are statistical models with hand-tuned smoothing and segmentation parameters rather than learned ones. Deep learning had largely stayed away, for a structural reason: a copy-number profile is defined by correlations spanning entire chromosomes, so the input sequence runs to roughly a million SNP loci, far beyond what quadratic self-attention can process.
araCNA, from Ellen Visscher and Christopher Yau at the University of Oxford, attacks that scaling barrier with subquadratic sequence-model architectures. It replaces the transformer backbone with bidirectional Mamba or Hyena blocks — the same long-context building blocks used by genomic language models such as HyenaDNA and Caduceus — which scale linearly with sequence length and can therefore read an entire genome's worth of read-depth and B-allele-frequency observations in one forward pass.
The second idea is where the training data comes from. Because ground-truth copy-number profiles do not exist for real tumors, araCNA is trained exclusively on synthetic genomes generated from an explicit generative model of copy number, purity, and sequencing noise. The resulting checkpoints are then applied zero-shot — with no fine-tuning on patient data — to 50 real whole-genome-sequenced tumors from The Cancer Genome Atlas. The work was posted to bioRxiv in January 2025 and published in NAR Genomics and Bioinformatics later that year.
Read depth and B-allele frequency at each SNP locus are projected through an MLP encoder, passed through a bidirectional Hyena or Mamba-2 backbone, and decoded by two heads: a softmax over discrete major/minor copy-number classes and a sigmoid for purity. Total copy number is capped at 8. The two variants differ strikingly in size — araCNA-mamba has 70K parameters, araCNA-hyena 12.09M, of which roughly 12M are the two positional-encoding tables sized to the 1M maximum sequence length. Training proceeds in stages, starting with low-complexity profiles and progressively adding purity and noise sampling, higher copy numbers, and longer sequences up to 650,000 loci; purity is sampled uniformly between 0.5 and 1.
On 100 held-out simulated genomes both variants recover copy number, purity, and ploidy accurately. On the 50 TCGA colorectal, breast, and ovarian samples, where no ground truth exists, the authors use reconstruction error of read depth and B-allele frequency as a proxy: araCNA is comparable to ASCAT, Battenberg, CNVkit, and HMMCopy while using similar segment counts. araCNA-mamba reaches a median 80% call concordance with ASCAT and 70% with Battenberg, with most disagreements being off-by-one copy-number differences. ASCAT and Battenberg occasionally assign as many as 100 copies to localized regions, mostly devoid of known cancer genes, which araCNA does not do. Notably, the 70K-parameter Mamba model outperformed the far larger Hyena variant.
araCNA suits cancer genomics groups profiling tumor whole genomes at scale, particularly in cohorts where matched normal samples were never collected or have been exhausted — a common situation in retrospective archives and clinical biobanks. Minutes-scale inference makes large cohort reprocessing and interactive re-analysis practical. The released repository runs inference from a BAM file plus a SNP loci list, emitting per-locus calls, segmented profiles, and purity/ploidy/whole-genome-doubling estimates as CSVs.
araCNA's broader argument is methodological: where labeled biological ground truth is unobtainable, a well-specified generative simulator can supply the training signal, and a long-context sequence model can then absorb structure that hand-tuned statistical callers encode as hyperparameters. It also shows that subquadratic architectures make genome-scale supervised learning tractable at a parameter budget orders of magnitude below contemporary genomic foundation models. Evaluation on real data rests on unsupervised proxy metrics and cross-method concordance rather than validated truth, across 50 samples and three cancer types, so generalization to other tumor types, sequencing depths, and low-purity samples remains to be established. The code is released under Apache 2.0 with pretrained checkpoints and a permanent Figshare archive.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.