Loading…
Genomic variant effect prediction from a bidirectional absorbing-diffusion model over human DNA, pretrained with entropy-guided span masking.
No providers recorded yet. Browse all providers
Most genomic reconstruction problems arrive as a gap: sequence is known on both sides, and the missing stretch has to be filled. An autoregressive generator can only read the left flank, which makes a bidirectional denoiser — conditioning on both flanks at once and revising uncertain positions over several passes — the more natural instrument. GenDA (Genomic Density-optimized Absorbing Diffusion) was built to test that instrument, plus a second idea layered on top: that pretraining corruption should be aimed at compositionally complex sequence rather than scattered uniformly, because masking a repetitive stretch asks the model to recover what short-range base statistics already give away.
GenDA is a 202M-parameter discrete absorbing-diffusion model over human genomic DNA, posted as arXiv preprint 2609.04861 in September 2026 by a group centred on the Department of Translational Surgical Oncology at the National Center for Tumor Diseases Dresden, with collaborators at ScaDS.AI Dresden/Leipzig, TIB Hannover and Gustave Roussy. The paper is deliberately framed as a partial refutation of its own design premise: the checkpoint performs, but not through the mechanism its name advertises.
After supervised fine-tuning, GenDA reaches a pooled ClinVar SNV AUROC of 0.774, ahead of a depth- and width-matched autoregressive baseline at 0.671 and of zero-shot Evo 2 at 0.733. But a GenDA variant that places spans at random positions rather than by entropy reaches 0.777, so density guidance is not the source of that gain. And in a zero-shot inpainting stress test scored by AlphaGenome, GenDA does not beat a control that shuffles the native gap while preserving its exact 3-mer composition — the same outcome the authors observe for Evo 2 and D3LM.
GenDA is a ModernBERT backbone — 12 layers, 1024 hidden dimension — trained with an absorbing
discrete diffusion objective on roughly 3.2 million human genomic anchors drawn from hg38 using
GENCODE and standard peak annotations: 33.7% exons, 30.0% enhancers, 17.8% introns, 16.0% repeats
and 2.5% promoters, split 80/10/10 by gene_id to prevent homologous leakage. A 1-mer tokenizer
over a 28-token vocabulary preserves single-nucleotide resolution and fixes the window at 4,096 bp.
Span lengths follow a Poisson distribution with λ = 100 clipped to 1–300 bp, and the final
configuration mixes density-guided spans on 70% of examples with position-uniform 1-mer masking on
the rest. Training runs 100,000 steps at a global batch of 128. Region-wise ClinVar AUROC is 0.778
for promoters, 0.773 for introns, 0.757 for exons and 0.746 for enhancers; a 24-layer GenDA-large
reaches 0.786 pooled.
The fine-tuned checkpoint scores single-nucleotide variants by an inverted log-likelihood ratio between reference and alternate alleles in a centred 256-bp window, which suits pathogenicity triage in clinical-genetics settings where a calibrated ordering of candidate variants matters more than a generative sample. The more durable use is methodological: the paper's evaluation protocol — report composition-preserving controls, separate absolute fidelity from control-normalized restoration, stratify by gap length, and evaluate discriminative and generative capability separately — is a template other groups can apply before committing to larger training runs.
GenDA's contribution is a clean negative control on a widely held assumption: that a pretraining prior which improves variant discrimination has therefore learned a useful conditional distribution over functional sequence. It has not. Strong fine-tuned ClinVar performance coexists here with functional inpainting that a 3-mer shuffle matches, and the entropy-guided masking the model is named for is indistinguishable from random span placement on the benchmark it was meant to improve. The authors attribute the generative failure to boundary conditions they did not isolate — 1-mer tokenization capping physical context at 4,096 bp, a 300-bp training-span limit against evaluation gaps up to 3,500 bp, and entropy standing in for functional importance — and no one factor is established as the sole cause. No code or weights accompany the preprint; the reproducibility statement commits to releasing code and processed evaluation metadata upon publication.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.