bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelTermsPrivacyContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / dna-gene / dnasora
DNA & Gene
Ecotone AIReleased January 2025

dnaSORA

Diffusion transformer for genetic mapping that generates and classifies bulk segregant point clouds to localize causal mutations at 0.3 Mb.

The short version

  • —Narrows the search space around a phenotype-causing mutation to 0.3 Mb
  • —Assigns a new phenotype from pooled genomes of several dozen individuals
  • —One diffusion transformer serves as both point cloud generator and classifier
  • —Synthetic Gaussian point clouds substitute for scarce real mapping genomes in pretraining
  • —Frozen latent representation feeds a linear, attention, or zero-shot classification head
63Openness

Where to run it

No providers recorded yet. Browse all providers

Bulk segregant analysis — known in the C. elegans community as the Hawaiian experiment — crosses a mutant in one genetic background to a divergent wild isolate, pools phenotypic offspring, and sequences them together. Meiotic linkage leaves the region around the causative mutation enriched for markers from the mutant parent, producing a roughly Gaussian bump in allele frequency whose peak points at the mutation. The interval it yields is nearly always correct, following from the physics of recombination rather than a statistical association. It is also wide: published C. elegans studies using LOESS smoothing report search spaces from 2 to 13.8 Mb.

dnaSORA, from Ecotone AI, treats that allele-frequency profile as a point cloud and models it with a diffusion transformer. The framing rests on one observation: every phenotype produces the same Gaussian-shaped feature and differs only in where the peak sits, so the representation is uniform across diseases and a model can learn the shape once. That is what lets it be pretrained on synthetic data rather than genomes, which for this assay are effectively unobtainable at scale — the authors had 35 real chromosomes.

The architectural claim is that one model does both jobs. Rather than training a generator and then a separate classifier on its output, dnaSORA freezes the trained transformer and reads a classification head off its internal latent representation, following work showing that diffusion latents are competitive classification features in image space.

#Key Features

  • Unified generator and discriminator: One diffusion transformer produces synthetic point clouds and, with weights and conditioning frozen, supplies the representation a classification head reads to predict the peak position.
  • Synthetic pretraining: Pretraining uses 120,000 sparse mock point clouds from tunable Gaussian functions, sidestepping the near-total absence of public bulk segregant data.
  • Whole-chromosome context: Rather than splitting a chromosome into micropatches, which left discontinuities between them, the model ingests an entire chromosome point cloud as one macropatch.
  • Low sample requirement: Assigning a new phenotype requires several dozen pooled admixed individuals across two to three whole-genome sequencing runs.
  • Three classification heads: Linear, attention-based, and a zero-shot head comparing predicted to scheduled noise, all evaluated against a conventional two-model baseline.

#Technical Details

The backbone follows the DiT formulation: 1D point cloud input is converted to 2D patches by a convolutional layer and projected into a high-dimensional feature space, then processed by transformer blocks with multi-head self-attention and adaptive layer normalization (adaLN-Zero), conditioned on a class label (the peak position) and a diffusion timestep from 0 to 1000. Real data comprised 35 human-curated chromosomes from whole-genome sequencing of 16 C. elegans strains, pooling 20–60 animals per strain; three — carrying vab-3(ot266), him-4(u924), and mec-1(u925) — have mutations confirmed by rescue experiments and serve as gold-standard ground truth. Mock point clouds were sparse (784 elements) for pretraining and dense (16,384 elements, matching real density) for validation. At inference, chromosomes are sliced into 300–600 phase-offset representations whose predictions are averaged, on RTX 3090 GPUs.

On aggregate accuracy at identifying the correct search space, the unified model with a linear head scored 96.57% against real-data ground truth peaks, versus 93.10% for the generator-plus-classifier baseline. The headline resolution comes from the per-allele table: the smallest offset between predicted and true peak was 0.30 Mb, from the linear head on vab-3(ot266) and the zero-shot head on mec-1(u925), with other head-allele pairs ranging up to 3.79 Mb. On vab-3(ot266), the Boost algorithm gave a 0.82 Mb interval.

#Applications

The stated target is rare disease genetics: identifying which variant in a mapped interval actually causes a phenotype, so a CRISPR correction can be designed against the right target. The same workflow serves basic genetics in any sexually reproducing diploid where bulk segregant crosses are routine — nematodes, plants, arthropods, zebrafish — by shrinking the inspection burden after a mapping cross. The paper also models the cost of coarse resolution downstream; those figures are the authors' scenario analysis, not measured outcomes.

#Impact

dnaSORA is an unusual entry in genomic sequence modeling: where the Nucleotide Transformer, DNABERT-2, and DNAGPT learn from raw sequence with masked or autoregressive objectives, dnaSORA models a derived statistical signal and applies diffusion to it. It is best read as a proof of concept with bounded evidence: all results are in C. elegans, the resolution claim rests on three experimentally validated alleles, and the human applications are prospective. Code is MIT licensed and complete — denoiser, unified and convolutional classifiers, configs for every trained variant, the analysis notebooks computing the reported metrics, and 31 real chromosome traces the README says suffice to reproduce the paper. Trained weights are not released: the repository ships no checkpoints and no tagged releases, so every result must be retrained. The work has not been peer reviewed, and the preprint is CC BY-NC.

At a glance

Released
January 2025
Category
DNA & Gene
Organization
Ecotone AI

Links

GitHub RepositoryResearch PaperOfficial Website

Tags

diffusiondnagenerativegenetic_mappinggenomicstransfer_learningtransformervariant_effect_prediction

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.