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 modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / dna-gene / bioreason
DNA & GeneLanguage model
Bowang LabUniversity of TorontoVector InstituteUniversity Health NetworkArc InstituteUCSFReleased May 2025

BioReason

DNA-LLM reasoning model fusing genome foundation model embeddings with an LLM to produce step-by-step pathway and variant effect explanations.

71Openness

Where to run it

No providers recorded yet. Browse all providers

BioReasonDNA & GeneBowang Lab

BioReason couples a DNA foundation model directly to a large language model so that the LLM can read genomic sequence as an input modality and reason over it in natural language. It was developed by the Bowang Lab at the University of Toronto, Vector Institute, and University Health Network, with collaborators at the Arc Institute and UCSF, and was published at NeurIPS 2025.

The motivation is a gap between two mature model families. Genome language models such as Evo 2 and Nucleotide Transformer learn excellent representations of DNA but emit scores, not explanations, and they do not chain several biological facts together. General-purpose LLMs can reason in multiple steps, but they cannot see a sequence — a variant has to be described to them in words first, which throws away exactly the information that determines its effect. BioReason keeps both: DNA embeddings are projected into the LLM's token space and concatenated with the user's query, so a single model attends to raw sequence and to the question about it at the same time.

The result is a system that answers genomic questions with an explicit derivation rather than a bare label — naming the affected gene, the mechanism of disruption, and the downstream pathway — and that transfers this behavior to biological entities it has not seen in training. A later sibling model, BioReason-Pro, carries the same DNA-LLM recipe into the protein domain by substituting a protein language model for the DNA backbone.

#Key Features

  • Native sequence input: A learnable linear projection maps DNA foundation model embeddings into the LLM's embedding dimension, so genomic context enters the prompt as tokens rather than as a textual description.
  • Interpretable reasoning traces: The model emits step-by-step biological deductions alongside its prediction, which can be inspected and challenged by a domain expert instead of taken on trust.
  • SFT plus GRPO reinforcement learning: Supervised fine-tuning on reasoning traces is followed by Group Relative Policy Optimization against a composite reward for correctness, formatting, and conciseness.
  • Generalization to unseen entities: Evaluation includes genes and pathways held out from training, testing whether the reasoning is genuine or memorized.
  • Open code and data: Code is Apache-2.0 licensed and the KEGG and variant effect datasets are published on HuggingFace with data cards.

#Technical Details

BioReason is evaluated in four configurations, pairing either Evo 2 (1B) or Nucleotide Transformer (500M) as the DNA encoder with Qwen3 at 1.7B or 4B parameters. During supervised fine-tuning the DNA model is frozen and the LLM is adapted with LoRA (rank 32, alpha 64) using AdamW at a learning rate of 5e-5; the reinforcement learning stage applies GRPO with a clipped surrogate objective and KL regularization. Reported gains are substantial: KEGG-based disease pathway prediction accuracy rises from 86% to 98%, and variant effect prediction improves by an average of 15% over strong baselines. On coding variant effect prediction the Evo 2 + Qwen3-4B configuration reaches 80.2% accuracy against 70.1% for the DNA-model-only baseline and 49.0% for the LLM alone.

#Applications

The intended users are clinical and research geneticists who need a mechanistic account, not just a pathogenicity score: interpreting variants of uncertain significance, connecting a mutation to the pathway it perturbs, and triaging candidate variants where the rationale determines whether a finding is worth following up. Because the reasoning is exposed in text, the output can be audited against literature rather than accepted as a black-box call.

#Impact

BioReason demonstrates that DNA foundation models and LLMs are complementary rather than competing, and that fusing them yields both accuracy gains and interpretability that neither achieves alone. The repository has drawn substantial community attention, and the approach was carried forward into BioReason-Pro for protein function annotation. Two limitations are worth stating plainly. Trained checkpoints have not been released — only code, datasets, and a project site — so reproduction requires training from scratch. And the authors note that reliance on curated resources such as KEGG introduces bias toward well-characterized genomic regions, that encoding long DNA sequences is computationally expensive, and that the model has no uncertainty quantification over the reasoning paths it produces.

At a glance

Released
May 2025
Category
DNA & Gene
Organizations
Bowang Lab / University of Toronto / Vector Institute / University Health Network / Arc Institute / UCSF

Links

GitHub RepositoryResearch PaperOfficial WebsiteDataset

Tags

genomicsmultimodalpathway_predictionreasoningreinforcement_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.