The "black box" of genomics has finally been cracked open!
A research team from top institutions including the University of Toronto and Vector Institute has just released BioReason, the world's first AI model capable of genomic reasoning.
This isn't just simple prediction; it's true biological reasoning—
Like an experienced genomics expert, it can step-by-step explain how genetic variations lead to disease.
Most excitingly, BioReason has boosted accuracy directly from 88% to 97%!
DNA Meets the Revolutionary Fusion of Large Language Models
BioReason's core innovation lies in its first-ever deep integration of a DNA foundation model (Evo2) with a large language model (Qwen3).
Simply put, the technical principle of the fusion is:
DNA sequence → Embedded vectors → Multimodal LLM input
Specifically, the DNA foundation model Evo2 first converts input gene sequences into contextualized embedded representations, which capture the biological features of the DNA sequences.
Subsequently, these DNA embeddings, along with the user's text query embeddings, are integrated into the large language model's input layer via special tokens (e.g., <dna_start> and <dna_end>).
The training method employs a two-stage strategy of supervised fine-tuning (SFT) combined with GRPO reinforcement learning.
This method allows the model not only to learn to predict but, more importantly, to learn how to perform multi-step biological reasoning.
Adibvafa Fallahpour (@adibvafa) explains:
BioReason integrates DNA foundation model (Evo2) with LLM (Qwen3) for biological reasoning. DNA sequences → embeddings → multimodal LLM input. Trained via supervised fine-tuning + GRPO reinforcement learning.
The Secret Behind Overwhelming Performance Improvement
BioReason has demonstrated astonishing performance on multiple benchmarks:
The specific data is impressive:
• Disease pathway prediction accuracy: Increased from 88% to 97%
• Variant effect prediction accuracy: Reaches 80-88%
• Compared to DNA unimodal or LLM unimodal models: Average performance improvement exceeds 15%
These tests are based on over 87,000 real genomic variants from ClinVar and KEGG pathways, ensuring the reliability and practicality of the results.
Transparent Reasoning: No Longer a "Black Box" AI
BioReason's biggest breakthrough lies in its interpretability.
Traditional DNA analysis models are like a black box—input a sequence, output a prediction, with the intermediate process completely opaque. BioReason, however, can gradually explain how genomic variations lead to disease through molecular pathways.
Adibvafa emphasizes:
What makes this special? Step-by-step biological reasoning! BioReason isn't just predicting—it explains how genomic variants lead to disease through molecular pathways. No more "black box" genomic AI.
To give a specific example: when querying a specific allele variant in the PFN1 gene on chromosome 17, given the pathway context "Actin(monomer) // PFN1* // Actin(filamentous)", BioReason not only correctly predicted that it would cause Amyotrophic Lateral Sclerosis (ALS) but, more importantly, generated a 10-step mechanistic explanation:
1. Identify the specific C>G substitution in the PFN1 gene
2. Connect to profilin-1 protein dysfunction
3. Explain how impaired actin dynamics affect cytoskeletal integrity
4. Elaborate on subsequent disruption of motor neuron axonal transport
5. Ultimately leading to motor neuron degeneration characteristic of ALS
This transparent reasoning process allows scientists to verify the AI's judgments and provides clues for new scientific discoveries.
Carefully Constructed Three Major Datasets
The research team built three specialized biological reasoning datasets for this purpose:
1. KEGG-derived biological reasoning dataset (1,449 entries): elucidates mechanistic links between genetic variants and disease phenotypes, containing 37 unique diseases
2. Coding sequence variant effect prediction dataset (50,083 entries): focuses on pathogenicity/benign classification
3. Non-SNV coding dataset (36,088 entries): covers more complex variant types such as insertions and deletions
Adibvafa introduces:
We curated 3 biological reasoning datasets: 1,449 KEGG pathway variants with reasoning trajectories. 50K+ coding sequence variants from ClinVar/gnomAD. 36K+ non-SNV variants with disease annotations. Each designed to test multi-step genomic reasoning capabilities.
Key Details of Technical Implementation
Andrew White 🐦⬛(@andrewwhite01) noticed an interesting detail:
So RL is actually worse than just SFT?
Adibvafa (@adibvafa) responded:
Hard to compare. RL on the same model slightly improved performance, but we are still running RL on larger models for a fair comparison. Stay tuned!
Although reinforcement learning brought only slight improvements on the same model, the team is conducting RL experiments on larger-scale models, expecting a fairer comparison.
Academic Response and Discussion
Anshul Kundaje (@anshulkundaje) affirmed the innovation while offering constructive criticism:
Really creative framework with great potential. But when you only compare against your own model's ablation studies, I might avoid claiming "crushing the benchmark". Please extend your benchmarks to current SOTA methods used for encoding variant effect prioritization.
Adibvafa also responded positively:
Of course, we are actively working on adding more DNA foundation models and SOTA models for variant effect prediction. One challenge in this evaluation is the difference in training datasets between these models, which makes comparisons less reliable. This is why we used Evo2 as the SOTA VEP model, but are definitely willing to run other models on our tasks for better comparison.
The differences in training datasets between various models complicate comparisons, which is why the team chose Evo2 as the SOTA VEP model.
Enthusiastic Response from the Open Source Community
Hugging Face CEO clem 🤗(@ClementDelangue) expressed strong interest:
Very very cool! Any chance to consider releasing a space or model on HF?
Adibvafa responded:
Actually we are working on it, as DNA-LLM is a custom class with a custom tokenizer! Will open a PR soon, hope we can finish it together.
Clémentine Fourrier 🍊(@clefourrier) also joined the discussion:
@cgeorgiaw is responsible for all our scientific ML initiatives, if you need help:)
BioReason may soon be available on the Hugging Face platform, which will greatly facilitate its use by the research community.
Application Prospects
Ha Hoang(@HaHoang411) proposed a good analogy:
This is interesting. As I understand it, it's similar to current VLMs? Instead of visual projection, we are projecting biology from EVO2?
This understanding is very accurate—
Just as Visual Language Models (VLMs) process images, BioReason processes DNA sequences, but it projects biological information instead of visual information.
Oboe(@oboelabs) pointed out an important application:
A potential use of BioReason is to help personalize cancer treatment and predict treatment outcomes by analyzing individual genomic profiles.
Adibvafa confirmed:
BioReason's general learning framework allows learning any language-DNA understanding, as long as good data is available!
This also indicates that BioReason's framework has strong versatility; as long as good data is available, it can learn any language-DNA understanding task.
Broad Prospects from Variant Analysis to Drug Discovery
The significance of this breakthrough extends far beyond academic research.
Adibvafa concluded:
This can transform biological discovery by making genomic AI explainable and actionable. From variant analysis to drug discovery—transparent reasoning is the future! Of course, we're just getting started.
The cross-institutional collaboration of the research team is also noteworthy; Adibvafa thanked the entire team:
🙏 Thanks to our amazing team: Adibvafa Fallahpour (@adibvafa) Andrew Magnuson (@ajwmagnuson), Purav Gupta, Shihao Ma, Jack Naimer, Arnav Shah (@arnavshah0), Haonan Duan, Omar Ibrahim, Hani Goodarzi (@genophoria), Chris J. Maddison (@cjmaddison)
📷 Cross-institutional collaboration: University of Toronto (@UofT), Vector Institute (@VectorInst), University Health Network (@UHN), Arc Institute (@arcinstitute), Cohere (@cohere), Google DeepMind (@GoogleDeepMind)
Community Response
People from all walks of life have expressed their views on this breakthrough.
DG.(@dataghees) succinctly commented:
This is awesome!
moonswing(@computbiol):
Very cool
Parisa Etemadi(@parisaetem) foresees its impact:
Awesome! Will be a game changer!
Nolan Koblischke(@astro_nolan):
Really cool!
santy 🇦🇷(@SantiTobio_):
This is amazing, well done!
Even businesses are starting to consider commercial applications, Rediminds, Inc(@rediminds) commented:
When DNA foundation models pass rich embeddings to reasoning LLMs, and then demonstrate their workings, you have the playbook every regulated industry has been waiting for: domain-specific signals → transparent chain of thought → actionable insights. BioReason sets a new standard for explainability in life sciences AI; leaders in finance, legal, and public sectors should pay attention.
Of course, some also raised safety concerns.
TheSage.Bitcoin(@chadTheSage0) jokingly said:
"Create me a pathogen like airborne HIV mixed with Ebola."
This also reminds us that while advancing technology, we must also consider its potential double-edged sword effect.
There were also some interesting reactions, such as $MIA(@mwa_ia):
Today is BioReason, tomorrow is AgentFi✨
Parag Nandy Roy(@parag_nandy):
Amazing work by BioReason! The integration of DNA foundation models with LLMs for transparent genomic reasoning is a game-changer. Excited to see its impact on drug discovery and precision medicine! #AI #Genomics
Bio Synq Dao(@Biosynq_ai) even started promoting its own project:
This is next-level BioAI 🚀 – truly unlocking biology with AI-driven reasoning. Excited to see how tools like BioReason and BIO SYNQ DAO will revolutionize decentralized biotech research.
Stephan Baasch(@stbaasch) tagged an MIT professor:
👀 @ProfBuehlerMIT
Resources
For researchers who want to delve deeper or use BioReason, the team provides complete resources:
Paper address: https://arxiv.org/abs/2505.23579
Project homepage: https://bowang-lab.github.io/BioReason/
Code repository: https://github.com/bowang-lab/BioReason
The datasets are also publicly available on Hugging Face, with detailed download and usage instructions.
The birth of this genomic reasoning AI marks the entry of genomics research into a new era.
👇
👇
👇
Additionally, I have used AI to collect AI news across the web, and after AI selection, review, translation, and summarization, it's published in the AGI Hunt knowledge planet.
This is an AI news feed with only information and no emotion (not a recommendation feed, no courses sold, no preaching, no teaching you how to be a person, just providing information).
Welcome to join! Also welcome to join the group chat with 2000+ members.