BioReason: Revolutionizing Genomic AI with Expert-Level Biological Reasoning

Bridging the Gap Between DNA Data and Biological Insight

A key challenge in applying AI to genomics is the absence of interpretable, step-by-step reasoning when analyzing complex DNA data. While DNA foundation models excel at recognizing sequence patterns for tasks like variant prediction and gene regulation, they typically function as black boxes, providing limited understanding of the biological mechanisms at play. On the other hand, large language models (LLMs) demonstrate strong reasoning abilities in many domains but are not designed to process raw genomic sequences. This disconnect between advanced DNA representation and deep biological reasoning has hindered AI from achieving expert-level comprehension and limited its ability to foster scientific discovery through hypothesis-driven explanations.

Advances and Limitations in Genomic AI Models

DNA foundation models have made considerable progress by learning rich representations directly from genomic sequences, achieving strong results across various biological tasks. For example, Evo2 demonstrates impressive long-range sequence modeling capabilities, yet its lack of interpretability restricts deeper biological understanding. Concurrently, LLMs excel at reasoning with biomedical text but do not engage directly with raw genomic data. Early attempts like GeneGPT and TxGemma have sought to bridge this gap, but current genomic benchmarks mainly assess task performance without adequately evaluating reasoning or hypothesis generation.

Introducing BIOREASON: A Hybrid AI Model

A collaborative team from the University of Toronto, Vector Institute, University Health Network, Arc Institute, Cohere, University of California San Francisco, and Google DeepMind developed BIOREASON, a pioneering AI system that integrates a DNA foundation model with an LLM. This fusion enables BIOREASON to analyze raw genomic sequences while applying LLM-based reasoning to produce clear, biologically grounded insights. Through supervised fine-tuning and reinforcement learning, BIOREASON achieves over 15% performance improvement compared to traditional models, reaching up to 97% accuracy in KEGG-based disease pathway prediction. Its interpretable, stepwise outputs significantly advance biological understanding and facilitate hypothesis generation.

How BIOREASON Works

BIOREASON is a multimodal framework combining genomic sequences with natural language queries to support deep, interpretable biological reasoning. It utilizes a DNA foundation model to extract rich, contextual embeddings from raw DNA inputs, which are then integrated with tokenized textual queries to create a unified input for an LLM called Qwen3. The system is trained to generate step-by-step explanations of biological processes. DNA embeddings are projected into the LLM’s embedding space through a learnable layer, and the combined input is enhanced with positional encoding. Reinforcement learning using Group Relative Policy Optimization further refines its reasoning abilities.

Performance and Case Studies

BIOREASON was evaluated on three datasets focused on DNA variant interpretation and biological reasoning. It outperformed models that relied solely on DNA data or LLMs, excelling in predicting disease outcomes from genomic variants. The top-performing configuration combining Evo2 and Qwen3-4B achieved high accuracy and F1 scores across all tasks. A notable example includes the PFN1 mutation associated with ALS, where BIOREASON accurately predicted the disease and generated a detailed 10-step explanation outlining how the variant impacts actin dynamics and leads to motor neuron degeneration. This demonstrates not only its predictive accuracy but also its ability to provide transparent, biologically grounded reasoning.

Future Directions and Impact

BIOREASON uniquely combines DNA encoders with large language models to enable detailed, interpretable reasoning over genomic data. Unlike conventional AI models, it explains the biological rationale behind predictions through stepwise outputs, assisting scientists in understanding disease mechanisms and generating new research questions. Despite its power, BIOREASON faces challenges such as high computational demands and limited uncertainty quantification. Future work aims to improve scalability, integrate additional biological data types like RNA and proteins, and expand applications to broader tasks including genome-wide association studies (GWAS). BIOREASON holds significant promise for advancing precision medicine and genomic research.

For more information, see the Paper, GitHub Page, and Project Page linked in the original publication.