Research map / Large language models / Agents, code and math
BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces BioReason, a multimodal architecture that feeds frozen DNA foundation model embeddings (Evo2, Nucleotide Transformer) into a Qwen3 LLM, trained with SFT and GRPO for step-by-step biological reasoning and variant effect prediction.
Problem
DNA foundation models produce sequence representations but lack interpretable, multi-step biological reasoning; LLMs cannot natively process raw genomic sequences. The paper targets reasoning over genomic variants to disease phenotypes.
Approach
Project frozen DNA foundation model token embeddings through a learnable linear layer into the LLM's input space, prepend them to the text query, and fine-tune the LLM (LoRA SFT, then GRPO) to generate reasoning traces and predictions.
Key results
- Evo2+Qwen3-4B with GRPO reaches 98.28% accuracy and 93.05% F1 on the KEGG reasoning task, versus 90.00% accuracy for Qwen3-4B alone and 88.28% for Evo2-1B alone.
- On VEP-Coding, Evo2+Qwen3-4B achieves 80.21% accuracy and 80.00% F1, compared with 70.07% accuracy for Evo2-1B alone.
- On VEP-Non-SNV, Evo2+Qwen3-1B reaches 88.20% accuracy and 89.91% F1, above Evo2-1B alone (76.17%/66.51%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Integrate Evo2-1B embeddings with Qwen3-4B and apply GRPO | BioReason · KEGG-derived reasoning | Accuracy: 95.17 to 98.28 | Helped |
| Add GRPO fine-tuning to Evo2-1B + Qwen3-4B (ablation) | BioReason · KEGG-derived reasoning | F1-score: 86.14 to 93.05 | Helped |
| Add GRPO fine-tuning to NT-500M + Qwen3-4B (ablation) | BioReason · KEGG-derived reasoning | Accuracy: 95.86 to 98.28 | Helped |
| Add GRPO fine-tuning to NT-500M + Qwen3-1B (ablation) | BioReason · KEGG-derived reasoning | F1-score: 81.46 to 75.06 | Hurt |
| Add GRPO fine-tuning to NT-500M + Qwen3-1B (ablation) | BioReason · KEGG-derived reasoning | Accuracy: 89.31 to 91.72 | Helped |
| Combine Evo2-1B DNA embeddings with Qwen3-4B | BioReason · VEP-Coding | Accuracy: 70.07 to 80.21 | Helped |
| Combine Evo2-1B DNA embeddings with Qwen3-1B | BioReason · VEP-Non-SNV | Accuracy: 76.17 to 88.20 | Helped |
| Combine Enformer embeddings with Qwen3-4B | BioReason · Chromatin accessibility | Macro F1: 17.18 to 33.70 | Helped |
Surprising findings
- On the VEP-Non-SNV task, Qwen3-1B alone outperforms Qwen3-4B alone (70.67% vs 61.86% accuracy), and NT+Qwen3-4B underperforms NT+Qwen3-1B on coding VEP F1 gaps are mixed.
- GRPO improved accuracy for the 1B NT hybrid while F1 declined (81.46% to 75.06%).
Limitations
- Reliance on curated KEGG data introduces biases and limits coverage of less-characterized regions.
- Computational cost of encoding long DNA sequences and running GRPO limits scalability; DNA inputs truncated to 2048 tokens.
- Lacks robust uncertainty quantification for high-stakes decisions.
Experimental setup
Datasets: KEGG Network Variants / KEGG pathway database ClinVar OMIM dbSNP COSMIC gnomAD v3.1.2 GPN-MSA variant set DeepSEA DNase-seq data
Benchmarks: KEGG-derived biological reasoning dataset VEP-Coding VEP-Non-SNV DeepSEA DNase-seq chromatin accessibility
Metrics: accuracy F1-score macro precision macro recall
Baselines: Evo2-1B Nucleotide Transformer NT-500M Qwen3-1.7B Qwen3-4B Enformer
Builds on: GRPO Dr. GRPO Qwen3 Evo2 Nucleotide Transformer
Closest papers by meaning
- Contextualizing biological perturbation experiments through language (ICLR 2025 · similarity 0.83)
- BioXArena: Benchmarking LLM Agents on Multi-Modal Biomedical Machine Learning Tasks (NeurIPS 2026 · similarity 0.83)
- BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments (ICLR 2025 · similarity 0.82)
- BioBridge: Bridging Biomedical Foundation Models via Knowledge Graphs (ICLR 2024 · similarity 0.82)
- CGBench: Benchmarking Language Model Scientific Reasoning for Clinical Genetics Research (NeurIPS 2025 · similarity 0.82)
- Bio-xLSTM: Generative modeling, representation and in-context learning of biological and chemical sequences (ICLR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.