PaperlayerView exact revision on arXiv

Source-linked AI summary

BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model

Adibvafa Fallahpour, Andrew Magnuson, Purav Gupta, Shihao Ma, Jack Naimer, Arnav Shah, Haonan Duan, Omar Ibrahim, Hani Goodarzi, Chris J. Maddison, Bo Wang

arXiv:2505.23579v2cs.LG

TL;DR

BioReason addresses the gap between genomic sequence representations and transparent, multi-step biological reasoning by integrating a DNA foundation model with an LLM. Across the reported tasks, the hybrids outperform single-modality models, with average baseline gains above 15%.

  • Problem

    DNA foundation models can represent genomic sequences but often lack transparent mechanistic explanations, and standard genomics benchmarks inadequately assess higher-order reasoning.

    Source: §1 · p. 2, §2.1 · p. 3, §2.3 · p. 3

  • Method

    BioReason integrates DNA-model embeddings with textual queries for LLM reasoning and trains the system using supervised fine-tuning and reinforcement learning.

    Source: §1 · p. 2, §3 · p. 3, §3.3 · p. 4

  • Results

    15%+ average performance gains over baseline are reported, and the DNA–LLM hybrids consistently outperform single-modality models across KEGG reasoning tasks and variant-effect benchmarks.

    Source: §1 · p. 2, §7 · p. 10

  • Takeaways & Limitations

    The framework generates transparent, stepwise explanations intended to facilitate expert validation of genomic analysis.

    Source: §7 · p. 10

    Curated datasets may bias coverage, long-sequence encoding and GRPO reduce scalability, DNA truncation may omit distal context, and uncertainty quantification is limited.

    Source: §6 · p. 9

Abstract

from arXiv · show

Unlocking deep and interpretable biological reasoning from complex genomic data remains a major AI challenge limiting scientific progress. While current DNA foundation models excel at representing sequences, they struggle with multi-step reasoning and lack transparent, biologically meaningful explanations. BioReason addresses this by tightly integrating a DNA foundation model with a large language model (LLM), enabling the LLM to directly interpret and reason over genomic information. Through supervised fine-tuning and reinforcement learning, BioReason learns to produce logical, biologically coherent deductions. It achieves major performance gains, boosting KEGG-based disease pathway prediction accuracy from 86% to 98% and improving variant effect prediction by an average of 15% over strong baselines. BioReason can reason over unseen biological entities and explain its decisions step by step, offering a transformative framework for interpretable, mechanistic AI in biology. All data, code, and checkpoints are available at https://github.com/bowang-lab/BioReason

1 Introduction

DNA foundation models represent genomic sequences well but often lack transparent, mechanistic explanations, while LLMs struggle to process raw DNA. BioReason combines the two and trains for multi-step biological reasoning, reporting improved performance and interpretable reasoning traces.

Motivation

  • DNA foundation models can represent sequences effectively, but often act as black boxes without biologically intuitive explanations for complex mechanistic tasks [4] [26].The introduction names gene pathway analysis, phenotype prediction, and disease mechanism elucidation as examples.Source: §1 · p. 2
  • LLMs offer multi-step reasoning but lack specialized architecture for raw genomic sequences and may miss nuanced genetic patterns [32] [12] [34].The paper frames the gap as a disconnect between sequence representation and reasoning capabilities.Source: §1 · p. 2

Approach

  • BioReason integrates a DNA foundation model with an LLM so the LLM can reason directly over genomic information.The model combines supervised fine-tuning and reinforcement learning to generate biologically coherent, multi-step explanations.Source: §1 · p. 2

Contributions

  • The paper contributes an annotated KEGG-derived benchmark for gene pathway and disease prediction, alongside interpretable, step-by-step reasoning traces.The traces are intended to provide interpretable predictions that can support scientific insight and hypothesis generation.Source: §1 · p. 2

Empirical performance improvements

  • 15%+ average performance gains over baseline are reported for BioReason against DNA foundation models and LLMs.The paper also identifies step-by-step biological reasoning traces as a contribution.Source: §1 · p. 2

2 Background & Related Work

Prior work supplies strong sequence representations and language-model reasoning, but existing genomics benchmarks do not adequately evaluate higher-order biological reasoning. BioReason is positioned as a multimodal reasoning framework and introduces a benchmark aimed at that gap.

2.1 DNA Foundation Models

  • DNA foundation models learn useful genomic representations and perform across genomics tasks, but their black-box outputs can obscure mechanistic insights [43] [31].The section highlights Evo2’s long-range context capability while noting the interpretability limitation.Source: §2.1 · pp. 2–3

2.2 Large Language Models for Biological Reasoning

  • LLMs reason over biomedical text, while prior genomic and multimodal systems had not been trained for complex biological reasoning [43] [25] [19].BioReason is presented as combining textual knowledge from LLMs with nucleotide-level representations from DNA foundation models.Source: §2.2 · p. 3

2.3 Genomics Benchmarks

  • Standard genomics benchmarks measure downstream prediction but inadequately test higher-order reasoning or hypothesis generation; BioReason curates KEGG data for mechanistic reasoning evaluation.The proposed evaluation targets multi-step logic and prediction of potential biological mechanisms.Source: §2.3 · p. 3

3 BioReason Model

BioReason forms a unified input from genomic sequences and textual queries so an LLM can produce explanations grounded in genomic information. Its pipeline uses DNA-model embeddings, an LLM backbone, multimodal integration, and reinforcement learning.

3 BioReason Model

  • BioReason contextualizes genomic sequences with a DNA foundation model and combines those embeddings with tokenized text queries to generate genomic-grounded explanations.The model’s output includes both biological reasoning and a final response.Source: §3 · p. 3, Figure 1 · p. 4

3.1 DNA Foundation Model (fDNA) Encoder

  • The DNA foundation model converts each sequence into contextualized per-token embeddings, truncating inputs that exceed the defined context length; its weights remain frozen during training and inference.The passage gives 2048 DNA tokens as an example context length.Source: §3.1 · p. 4

3.2 Large Language Model (fLLM) Backbone

  • Qwen3 serves as the autoregressive LLM backbone, receiving the multimodal input and predicting the explanatory output token by token.Special tokens structure conversational roles and reasoning steps in the output.Source: §3.2 · p. 4

3.3 Multimodal Genomic Integration

  • DNA embeddings are projected to the LLM embedding dimension and stacked with query and special-token embeddings to form the multimodal input sequence.RoPE supplies positional information across the resulting genomic and textual components.Source: §3.3 · pp. 4–5

3.4 Group Relative Policy Optimization (GRPO)

  • GRPO refines reasoning generation using reward signals across groups of sampled outputs, without an explicit value estimator; BioReason implements the Dr. GRPO variant.The paper describes Dr. GRPO as an unbiased variant that improves token efficiency while maintaining performance.Source: §3.4 · p. 5

4 Datasets

BioReason combines a novel KEGG-derived reasoning dataset with adapted ClinVar- and OMIM-based variant-effect datasets. The datasets support mechanistic variant-to-disease reasoning and coding-sequence variant prediction tasks.

4 Datasets

  • The three datasets comprise a novel KEGG Network Variants reasoning dataset and two adapted datasets for variant-effect prediction derived from ClinVar and OMIM.The KEGG data uses curated pathway descriptions and downstream phenotypic effects.Source: §4 · p. 5

4.1.1 Dataset Integration and Statistics

  • The KEGG-derived dataset contains 1,449 entries linking genetic variants to disease phenotypes through mechanistic connections.Pathway networks were augmented with variant information and represented using symbolic interaction notation.Source: §4.1.1 · p. 5

4.1.2 Reasoning Path Construction and Curation

  • KEGG reasoning examples pair variant, network, and gene information in questions with mechanism–disease answers and accompanying reasoning paths.The paths were constructed using Claude 3.7 Sonnet and grounded in KEGG disease information; their mean length is 303.8 words.Source: §4.1.2 · pp. 5–6, Figure 2 · p. 6

4.2 Variant Effect Prediction of Coding Sequences

  • Coding-sequence variant examples originate from GPN-MSA; chromosome-based splitting holds chromosome 8 for testing, and GPT-4o generates 50 semantically equivalent question variations per sample [32].Benign variants require allele number ≥25,000 and minor allele frequency > 5%.Source: §4.2 · p. 6

4.3 Variant Effect Prediction of Coding Non-SNVs

  • Coding non-SNVs are filtered from the 2024-02-28 ClinVar release for nuclear-genome variants affecting ≤64 base pairs, with at least two-star review status and GRCh38.p14 transcript matches [32].A custom algorithm balances disease representation across train/test splits, and GPT-4o generates 50 question variations per entry.Source: §4.3 · p. 6

5 Experiments

Experiments evaluate BIOREASON on KEGG-derived pathway reasoning and coding and non-SNV variant-effect prediction, comparing DNA–LLM hybrids with single-modality models. The hybrid models lead the reported comparisons, and a PFN1 case study illustrates a stepwise variant-to-disease rationale.

5.1 Datasets

  • The KEGG-derived dataset tests multi-step reasoning over 1,449 variants and 37 diseases; coding and non-SNV datasets assess pathogenicity classification and associated disease prediction.The datasets contain 50,083 core coding-variant entries and 36,088 core non-SNV entries; the coding test split holds out chromosome 8.Source: §5.1 · p. 7

5.2 Models and Baselines

  • BIOREASON pairs Evo2-1B or NT-500M encoders with Qwen3-1.7B or Qwen3-4B backbones, using supervised fine-tuning and applying GRPO to selected models.Comparators include pretrained DNA foundation models with task-specific attention heads and fine-tuned Qwen3 models receiving DNA as text.Source: §5.2 · p. 7

5.3 Experimental Setup

  • Training adapts BIOREASON and LLM-only models with LoRA, freezes core DNA-model weights for fDNA-only baselines, and trains responses for reasoning or variant classification tasks.A specialized attention mask restricts loss to the generated response; evaluation uses task-specific accuracy and F1 metrics.Source: §5.3 · p. 7

5.4 Quantitative Results

  • 98.28% accuracy and 93.05% F1: Evo2+Qwen3-4B with GRPO outperforms the standalone Qwen3-4B and Evo2 models on KEGG-derived reasoning.For Evo2+Qwen3-4B, GRPO raises accuracy from 95.17% to 98.28% and F1 from 86.14% to 93.05%; NT+Qwen3-4B with GRPO reaches 98.28% accuracy and 90.15% F1.Source: §5.4 · p. 8, Table 1 · p. 8
  • 80.21% accuracy and 80.00% F1 for coding variants, and 88.20% accuracy and 89.91% F1 for non-SNVs, exceed the reported single-modality baselines.Evo2+Qwen3-4B leads the coding comparison; Evo2+Qwen3-1B leads the non-SNV comparison.Source: §5.4 · p. 8, Table 2 · p. 8

5.5 Case Study

  • In the PFN1 case study, BIOREASON predicts ALS and gives a 10-step rationale linking a C>G substitution to profilin-1 dysfunction, impaired actin dynamics, disrupted axonal transport, and motor-neuron degeneration.The example follows the pathway Actin(monomeric) // PFN1* // Actin(filamentous).Source: §5.5 · p. 9

6 Discussion

BIOREASON combines contextualized DNA embeddings with LLM reasoning to produce predictions and stepwise explanations. The discussion highlights potential applications alongside limitations in dataset coverage, computational cost, sequence length, and uncertainty estimation.

Framework

  • BIOREASON integrates DNA foundation models and LLMs, aiming to address opaque DNA-model outputs and LLMs’ limited ability to process raw DNA sequences.The framework is trained for multi-step biological reasoning and is presented as outperforming single-modality approaches.Source: §6 · p. 9

Interpretability

  • Stepwise mechanistic explanations accompany predictions, allowing researchers to scrutinize model logic and formulate testable scientific hypotheses.The discussion describes this interpretability as a core strength of processing contextualized DNA embeddings within the LLM.Source: §6 · p. 9

Limitations

  • The work identifies curated-dataset bias, high training and inference costs, truncation to 2048 tokens, and limited uncertainty quantification as constraints.These constraints may limit coverage of less-characterized regions, scalability to whole-genome or real-time settings, and reliability in high-stakes decisions.Source: §6 · p. 9

Future Work

  • The authors propose adding orthologous sequences and adapting BIOREASON to RNA and protein, with GWAS and clinical mutation interpretation as possible applications.These directions are framed as future work to broaden the framework’s scope and applicability.Source: §6 · p. 9

7 Conclusion

The conclusion presents BIOREASON as a multimodal framework combining DNA sequence encoders with LLM reasoning for pathway inference and variant pathogenicity prediction. It reports stronger performance than single-modality models and transparent stepwise explanations across the evaluated benchmarks.

Conclusion

  • BIOREASON’s DNA–LLM hybrids outperform single-modality models across KEGG-derived reasoning and VEP benchmarks while generating transparent, stepwise explanations.The conclusion identifies future directions including larger models and datasets, leaner architectures, and protein and RNA modalities.Source: §7 · p. 10

A Training Details

The appendix reports training settings and compute resources, then describes GRPO rewards and optimization. Its composite reward emphasizes answer correctness and adherence to a specified reasoning format.

A.1 Hyper-Parameters

  • The reported configurations use gradient accumulation of 8 steps and 4 steps, while task runs cap DNA input at 2048 tokens; KEGG uses batch size 1 for 5 epochs, and variant prediction uses batch size 2 for 3 epochs.Source: §A.1 · pp. 17–18

A.2 Computational Resources

  • Experiments used NVIDIA A100 and H100 GPU clusters, with four A100 GPUs assigned to reinforcement learning and other experiments run on single H100 GPUs.The appendix also reports CPU and memory configurations for A100 systems.Source: §A.2 · p. 18

A.3 Reward Details

  • GRPO uses a deterministic composite reward for correctness, concise final answers, and the required reasoning format; the components sum to a reward in [0, 2.5].The components are non-differentiable and used for group-normalized advantages; no learned critic or value function is used.Source: §A.3 · p. 18

A.4 GRPO Details

  • GRPO samples multiple outputs per query, scores them with the composite reward, and optimizes policy parameters using a clipped surrogate objective.The described process assigns each candidate a reasoning trace and final response before reward evaluation.Source: §A.4 · p. 18

A.5 GRPO Learning Curve

GRPO learning curves show early reward gains, faster stabilization for larger models, and convergence near the reward maximum. The appendix also frames a test of generalization to a supervised DNA encoder and chromatin-accessibility prediction.

A.5 GRPO Learning Curve

  • 4B models stabilize around step 400 versus step 800 for the 1.7B model and show lower late-training reward variance.Source: §A.5 · p. 19, Figure 4 · p. 19
  • NT and Evo2 encoders paired with Qwen3-4B follow nearly identical curves, indicating GRPO generalizes across these DNA encoder architectures.Source: §A.5 · p. 19, Figure 4 · p. 19
  • All configurations eventually approach reward 2.5, reflecting acquisition of correct answers and the required reasoning format.Source: Figure 4 · p. 19, §A.5 · p. 19

B Generalization to Supervised DNA Models

B.1 Experimental Setup

The experiment formulates chromatin accessibility as multi-label classification over genomic bins and compares standalone, LLM-only, and hybrid models on held-out data.

B.1 Experimental Setup

  • The DeepSEA-derived dataset contains 60,000 examples from 200bp human-genome bins, with open/closed accessibility labels across 20 DNase-seq tracks.Source: §B.1 · p. 20
  • Classification matches DNase-seq’s binary accessibility states, while the autoregressive architecture is optimized for discrete predictions despite Enformer’s regression training.Source: §B.1 · p. 20

B.2 Model Configurations

  • The comparison includes frozen Enformer, Qwen3-1.7B and Qwen3-4B LLM-only models, and matching Enformer-Qwen3 hybrids evaluated on a held-out test set.Source: §B.2 · p. 20

B.3 Results

On chromatin-accessibility prediction, BIOREASON hybrids outperform standalone DNA and language models, including when paired with supervised Enformer. The results support encoder flexibility while distinguishing task-specific from general-purpose representations.

B.3 Results

  • Enformer’s hybrid gains support the authors’ view that combining sequence embeddings with LLM reasoning is not limited to unsupervised DNA encoders.Source: §B.3 · p. 20
  • 33.70% Macro F1 for Enformer-Qwen3-4B, nearly double standalone Enformer’s 17.18%, with hybrids outperforming both DNA-only and LLM-only baselines.Source: §B.3 · p. 20, Table 3 · p. 21
  • Enformer’s specialized embeddings may be less transferable to novel tasks than unsupervised models’ general-purpose representations, motivating the latter for broad reasoning.Source: §B.3 · pp. 20–21
  • BIOREASON accommodates supervised and unsupervised DNA encoders, allowing users to select a foundation model for their application domain.Source: §B.3 · p. 21

C Per-Disease Performance on the KEGG-Derived Reasoning Benchmark

BioReason reports consistently strong KEGG reasoning performance across disease classes, including beyond high-frequency classes. The accompanying checklist reports single-run statistical limits and supporting reproducibility and release practices.

C Per-Disease Performance on the KEGG-Derived Reasoning Benchmark

  • BioReason shows consistently strong performance across diverse KEGG disease classes, including beyond high-frequency classes and across varied mechanistic contexts.Table 4 covers per-disease metrics for NT + Qwen3-4B and Evo2 + Qwen3-4B, with metrics reported as percentages.Source: C Per-Disease Performance on the… · p. 21, Table 4 · p. 21

4. Experimental result reproducibility

  • The authors report that Appendix A provides information needed to reproduce the main results and understand the experimental setting.The reproducibility response also reports that the source code and data are provided with instructions for reproducing the main experiments.Source: §4 · p. 23, §6 · p. 24, §5 · p. 24

5. Open access to data and code

  • The authors report releasing code and data with instructions for reproducing the main experimental results.They also state that new assets are accompanied by documentation and experiment scripts.Source: §5 · p. 24, §13 · p. 27

7. Experiment statistical significance

  • Comprehensive statistical analyses were limited by compute resources; deterministic generation at temperature 0 provides point estimates, with inter-run variability deferred to future work.The authors report that text generations from BioReason and LLM-only models used temperature 0.Source: §7 · p. 24

9. Code of ethics

  • The authors report that the research conforms to the NeurIPS Code of Ethics and that the creators of used assets are appropriately credited.Source: §9 · p. 25, §12 · p. 26

10. Broader impacts

  • The authors describe potential positive impacts on biological discovery and disease understanding, while flagging curated-data bias and the need to validate interpretable outputs.They frame the impacts as potential and call for cautious scientific application.Source: §10 · p. 25

11. Safeguards

  • The authors report that the research did not use human subjects or crowdsourcing and released no data or models they considered high risk for misuse.Source: §11 · p. 26, §14 · p. 27
Loading 2505.23579v2…