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

We thank the reviewer for their thorough review, detailed feedback, and especially the thoughtful suggestions! Below, we address the identified weaknesses.

Interpretability. Thank you for raising this important point and suggestion. We will revise our writing in the form of a new subsection of the paper, touching on the scope and limitations of our framing of interpretability.

Questions.

1. We appreciate this question about the effects of truncation. We used a 2048 nucleotide context window centered on each variant, following established practices from prior Nucleotide Transformer and Evo2 studies. As the context window was chosen after data curation, we truncated our sequences during batching to fit sequences to this 2048 nucleotide window.

   We acknowledge that truncation may result in the loss of distal regulatory elements or long-range genomic context that could be relevant for variant interpretation. However, we have not systematically evaluated how increasing the context window would affect performance, as this was constrained by our computational resources and model limitations. This represents an important direction for future work - understanding the trade-offs between computational efficiency and the inclusion of longer-range genomic context.

   We will clarify the truncation methodology in section 3.1 of the final manuscript and acknowledge this limitation in our discussion.

2. We thank the reviewer for pointing out this space for improved communication. We will update Figure 1 with labels matching the steps from 3.3, as well as specify that we concatenate the embeddings along the token dimension.

3. We utilized the established chromosome-based split strategy from the InstaDeepAI/genomics-long-range benchmark, which is a processed version of the data from the GPN-MSA paper. This splitting approach tests model generalizability to an entirely new genomic context and represents a rigorous evaluation strategy. We maintained this split in our dataset to ensure direct comparability with existing benchmarks and to provide a fair assessment of BioReason’s generalization capabilities across chromosomal regions.

4. Please see Tables 1-2 below for results.

Performance Across Disease Representations.

Table 1: NT + Qwen3 - 4B

Table 2: Evo2 + Qwen3 - 4B

5. We thank the reviewer for this insightful question. Table 2 of the paper provides strong evidence that our performance gains are not simply due to increased model size, but due to a higher quality DNA embedding being passed into the LLM.

   A particularly insightful comparison is between Qwen3-4B and NT + Qwen3-1B. The BioReason model (NT + Qwen3-1B) has fewer total parameters than the standalone Qwen3-4B, yet performs significantly better on variant effect prediction tasks. We believe this is due to the quality of the cross-modal embeddings from the DNA model, which is then integrated as part of the LLM’s prediction context. The LLM can then leverage these representations to achieve superior performance on biological tasks, even with fewer parameters. This suggests that BioReason's improvements result from the combination of domain-specific DNA representations and LLM reasoning capabilities.

6. We thank the reviewer for this suggestion to tie the discussion directly to the figure, and will make this change in the final manuscript.

Dataset Creation Hallucination Risk. We acknowledge that this is an important concern. To mitigate this risk, we curated ground truth data on mutation effects and downstream phenotypic impacts from the KEGG database, which contains expert-curated summaries with citations on mechanistic effects. This served as our gold standard, and we used Claude solely to reformat these verified paragraphs into chain-of-thought reasoning traces rather than generating novel content.

To assess quality, we conducted manual validation on a random sample of generated reasoning traces to verify biological accuracy and consistency. While quantifying exact confidence scores remains an open issue for LLM-generated content, our approach of providing comprehensive context and restricting the model's role to reformatting ground truth data mitigates hallucination risk.

This methodology, combined with our validation sampling, provides confidence that our reasoning traces are biologically accurate and constitute reliable training data.

User Study. We thank the reviewer for this excellent suggestion regarding user studies to evaluate our interpretability claims. We wholeheartedly agree that a comprehensive evaluation with domain experts would provide invaluable insights into BioReason's real-world utility and guide improvements for practical use with biologists studying unknown mutations.

Such a study would require extensive coordination with domain experts, such as clinical geneticists. We believe this evaluation is important, but would be better suited as dedicated follow-up work that can properly address the complexities of interpretability in genomics.

Thank you, Authors, for the thorough rebuttal and additional details.

Interpretability : What definition of interpretability will be used?

Question Rebuttals :

1. Thanks for the elaboration and thoughtfulness about incorporating the explanation into the paper. 2 & 6. Already a well written paper, and looking forward to seeing the even clearer version. Thanks!
2. Great, thank you for the clarification and attention to methodological detail.
3. Thank you for the additional details - great to see fairly consistent performance across classes.
4. Convincing, thank you for pointing out the data that supports this strength of the model.

Hallucinations : Thank you for your arguments. **User Studies **: Yes, looking forward to this dedicated follow up work.

We thank the reviewer for their positive assessment and insightful feedback. We are encouraged that they found our approach innovative and the results convincing. In response to the constructive weaknesses identified, we have conducted a new experiment that addresses both suggestions: evaluating our framework with a supervised DNA foundation model on a new task beyond sequence-to-disease reasoning.

---

New Experiment: Chromatin Accessibility Prediction

We evaluated the BioReason framework on a chromatin accessibility prediction task using the supervised DNA model, Enformer.

• Model Choice: We specifically chose Enformer as it is a powerful, state-of-the-art supervised model explicitly trained to predict genomic annotations like chromatin accessibility, directly aligning with the reviewer's suggestions.

• Data and Task: We utilized a public dataset for long-range genomics containing 60,000 data points. We trained all models on a subset of this data and evaluated their performance on a held-out test set. The task was formulated as a multi-label binary classification problem to predict chromatin accessibility across 20 different tracks.

Results

Our results demonstrate that the BioReason framework's core advantages hold true. The hybrid Enformer-Qwen models significantly outperformed both the standalone Enformer model and the Qwen LLMs.

The Enformer-Qwen3-4B model doubled the F1 Macro score of the standalone Enformer model. This clearly shows that integrating sequence embeddings with an LLM via the BioReason architecture provides a substantial performance boost, even when using a supervised DNA encoder.

---

Discussion

This experiment provides two key insights:

1. The BioReason framework is robust and generalizable. The significant performance lift demonstrates that the synergy between a DNA model and an LLM is a general principle, not an artifact of using unsupervised DNA encoders like Evo2 or Nucleotide Transformer (NT). This confirms the fundamental value of combining a specialized sequence encoder with the flexible reasoning engine of an LLM, regardless of the encoder’s training paradigm.

2. Our initial choice of unsupervised DNA foundation models in the manuscript was deliberate. While supervised models like Enformer are powerful, they are highly specialized for the tasks they were trained on, such as predicting a predefined set of epigenetic marks. In contrast, unsupervised models like Evo2 and NT learn more general-purpose representations of genomic syntax and semantics by being trained on vast, diverse genomic data without being constrained to a specific set of labels.

In our new experiment, the standalone Enformer model's performance was modest, suggesting its specialized embeddings, while potent, may not be as rich or transferable for novel downstream tasks. While our framework still extracts significant value, the overall performance ceiling is ultimately influenced by the quality and generalizability of the initial DNA embeddings. For our paper's central goal of creating a system for broad, multi-step biological reasoning, we believe the rich, flexible embeddings from unsupervised models provide a more robust foundation, better suited for the diverse and complex queries we aim for BioReason to handle.

We thank the reviewer again for their valuable suggestions, which have helped us strengthen our work and further validate the robustness of the BioReason framework.

We thank the reviewer for their thorough evaluation and insightful feedback on our work. We appreciate the recognition of our contributions in multimodal innovation, interpretable reasoning, and benchmark development. Below, we address the identified weaknesses.

Dataset Bias and Generalizability. We selected the KEGG dataset, despite its bias towards well-studied pathways, because its expert-curated nature was essential for our primary goal of training a model to perform deep, multi-step reasoning on genomic sequence. This required a foundation of ground-truth causal chains linking variants to disease phenotypes. KEGG provided the ideal, high-quality data necessary to rigorously benchmark this novel capability, which is a core contribution of our work.

To prove the framework generalizes beyond these curated examples, we trained and evaluated it on two large-scale Variant Effect Prediction (VEP) benchmarks. These tasks challenge the model with over 86,000 coding variants sourced from broader databases like ClinVar and gnomAD, covering diverse mutation types including SNVs and non-SNVs. Many of these variants are under-characterized compared to the KEGG examples. BioReason’s high accuracy across these benchmarks confirms its ability to generalize, effectively classifying a wide array of variants without relying on the detailed pathway information used during its reasoning training.

Computational Overhead. We have made several deliberate design choices to ensure our framework remains efficient. During training, the large DNA foundation model is kept frozen, and we employ Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning of the language model, which significantly reduces the training burden. Additionally, we truncate input DNA sequences to a fixed context length during training to save memory. For inference, our publicly available framework is designed for compatibility with optimized inference servers like vLLM, which dramatically reduces latency and enhances scalability for practical applications through efficient batching and memory management. This also drives down the computational cost of GRPO stage, which is mostly bottlenecked by compute required for generations.

Lack of Uncertainty Quantification. We thank the reviewer for highlighting this critical issue. The lack of robust uncertainty quantification is indeed a significant limitation and a well-recognized, field-wide challenge for LLMs. Current LLMs are fundamentally optimized to predict the next most probable sequence of tokens, which often results in confident-sounding, definitive outputs, even when the underlying information is ambiguous. Because BioReason leverages an LLM as its core reasoning engine, it naturally inherits this limitation. However, we designed BioReason's core mechanism of generating interpretable reasoning traces as a practical step towards addressing this trust deficit. While the model does not output a formal confidence score, it produces transparent, step-by-step biological explanations that allow domain experts to scrutinize the model's logic, evaluate its deductions, and ultimately build confidence in its conclusions in a way that is impossible with conventional black-box models.

Implementation Complexity and Reproducibility. To ensure our results are fully reproducible, we will make all data, training code, and model checkpoints publicly available. For deployment, we are preparing our models for release on standard platforms like HuggingFace and have designed the framework to be compatible with high-throughput inference engines like vLLM, ensuring fast and efficient operation. Importantly, our approach is not limited to large-scale systems; we have demonstrated that the BioReason architecture provides significant performance gains even when using smaller 1.7B parameter models, making the benefits of our framework accessible across different computational budgets.

Limited Modal Diversity and Scope. We thank the reviewer for this elegant future work suggestion. Our primary goal with this work was to establish BioReason as a foundational proof-of-concept. We focused on the deep integration of a DNA foundation model with a large language model to prove that our architecture could successfully unlock interpretable, multi-step reasoning from fundamental genomic sequences. This focused approach allowed us to rigorously validate this novel architectural paradigm. Adapting the core framework to other modalities, such as RNA and protein sequences, is a key priority. This extension is a central part of our upcoming BioReason2 model and in line with our vision of training multimodal biological LLMs that can reason across DNA, RNA, and Protein.