acceptodds
Under review as a conference paper at ICLR 2027

BioBLIP: A Multimodal Architecture for Zero-Shot Task Transfer in Genomic Variant Interpretation

Abstract

Developing scientific hypotheses in biology requires integrating heterogeneous evidence from DNA sequence, gene context, protein function, and prior literature. Existing multimodal AI systems expose biological evidence to LLMs through textification or by projecting biological embeddings into fine-tuned language models. However, these models' performance is limited to the specific tasks for which they are fine-tuned. Here we present BioBLIP, a multimodal Q-former based architecture which leverages biological embeddings and a LLM to generalize to unseen genomic tasks without fine-tuning. The key to BioBLIP is a new model architecture that integrates four data modalities – DNA, genes, proteins, and text – through a master Qformer model, which consolidates the modality-specific information into multimodal tokens for a frozen LLM. BioBLIP is pretrained on the task of human single nucleotide variant annotation (85.2% overall accuracy). Without fine tuning and only 66M trained parameters, pretrained BioBLIP achieves 70.2% Top-1 accuracy in variant prioritization, as opposed to 64.9% accuracy for LLMs given matched biological evidence in text and 59.3% for 40B parameter genomic foundation models. In target gene prediction, BioBLIP outperforms matched-text LLM baselines (60.50% versus 43.70%) while producing rich, transparent traces. Overall, BioBLIP demonstrates strong capabilities to leverage learned multimodal representations for robust zero-shot transfer to multi-step inference tasks in genomics.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multimodality, LLMs, genomics, Biological foundation models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NxXkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces BioBLIP to investigate whether pretraining on variant annotation enables a multimodal model to transfer to genomic tasks without task-specific training. The model uses AlphaGenome, GenePT, and ESM-2 to obtain DNA, gene-neighborhood, and protein representations, respectively. Modality-specific Q-Formers and a Master Q-Former fuse these representations into soft prompts for a frozen Qwen language model. Training first aligns each modality with annotation text, then uses a structured variant-annotation generation loss to update the modality-specific Q-Formers, Master Q-Former, and projection layer, while keeping the biological encoders and Qwen frozen. In downstream evaluations, BioBLIP performs variant prioritization and target-gene prediction without additional fine-tuning, outperforming the paper’s Qwen3-32B text baseline on both tasks. The largest improvement occurs on the target-gene prediction subset where the target is not the gene closest to the variant, with accuracy increasing from 43.70% to 60.50%.

Critical strengths
  1. BioBLIP achieves a significant improvement in identifying target genes that are not the closest to the variant. On this subset, BioBLIP outperforms the equally sized Qwen3-32B text baseline by 16.80 percentage points. The advantage persists on the stricter subset where each test variant is more than 1 kb from every training variant.

  2. The paper demonstrates the feasibility of using variant-annotation supervision to integrate multimodal biological information and support cross-task transfer. A single model trained on variant annotation transfers to multi-candidate variant prioritization and target-gene selection without updating the LLM parameters or performing downstream fine-tuning. Connecting multiple biological encoders to a frozen LLM through hierarchical querying modules provides a useful design for reusing biological representations.

Critical weaknesses
  1. BioBLIP and the text baseline receive different information and training, so the source of the gains is unclear.

    • GenePT contains gene-function descriptions, while the text baseline mainly receives gene names and distances. BioBLIP also receives OpenTargets annotation training and uses generated annotations for variant prioritization. The gains may therefore come from additional knowledge or training, rather than multimodal representations alone.
  2. The ablations do not isolate the sources of downstream gains.

    • Joint embedding shuffling disrupts multiple information sources.
    • Table 5 does not clarify whether reduced-modality models were retrained.
    • The Master Q-Former ablation reports only annotation results with Qwen3-4B trained for 15 epochs.
    • These results leave the downstream contributions of individual inputs and the fusion module unclear.
  3. Gene-prediction accuracy alone does not support the paper’s stronger claims about mechanistic reasoning.

    • Appendix G acknowledges correct gene predictions with incorrect intronic placements in the ACADL, CPT2, ASIP, and LCT examples.
    • The main evaluation scores the final gene without systematically assessing the accuracy of the explanations.
  4. Generalization across LLM backbones remains untested. Experiments only cover Qwen3-4B and Qwen3-32B, leaving it unclear whether BioBLIP’s downstream gains extend to other LLM families.

Questions for authors

See weaknesses.

5 comments

We thank the reviewer for thorough and constructive comments. Based on reviewer’s valuable feedback we conducted a number of additional experiments and provided additional analyses, which further validate the efficacy of our BioBLIP architecture. We believe these together further strengthen our work.

Below please find responses to individual comments/questions:

W1. BioBLIP and the text baseline receive different information and training.

A1. Thanks for the comment. As mentioned in Section 2.2, the LLM backbone of BioBLIP is not trained. Neither is the LLM in the text baseline.

The main contribution of BioBLIP is a new multimodal architecture which is pretrained to align biological foundation model embeddings with variant annotations. We show that the model performs well at variant annotation task and then demonstrate zero-shot task transfer. Pretraining the Qformers on OpenTargets variant annotation is an important part of the contribution and should not be viewed as a confound. To control for training-test set leakage, we report results on held-out loci for both downstream tasks.

BioBLIP and the text baseline receive different information by necessity, because it is impossible to completely textify the latent information contained in a biological embedding such as GenePT or AlphaGenome. For example, GenePT was constructed with GPT-3.5 embeddings (Chen et al, 2024). A major purpose of BioBLIP's architecture is to make such embedded knowledge available to an LLM.

To construct a strong text baseline, we provide textified AlphaGenome predictions and the same nearest-gene identities and distances used by BioBLIP and show the results in Section 3.2.

Variant prioritization on Traitgym is already a task that benefits greatly from accurate variant annotations. BioBLIP produces more accurate variant annotations through its multimodal tokens. However, we conducted an additional experiment to identify whether multimodal tokens add any predictive capabilities beyond BioBLIP's improved variant annotations in this specific Traitgym benchmark.

Takeaways: this experiment suggests that BioBLIP's multimodal tokens add predictive information beyond its generated variant annotations. For difficult, poorly sampled regions of the genome, adding multimodal tokens plus the annotations provides 15.05% Top-1 accuracy improvement and 26.16% Top-2 accuracy improvement.

In this experiment, Qwen3-32B was used to rank the variants with exactly BioBLIP's generated variant annotations, but no multimodal tokens. We also report the results on a more challenging subset of n=51 groups with variants from poorly sampled regions of the genome in OpenTargets (>1kb from any other variant).

BioBLIP Improvement, Relative

SubsetTop-3 AccuracyTop-2 AccuracyTop-1 Accuracy
All Held-out Loci3.31%5.76%1.93%
Held Out & Distance >1kb11.15%26.16%15.05%

W2. Isolating Sources of Downstream Performance Gains

A2. Thanks for this question. We conducted an additional experiment ablating the master Q-former and measuring performance on the downstream tasks. We trained a BioBLIP-style Qwen3-32B model with no master Q-former and the same training protocol.

Takeaways: The experiments demonstrate that the Master Q-former provides a large benefit in downstream task performance. This key component of the architecture provides ~22% improvement in accuracy on regulatory variant prioritization. It provides ~28% improvement in accuracy on target gene prediction on the challenging subset of non-closest gene cases. This Master Q-former is critical for zero-shot transfer performance.

Task 1: Regulatory Variant Prioritization

Top-1 AccuracyTop-2 AccuracyTop-3 Accuracy
Master Qformer70.22%76.49%79.62%
No Master Qformer47.93%53.55%56.21%
Delta (No Master – Master)-22.29%-22.94%-23.41%

Task 2: Target Gene Prediction

Non-Closest Gene & Held Out Loci (n=119)Non-Closest & Distance Cutoff (n=92)Held-Out Loci (n=369)
Master Qformer60.50%59.78%83.47%
No Master Qformer32.77%31.52%77.78%
Delta (No Master – Master)-27.73%-28.26%-5.69%

We have added these results to Section 4 of the paper.

W2.3 The Master Q-Former ablation reports only annotation results with Qwen3-4B trained for 15 epochs.

A2.3 Please see A2.


(contd. below)

W2.1 Clarifying Embedding Permutation Experiment

W2.1 Joint embedding shuffling disrupts multiple information sources.

A2.1 Thanks for raising this. The purpose of the embedding permutation experiment is to answer the question: does the model use its biological inputs at all? This is not guaranteed. MIRAGE (Asadi et al., 2026) shows models scoring well on multimodal benchmarks while ignoring the images, and concurrent work (arXiv:2610.00898) finds most multimodal models don’t rely on their inputs. Permuting embeddings drops BioBLIP’s performance by 70–89% on all three tasks (Table 4), confirming that BioBLIP does rely on the multimodal data.

W2.2 Reduced Modality Ablations

W2.2 Table 5 does not clarify whether reduced-modality models were retrained.

A2.2 Table 5 shows ablations without retraining. The ablations were applied at inference to the trained full model, with removed modalities replaced by [noise]. This study measures what the fully-trained and deployed model depends on.

Retraining the models from scratch with fewer modalities shows less of a gap (see results below reported for Qwen3-4B backbone; sampling done with three random seeds). The same modality trends are clear in the retrained models. AlphaGenome modality contributes the most to Chromosome, Variant Type, and Variant Consequence. GenePT contributes the most to the nearest genes field. ESM protein embeddings contribute least overall - which is expected because these are mostly noncoding variants which do not affect the protein sequence.

We have qualified in the paper that the nonlinear effects observed with modalities is at inference-time. Full results for the retrained models are below, with the greatest gap bolded.

Field3 modalityAFG ablationESM ablationGenePT ablation
Chromosome90.1 ± 0.484.8 ± 0.386.8 ± 0.389.3 ± 0.3
Variant Type89.7 ± 0.486.9 ± 0.389.2 ± 0.287.6 ± 0.6
Variant Consequence67.7 ± 0.863.0 ± 0.967.2 ± 0.565.0 ± 0.6
Nearest Genes64.1 ± 0.364.9 ± 0.365.4 ± 0.249.3 ± 0.3

Takeaways: The large gaps are due in part to inference-time ablations. With retrained reduced-modality models, the gaps are less, but the field-level contribution of each modality remains the same.

We have added these results and explanations to the Appendix.

W3 Factual Precision of Target Gene Reasoning Traces

Gene-prediction accuracy alone does not support the paper’s stronger claims about mechanistic reasoning.

A3. Thanks for the insightful comments. Such an analysis is indeed helpful.

We selected 100 random target gene identification traces, and analyzed them for the correctness of the claims using LLM-as-a-judge (Opus 5.5) and VEP Ensembl annotations and prior literature as evidence sources. We also hand checked a subset of 10 traces for agreement and found 100% alignment with the LLM judge's decisions.

Below, we demonstrate that on target gene identification (1) precision of the claims made is 90.5% overall (2) 27.5% of claims made are hypotheses. Hypotheses are defined as statements that need additional eQTL or experimental data to be verified. We have added this analysis to Appendix I.

StagePrecisionCorrect claimsHypothesesIncorrect claimsNumber of Claims
S1 Variant annotation82.10%82.10%0.00%17.90%285
S2 Gene-level effect89.40%58.40%34.60%7.00%762
S3 Cellular process94.40%69.80%26.10%4.10%510
S4 Tissue-level effect92.90%66.70%28.30%5.10%435
S5 Disease mechanism92.80%61.70%33.50%4.80%457
OVERALL90.50%65.60%27.50%6.90%2449

Precision = (Correct claims) / (Incorrect claims + Correct claims)

Takeaways: BioBLIP provides 90.5% precision on 100 randomly selected traces, using LLM-as-a-judge and humans manually checking a subset of the traces. Precision is defined as the number of correct claims over the number of verifiable claims.

However, BioBLIP does not presume to guarantee step-by-step correctness in the traces produced. The primary contribution of the paper is a new multimodal architecture for zero-shot transfer in genomic tasks, and the step-by-step reasoning traces are useful because they allow an expert to audit them and investigate biological hypotheses raised. These are strengths over a predictive-only architecture.

Improving groundedness, faithfulness, and other properties of LLM traces is an active area of research with tools such as VERDICT (Zhou et al, 2026). We leave applying this research to BioBLIP as future work. We have updated this in the Discussion section of the paper.

(contd. below)

W4 Generalization across LLM backbones

A4. No component of BioBLIP is specific to the LLM. The LLM is frozen and receives soft tokens through a learned projection. We also provide results that performance improves across two model scales (4B and 32B, Table 1). Qwen3-32B is among the strongest open source LLMs for reasoning (Yang et al., 2025) and it would be nontrivial for the experiments to swap it out at this stage in terms of time and compute. We leave it to future work and have added it in the Discussion section of the paper.

Please let us know if you have any further questions or comments!


References

[1] Mohammad Asadi, Jack W O’Sullivan, Fang Cao, Tahoura Nedaee, Kamyar Fardi, Fei-Fei Li, Ehsan Adeli, and Euan Ashley. Mirage the illusion of visual understanding. arXiv preprint arXiv:2603.21687, 2026.

[2] Fang, Ada, et al. "When Do Biological Reasoning Models Use Their Biological Inputs?." arXiv preprint arXiv:2610.00898 (2026).

[3] Yang, An, et al. "Qwen3 technical report." arXiv preprint arXiv:2505.09388 (2025).

[4] Zikai Zhou, Yufei Jin, Yilin Xu, Yu-Chiang Wang, Chieh-Ju Chao, and Monica Lam. Accountable ai with grounded, faithful, consistent, actionable rationales: A case study in clinical trial matching with verdict. pp. 58987–59032, 2026.

Reviewer_GCpJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BioBLIP connects frozen biological encoders and a frozen Qwen3 language model through modality-specific Q-Formers and a master Q-Former. It combines AlphaGenome DNA embeddings, GenePT gene-neighborhood embeddings, and ESM-2 protein embeddings. Training first aligns each modality to variant-annotation captions and then learns a fused soft prompt through answer-token prediction, with approximately 66 million trainable parameters. After training on OpenTargets variant annotation, the model is evaluated without downstream task fine-tuning on regulatory variant prioritization and target-gene prediction. It reports 85.16% average annotation-field accuracy, 70.22% prioritization Top-1 accuracy versus 64.89% for a matched-text Qwen3-32B baseline, and 60.50% versus 43.70% target-gene accuracy on 119 held-out loci whose causal gene is not the closest gene. The paper includes stricter held-out subsets, embedding and architecture ablations, and qualitative reasoning traces.

Critical strengths
  1. Meaningful downstream gains with a frozen language model and an annotation-trained interface. The non-closest-gene evaluation is particularly valuable because it removes the trivial nearest-gene solution: BioBLIP achieves 60.50% versus 43.70% for Qwen3-32B with matched gene-neighborhood context on 119 held-out loci (Table 3, p. 8). Appendix G, Table 8 (p. 15), shows 21 cases correct only for BioBLIP versus one correct only for the baseline. The advantage remains on the 92 examples more than 1 kb from an annotation-training variant, at 59.78% versus 43.48%. These results are a persuasive part of the case for task transfer, within the paper's stated scope of broad genomic pretraining rather than generalization to entirely unseen genomic regions.

  2. A modular architecture supported by several relevant checks. The hierarchical Q-Former design integrates complementary encoder outputs while retaining a frozen LLM. The paper tests embedding permutations, individual modality configurations, and removal of the master Q-Former (Section 4 and Appendices C–D). It also reports annotation accuracy on held-out loci and stricter neighborhoods rather than relying exclusively on the original random split: average field accuracy is 85.13% and 84.83%, respectively (Table 7, p. 15). These checks strengthen the empirical contribution, even though they do not fully isolate the source of downstream gains.

Critical weaknesses
  1. The downstream evidence does not cleanly establish the claimed benefit of multimodal tokens over generated textual annotations. Section 3.2 (p. 6, lines 295–297) says that variant annotations and multimodal tokens are interleaved for ranking. Appendix H.1 (p. 24, lines 1251–1254) instead says that the ranking model sees only BioBLIP's annotations. It is therefore unclear which protocol produced Table 2. This matters because an annotation-generation system followed by text-based gene–trait reasoning is a different evidential claim from demonstrating a benefit of continuous multimodal tokens during ranking. The embedding-permutation experiment in Table 4 disrupts upstream annotation information as well, so it cannot resolve this distinction. A control that keeps the same generated annotations and varies only the added soft tokens is missing. The result may still demonstrate a useful end-to-end pipeline, but its interpretation as a representation-format advantage is currently underdetermined.

  2. Correct terminal gene predictions are not sufficient support for the claims about mechanistic reasoning. The manuscript usefully discloses errors in its own examples: Appendix G's Cases A, B, D, and F place a variant inside the predicted causal gene even though the provided context places it elsewhere (pp. 16–18, 20–21, and 22–23). Cases A and B also mischaracterize the metabolites, as the authors acknowledge. Thus, even successful examples can reach the correct gene through factually incorrect intermediate explanations. The qualitative analysis focuses on the 21 cases where BioBLIP wins over Qwen3-32B (Section 3.3 and Table 9), and does not establish a systematic rate of valid variant-to-gene mechanisms. This limits the stronger conclusion about transparent, biologically grounded multi-step inference, though it does not invalidate the reported gene-selection accuracy.

Questions for authors
  1. Protocol clarification: Which exact ranking input produced Table 2: generated annotations alone, or annotations interleaved with multimodal tokens? How should the descriptions in Section 3.2 and Appendix H.1 be reconciled, and are the Appendix H traces generated by the same configuration as the reported benchmark result?

  2. Attribution of the transfer gains: What is the downstream performance when Qwen3-32B receives exactly BioBLIP's generated annotations but no soft tokens, with prompts and decoding held fixed? Please also clarify whether the Table 4 permutations regenerate the annotations. This comparison would distinguish the contribution of annotation extraction from an additional contribution of multimodal tokens to downstream inference.

  3. Scope of the reasoning claim: Is there an assessment of intermediate factual and mechanistic correctness across the evaluated target-gene cases, beyond the selected 21 wins and their LLM-based characterization? In particular, how frequently are correct final genes accompanied by incorrect variant locations or unsupported effect directions, as in Appendix G? If this was not assessed, would the authors restrict the conclusion to predictive task transfer and treat the traces as unvalidated explanations?

2 comments

We thank the reviewer for the positive and detailed feedback. We are glad that the reviewer likes our approach and our modular architecture. Below please find responses to individual comments/questions:

Q1. Protocol clarification

Which exact ranking input produced Table 2?

Table 2 (Regulatory variant prioritization) was produced with generated annotations interleaved with soft tokens, as described in Section 3.2 and shown in Figure 1b. The Appendix H traces were generated with the same configuration (soft tokens + annotations). Thanks for catching the potentially confusing wording, and we have updated it in the PDF.

Q2 Attribution of the transfer gains:

Q2.1 What is the downstream performance when Qwen3-32B receives exactly BioBLIP's generated annotations but no soft tokens?

Thank you for this question. Variant prioritization on Traitgym is already a task that benefits greatly from accurate variant annotations. BioBLIP produces more accurate variant annotations through its multimodal tokens. However, we conducted an additional experiment to test if the multimodal tokens add predictive capabilities in this benchmark beyond BioBLIP's improved variant annotations.

Takeaways: In difficult, poorly sampled regions of the genome, adding multimodal tokens plus the annotations provides 15.05% Top-1 accuracy improvement and 26.16% Top-2 accuracy improvement. This experiment suggests that BioBLIP's multimodal tokens add predictive information beyond generated variant annotations in this benchmark.

We ran this experiment using Qwen3-32B to rank the variants with exactly BioBLIP's generated variant annotations, but no multimodal tokens. We also report results on a more challenging subset of n=51 groups with top-ranked variants from poorly sampled regions of the genome in OpenTargets (>1kb from any other variant). Results are shown below. While the total improvement is moderate, the increase in accuracy is highest on the more scientifically challenging subset of held-out loci with a strict distance cutoff. More experiments would be needed on harder benchmarks to systematically probe this.

BioBLIP Improvement over Qwen3-32B with BioBLIP's annotations, Relative

SubsetTop-1 AccuracyTop-2 AccuracyTop-3 Accuracy
All Held-out Loci+1.93%+5.76%+3.31%
Held Out & Distance >1kb+15.05%+26.16%+11.15%

W2: Table 4 disrupts upstream annotation information

Thanks for this point. We agree that the permutation experiment disrupts annotation information. The permutation experiment was designed to answer a different question: does the model use its biological inputs at all? That is not guaranteed. MIRAGE (Asadi et al., 2026) shows models scoring well on multimodal benchmarks while ignoring the images, and concurrent work (Fang et al, 2026) finds most multimodal models don’t rely on their inputs. Permuting embeddings drops BioBLIP’s performance by 70–89% on all three tasks (Table 4), confirming that BioBLIP does rely on the multimodal data.

Q2.2: Whether the Table 4 permutations regenerate the annotations?

This depends on the task. The variant annotation experiment by definition would regenerate annotations. The variant prioritization task also does so. Target gene prediction does not regenerate the annotations as this is not a separate step in the setup.

(contd. below)

Q3 Factual Precision of Target Gene Reasoning Traces

Q3.1: Scope of the reasoning claim: Is there an assessment of intermediate factual and mechanistic correctness across the evaluated target-gene cases?

Thanks for the insightful comments. Such an analysis is indeed helpful. We selected 100 random target gene identification traces from the OpenTargets Genetics Gold Standard dataset, and analyzed them for the correctness of the claims using LLM-as-a-judge (Claude Opus 5.5). The LLM judge verified the claims using Ensembl VEP annotations and prior literature as evidence sources. We also hand checked a subset of 10 traces for agreement and found 100% alignment with the LLM judge's decisions.

Below, we demonstrate that on target gene identification (1) precision of the claims made is 90.5% overall (2) 27.5% of claims made are hypotheses. Hypotheses are defined as statements that need additional eQTL or experimental data to be verified. We have added this analysis to Appendix I.

StagePrecisionCorrect claimsHypothesesIncorrect claimsNumber of Claims
S1 Variant annotation82.10%82.10%0.00%17.90%285
S2 Gene-level effect89.40%58.40%34.60%7.00%762
S3 Cellular process94.40%69.80%26.10%4.10%510
S4 Tissue-level effect92.90%66.70%28.30%5.10%435
S5 Disease mechanism92.80%61.70%33.50%4.80%457
OVERALL90.50%65.60%27.50%6.90%2449

Precision = (Correct claims) / (Incorrect claims + Correct claims)

Takeaways: BioBLIP provides 90.5% precision on 100 randomly selected traces, using LLM-as-a-judge and humans manually checking a subset of the traces. Precision is defined as the number of correct claims over the number of verifiable claims.

However, BioBLIP does not presume to guarantee step-by-step correctness in the traces produced. The primary contribution of the paper is a new multimodal architecture for zero-shot transfer in genomic tasks, and the step-by-step reasoning traces are useful because they allow an expert to audit them and investigate biological hypotheses raised. These are strengths over a predictive only architecture.

Improving groundedness, faithfulness, and other properties of LLM traces is an active area of research with tools such as VERDICT (Zhou et al, 2026). We leave applying this research to BioBLIP as future work.

We have updated this in the Discussion section of the paper.

Thanks again for your comments, which have helped strengthen the paper. Please let us know if you have any further questions or comments!

References

[1] Mohammad Asadi, Jack W O’Sullivan, Fang Cao, Tahoura Nedaee, Kamyar Fardi, Fei-Fei Li, Ehsan Adeli, and Euan Ashley. Mirage the illusion of visual understanding. arXiv preprint arXiv:2603.21687, 2026.

[2] Fang, Ada, et al. "When Do Biological Reasoning Models Use Their Biological Inputs?." arXiv preprint arXiv:2610.00898 (2026).

[3] Zikai Zhou, Yufei Jin, Yilin Xu, Yu-Chiang Wang, Chieh-Ju Chao, and Monica Lam. Accountable ai with grounded, faithful, consistent, actionable rationales: A case study in clinical trial matching with verdict. pp. 58987–59032, 2026.

Reviewer_T9Y3OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a multimodal architecture for zero-shot transfer to genomic interpretation tasks. BioBLIP integrates heterogeneous biological representations and uses modality-specific Q-Formers, which fuses the resulting representations into multimodal tokens for a frozen LLM. Without further fine-tuning, BioBLIP is evaluated on two downstream tasks: variant prioritization and target gene prediction. Extensive experiments compare BioBLIP with genomic foundation models, including Evo2 and GPN, as well as LLM-based baselines. The results show promising zero-shot transfer performance.

Critical strengths
  • Comprehensive experiments. The paper conducts extensive experiments, comparing BioBLIP against genomic foundation models (e.g., Evo2 and GPN) and LLM-based baselines. The results demonstrate the effectiveness of BioBLIP and highlight the advantages of multimodal biological representations for zero-shot transfer to downstream genomic tasks.

  • Clear presentation. The paper is well-written and clearly organized. The proposed framework and experimental results are presented in a clear and easy-to-follow manner.

  • Novel multimodal framework. The paper introduces a Q-Former architecture that integrates heterogeneous biological representations, including DNA, gene, and protein embeddings. This design enables zero-shot transfer to downstream genomic tasks without further fine-tuning.

Critical weaknesses
  • Insufficient ablation studies. Although the paper evaluates the Q-Former, the ablation is limited to the variant annotation task and does not examine its impact on zero-shot downstream tasks.
  • Unexplained performance gaps in modality ablations. The modality ablation results exhibit surprisingly large performance gaps between different modality combinations. Further analysis is needed to explain these results and clarify the individual contributions and interactions of different modalities.
Questions for authors
  • The Q-Former ablation is limited to the variant annotation task. Could the authors provide ablation results on the two zero-shot downstream tasks, namely variant prioritization and target gene prediction, to clarify the contribution of the Q-Former to zero-shot transfer?
  • The modality ablation results exhibit surprisingly large performance gaps between different modality combinations. Could further analyses be provided to clarify the interactions among modalities and verify the robustness of these results?
5 comments

We thank the reviewer for the positive and detailed feedback. We are glad that the reviewer likes our architecture and approach. The reviewer asks great questions, and we provide detailed answers below.

Q1. Master Q-former Ablation on the two downstream tasks

A1. Thanks for this question. It is good to understand whether the Master Q-former's benefit persists on the downstream tasks. To answer this, we trained a BioBLIP Qwen3-32B model with no master Qformer to provide a head-to-head comparison on both downstream tasks.

Takeaways: The Master Q-former provides a large benefit in downstream task performance. This key component of the architecture provides ~22% improvement in accuracy on regulatory variant prioritization. It also provides ~28% improvement in accuracy on target gene prediction for non-closest gene cases. The Master Q-former is critical for zero-shot transfer performance.

Task 1: Regulatory Variant Prioritization

Top-1 AccuracyTop-2 AccuracyTop-3 Accuracy
Master Qformer70.22%76.49%79.62%
No Master Qformer47.93%53.55%56.21%
Delta (No Master – Master)-22.29%-22.94%-23.41%

Task 2: Target Gene Prediction

Non-Closest Gene & Held Out Loci (n=119)Non-Closest & Distance Cutoff (n=92)Held-Out Loci (n=369)
Master Qformer60.50%59.78%83.47%
No Master Qformer32.77%31.52%77.78%
Delta (No Master – Master)-27.73%-28.26%-5.69%

We have added these results to Section 4 of the paper.

Q2. Why do the modality ablation results have large performance gaps?

A2. This is a good question. The modality ablation results have large performance gaps because we report the ablations on the final trained model. Some of the gap is due to distribution shift, as the model has never seen soft tokens ablated.

Retraining the models from scratch with fewer modalities shows less of a gap (see results below reported for Qwen3-4B backbone; sampling done with three random seeds). The same modality trends are clear in the retrained models. AlphaGenome modality contributes the most to Chromosome, Variant Type, and Variant Consequence. GenePT contributes the most to the nearest genes field. ESM protein embeddings contribute least - which is expected because these are mostly noncoding variants which do not affect the protein sequence.

Field3 modalityAFG ablationESM ablationGenePT ablation
Chromosome90.1 ± 0.484.8 ± 0.386.8 ± 0.389.3 ± 0.3
Variant Type89.7 ± 0.486.9 ± 0.389.2 ± 0.287.6 ± 0.6
Variant Consequence67.7 ± 0.863.0 ± 0.967.2 ± 0.565.0 ± 0.6
Nearest Genes64.1 ± 0.364.9 ± 0.365.4 ± 0.249.3 ± 0.3

Takeaways: The large gaps are due in part to inference-time ablations. With retrained reduced-modality models, the gaps are less, but the field-level contribution of each modality remains the same.

We shall add these results and explanations to the final version of the paper.

Thank you for the detailed response and additional experiments, which address most of my concerns. Could the authors provide more details on how the modality ablations in the original Table 5 and the newly reported modality ablation experiments in Q2 were conducted, particularly the differences in their experimental procedures?

Thank you for the follow up. The two ablation experiments answer different questions: Table 5 measures the trained model's dependence on each modality at inference time. The Q2 table measures the model's performance if it is retrained without that modality. Absolute accuracies differ between tables because of retraining and the backbone size.

Experimental details for both are provided below:

Table 5 - Inference-time Ablations.

  • There is no retraining of the Master Q-former or LLM Projection.
  • BioBLIP with the Qwen3-32B backbone is loaded with its trained weights.
  • Each dropped modality has its Qformer tokens replaced with i.i.d. Gaussian noise before the tokens enter the Master Q-former.
  • These experiments evaluate the full model, each two-modality and single-modality combination, and a noise-only baseline.
  • Inference-only ablations, cheap to run.
  • All comparisons are done against the first column (3 modality, Qwen3-32B backbone). Mean is reported over three decoding seeds.

Q2 Table - Retrained Ablations.

  • Retraining was done of the Master Q-former and LLM projection without the ablated modality.
  • Training is conducted for 15 epochs on the OpenTargets Dataset following the Stage 2 details in Appendix B, using the BioBLIP Qwen3-4B architecture.
  • The dropped modality Qformer and tokens are removed during training and inference.
  • These experiments measure the effects of leave-one-modality-out training.
  • Training and inference ablations, expensive to run.
  • All comparisons are done against the first column (3 modality, Qwen3-4B backbone). Mean ± standard deviation is reported over three decoding seeds.
  • The Qwen3-4B backbone was used for all the retrained ablations because training each model takes several days.

Referring back to the reviewer's original question, the inference time ablations in Table 5 have large drops for two reasons: (1) information from the dropped modality is missing, and (2) the Master Q-former never saw noise tokens during training, so the inputs are out-of-distribution.

Importantly, both sets of experiments show the same modality-level contributions, which strengthens the result: AlphaGenome drives chromosome, type, and consequence accuracy, and the GenePT modality drives nearest-gene accuracy. We will update these details in the paper as well.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overview and Summary of our Rebuttal

We thank all reviewers for their thoughtful comments, which have helped us strengthen the paper. We are encouraged that the reviewers recognize the novelty and "useful design" of BioBLIP's hierarchical Q-Former architecture for integrating heterogeneous biological representations into a frozen LLM (@T9Y3, @GCpJ, @NxXk). Reviewers concur that the frozen LLM is a strong and novel advantage of BioBLIP's architecture (@GCpJ, @NxXk), that the paper's "comprehensive experiments" (@T9Y3) put forth "a persuasive part of the case for task transfer" (@GCpJ), and our model outperforms "genomic foundation models, including Evo2 and GPN, as well as LLM-based baselines" (@T9Y3) with "meaningful downstream gains" (@GCpJ, @NxXk) on genomic tasks.

We are also thrilled that the paper itself received reviews of well-written (@T9Y3) and well-organized (@T9Y3) with extensive experiments (@GCpJ).

Following the reviewers' valuable suggestions, we performed four new experiments that are summarized below. Updates have also been provided in the revised PDF and are shown in blue.

We believe these results and clarifications support the paper's primary contribution: as a new multimodal architecture that integrates biological representations into a frozen LLM, BioBLIP enables generalization to multi-step genomics tasks and outperforms strong text and biological baselines without task-specific fine-tuning.

Summaries of the additional results are as follows:

E1. Master Q-former Ablation in zero-shot task transfer.

We ablated the Master Q-former in the two downstream tasks and observed that it provides ~22% improvement in accuracy (abs.) on regulatory variant prioritization and ~28% improvement in accuracy (abs.) on target gene prediction for non-closest gene cases (Table E1). Thus the Master Q-former, a new part of the architecture, is critical for zero-shot transfer. We have added this result in Section 4 (reviewers T9Y3, NxXk).

E2. Retrained Single Modality Ablations for 15 Epochs

Leave-one-out modality ablations for AlphaGenome, ESM, and GenePT were performed by training each reduced-modality model for 15 epochs following Stage 2 training details. The retrained models show the same trends as the inference-time ablations already in the paper, which strengthens the results: AlphaGenome as a modality drives chromosome, type, and consequence accuracy, and the GenePT modality drives nearest-gene accuracy. We have added these results in Appendix E (reviewers T9Y3, NxXk).

E3. Disentangling Multimodal Tokens on TraitGym

We conducted an experiment to show that adding multimodal tokens plus the annotations improves Top-1 accuracy by 15.05% and Top-2 accuracy by 26.16% (relative) on a challenging subset of groups in the Traitgym benchmark. This experiment demonstrates that BioBLIP's multimodal tokens add information over improved annotations, particularly in regions of the genome that were poorly sampled in training (reviewers GCpJ, NxXk).

E4. Factual Precision of Biological Reasoning Traces

We performed a systematic evaluation of the factual precision of 2449 claims from 100 biological reasoning traces of BioBLIP using LLM-as-a-judge, and 196 claims checked by hand to evaluate agreement with the LLM's decisions. BioBLIP has 90.5% precision on the verifiable claims, with a per-stage breakdown provided. Guarantees of mechanistic correctness are left to future work. We have added the new analysis to Appendix I (reviewers GCpJ, NxXk).

Clarifications and experiments for individual reviewers are addressed in per-reviewer responses and the updated PDF. In summary:

  • Clarified that training the Q-formers is a key part of the BioBLIP architecture and not a confound (@NxXk)
  • Clarified how the variant prioritization traces were generated in the Appendix (@GCpJ)
  • Section 3.2 clarifies the evidence sources of Qwen3-32B and BioBLIP (@NxXk)
  • Section 6 contains an expanded discussion of limitations and future directions such as guaranteeing mechanistic correctness of the traces (@GCpJ, @NxXk) and emperical evaluation on different LLM model families (@NxXk).

We have updated the manuscript where noted and will further incorporate these changes for the camera-ready version. We would be happy to provide any additional clarifications or address any remaining concerns. We thank the Area Chair and the reviewers again and look forward to their engagement.

AuthorsOpenReview ↗

Experiment result details

E1. Master Q-former Ablation in zero-shot task transfer.

We trained a BioBLIP Qwen3-32B model with no master Qformer to provide a head-to-head comparison on both downstream tasks, and understand the contribution of the Master Q-former in zero-shot task transfer.

Takeaways: The Master Q-former provides a critical benefit in downstream task performance. This component of the architecture provides ~22% improvement in accuracy (abs.) on regulatory variant prioritization. It also provides ~28% improvement in accuracy (abs.) on target gene prediction for non-closest gene cases.

Task 1: Regulatory Variant Prioritization

Top-1 AccuracyTop-2 AccuracyTop-3 Accuracy
Master Qformer70.22%76.49%79.62%
No Master Qformer47.93%53.55%56.21%
Delta (No Master – Master)-22.29%-22.94%-23.41%

Task 2: Target Gene Prediction

Non-Closest Gene & Held Out Loci (n=119)Non-Closest & Distance Cutoff (n=92)Held-Out Loci (n=369)
Master Qformer60.50%59.78%83.47%
No Master Qformer32.77%31.52%77.78%
Delta (No Master – Master)-27.73%-28.26%-5.69%

E2. Retrained Single Modality Ablations for 15 Epochs.

Table 5 in the paper measures the trained model's dependence on each modality at inference time. We conducted additional ablations by training the BioBLIP model without the ablated modality. Training for all models was conducted using BioBLIP with the Qwen3-4B backbone, following the Stage 2 training details but for 15 epochs (Appendix B). The ablated modality has its Qformer and tokens removed during training.

The same modality trends are clear in the retrained models. AlphaGenome modality contributes the most to Chromosome, Variant Type, and Variant Consequence. GenePT contributes the most to the nearest genes field. ESM protein embeddings contribute least overall - which is expected because these are mostly noncoding variants which do not affect the protein sequence.

Field3 modalityAFG ablationESM ablationGenePT ablation
Chromosome90.1 ± 0.484.8 ± 0.386.8 ± 0.389.3 ± 0.3
Variant Type89.7 ± 0.486.9 ± 0.389.2 ± 0.287.6 ± 0.6
Variant Consequence67.7 ± 0.863.0 ± 0.967.2 ± 0.565.0 ± 0.6
Nearest Genes64.1 ± 0.364.9 ± 0.365.4 ± 0.249.3 ± 0.3

E3. Disentangling Multimodal Tokens on TraitGym.

Qwen3-32B was used to rank the variants with exactly BioBLIP's generated variant annotations, but no multimodal tokens. We also report the results on a more challenging subset of n=51 groups with variants from poorly sampled regions of the genome in OpenTargets (>1kb from any other variant).

Takeaways: this experiment suggests that BioBLIP's multimodal tokens add predictive information beyond its generated variant annotations. For difficult, poorly sampled regions of the genome, adding multimodal tokens plus the annotations provides 15.05% Top-1 accuracy improvement and 26.16% Top-2 accuracy improvement.

BioBLIP Improvement, Relative

SubsetTop-3 AccuracyTop-2 AccuracyTop-1 Accuracy
All Held-out Loci3.31%5.76%1.93%
Held Out & Distance >1kb11.15%26.16%15.05%

E4. Factual Precision of Biological Reasoning Traces.

We selected 100 random target gene identification traces and analyzed them for the correctness of the claims using LLM-as-a-judge (Opus 5.5) and VEP Ensembl annotations and prior literature as evidence sources. We also hand checked a subset of 10 traces with 196 claims for agreement and found 100% alignment with the LLM judge's decisions.

On target gene identification (1) precision of the claims made is 90.5% overall (2) 27.5% of claims made are hypotheses. Hypotheses are defined as statements that need additional eQTL or experimental data to be verified. We have added this analysis to Appendix I.

StagePrecisionCorrect claimsHypothesesIncorrect claimsNumber of Claims
S1 Variant annotation82.10%82.10%0.00%17.90%285
S2 Gene-level effect89.40%58.40%34.60%7.00%762
S3 Cellular process94.40%69.80%26.10%4.10%510
S4 Tissue-level effect92.90%66.70%28.30%5.10%435
S5 Disease mechanism92.80%61.70%33.50%4.80%457
OVERALL90.50%65.60%27.50%6.90%2449

Precision = (Correct claims) / (Incorrect claims + Correct claims)

Takeaways: BioBLIP provides 90.5% precision on 100 randomly selected traces. Precision is the number of correct claims over the number of verifiable claims.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.