acceptodds
Under review as a conference paper at ICLR 2027

OmicsLM: Integrating Quantitative Omics and Biological Knowledge in a Multimodal Language Model

Abstract

Large language models can reason over biological knowledge, but they are not naturally equipped to work with quantitative omics measurements. Existing approaches either convert expression profiles into long textual representations, often losing quantitative information, or connect omics encoders to language models for relatively narrow tasks. We introduce OmicsLM, a 9B multimodal language model that represents each bulk or single-cell transcriptome as a compact continuous embedding that can be interleaved with natural-language instructions and with other samples in the same context. An important contribution is the construction of a training corpus that connects experimental measurements with biological knowledge and task supervision. We combine templates linking transcriptomic profiles to annotations and experimental outcomes, reasoning examples connecting evidence across sources, and LLM-generated supervision with task-specific quality checks. The instruction-tuning corpus spans 157 biological and general-purpose task groups. Training includes continued pretraining on 20 billion tokens and supervised fine-tuning on 5.21 million conversations. We further introduce GEO-OmicsQA, a publication-disjoint benchmark of 3,000 questions requiring models to interpret and compare expression profiles from real GEO studies. OmicsLM achieves the best results among the evaluated models on three of four held-out annotation tasks and on all six GEO-OmicsQA variants, and the best average performance among compared methods on PerturbQA perturbation-response prediction. Ablations show a clear advantage for continuous omics representations over a text-only gene-list representation. Together, these results show that quantitative transcriptomic measurements and biological language can be integrated within a general conversational model, enabling both predictive tasks and language-guided interpretation of experimental samples.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
llm, biology, omics, rl, mulitmodal, modality fusion, cell

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.40 over 5 reviews, above 68% of the others.
Reviewer_MFKBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

OmicsLM integrates transcriptomic profiles into an LLM through continuous embeddings, allowing multiple samples to be included alongside natural-language instructions. The authors construct a large-scale omics–language training corpus and introduce GEO-OmicsQA benchmark for evaluation. The model is evaluated on annotation, expression interpretation, and perturbation-response direction prediction.

Critical strengths
  • The construction of a large-scale omics–language training corpus and a new evaluation benchmark is a valuable contribution to the community.
  • Representing expression profiles as continuous embeddings and interleaving them with natural-language instructions makes sense. This provides a compact way to include quantitative measurements and multiple samples without serializing the raw tabular expression data.
Critical weaknesses
  • The authors motivate continuous omics inputs partly by noting that ranked gene lists omit expression magnitudes. However, the perturbation evaluation only predicts directions. This leaves it unclear how well the model captures DEG or post-perturbation expression beyond their direction. Moreover, the evaluation coverage seems narrow relative to the diversity of the training tasks. For instance, drug sensitivity and pathway reasoning are not evaluated. Additional evaluations would help demonstrating the broad biological capabilities of OmicsLM.
  • The paper does not compare OmicsLM with a tool-augmented agent setup. Tool-augmented systems including Claude Science and Biomni have provided practical alternatives for solving biological tasks. A comparison with such a setup would help clarify why specialized LM training is needed and what additional benefits it provides.
Questions for authors

GEO-OmicsQA free-text answers are evaluated using a Gemini judge. How well do its scores agree with expert judgments and the actual expression data?

4 comments

Thank you for recognizing the value of our training data, benchmark, and compact interface for multiple profiles.

1. Input values and predicted changes. An up/down prediction can still benefit from the expression values in the input. OmicsLM receives continuous transcriptomic profiles. PerturbQA tests whether a gene's expression increases or decreases after a perturbation. In our controlled 0.8B ablation, the full continuous input achieves 0.607 AUROC on PerturbQA. A ranked list of the 1,024 highest-expressed genes gives 0.488. Both scores are averages over three training runs (Appendix A.7). This supports our input representation even for up/down predictions. PerturbQA includes readout genes already known to change after perturbation. It does not test which genes change or by how much.

2. Task coverage. Our evaluation focuses on tasks that use measured expression profiles: annotation, biological interpretation, and perturbation prediction. Text-only benchmarks could test additional biological abilities, such as pathway reasoning, but they would not directly evaluate our main focus: using measured omics profiles to answer biological questions.

3. Tool-using agents. A central goal of OmicsLM is to learn to interpret omics profiles directly through multimodal training. Our GEO-OmicsQA comparison tests this intended capability: each model answers from the supplied inputs in one call, without external tools or code execution (Appendix A.6.1).

4. Judge reliability. Experts reviewed 50 GEO-OmicsQA model-response traces as a qualitative check and found no errors that would affect scoring in those cases. The judge checks agreement with reference answers supported by the source publications.

Separately, experts reviewed 300 GEO-OmicsQA questions for question–answer validity, sample grounding, and biological coherence. Over 90% of the checked reference claims agreed with the linked expression data.

Thanks for the response. I now understand that the PerturbQA direction task is restricted to genes already known to respond. But I still think OmicsLM would have been more useful in practice if its training objectives had included predicting gene-level expression after perturbation. TxPert [1], for example, uses the same underlying Replogle and Nadig datasets for this kind of prediction.

Regarding W2, a general-purpose LLM agent given the same expression profiles and basic python or deep learning model-based tools is a practical and relevant baseline, but the authors do not provide additional comparison with agents.

Given these two remaining concerns, I’m keeping my score at 2.


References

[1] TxPert: using multiple knowledge graphs for prediction of transcriptomic perturbation effects, Nature Biotechnology '26

Thank you for your continued engagement and for clarifying the remaining concerns.

1. Predicting post-perturbation expression. We agree that predicting full post-perturbation expression profiles would be useful. However, we do not consider its absence a critical weakness given OmicsLM’s focus on flexible interpretation of omics profiles. TxPert is a specialized model for predicting transcriptomic perturbation effects. OmicsLM is designed for interpreting omics profiles through flexible natural-language questions and responses across many tasks. These models therefore address different objectives.

Our training already includes perturbation-response prediction beyond direction: predicting affected genes, the most up- and downregulated genes, and the top-expressed genes after knockout or treatment, as documented in the appendix’s training-task table. These outputs describe the response at the gene level, without generating a complete numerical expression vector. The direction-only PerturbQA evaluation should therefore not be taken to describe the full scope of our training objectives.

The distinction is analogous to image understanding versus image generation: a vision-language LLM can interpret images without generating them, while a specialized diffusion model is trained for generation. The absence of image generation is not, by itself, a critical weakness of an image-understanding model. Likewise, full transcriptome generation would be a useful extension, but it is a different objective from the flexible omics–language interface studied here.

2. Tool-using agents. Thank you for this suggestion. We evaluated Gemini 3.8 Flash with medium reasoning effort in two agent setups on all 1,000 binary and 2,000 free-text GEO-OmicsQA questions. Each agent received the complete expression profiles. In Gemini + Python, the Python tool could read and analyze all measured genes’ numerical expression values (TPM) for every supplied sample. Evaluation used the original references and scoring. Biomni A1 used a basic Python environment. We allowed up to 32 model calls and 65,536 output tokens per call.

On all 1,000 binary and 2,000 free-text questions:

ModelBinary accuracyFree-text scoreAvg. generated tokens/questionAvg. tool calls/question
OmicsLM (without tools)76.7%0.651—0
Gemini + Python75.3%0.6895,21010.81

For Biomni, we compare both models on the same 788 binary and 1,989 free-text questions where Biomni returned a valid answer. The remaining 223 attempts did not produce a clear final answer.

ModelBinary accuracyFree-text scoreAvg. generated tokens/questionAvg. tool calls/question
OmicsLM (without tools)77.0%0.651—0
Biomni A1 + Gemini78.2%0.5215,3256.43

Token and tool-call means cover the questions in each table. Generated tokens include output and reasoning across model turns, excluding input tokens, judging and recovery attempts. Tool calls count Python executions.

OmicsLM generates answers directly without external tool calls and is competitive with the proprietary Gemini model in both setups: it scores higher on one metric and lower on the other in each comparison.

Thank you for your detailed explanation and providing additional experiments. Reconsidering again, the authors' response regarding the perturbation response prediction makes sense, and the performance seems comparable with a tool-using agent. I raise my score to 3.

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

This paper aims to bridge transcriptomic data and general-purpose LLMs by projecting gene expression profiles into the LLM's token embedding space. The work focuses on integrating diverse omics tasks through large-scale instruction tuning and introduces GEO-OmicsQA, a benchmark for evaluating multi-sample transcriptomic reasoning. Moreover, the model is trained through a three-stage pipeline consisting of multimodal pretraining, supervised instruction tuning, and reinforcement learning with GRPO.

Critical strengths

Overall, its main contributions lie in data integration, instruction-tuning pipelines, and benchmark construction rather than fundamental architectural innovation.

Critical weaknesses
  1. My main concern is the reliance on a fixed panel of 19,260 genes. Although the authors state that this panel is derived from GENCODE annotations, the rationale behind the specific gene selection is not sufficiently justified. Different transcriptomic datasets and downstream tasks may rely on different sets of biologically informative genes. Consequently, a predefined gene panel may disproportionately affect certain datasets or tasks by excluding relevant biological signals, introducing potential gene selection bias. This also raises concerns about the fairness of comparisons with existing methods that use different gene sets and may therefore have access to different biological information. Furthermore, the paper does not clearly explain how genes absent from individual datasets are handled or whether missing measurements are distinguished from genuine zero expression. The lack of gene coverage statistics and sensitivity analyses across alternative gene panels makes it difficult to assess the robustness of the reported results and the comparability of the benchmarks.

  2. One potential advantage of integrating transcriptomic data with LLMs is the ability to support flexible, gene-aware representations that accommodate varying gene sets and input lengths. However, OmicsLM retains a fixed-dimensional expression vector, similar to conventional transcriptomic models, while the introduced gene-specific tokens are primarily used for language processing rather than explicitly representing individual genes alongside their expression values. This design limits the flexibility of the input interface and may require modifications or additional training when adapting to new gene sets or updated annotations. Given the substantial computational resources invested in multimodal training, the reliance on a rigid input representation raises concerns about the scalability and extensibility of the proposed framework. Moreover, the paper does not provide sufficient analysis of how this design affects adaptation to heterogeneous datasets or what advantages it offers over conventional fixed-input transcriptomic models.

Questions for authors
  1. Could the authors provide more details on how the fixed panel of 19,260 genes was selected from GENCODE annotations and how genes absent from individual datasets are handled? In particular, it would be helpful to clarify whether missing measurements are distinguished from true zero expression and how variations in gene coverage across datasets might affect model performance.

  2. Could the authors elaborate on the motivation for adopting a fixed-dimensional gene representation, similar to conventional transcriptomic models, rather than leveraging gene-specific tokens to accommodate varying gene sets and input lengths? It would be interesting to understand the advantages and trade-offs of this design choice, particularly in the context of LLM-based transcriptomic modeling.

  3. Could the authors provide some insights into why GRPO leads to only a modest improvement in overall performance (Appendix A.7)? Could this be related to the quality, diversity, or difficulty of the training data and reward signals used during the RL stage?

  4. Could the authors clarify how the baseline comparisons were conducted, particularly whether the baseline models were evaluated using the same gene set and preprocessing procedures as OmicsLM? Were these models retrained or fine-tuned on the datasets used in this work, or were their original pretrained checkpoints used directly?

4 comments

Thank you for recognizing the value of the data integration, training pipeline, and benchmark. We address your four questions below.

1. Gene panel and missing measurements. Humans have approximately 20,000 protein-coding genes. Our panel includes 19,223 protein-coding genes plus 37 mitochondrial genes, covering the large majority of human protein-coding genes. We consider this broad, annotation-based selection well justified for use across datasets and tasks. We retain all genes of these types in a human reference that reconciles Ensembl, HGNC, and NCBI identifiers. Section 3.1 and Appendix A.2 state the gene categories and their GENCODE basis. The complete gene list, including Ensembl IDs and symbols, is available in the released expression files linked from Appendix A.1.

2. Why a fixed vector? Our focus is transcriptome-wide RNA-seq, using the broad protein-coding and mitochondrial panel described above. These data are quantified against a reference gene annotation, making a shared gene space practical. Each profile enters the LLM as one projected embedding plus two delimiter tokens. This keeps the LLM sequence cost low for multi-sample analysis; a gene/value sequence would grow with both gene and sample counts. Gene-specific tokens support biological language, while the projected profile carries the expression measurements. However, a direct representation with one token per gene would already require 19,000–20,000 tokens per full profile, making it impractical for our multi-sample setting.

The input representation can also be changed without redesigning the LLM. We already trained a 0.8B variant without direct expression values, retaining the Funomics and Geneformer embeddings. Its mean across seven benchmark metrics is 76.2 versus 77.6 for the full input, over three training runs (Appendix A.8). This shows that the fixed gene-by-gene branch is optional, with a measured performance trade-off. Removing it shifts gene-set handling to the encoders while preserving the same LLM interface. This offers a route to different gene sets, subject to the encoders' supported vocabularies and validation on those inputs.

3. Contribution of GRPO. The model already performs well after SFT. Our SFT corpus covers a wider range of tasks than the RL data, so we limit RL training to reduce the risk of forgetting capabilities learned during SFT. The additional gains may seem modest, but they are broadly consistent across tasks: Table 16 shows that GRPO improves or maintains performance on all but one reported metric.

4. Baseline inputs and training. For transparency, our reproducibility package includes the code and protocols for all our evaluations, including baseline models. We did not fine-tune or train any baseline model from scratch: we used released checkpoints, APIs, or released predictions for rescoring. Results taken directly from prior papers are explicitly marked. Where applicable, we used the same source profiles, preprocessed as expected by each baseline model. Where supported, we supplied full expression profiles; models using "cell sentences" received gene names ranked by expression within their input limits. All required samples were retained.

Reviewer_jqoMOpenReview ↗

Thank you for the response

Thank you for the response. The computational advantage of a fixed-dimensional representation for multi-sample analysis is reasonable, and the explanations regarding the gene panel and GRPO help clarify some of my concerns. However, I still have some concerns remain:

  1. The response explains which genes are included in the reference panel, but does not clarify how individual datasets are mapped into this shared gene space. How are panel genes that are unmeasured or absent from a dataset handled? Are missing measurements distinguished from genuine zero expression? A reference panel covering most human protein-coding genes does not imply that every input dataset covers those genes. Please report dataset-level gene coverage and discuss or evaluate how differences in coverage affect model performance.

  2. The variant without the direct expression branch demonstrates that an alternative input configuration can achieve comparable performance after training. However, this does not establish whether the existing trained OmicsLM can accommodate different gene sets or updated annotations without additional training. Could the authors clarify how such inputs would be handled, which components would require modification or retraining, and what constraints remain due to the encoders’ supported gene vocabularies?

  3. The response explains the computational benefit of projecting each expression profile into a compact embedding, but does not directly clarify what additional capabilities the LLM backbone provides over conventional fixed-input transcriptomic models.

  4. Comparability of baseline evaluations. Thank you for clarifying that the baselines use released checkpoints, APIs, or existing predictions without retraining or fine-tuning. However, different gene vocabularies, coverage, and truncation strategies mean that models may receive different biological information even from the same source samples. OmicsLM’s multi-stage training also introduces differences in training conditions. I recognize that fully aligning these conditions may be impractical and am not requesting identical gene sets or baseline retraining. It would be better if the authors could document each baseline’s gene set/coverage, preprocessing, truncation rules, and task-specific training or adaptation in the manuscript, and acknowledge these differences when interpreting the reported performance comparisons.

Thank you for your thorough feedback and continued engagement with our work!

1. Gene coverage and missing measurements. We map each dataset to our shared 19,260-gene panel. In the direct expression branch, absent genes are filled with zero before normalization, so missing measurements and measured zeros are not distinguished there. Geneformer and Funomics both exclude missing genes when encoding a profile. Both encoders were trained with random gene masking, so we expect them to generalize to profiles with missing genes in practice.

2. Different gene sets and updated annotations. The final 9B model's expression branch expects our fixed gene panel. New datasets can be mapped to these gene IDs, with missing entries handled as above, without retraining. Adding gene dimensions would require adapting and training the input projection. Our separately trained 0.8B no-expression variant removes this dependence on a fixed gene-by-gene input. It could, in principle, generalize to different gene sets within the encoders' supported vocabularies, although this has not been tested directly.

Empirically, the full 9B model achieves strong zero-shot results on held-out Tabula Sapiens and Human Diseases. Neither dataset's exact gene coverage pattern occurs in OmicsLM's training data. Their original source files cover 19,221 panel genes (99.80%) and 19,169 (99.53%), respectively; these counts describe source gene availability, not per-sample expression detection. The 0.8B ablation shows that direct expression does not improve every annotation task: removing it improves Tabula Sapiens cell-type prediction, while the full input performs better on tissue prediction and both Human Diseases tasks (Appendix A.8).

3. What the LLM adds. Our motivation is multitask learning, a flexible language interface, and combining knowledge learned across tasks. One model can annotate samples, predict perturbation effects, and answer questions comparing profiles with biological context. Specialized models can excel at their target tasks. Our aim is competitive performance across tasks while retaining the flexibility to follow new questions and connect observations, rather than general superiority over specialized models.

4. Baseline inputs and adaptation.

a) PerturbQA. Both OmicsLM and AROMA receive a perturbation question and biological annotations, using model-specific prompts. OmicsLM additionally receives each cell line's unperturbed control profile from the Replogle/Nadig screens, aligned to our 19,260-gene panel. AROMA additionally receives protein/pathway/gene-graph features. Its released interface does not support direct expression profiles, so we retain its native inputs and preprocessing, without prompt truncation. Other baselines use released predictions rescored on the official direction subset, while SynthPert uses published results.

b) Tabula Sapiens and Human Diseases. Locally evaluated baselines use the same 19,260-gene source matrices. Ranked gene lists omit expression magnitudes, as expected by these models' input format.

BaselineGene inputs, preprocessing and limitsEvaluation/adaptation
Cell-o1 / C2S-Scale-27BUp to 1,000 highest-expressed genes, ranked by expression.Candidate-label scoring. C2S uses its annotation templates, with a disease-subtype adaptation.
CellWhisperer LMGeneformer recognizes 18,938/19,260 panel genes. Uses CellWhisperer's native preprocessing.Released encoder and chat model; task-specific answer prefixes and candidate-label scoring.
CellWhisperer CLIPOriginal published protocol; effective gene coverage was not independently audited.Published results; no local rerun.

c) GEO-OmicsQA. Baselines receive gene symbols ranked by expression from the same 19,260-gene source matrices, without magnitudes. All lists are capped at 1,000 genes per profile and shortened equally across profiles to fit each model's context window; all samples and the question are retained.

BaselineGenes per profileEvaluation/adaptation
Gemini 3.8 Flash / 3.1 Pro, GPT-6 Sol / Luna, Qwen3.5-9B / Qwen3.8-27B1,000APIs or released checkpoints; QA prompts and answer generation.
TxGemma-27BUp to 1,000, reduced to fit contextReleased chat checkpoint; QA prompts and answer generation.
Cell-o1Up to 1,000, reduced to fit contextReleased checkpoint; QA prompts and answer generation.
C2S-Scale-27BUp to 1,000, reduced to fit contextNative prompt format; Yes/No logits for binary questions and generation for free text.

These results compare complete systems, so differences in inputs, preprocessing and training can affect the scores. We believe that preserving each model's intended inputs and preprocessing provides the closest practical comparison while remaining faithful to the released models.

Reviewer_jqoMOpenReview ↗

Thank you for the response

Thank you for the further clarification. Your responses have resolved some of my concerns, particularly regarding missing-gene handling, adaptation requirements, and baseline evaluation settings. However, I would like to clarify why the input representation remains my main reservation about the contribution.

The flexibility I had in mind is not only the ability to follow instructions or perform multiple tasks, but also a more flexible way of representing heterogeneous transcriptomic measurements. The current expression branch still relies on a fixed 19,260-gene coordinate system, however, different datasets are mapped into this panel, absent genes are filled with zeros, and incorporating additional gene dimensions requires adapting and training the projection. This provides compatibility through preprocessing, but does not fundamentally address heterogeneous gene sets or distinguish unmeasured genes from observed zeros.

I understand the computational motivation for compressing each profile into a compact embedding. However, flexible gene-aware encoding and compact LLM inputs are not mutually exclusive. An encoder could, in principle, process variable-sized gene–expression inputs, preserve the distinction between missing and observed measurements, and compress them into a small number of latent tokens. I am not claiming that such an alternative would necessarily perform better, but the cost of directly passing approximately 20,000 gene tokens does not fully justify retaining a fixed-panel representation.

My concern is therefore not simply that the model lacks an additional feature, but that the interface connecting transcriptomic data to the LLM, a central opportunity for methodological innovation, largely retains the fixed-coordinate formulation of conventional transcriptomic models. The multitask and language-interface benefits described in the response are valuable, but address a different aspect of the framework.

I recognize the substantial effort in data integration, training, and benchmark construction. Nevertheless, these strengths do not by themselves resolve my concern about whether the work sufficiently advances the representation paradigm for transcriptome–LLM integration.

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

OmicsLM integrates bulk and single-cell transcriptomes into a Qwen3.5-9B language model by projecting normalized expression values and frozen the feature of the pretrained cell foundation model into continuous embedding per profile. Training combines 20B-token continued pretraining, 5.21M instruction conversations, and GRPO. The paper introduces GEO-OmicsQA, containing 3,000 questions from 235 studies and 183 publications excluded from the authors’ training corpus. OmicsLM leads on three of four annotation tasks, achieves the strongest evaluated baseline, and reports the results of perturbation setting following PerturbQA.

Critical strengths
  1. Important and practically relevant research problem. Enabling language models to interpret quantitative transcriptomic data and compare multiple experimental samples addresses a meaningful need in biological research. The focus on comparisons across experimental conditions and cellular states is well motivated by common biological workflows and gives the work clear practical relevance.

  2. Substantial integration of biological data and supervision. The training corpus connects experimental measurements, biological knowledge, and diverse tasks.

Critical weaknesses
  1. Limited reproducibility of the central training-data contribution. The public bundle provides held-out evaluation datasets, precomputed inputs, model weights, and inference code, but I could not locate the processed training corpus or a complete construction pipeline. Given the emphasis on data construction, access to either is important for independent auditing and community reuse. Existing ablations also do not isolate the benefits of the proposed reasoning data and quality filters.

  2. Unresolved discrepancies in PerturbQA results. Although Appendix A.6.4 specifies the official direction test split and per-gene AUROC, baseline scores differ from the original paper. For example, retrieval-only K562 AUROC changes from 0.50 to 0.6913. These discrepancies require reconciliation. The decision rules underlying the newly computed diagnostic macro-F1 also need justification before that comparison can be fully interpreted.

Questions for authors
  1. Is there any plan for releasing the pretrained data or the whole pre-processing pipeline?

  2. Can the authors reconcile the original and rescored PerturbQA baseline results, specifying the prediction files, replicate aggregation, and score transformations responsible for the discrepancies?

  3. How many GEO-OmicsQA questions were expert-reviewed, and what fraction of reference claims was verified directly against the linked profiles? What do the existing profile-replacement results show separately for single-profile questions and between-group comparisons?

2 comments

Thank you for recognizing the practical value of multi-profile interpretation and the breadth of our biological training data.

1. Reproducibility and training-data release. We believe our release provides substantial support for independent verification. It includes model weights, evaluation datasets, precomputed inputs, and the evaluation pipeline, including inference and scoring code for OmicsLM and baseline comparisons. These materials enable others to evaluate the model and reproduce our benchmark comparisons. The processed training corpus and its construction pipeline are commercially sensitive company assets and cannot be released. We document the data sources, task definitions, training mixture, quality checks, and benchmark exclusions in the appendix. Our ablations test input representations and training stages.

2. PerturbQA scores and decision rules. We reproduced 0.6913 for K562 retrieval using the authors' evaluation code and released predictions (results/retrieve_no_llm/k562.pkl in results.zip, Zenodo 14915313). All 3,119 test cases and labels match the official split. No score transformation or additional averaging was applied. The difference from the original paper's 0.50 remains unexplained.

We report macro-F1 to assess final Up/Down decisions, giving both directions equal weight. Majority voting combines three generated answers into one prediction. With Up=1 and Down=0, this is exactly a 0.5 threshold on their mean. For released scalar baselines, we mark this F1 as diagnostic because their original decision rules are unverified (Appendix A.6.4).

3. Expert review and profile dependence. Biological experts reviewed 300 GEO-OmicsQA questions for question–answer validity, sample grounding, and biological coherence. Over 90% of the checked reference claims agreed with the linked expression data. Separately, experts reviewed 50 model-response traces and found no errors that would affect scoring in those cases.

In the reported profile-replacement experiment, binary accuracy falls from 76.7% to 69.5%, and the free-text score from 0.65105 to 0.61445. Replacement profiles are sampled from the lowest-cosine 5% of eligible profiles. This shows sensitivity to the supplied profiles, but does not establish that every answer depends on the correct experimental comparison.

Reviewer_QGuPOpenReview ↗

Reply to the authors

Thank you for the clarifications. Reproducing the K562 retrieval score from the released predictions resolves my concern about the PerturbQA baselines, and the expert-review numbers are helpful.

My main reservation remains the value of this work to the community. The paper presents the training corpus and its construction as a central contribution, yet neither can be released, so other researchers cannot reuse, audit, or extend this part. In addition, in the profile-replacement experiment, accuracy only drops from 76.7% to 69.5% when the supplied profiles are replaced with highly dissimilar ones. This suggests that a substantial share of answers can be obtained without the correct expression data. Could the authors comment on how much of the GEO-OmicsQA performance reflects use of the expression profiles rather than textual context and prior knowledge?

Given these points, I will keep my score at 2.

Reviewer_5GcWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces OmicsLM, a 9B-parameter multimodal language model that integrates quantitative transcriptomic measurements with natural language instructions for biological reasoning. This approach aims to overcome the limitations of text-based representations of gene expression, which can lose information regarding expression magnitude and consume excessive context window space when multiple biological samples are involved.

Critical strengths
  1. Combining quantitative transcriptomic data with natural language-based reasoning holds significant value for biological research. The ability to jointly analyze multiple expression profiles is particularly important for experimental comparisons, perturbation analyses, and disease-related studies.

  2. By projecting transcriptomic features into compact embedding vectors, OmicsLM avoids the substantial token overhead associated with long sequences of gene names. Ablation experiments further demonstrate that this continuous omics representation outperforms the text-based, ranked-gene representation methods tested.

  3. The integration of heterogeneous biological measurements, experimental results, curated knowledge, and reasoning supervision—drawn from 157 task groups—represents a substantial undertaking. The combination of multimodal pre-training and instruction tuning offers a viable pathway for developing general-purpose biological language models.

Critical weaknesses
  1. The methodological innovation at the architectural level is relatively limited. OmicsLM primarily employs a well-established multimodal projection approach, in which pre-trained continuous representations are mapped into the LLM’s embedding space via a linear layer. While effective, this design has already been explored in existing multimodal language models, including prior frameworks for aligning omics data with language. Consequently, the primary contributions appear to lie in the construction of training data, the breadth of task coverage, and the scale of the implementation, rather than in a novel modeling mechanism. The paper could more clearly distinguish between specific methodological advancements and elements inherited from existing multimodal architectures.
  2. The evaluation fails to sufficiently demonstrate that the model relies on quantitative omics data when performing biological reasoning. The questions in GEO-OmicsQA are derived from published studies, which often contain experimental context, biological entities, and task-specific information; this allows language models to generate plausible-sounding answers even without fully interpreting the provided expression profile data. Although the authors demonstrate the benefits of continuous omics inputs through ablation studies, more direct evidence is required to confirm whether the final model truly leverages sample-specific quantitative differences when answering biological questions.
  3. There are significant differences in training methodologies and input protocols when compared to existing models. OmicsLM benefits from extensive domain-specific training across a wide range of biological tasks, whereas many general-purpose large language models (LLMs) typically receive ranked gene lists rather than continuous omics embeddings. Consequently, the performance gains reported in the study may stem from differences in input representation, biological supervision, and training scale, rather than solely from the proposed modeling framework itself. Notably, Table 2 lacks a directly comparable "omics-language model" evaluated using the same multi-sample interface. A more rigorous comparison—utilizing unified omics representations or comparable training conditions—would provide stronger support for the claims regarding the model's advantages.
  4. The reliability of the GEO-OmicsQA evaluation warrants further validation. While it is commendable that the benchmark is constructed from real-world studies—with content filtered to meet publication standards—the reference answers are drawn directly from the original literature. Furthermore, the evaluation of free-text predictions relies on an "LLM-based judge." This approach risks causing the model to over-rely on the phrasing and biological conclusions found in the source literature, rather than independently verifying whether the generated explanations are genuinely supported by the underlying expression profile measurements. Although parts of this benchmark have been reviewed by experts, the study would be more compelling if it included a quantitative analysis of the agreement between human evaluators and model-generated assessments, as well as a more detailed evaluation of instances where the model provides biological explanations that appear plausible but lack sufficient evidence.
  5. Regarding the ablation studies, the research fails to fully isolate the specific contributions of individual components within the final 9B-parameter model. While Table 4 evaluates several design choices, most of these controlled ablations were conducted using a 0.8B-parameter base model; the final model, by contrast, is significantly larger and has undergone additional reinforcement learning. Although these small-scale experiments offer some insight, their conclusions may not necessarily generalize to the final system. Furthermore, since the model’s input representation combines normalized expression values ​​with two pre-trained omics encoders, it is essential to clarify the relative contributions of the raw quantitative measurements versus the pre-trained biological representations at the scale of the final model. Incorporating more targeted control experiments would provide stronger evidence for the necessity of each component.
  6. The current results for perturbation prediction are insufficient to fully demonstrate the model’s generalization capability in novel biological contexts. On the PerturbQA benchmark, OmicsLM achieves an average AUROC of 0.7635, outperforming other baseline methods. However, this evaluation covers only four cell lines and focuses primarily on predicting the direction of change for differentially expressed genes (i.e., readout genes). The authors also acknowledge that the baseline systems differ in terms of their training data and input schemes. Consequently, while these results are strong within the context of this specific benchmark, they do not establish the model’s ability to generalize to unseen cell lines, novel perturbation targets, or a broader range of quantitative perturbation responses. Conducting further controlled generalization experiments would help to better substantiate the practical significance of these findings.
Questions for authors

Critical Weaknesses

1 comment

Thank you for recognizing the value of multi-profile analysis, compact inputs, and broad biological supervision.

1. Contribution and architectural novelty. We agree that the architectural novelty is limited. We see the main contribution in how we construct the data and train the model: connecting quantitative expression profiles with experimental results, biological knowledge, and reasoning supervision across 157 task groups. Bringing these sources into a common training setup enables one model to address diverse biological tasks and interpret multiple profiles together.

2. Does the final model use the profiles? We believe the existing results provide complementary evidence of profile use. On GEO-OmicsQA, replacing only the profiles in the final 9B model reduces both binary accuracy and free-text scores (Appendix A.6.1). Since the question text is unchanged, this links answer quality to the supplied profiles. The controlled 0.8B ablation further supports the value of the full omics representation: it achieves higher mean scores than ranked gene names on all seven benchmark metrics, with the same backbone and training stages (Appendix A.8).

Tabula Sapiens and Human Diseases provide a separate check. Within each annotation task, the question and candidate labels are fixed; only the profile varies. The strong annotation results therefore demonstrate use of sample-specific information. The improvements across different task categories further suggest that this ability is useful beyond a single benchmark.

3. Fair comparisons and omics-language baselines. Our contribution is the complete OmicsLM system, including its training data, training procedure, and compact omics interface. The benchmark comparisons evaluate this combined contribution under the reported input protocols. The controlled 0.8B ablations further test design choices within the same backbone and training setup.

4. Reference validity and expert agreement. Free-text comparisons use the same LLM judge and rubric. It accepts synonyms and equivalent wording and checks entities, directions, conclusions, and completeness. Publication passages make references traceable. However, the judge does not inspect expression values, so a plausible explanation can agree with the reference without being established by the supplied measurements.

Experts reviewed responses to 50 distinct questions and found no errors that would affect scoring in those cases.

5. Components at 9B scale. We agree that component ablations on the final 9B model would provide stronger evidence. A single 9B pretraining and SFT run required about 2,700 H100 GPU-hours, before RL. Repeating each 9B variant across three seeds was beyond our compute budget, so we used 0.8B models for the controlled component ablations. We performed RL on the 9B model and report its performance before and after RL.

6. Generalization. Beyond held-out perturbation–cell-line combinations in PerturbQA, we show transfer to held-out studies in GEO-OmicsQA and the held-out Tabula Sapiens dataset. We compare complete pipelines under the reported protocols, including their different training data and input designs, as our contributions include data construction and training alongside the model.

Reviewer_4Jn8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces OmicsLM, a 9B multimodal language model for transcriptome-conditioned prediction and biological question answering. Each bulk or single-cell expression profile is represented by a concatenation of normalized expression values, frozen Funomics and Geneformer embeddings, and an input-scale indicator. An affine projection maps this representation into a single language-model embedding, enabling multiple profiles to be interleaved with natural-language instructions. The model also incorporates explicit gene tokens. Evaluation covers four annotation tasks, six GEO-OmicsQA variants, and PerturbQA direction-of-change prediction. OmicsLM leads on three annotation tasks, all six QA variants, and mean PerturbQA AUROC. Ablations investigate representations, training stages, model size, and gene tokenization.

Critical strengths
  1. A substantial integration of biological measurements and language supervision. The contribution goes beyond attaching an encoder to an LLM. The paper describes a broad corpus connecting expression profiles to annotations, perturbations, biological knowledge, and evidence-based explanations. The compact interface also makes comparative questions involving multiple profiles practical. This combination could be a useful foundation for subsequent omics–language research, even though the projection architecture itself follows an established design.

  2. Evidence of utility across complementary tasks, with unusually transparent evaluation details. OmicsLM performs well on annotation, experimental interpretation, and perturbation-response direction prediction. For example, Table 1 reports disease-subtype AUROC of 0.906 versus 0.839 for CellWhisperer’s chat model, while Table 3 reports mean PerturbQA AUROC of 0.7635 versus 0.6547 for SUMMER. The publication-level GEO split, documented benchmark exclusions, and detailed baseline-scoring audit strengthen these results. The controlled 0.8B representation ablation also supports the usefulness of the continuous-input approach relative to the tested ranked-gene alternative.

Critical weaknesses
  1. The evidence does not yet sufficiently isolate quantitative, sample-specific reasoning from other sources of advantage. The incorrect-profile experiment in Appendix A.6.1 is valuable and demonstrates input sensitivity. Nevertheless, replacing the profiles only reduces binary accuracy from 76.7% to 69.5% and free-text score from 0.65105 to 0.61445, leaving substantial performance intact. This does not invalidate the model, but it leaves unclear which questions genuinely require the supplied samples, particularly their comparative relationships. The lowest-cosine replacement control also does not directly test whether the model distinguishes closely related experimental conditions. Moreover, the GEO-OmicsQA baselines receive truncated ranked gene names without expression magnitudes, whereas OmicsLM receives full expression information and pretrained embeddings. CellWhisperer is omitted from this benchmark. The within-recipe ranked-gene ablation partially addresses this concern, but does not separate representation format from information content. These limitations matter because quantitative multi-sample interpretation is the paper’s central motivation.

  2. The principal QA evaluation lacks sufficiently quantified independent validation and uncertainty. GEO-OmicsQA is generated from publications, and its free-text evaluation uses a single Gemini judge comparing answers with literature-derived references (Appendix A.6.3). The judge does not directly inspect the expression measurements. Reference agreement therefore need not establish that an explanation is supported by the supplied profiles. Expert review is mentioned, but the reviewed subset’s size, selection procedure, reviewer agreement, and observed error rates are not reported. In addition, the 3,000 questions come from 235 studies and 183 publications, with up to 57 questions per study, yet the main comparisons lack uncertainty estimates accounting for this clustering. This is consequential for interpreting the free-text improvement of 0.651 versus 0.613. The reported advantage is promising, but its robustness and correspondence to independently assessed biological correctness remain uncertain.

Questions for authors
  1. Could you report final-9B performance with no profiles and with within-study, condition-matched profile swaps, keeping question wording unchanged? Separating single-profile from multi-profile questions would help establish whether the model uses the particular samples and their comparative relationships.
  2. Could you evaluate a strong language baseline supplied with expression magnitudes for question-relevant genes, rather than only ranked gene names? A CellWhisperer comparison on compatible single-profile questions would also help distinguish the benefit of the proposed training recipe from the input-interface advantage.
3 comments

Thank you for recognizing the biological integration, compact multi-profile interface, and evaluation details.

1. Profile use. We evaluate the final 9B model. Only 40 binary and 25 free-text questions contain a single profile, making a separate comparison sensitive to a few answers. We therefore pool single- and multi-profile questions, using all 1,000 binary and 2,000 free-text questions where the control is applicable. Tables show correct profiles → control on the same questions; n is the number of questions.

a. Removing profiles. We do not consider removing profiles without adjusting the question text a clean test of profile use. The text still refers to profiles as if they were available, creating a mismatch between the instructions and the actual input. This may confuse the model and makes the resulting scores difficult to interpret as a measure of reliance on expression data, assuming that the model can even output answers in expected format.

We therefore also tested an adjusted text-only prompt: we removed profile inputs and markers, replaced sample IDs with neutral labels, and explicitly instructed the model to answer from the question text and prior biological knowledge, without access to expression measurements. We retained the biological question and original scoring.

QuestionsBinary accuracy: profiles → text-onlyFree-text score: profiles → text-only
All76.7% → 69.6% (n=1,000)0.651 → 0.435 (n=2,000)

Many questions still provide biological context, so we would not expect text-only predictions to fall to chance. The 7.1-point binary accuracy drop matches the paper’s dissimilar-profile result (76.7% → 69.5%), providing consistent evidence that expression profiles contribute to the answers.

b. Whole-group swaps within studies. We randomly exchange complete profile groups within each study. We keep biological labels and instructions unchanged, adjusting sample-list lengths when group sizes differ.

ControlBinary accuracyFree-text score
Whole-group swaps76.0% → 66.3% (n=793)0.659 → 0.636 (n=1,581)
Dissimilar profiles76.0% → 67.0% (n=793)0.659 → 0.633 (n=1,581)

We also report the paper’s dissimilar-profile control on these same questions, replacing each profile with a random donor from the lowest-cosine 5% of eligible profiles.

We include all 2,374 questions where complete groups can be exchanged unambiguously. Of the remaining 626, 592 lack identified comparison groups (including single-profile questions), and 34 have overlapping, incomplete or interleaved sample lists, or explicit sample-count wording that would become inconsistent. Eligibility was fixed before inference. Scores measure agreement with the original references after swapping.

Both controls cause similar score drops, suggesting that OmicsLM uses the correspondence between expression profiles and the conditions described in the question.

2. Strong baseline.

a. LLMs with expression values. Thank you for these suggestions. We evaluate Gemini 3.8 Flash, the strongest baseline across all six variants in our GEO-OmicsQA table, and Qwen3.8-27B. We select question-relevant genes with a heuristic using HGNC names/aliases and keyword matching to Reactome pathways and HGNC gene groups. Each model receives the same selected genes’ expression values (TPM) for each sample. Neither gene selection nor answering uses reference answers.

The new evaluations cover all 1,000 binary and 2,000 free-text questions. Across these questions, the heuristic selects a median of 97 genes and a mean of 69.1 genes per question. Each cell reports binary accuracy / free-text score.

ModelBefore: ranked genesHeuristic + expression valuesNative expression profiles
Gemini 3.8 Flash66.2% / 0.61373.8% / 0.642—
Qwen3.8-27B54.6% / 0.40365.0% / 0.527—
OmicsLM——76.7% / 0.651

Selecting relevant genes and providing their expression values substantially improves baseline scores, but both models still score below OmicsLM on both metrics.

b. CellWhisperer. GEO-OmicsQA contains 65 single-profile questions. We evaluated CellWhisperer using its released checkpoint and native preprocessing, and compared both models on these same questions.

ModelBinary accuracy (40 questions)Free-text score (25 questions)
CellWhisperer (Yes/No likelihoods)57.5%-
CellWhisperer (full generated answers)50.0%0.488
OmicsLM75.0%0.792

The likelihood method selects Yes or No by comparing their average token log-probabilities, using CellWhisperer's released scoring code. For the full-answer check, we generated all 40 binary responses with a 1,024-token limit; none was truncated. We manually assigned 35 Yes, 2 No and 3 Unknown labels, with Unknown counted as incorrect. Free-text answers use the same judge for both models.

Hi reviewer 4Jn8,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

Thank you for the thorough rebuttal. The new profile-swap controls, expression-informed baselines, and CellWhisperer comparison address my main concerns about whether OmicsLM uses the input data and how fair the comparisons are. While some evaluation limitations remain, the added evidence is sufficient for me to raise my score to 3 (weak accept).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.