acceptodds
Under review as a conference paper at ICLR 2027

Beyond Fitting: Knowledge-Guided LLM Reasoning for Combinatorial Gene Perturbation Prediction

Abstract

Combinatorial perturbation prediction aims to estimate transcriptional responses induced by simultaneous perturbation of multiple genes, which is essential for advancing a mechanistic understanding of biochemical processes. Existing deep learning methods primarily rely on learned representations to predict perturbation responses, often lacking explicit reasoning and mechanistic interpretability. Meanwhile, emerging LLM-enhanced methods still struggle with continuous-response prediction and inductive generalization to unseen perturbation genes. To address these limitations, we propose **CombPert**, a biological knowledge-guided LLM reasoning framework for **Comb**inatorial **Pert**urbation prediction. Specifically, CombPert first constructs a task-specific biological knowledge graph from curated biological resources and dataset-specific co-expression statistics, and retrieves mechanistic paths connecting the perturbation genes and the target gene. To improve inductive generalization and response magnitude estimation, CombPert further incorporates similarity-based evidence from similar perturbation genes and conditional statistical priors. Finally, CombPert integrates these sources into a structured CoT prompt, supporting Reasoning Elicitation for training-free prediction and mechanistic explanation, and Reasoning Synthesis for CoT data generation and downstream fine-tuning. Extensive experiments on representative benchmarks demonstrate strong predictive performance and inductive generalization to unseen perturbation genes. Our code is available at https://anonymous.4open.science/r/CombPert-submit-FBA5.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Combinatorial Perturbation Prediction, Large Language Models, Biological Knowledge Graphs

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vXVAOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors proposed CombPert, a biological knowledge-guided LLM reasoning framework for predicting perturbation responses combinatorially. The work builds on LLM-enhanced methods, for example, Scouter, and argues that LLM-enhanced methods have limited generalisation power for genes partially seen or unseen in the training data. They instead propose to use prior knowledge through constructing a task-specific biological knowledge graph from curated databases and dataset-specific co-expression statistics. Gene similarity and conditional priors are bundled into a CoT prompt for prediction and reasoning (reasoning elicitation) and distillation onto smaller language models (reasoning synthesis). CombPert with GPT 5.4 outperforms LLM-only, GEARS, scLAMBDA, and Scouter on Norman et al. and Replogle et al. datasets for the top 20 DEGs in most settings.

Critical strengths
  1. Combinatorial generalisation to unseen perturbations is an important challenge, and the question is well motivated by observation, and the writing is generally clear.

  2. The paper explored the idea of whether having task-specific prior knowledge would help LLMs to predict combinatorial perturbation direction and response value accurately, with the added advantage of explicit reasoning.

  3. The datasets and baselines are appropriately selected, with clear performance gains on the 20 DEGs subset of the Norman et al. dataset.

Critical weaknesses
  1. The framework is mainly an assembly of retrieval from KG priors and prompting techniques (CoT, CoT distillation) applied to a new task, with most gains over GEARS on unseen perturbations coming from a strong LLM only (GPT 5.4).

  2. CombPert is given the (pair, gene), and operates only on the DEGs in the main result section, unlike GEARS which works on the full gene set. This requires the model in deployment to predict the DEGs first.

  3. The effect of reasoning synthesis is weak to marginal compared to reasoning elicitation. The student model retains the full retrieval context, and it is unclear if there is any gain coming from the CoT traces.

  4. The results section does not report run-to-run variations with a non-zero temperature. This makes it difficult to separate the real gains from LLM stochasticity.

  5. The paper argues that unseen perturbation performance is due to knowledge-guided reasoning, but the evidence does not isolate this. Open-weight models generally lose to GEARS on unseen perturbation pairs on MSE. The effect of the approach over GEARS could be due to a strong and generalisable LLM (GPT5.4) instead.

  6. The method relies on the quality of curated prior knowledge, which might be biased or contain errors. This should be highlighted.

Questions for authors
  1. Can the LLM reliably predict the DEGs before predicting the responses?

  2. Do the fine-tuned student models still receive the prior knowledge paths/priors at inference?

  3. If noise is purposefully introduced into the prior knowledge KG, would CombPert performance degrade?

  4. What are the test set sizes for the seen/unseen perturbation settings? They appear to be relatively small. Then the LLM stochasticity could weigh more.

7 comments
AuthorsOpenReview ↗

Response to Reviewer vXVA (Part 1)

Dear Reviewer vXVA:

We sincerely thank you for your review and invaluable feedback. To better explain our work, we will make the following clarifications about your concerns:

Concern 1: The framework is mainly an assembly of retrieval from KG priors and prompting techniques (CoT, CoT distillation) applied to a new task, with most gains over GEARS on unseen perturbations coming from a strong LLM only (GPT 5.4).

We thank the reviewer for raising these points. We address them in two parts.

  1. Task-specific biological guidance. CombPert organizes mechanistic paths, evidence from analogous perturbations, and statistical priors into a scaffold for combinatorial perturbation prediction. The guidance directs the LLM to assess the individual effects of the two perturbations, use cross-interaction and convergence paths to assess their joint effect, and estimate the resulting DE with the conditional prior as a quantitative reference. This scaffold supports Reasoning Elicitation for training-free prediction and mechanistic explanation, and Reasoning Synthesis for generating CoT trajectories for downstream fine-tuning.

  2. Performance with open-weight backbones. CombPert also performs strongly with open-weight backbones. On Norman Seen 0/2, Reasoning Elicitation reduces Qwen3-32B’s MSE from 0.540 under few-shot prompting to 0.321 with the same backbone. With Reasoning Synthesis, CombPert with Qwen3-32B achieves lower MSE (0.245 vs. 0.254) and higher direction accuracy (92.1% vs. 88.2%) than GEARS, with comparable PCC (0.791 vs. 0.800). In this setting, GPT-5.4 generates the training trajectories, and Qwen3-32B performs inference.In the overall Norman evaluation, both Qwen3-32B variants and the Qwen3-8B Synthesis variant outperform GEARS on all three metrics. The Qwen3-32B Synthesis variant also outperforms Scouter and scLAMBDA on all three overall metrics. On Replogle Seen 0/2, the open-weight CombPert variants achieve higher direction accuracy than GEARS, while GEARS achieves lower MSE.

MethodOverall MSE ↓Overall PCC ↑Overall ACC ↑Seen 0/2 MSE ↓Seen 0/2 PCC ↑Seen 0/2 ACC ↑
GEARS0.2330.7990.8950.2540.8000.882
Scouter0.2650.8050.9130.4460.6070.809
scLAMBDA0.2780.8340.9370.4140.6830.850
Qwen3-32B (Few-shot)0.5710.2880.6060.5400.2810.606
CombPert + Qwen3-8B (Synthesis)0.2160.8320.9280.2840.7510.885
CombPert + Qwen3-32B (Elicitation)0.2070.8570.9370.3210.7650.908
CombPert + Qwen3-32B (Synthesis)0.1940.8510.9410.2450.7910.921
CombPert + GPT-5.4 (Elicitation)0.1590.8770.9550.2320.7960.917

Concern 2: CombPert is given the (pair, gene), and operates only on the DEGs in the main result section, unlike GEARS which works on the full gene set. This requires the model in deployment to predict the DEGs first.

Following the evaluation protocol used in prior work [1], our main results report prediction accuracy on the 20 genes with the largest observed absolute DE for each perturbation. This evaluation emphasizes the strongest transcriptional responses, whose direction and magnitude help characterize the effects of the perturbation.

At inference, CombPert takes a perturbation pair and a target gene as input. It can predict responses across the full measured panel by querying each gene directly, without prior DEG screening. Appendix B.9 and Table 14 of the submitted manuscript already report this full-panel evaluation on Norman. Using GPT-5.4-mini, CombPert achieves an overall PCC of 0.711, compared with 0.593 for GEARS and 0.575 for few-shot prompting with the same backbone. On Seen 0/2, CombPert and GEARS achieve comparable PCC (0.550 vs. 0.549). We will highlight this result in the main text and clarify that the top-20 restriction applies to the primary evaluation subset.

MethodSeen 2/2Seen 1/2Seen 0/2Overall
CPA0.563 ± 0.0000.583 ± 0.0010.520 ± 0.0020.571 ± 0.001
GEARS0.649 ± 0.0020.592 ± 0.0010.549 ± 0.0020.593 ± 0.001
GPT-5.4-mini (Few-shot)0.793 ± 0.0010.562 ± 0.0020.215 ± 0.0000.575 ± 0.001
CombPert (GPT-5.4-mini)0.894 ± 0.0010.714 ± 0.0020.550 ± 0.0020.711 ± 0.001

[1] Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nature Biotechnology, 2024.

AuthorsOpenReview ↗

Response to Reviewer vXVA (Part 2)

Concern 3: The effect of reasoning synthesis is weak to marginal compared to reasoning elicitation. The student model retains the full retrieval context, and it is unclear if there is any gain coming from the CoT traces.

We agree that the gains over Reasoning Elicitation are modest and vary by metric. On Norman, Reasoning Synthesis reduces overall MSE from 0.241 to 0.216 for Qwen3-8B and from 0.207 to 0.194 for Qwen3-32B. The students retain the retrieval context at inference because the objective is to improve prediction given that evidence. Appendix B.8 (Table 13) of the submitted manuscript also compares training data generated by the same GPT-5.4 teacher using either CombPert or few-shot prompting. Under matched fine-tuning and evaluation settings, training on CombPert-generated data reduces overall MSE from 0.562 to 0.216 for Qwen3-8B and from 0.394 to 0.194 for Qwen3-32B. This comparison supports the utility of CombPert-generated training data.

Besides, to isolate the contribution of CoT traces, we fine-tune Qwen3-8B on the same CombPert-generated examples from Norman. Answer-only SFT uses only the teacher’s DE value, while CoT SFT uses the same value and its reasoning trace. Both variants use the same initialization, fine-tuning settings, and retrieved evidence, and are evaluated on identical test queries.

Training variantOverall MSE ↓Overall PCC ↑Overall ACC ↑
Answer-only SFT0.2640.7930.901
CoT SFT (Reasoning Synthesis)0.2160.8320.928

Adding CoT trace supervision reduces overall MSE from 0.264 to 0.216. With the training examples, numerical targets, and retrieval context held fixed, this improvement quantifies the additional predictive value of the CoT traces.

Concern 4: The results section does not report run-to-run variations with a non-zero temperature. This makes it difficult to separate the real gains from LLM stochasticity.

Appendix B.5 and Tables 9–10 of the submitted manuscript already report the mean and standard error for MSE, PCC, and direction accuracy. These estimates summarize variation across test perturbation conditions at a decoding temperature of 0.6.

Concern 5: The paper argues that unseen perturbation performance is due to knowledge-guided reasoning, but the evidence does not isolate this. Open-weight models generally lose to GEARS on unseen perturbation pairs on MSE. The effect of the approach over GEARS could be due to a strong and generalisable LLM (GPT5.4) instead.

Thank you for raising this point. Table 3 of the submitted manuscript already reports a comparison with the same GPT-5.4 backbone and input evidence. In the w/o Biological Guidance variant, the LLM directly predicts DE from the assembled evidence after we remove the task-specific biological instructions. The comparison below measures the contribution of those instructions on Seen 0/2.

MethodNorman MSE ↓Norman PCC ↑Norman ACC ↑Replogle MSE ↓Replogle PCC ↑Replogle ACC ↑
GEARS0.2540.8000.8820.0610.5830.500
w/o Biological Guidance0.2830.7600.9030.3180.4820.700
CombPert0.2320.7960.9170.1760.5880.790

With the backbone and evidence held fixed, adding biological guidance reduces MSE from 0.283 to 0.232 on Norman and from 0.318 to 0.176 on Replogle. On Norman, GEARS’s MSE of 0.254 lies between the two GPT-5.4 variants, showing that biological guidance contributes to CombPert’s lower MSE in this setting. On Replogle, the guidance also improves prediction, while GEARS retains a lower MSE. As discussed in our response to Concern 1, the framework also improves prediction with an open-weight backbone. On Norman Seen 0/2, training-free CombPert reduces Qwen3-32B’s MSE from 0.540 to 0.321 relative to few-shot prompting with the same backbone.

Together, these results support a contribution from task-specific biological guidance under a fixed backbone and show that performance relative to GEARS varies across datasets and backbones.

Concern 6: The method relies on the quality of curated prior knowledge, which might be biased or contain errors. This should be highlighted.

We agree that the quality of curated biological knowledge matters. Appendix D (Limitations) of the submitted manuscript explicitly notes that incomplete or noisy biological knowledge can affect evidence retrieval and prediction quality. Besides, Appendix B.4 illustrates this sensitivity on Norman. With the numerical prior and evidence from analogous perturbations held fixed, reversing a key path increases MSE from 0.159 to 0.256.

AuthorsOpenReview ↗

Response to Reviewer vXVA (Part 3)

Question 1: Can the LLM reliably predict the DEGs before predicting the responses?

CombPert can predict DE for each gene in the measured panel directly, without prior DEG screening. Appendix B.9 (Table 14) of the submitted manuscript reports this full-panel evaluation on Norman. As discussed in our response to Concern 2, the top-20 restriction applies to the primary evaluation subset.

Question 2: Do the fine-tuned student models still receive the prior knowledge paths/priors at inference?

Yes. The fine-tuned students receive the same types of evidence as in Reasoning Elicitation, including retrieved KG paths, evidence from analogous perturbations, and the conditional statistical prior. As discussed in our response to Concern 3, Reasoning Synthesis trains the student to make predictions given this evidence.

Question 3: If noise is purposefully introduced into the prior knowledge KG, would CombPert performance degrade?

Appendix B.4 (Figure 9) tests sensitivity to errors in the retrieved paths. On Norman, reversing a key path increases MSE from 0.159 to 0.256 while holding the numerical prior and evidence from analogous perturbations fixed. This experiment tests corruption of a retrieved path at inference. Our response to Concern 6 discusses the resulting limitation.

Question 4: What are the test set sizes for the seen/unseen perturbation settings? They appear to be relatively small. Then the LLM stochasticity could weigh more.

We evaluate CombPert on two established Perturb-seq datasets, Norman and Replogle. Tables 6 and 7 of the submitted manuscript report the test set sizes. Each double-perturbation condition contributes 20 target-gene queries to the main evaluation. We report both counts below.

DatasetSeen 2/2 conditions / queriesSeen 1/2 conditions / queriesSeen 0/2 conditions / queries
Norman14 / 28048 / 96017 / 340
Replogle5 / 10012 / 2403 / 60

Individual perturbation conditions can have a greater influence on aggregate scores in smaller splits, particularly Replogle Seen 0/2. To quantify uncertainty in the reported mean performance, Appendix B.5 (Tables 9–10) reports standard errors for MSE, PCC, and direction accuracy across test perturbation conditions. As discussed in our response to Concern 4, these estimates describe uncertainty across conditions.

Reviewer_Gd2zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CombPert integrates knowledge-graph mechanistic paths, evidence from homologous/similar genes, and conditional statistical priors into a structured cot. This enables an LLM to directly regress continuous de values of target genes under double-gene perturbations while providing mechanistic explanations.

Critical strengths
  1. The problem is well-motivated. Previous LLM-based reasoning methods were limited to binary classification or single-gene perturbations,

  2. The paper is easy to read and follow.

Critical weaknesses
  1. Inductive generalization. On the Replogle Seen 0/2 split, CombPert yields an MSE of 0.176—roughly three times higher than GEARS (0.061), worse than both Linear (0.078) and scLAMBDA (0.112), and only marginally outperforming the Mean baseline (0.196). While the method captures the direction of change well, its magnitude calibration remains poor.

  2. Disentangling the sources of performance gains. It remains unclear to what extent the improvements stem from the LLM's reasoning versus the injected statistical priors. An ablation study is necessary.

  3. Baselines. The evaluation lacks comparisons against more recent state-of-the-art perturbation-prediction methods, such as CellFlow, PerturbDiff, and scDFM.

Questions for authors

see weakness.

4 comments
AuthorsOpenReview ↗

Response to Reviewer Gd2z (Part 1)

Dear Reviewer Gd2z:

We sincerely thank you for your review and invaluable feedback. To better explain our work, we will make the following clarifications about your concerns:

Concern 1: Inductive generalization. On the Replogle Seen 0/2 split, CombPert yields an MSE of 0.176—roughly three times higher than GEARS (0.061), worse than both Linear (0.078) and scLAMBDA (0.112), and only marginally outperforming the Mean baseline (0.196). While the method captures the direction of change well, its magnitude calibration remains poor.

Thank you for highlighting this result. We interpret the Seen 0/2 results from two perspectives:

1. Performance on unseen perturbation genes varies across datasets. We agree that estimating response magnitudes remains challenging on Replogle Seen 0/2. CombPert achieves an MSE of 0.176, compared with 0.061 for GEARS, 0.078 for Linear, 0.112 for scLAMBDA, and 0.196 for the Mean baseline. One suitable explanation for this difference is that the Replogle test split is smaller than Norman’s (60 vs. 340 target-gene queries), making its aggregate scores more sensitive to individual perturbation conditions. Howerver, on the larger Norman Seen 0/2 split, CombPert achieves both lower MSE and higher direction accuracy than GEARS, with comparable PCC. These gains provide evidence of inductive generalization to unseen perturbation genes across a larger set of held-out conditions.

2. CombPert’s clearest advantage on Replogle is in predicting the direction of expression changes. On Replogle Seen 0/2, CombPert achieves higher direction accuracy than GEARS and Linear (79.0% vs. 50.0% and 63.3%). Its PCC is comparable to GEARS and higher than Linear. These results, together with the higher MSE, illustrate why the three metrics should be interpreted jointly. For example, when the observed DE is +0.1, predicting −0.1 incurs a smaller squared error than predicting +0.4 (0.04 vs. 0.09), even though only the latter correctly predicts upregulation. Direction accuracy captures whether the predicted change has the correct sign, and PCC assesses agreement in response patterns across target genes.

Taken together, the results on Norman and Replogle support CombPert’s ability to generalize to unseen perturbation genes, with its most consistent advantage in predicting the direction of expression changes.

Concern 2: Disentangling the sources of performance gains. It remains unclear to what extent the improvements stem from the LLM's reasoning versus the injected statistical priors. An ablation study is necessary.

We thank the reviewer for raising this point. To complement the existing ablations, we evaluate Conditional Prior-only, which predicts DE using the query-specific prior mean supplied to CombPert. On the same test queries, we compare this baseline with full CombPert and the w/o Biological Guidance variant already reported in Table 3. This variant retains the same LLM backbone and input evidence, removes the task-specific biological instructions, and asks the model to predict DE directly.

Norman et al. dataset

MethodOverall MSE ↓Overall PCC ↑Overall ACC ↑Seen 0/2 MSE ↓Seen 0/2 PCC ↑Seen 0/2 ACC ↑
Conditional Prior-only0.2860.7510.8420.2440.7560.844
w/o Biological Guidance0.1820.8350.9350.2830.7600.903
CombPert0.1590.8770.9550.2320.7960.917

Replogle et al. dataset

MethodOverall MSE ↓Overall PCC ↑Overall ACC ↑Seen 0/2 MSE ↓Seen 0/2 PCC ↑Seen 0/2 ACC ↑
Conditional Prior-only0.1090.4630.6580.1220.2560.550
w/o Biological Guidance0.1120.6830.8800.3180.4820.700
CombPert0.0760.7950.9000.1760.5880.790

Relative to Conditional Prior-only, CombPert reduces overall MSE from 0.286 to 0.159 on Norman and from 0.109 to 0.076 on Replogle, quantifying the predictive gain of the full framework over the prior alone. The matched ablation isolates the additional contribution of biological guidance. With the LLM backbone and input evidence held fixed, removing the guidance increases overall MSE from 0.159 to 0.182 on Norman and from 0.076 to 0.112 on Replogle.

AuthorsOpenReview ↗

Response to Reviewer Gd2z (Part 2)

Concern 3: Baselines. The evaluation lacks comparisons against more recent state-of-the-art perturbation-prediction methods, such as CellFlow, PerturbDiff, and scDFM.

We thank the reviewer for suggesting these baselines. We compare CombPert with CellFlow, PerturbDiff, and scDFM on Norman and Replogle using the same data splits and evaluation queries as in the main experiments. For each baseline, we average the predicted expression profiles within each perturbation condition and subtract the control mean to obtain predicted DE. We then compute MSE, PCC, and direction accuracy using the same target genes and aggregation protocol as in the main tables.

Norman et al. dataset

MethodMSE↓PCC↑ACC↑
GEARS0.2330.7990.895
CellFlow0.3200.7420.895
PerturbDiff0.5310.6600.861
scDFM0.5950.6350.754
CombPert0.1590.8770.955

Replogle et al. dataset

MethodMSE↓PCC↑ACC↑
GEARS0.0850.5380.540
CellFlow0.0910.5420.788
PerturbDiff0.0770.6570.750
scDFM0.1040.5120.563
CombPert0.0760.7950.900

Compared with the three newly added baselines, CombPert achieves the lowest MSE and the highest PCC and direction accuracy on both datasets. On Norman, it reduces MSE from 0.233 for GEARS to 0.159. On Replogle, its MSE is close to PerturbDiff’s, with clearer gains over the strongest baselines in PCC (0.795 vs. 0.657) and direction accuracy (0.900 vs. 0.788).

Dear Reviewer Gd2z,

This is a gentle reminder that we have provided a detailed response to your concerns. We'd appreciate it if you could find time to check the response and further consider your evaluations of our paper. We look forward to further discussions on these new clarifications.

Best wishes,

On behalf of the authors of Paper 24547

Thank you for your detailed response and additional clarifications. I carefully read your rebuttal and the other reviewer's review. I will maintain my original score.

Reviewer_kpH4OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies prediction of continuous target-gene responses to two-gene perturbations, with a focus on unseen genes. It proposes CombPert, which retrieves mechanistic paths from a task-specific knowledge graph, retrieves similar training genes, and computes a similarity-weighted mean and standard deviation of the target’s training responses as a conditional prior. These inputs form a CoT prompt for training-free LLM prediction, or are used to generate traces for fine-tuning smaller models. On the Norman and Replogle datasets, CombPert with GPT-5.4 is reported best on overall MSE, PCC and direction accuracy.

Critical strengths

The paper extends LLM-based perturbation reasoning from binary classification to continuous responses, which is a relevant step. The evidence sources and the full prompt are clearly described, and the generated traces improve smaller open models over few-shot prompting.

Critical weaknesses
  1. The gains may come mainly from the similarity-weighted prior rather than from LLM reasoning. The prior is a weighted average of the target’s training responses to similar perturbations, which is close to a nearest-neighbour predictor. Such predictors, alone or guided by an LLM, are strong for unseen single-gene perturbations (Märtens et al., 2025; Fawkes et al., 2026). On Norman, removing the prior causes the largest drop on all metrics and removing the biological guidance the smallest, while on Replogle the two have effects of similar size. Case 1 stays close to the prior mean, while Case 2 shows a useful correction of it. Without the prior alone as a predictor, it is hard to tell how much the LLM adds.

  2. The contribution specific to combinatorial perturbation is unclear. Only the cross-interaction and convergence paths and the interaction steps of the prompt (Steps 3 and 4) are specific to combinations, and CombPert also improves strongly on unseen single-gene perturbations, where these parts are removed (Appendix B.2). In Seen 0/2, the single effects of both genes appear to be held out as well, so this split mainly tests unseen-gene prediction. There, CombPert is similar to GEARS on Norman, and on Replogle it has better direction accuracy but higher MSE than GEARS and Linear. On the full gene panel (Table 14), the gain over GEARS also disappears in Seen 0/2. Figure 3 does not include the additive baseline, so it does not show whether non-additive effects are predicted.

Questions for authors
  1. How much does the LLM add over predictors given the same evidence? Useful baselines would be the conditional prior alone, a nearest-neighbour average over similar training genes (Fawkes et al., 2026), a ridge regressor on the provided numbers, LLM-guided neighbour selection as in LangPert (Märtens et al., 2025), GenePert, which the paper cites, and the additive baseline for Seen 2/2. It would also help to know what the few-shot baseline receives.

  2. Since the prior already captures how each target gene usually responds, could you evaluate whether the predictions are specific to each perturbation, for example with per-gene or centered metrics?

  3. What evidence shows that CombPert captures non-additive effects? Including the additive baseline in Figure 3 would help, and please clarify how the predicted GI score is computed.

  4. In Appendix B.4, editing a single path increases error more than removing all paths. How should this be interpreted, and how were the human evaluations blinded and checked for agreement?

  5. Could you clarify the evaluation details: the identical standard errors in Tables 9 and 10, variance across sampling runs, whether the KG anchors and hyperparameters used only training or validation data, and the source of the combinatorial Replogle data?

8 comments
AuthorsOpenReview ↗

Response to Reviewer kpH4 (Part 1)

Dear Reviewer kpH4:

We sincerely thank you for your review and invaluable feedback. To better explain our work, we will make the following clarifications about your concerns:

Concern 1: The gains may come mainly from the similarity-weighted prior rather than from LLM reasoning. The prior is a weighted average of the target’s training responses to similar perturbations, which is close to a nearest-neighbour predictor. Such predictors, alone or guided by an LLM, are strong for unseen single-gene perturbations (Märtens et al., 2025; Fawkes et al., 2026). On Norman, removing the prior causes the largest drop on all metrics and removing the biological guidance the smallest, while on Replogle the two have effects of similar size. Case 1 stays close to the prior mean, while Case 2 shows a useful correction of it. Without the prior alone as a predictor, it is hard to tell how much the LLM adds.

We thank the reviewer for these valuable comments. CombPert’s central contribution is a task-specific scaffold that guides the LLM in integrating mechanistic paths, evidence from analogous perturbations, and statistical priors. Within this scaffold, the conditional prior provides a quantitative reference for estimating response magnitudes. The scaffold supports Reasoning Elicitation for training-free prediction and mechanistic explanation, and Reasoning Synthesis for generating CoT trajectories for downstream fine-tuning. To quantify the predictive gain over the prior alone, we evaluate Conditional Prior-only, which uses the same query-specific prior mean supplied to CombPert as the predicted DE value. We compare this baseline with w/o Biological Guidance and full CombPert on the same test queries. In w/o Biological Guidance, the LLM directly predicts the response from the same assembled evidence after the task-specific biological instructions are removed.

Norman et al. dataset

MethodOverall MSE ↓Overall PCC ↑Overall ACC ↑Seen 0/2 MSE ↓Seen 0/2 PCC ↑Seen 0/2 ACC ↑
Conditional Prior-only0.2860.7510.8420.2440.7560.844
w/o Biological Guidance0.1820.8350.9350.2830.7600.903
CombPert0.1590.8770.9550.2320.7960.917

Replogle et al. dataset

MethodOverall MSE ↓Overall PCC ↑Overall ACC ↑Seen 0/2 MSE ↓Seen 0/2 PCC ↑Seen 0/2 ACC ↑
Conditional Prior-only0.1090.4630.6580.1220.2560.550
w/o Biological Guidance0.1120.6830.8800.3180.4820.700
CombPert0.0760.7950.9000.1760.5880.790

On the same test queries, CombPert reduces overall MSE relative to Conditional Prior-only from 0.286 to 0.159 on Norman and from 0.109 to 0.076 on Replogle. Removing task-specific biological guidance while keeping the LLM backbone and input evidence fixed increases MSE to 0.182 and 0.112, respectively. These comparisons show the predictive gain of the full framework over the conditional prior alone and the additional contribution of biological guidance when integrating the prior with mechanistic paths and evidence from analogous perturbations.

AuthorsOpenReview ↗

Response to Reviewer kpH4 (Part 2)

Concern 2: The contribution specific to combinatorial perturbation is unclear. […] Figure 3 does not include the additive baseline, so it does not show whether non-additive effects are predicted.

We thank the reviewer for these valuable comments. we solve these problem from four perspectives:

1. Contribution of the combination-specific reasoning steps. Step 3 assesses potential interactions through cross-interaction and convergence paths, and Step 4 integrates this assessment with individual effects to predict the joint response’s direction and magnitude, using the additive estimate as a reference when available. To quantify the contribution of each step, we compare full CombPert with variants that omit Step 3, Step 4, or both on the Norman et al. dataset.

MethodMSE ↓PCC ↑Direction ACC ↑
CombPert0.1590.8770.955
w/o Step 3 (Interaction Topology)0.1790.8410.934
w/o Step 4 (Quantitative Integration)0.1790.8380.932
w/o Steps 3 and 40.1810.8370.934

Removing either Step 3 or Step 4 increases MSE from 0.159 to 0.179, and removing both increases it to 0.181. All three ablations also reduce PCC and direction accuracy. These results indicate that both steps contribute to predicting joint perturbation responses.

2. Task and split definitions. Combinatorial perturbation prediction involves predicting the joint response to the simultaneous perturbation of two genes. In all three test splits, the test gene pair is held out from training. Seen 2/2, Seen 1/2, and Seen 0/2 indicate that two, one, or zero genes in the test pair appeared as perturbed genes in other training combinations, respectively. We hope this clarification helps distinguish generalization to new combinations of previously observed genes (Seen 2/2) from the additional challenge of generalizing to unseen perturbation genes (Seen 1/2 and Seen 0/2).

3. Performance on unseen perturbation genes. Following prior work [1], our primary evaluation focuses on the 20 genes with the largest observed absolute expression changes per perturbation. On this subset, CombPert’s clearest advantage is in predicting the direction of expression change. On Norman Seen 0/2, CombPert achieves lower MSE (0.232 vs. 0.254) and higher direction accuracy (91.7% vs. 88.2%) than GEARS, with comparable PCC (0.796 vs. 0.800). On Replogle Seen 0/2, CombPert achieves substantially higher direction accuracy than GEARS and Linear, while its MSE is higher. This difference should be interpreted in light of the smaller Replogle split (60 target-gene queries vs. 340 on Norman), which makes aggregate scores more sensitive to individual perturbation conditions. Moreover, MSE, PCC, and Direction ACC capture complementary aspects of prediction. For example, when the observed DE is +0.1, predicting −0.1 yields a smaller squared error than predicting +0.4 (0.04 vs. 0.09), yet it predicts downregulation where the gene is actually upregulated. Direction accuracy captures this distinction and reflects CombPert’s clearest advantage on Replogle, while PCC measures agreement in response patterns across target genes. Finally, on the full measured gene panel in Norman Seen 0/2, CombPert achieves a PCC comparable to GEARS (0.550 vs. 0.549, Table 14), demonstrating its applicability across the measured panel. Many genes show little or no expression change under a given perturbation, and their inclusion may make differences in predicting the strongest responses less visible in panel-wide metrics. For near-zero responses, even a small prediction error can reverse the predicted direction.

[1] Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nature Biotechnology, 2024.

4. Predicting non-additive effects. we add an Additive baseline to each of the four interaction types in Figure 3. For each query, Additive predicts the sum of the two observed single-perturbation DE values. We subtract this same sum from the observed and predicted double-perturbation DE values to obtain their respective GI scores. Additive therefore predicts a GI score of zero, and its GI Score MSE is the mean squared observed GI score within each interaction type. Appendix B.1 provides the calculation details.

MethodSynergy↓Suppression↓Neomorphism↓Redundancy/Epistasis↓
Additive0.3060.1710.1990.213
GEARS0.3590.2200.1910.511
CPA0.6280.5260.6210.598
Scouter0.3750.2430.3020.256
scLAMBDA0.4210.2630.3220.191
CombPert0.2390.1550.1800.116

Compared with this baseline, CombPert reduces GI Score MSE by 21.9% for synergy, 9.4% for suppression, 9.5% for neomorphism, and 45.5% for redundancy/epistasis. These reductions show that CombPert’s predicted deviations from additivity bring its joint response predictions closer to the observed responses in all four interaction categories.

AuthorsOpenReview ↗

Response to Reviewer kpH4 (Part 3)

Question 1: How much does the LLM add over predictors given the same evidence? Useful baselines would be the conditional prior alone, a nearest-neighbour average over similar training genes (Fawkes et al., 2026), a ridge regressor on the provided numbers, LLM-guided neighbour selection as in LangPert (Märtens et al., 2025), GenePert, which the paper cites, and the additive baseline for Seen 2/2. It would also help to know what the few-shot baseline receives.

We appreciate the reviewer’s detailed comment.We therefore conduct evidence-matched comparisons on the Norman et al. dataset, using the same retrieved evidence as CombPert. The GPT-5.4 few-shot baseline is already reported in Table 1.

MethodOverall MSE ↓Overall PCC ↑Overall ACC ↑
Conditional Prior-only0.2860.7510.842
Similar-gene KNN0.2610.7850.898
KG-nearest Neighbor0.2810.7610.895
Ridge Regression0.2470.7960.868
GPT-5.4 (Few-shot, Table 1)0.3140.7100.881
CombPert (GPT-5.4)0.1590.8770.955

CombPert consistently outperforms all non-LLM baselines, showing gains beyond numerical interpolation while also providing human-readable, evidence-grounded mechanistic explanations.

Question 2: Since the prior already captures how each target gene usually responds, could you evaluate whether the predictions are specific to each perturbation, for example with per-gene or centered metrics?

The conditional prior depends on both the queried perturbation and the target gene. For a fixed target gene, changing the perturbed gene pair can change which training perturbations are retrieved and how their observed responses are weighted. As reported in our response to Concern 1, CombPert improves overall PCC over Conditional Prior-only from 0.751 to 0.877 on Norman and from 0.463 to 0.795 on Replogle. These gains reflect better agreement with the observed response profile across target genes. We agree that computing PCC across held-out perturbations for each fixed target gene would directly assess how well CombPert tracks changes in that gene’s response across perturbations.

Question 3: What evidence shows that CombPert captures non-additive effects? Including the additive baseline in Figure 3 would help, and please clarify how the predicted GI score is computed.

Our Gene Interaction Analysis (RQ1, Section 5.4.1 and Figure 3) provides empirical evidence that CombPert captures non-additive effects. As defined in Appendix B.1 of the submitted manuscript, the observed GI score for query q=(A,B,t)q=(A,B,t) is the deviation of the observed double-perturbation DE of target gene tt from the sum of its two observed single-perturbation DE values.

δq=DEtA+B−(DEtA+DEtB).\delta_q = \mathrm{DE}_t^{A+B} - (\mathrm{DE}_t^A + \mathrm{DE}_t^B).

We compute the predicted GI score by subtracting the same observed single-perturbation sum from the predicted double-perturbation DE.

δ^q=DE^tA+B−(DEtA+DEtB).\hat{\delta}_q = \widehat{\mathrm{DE}}_t^{A+B} - (\mathrm{DE}_t^A + \mathrm{DE}_t^B).

A GI score of zero corresponds to an additive response. For each interaction category, GI Score MSE measures how closely the predicted deviations match the observed deviations across its evaluation queries Q\mathcal{Q}.

\frac{1}{\lvert\mathcal{Q}\rvert} \sum_{q\in\mathcal{Q}}(\hat{\delta}_q-\delta_q)^2.$$ CombPert achieves the lowest GI Score MSE among the five methods shown in Figure 3 across all four interaction categories. **Point 4 of our response to Concern 2** reports the comparison with the Additive baseline, which predicts a GI score of zero, on the same evaluation queries in each category.

Comments on the paper

1 comment

Dear reviewers,

Thank you for your reviews. Please read the authors’ response and follow up in the discussion, briefly indicating whether your main concerns have been addressed and highlighting any remaining questions. Please also update your review and rating where appropriate.

Many thanks for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.