acceptodds
Under review as a conference paper at ICLR 2027

Learning When Cellular Responses Transfer: A Factorized Memory-Guided Agent for Perturbation Prediction

Abstract

One main goal of virtual cell modeling is to predict cellular responses under unseen conditions, since the combinatorial space of perturbations and cellular contexts is too large to assay exhaustively. Existing perturbation prediction agents use gene-expression responses and knowledge graphs, but gene-level response tendencies can obscure perturbation-specific effects in their predictions. Retrieved evidence can also be unhelpful or misleading when incorporated without assessing its predictive value for the target query. We introduce the Multi-route Memory-guided Cellular Agent with Calibrated Transfer (MMCA-CT), which extracts perturbation-specific information and learns to assess the predictive value of response estimates to guide evidence selection and model reasoning. Its factorized response memory separates stable gene tendencies from perturbation-specific effects to estimate responses under unseen conditions. The agent learns to weight predictions by their estimated improvement over a gene-prior reference and to select response estimates expected to further improve predictions based on observed responses. Across five cancer cell lines, MMCA-CT improves response detection and direction prediction in both cross-perturbation and cross-cell-line settings. With QWEN3.5-9B, it yields relative gains of 15.3% in macro-per-gene AUPRC over VCWORLD for cross-perturbation response detection and 14.0% in macro-per-gene AUROC over CRISP for cross-cell-line direction prediction. Ablation studies further show that factorized response memory yields the largest accuracy gains, with additional improvements from conditional transfer estimation and utility-guided evidence selection.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Cell Modeling, Perturbation Response Prediction, Out-of-Distribution Generalization, Transfer Learning, Large Language Model Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rp1MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the Multi-route Memory-guided Cellular Agent with Calibrated Transfer (MMCA-CT) for differential expression (DE) detection and expression direction (DIR) prediction under unseen perturbations and cellular contexts. The framework employs factorized response memory to separate stable gene tendencies from perturbation-specific effects, generating response estimates based on molecular structure, drug targets, and mechanism similarity. Conditional transfer estimation then weights response predictions according to their expected predictive gain and transfer reliability, while utility-guided evidence selection retains response estimates expected to improve predictions. Finally, a frozen LLM uses the selected evidence and fused prediction to produce task decisions or biological rationales.

Critical strengths

1. A meaningful research problem and a well-motivated evidence transfer framework. Existing perturbation prediction methods can be dominated by stable gene-level response tendencies, while biological similarity does not necessarily imply predictive transferability. The combination of factorized response memory and conditional transfer estimation provides a reasonable approach to addressing these challenges.

2. MMCA-CT demonstrates improvements in both cross-perturbation and cross-cell-line prediction across five cancer cell lines. The results on DE detection and DIR prediction, together with the component ablations, provide encouraging evidence of its predictive effectiveness.

Critical weaknesses

1. The central factorization hypothesis lacks direct validation. A key contribution of this paper is the separation of stable gene tendencies from perturbation-specific effects through factorized response memory. However, assigning these biological interpretations to the decomposed components does not guarantee that the matrix factorization actually achieves such a separation. The current ablations demonstrate the predictive benefit of the memory module but do not independently validate the stability or perturbation specificity of the learned components. Without such evidence, the central claim underlying the proposed representation remains insufficiently substantiated.

2. The biological relevance of selected evidence remains unvalidated. Utility-guided evidence selection optimizes reductions in prediction log loss rather than the biological correctness or mechanistic relevance of the selected evidence. Although biological information is incorporated into candidate evidence construction, this does not establish that the final selection identifies biologically meaningful mechanisms.

3.Appendix F.4 reports that MMCA-CT underperforms Prior vote in all 128 eligible cross-perturbation gene-set–cell-line comparisons and underperforms same-drug response voting in 88 of 125 cross-cell-line comparisons. While these subgroup results do not invalidate the aggregate improvements, they raise important questions about the consistency and robustness of the claimed transfer benefits, particularly relative to simple response-based predictors.

Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer rp1M [Part 1]

Thank you for recognizing the motivation and predictive effectiveness of our framework. Below, we validate the factorized response memory, discuss the evidence supporting mechanistic hypotheses, and clarify the scope of the Hallmark analyses.

Q1. Direct validation of the factorized response memory

The added analysis supports reproducible cell–gene tendency estimates and a learned structured response that is predominantly perturbation-dependent and recovers held-out gene-by-perturbation variation. These component-level results complement the predictive ablations.

Design rationale

General gene-response tendencies can support strong aggregate performance without distinguishing perturbations for the same target gene (CORE). We therefore estimate the stable cell–gene tendency αcg\alpha_{cg} as the median training expression effect and hold it fixed during factorization. Broad perturbation effects account for shifts shared across genes, while the learned structured response WHc⊤WH_c^\top models gene-by-perturbation variation. Centering the gene loadings gives the structured response zero mean across genes for each perturbation, separating it from the broad perturbation effect (Section 2.2; Appendix A.2).

Direct validation and measures

We performed masked-response reconstruction with five random holdout splits within each cell line's response-memory fitting partition, using all 2,000 genes and recorded perturbations in each of the five cell lines. In each repetition, we randomly withhold 30% of the entries in each perturbation column of the gene-by-perturbation response matrix and refit the memory on the remaining 70% using the reported ranks and penalties.

  • Cell–gene tendency consistency (ρ\rho). For each gene, we compare αcg\alpha_{cg}, estimated from its visible responses, with its median expression effect over the held-out responses. Spearman correlation across genes, averaged over repetitions, measures the reproducibility of the estimated tendency rankings; ρ=1\rho=1 denotes identical rankings.
  • Held-out gene-by-perturbation recovery (R2R^2). We double-center the observed expression effects XcX_c and the fitted structured response WHc⊤WH_c^\top to remove additive gene and perturbation effects, then compare predicted and observed variation at the same held-out gene–perturbation entries. Squared errors are pooled across repetitions. The reference predicts no variation beyond the additive effects and has R2=0R^2=0; higher values indicate better recovery, with R2=1R^2=1 denoting exact recovery. Observed response means define only the evaluation target. Appendix F.5 provides the full definitions.

Table R2. Component-level validation of the factorized response memory.

Cell lineCell–gene tendency consistency (ρ\rho)Held-out gene-by-perturbation recovery (R2R^2)
C320.9010.643
HepG2/C3A0.8690.654
HOP620.8540.683
Hs 766T0.8090.561
PANC-10.8500.593

Perturbation dependence of the learned structured response. Before double-centering, we also measure the perturbation-independent component of WHc⊤WH_c^\top, formed by repeating each gene's mean structured response across perturbations. It accounts for 1.93%-4.22% of the total squared magnitude of WHc⊤WH_c^\top, with fractions averaged over repetitions within each cell line. This assesses the perturbation-independent component remaining in the learned structured response before evaluation removes additive effects.

Interpretation. Table R2 shows consistent cell–gene tendency rankings across disjoint response subsets. The structured response recovers gene-by-perturbation variation with less than half the held-out squared error of predicting no such variation. Together with the small perturbation-independent component measured before double-centering, these results support the intended approximate decomposition into stable cell–gene tendencies and perturbation-dependent structured responses.

We have added this validation, its protocol, and analysis to the revised manuscript (Appendix F.5, Table S14; Section 3.6).

AuthorsOpenReview ↗

Response to Reviewer rp1M [Part 2]

Q2. Biological relevance of selected evidence

Our evidence for biological relevance combines response recovery in a defined biological panel with an example of traceable mechanistic interpretation. MMCA-CT links query-specific response evidence to biological knowledge to formulate explicit, testable mechanistic hypotheses. Utility prioritizes memory and coexpression response estimates expected to improve prediction beyond observed responses. The LLM interprets them alongside retrieved drug-target and pathway relations and source-linked response records in the queried cellular context (Section 2.4).

  • Biologically relevant response recovery. In BRAF V600E-mutant C32, treated responses to the held-out BRAF inhibitors are excluded from fitting and retrieval. On the RAS-RAF-MAPK response panel, defined from Hallmark KRAS signaling sets and Reactome pathway participants, MMCA-CT correctly predicts 15 of 21 directions for dabrafenib and all 23 for vemurafenib. These results support response recovery within a biologically defined panel relevant to the perturbation and cell context (Section 3.5).
  • Traceable mechanistic interpretation. For AT7519-FOS in held-out C32, the evidence combines same-drug, same-gene responses from source cell lines with a coexpression estimate using source-cell responses of 17 neighbors identified from untreated C32 controls. The public KG supplies AT7519-CDK1 and an unsigned CDK1-FOS association. The correct up-regulation prediction cites the response packets and KG relation. Its explanation distinguishes directional support from response evidence and contextual support from the unsigned KG association, keeping observed responses, model estimates, and biological annotations identifiable (Appendix B.4).

Together, these analyses support biologically relevant response recovery and traceable mechanistic interpretation. Computational inference and targeted experimental validation play complementary roles in mechanism discovery, as illustrated by CellOracle (Kamimoto et al., Nature, 2023), which combines network-based perturbation predictions with experimental validation. Our contribution to mechanism discovery is an evidence-linked computational approach to formulating biologically grounded, query-specific hypotheses.

Q3. Hallmark comparisons and consistency of transfer benefits

The aggregate results and the two Appendix F.4 comparisons concern different evaluation scopes or comparators and need not yield the same ranking. The 128 entries concern Hallmark-restricted cross-perturbation DIR prediction. The 88/125 result concerns cross-cell-line prediction against drug-and-gene-matched voting, whereas Figure 3(a) uses perturbation-independent Prior vote.

  • Cross-perturbation comparison. The 128 entries are eligible Hallmark gene-set/cell-line pairs across five cell lines, selected using the query, gene, and drug coverage criteria in Appendix F.4. Each covers 10-75 distinct queried genes, potentially under multiple drugs. Hallmark sets overlap, so genes and queries contribute to multiple entries; the entries are neither independent experimental replicates nor an exhaustive, disjoint partition of the test set. Their lower DIR accuracy relative to Prior vote concerns these specified subsets. Section 3.2 instead evaluates overall and macro-per-gene performance across the complete test set. These different scopes need not produce the same ranking: the gene-set comparisons characterize performance within particular biological processes.
  • Cross-cell-line comparison. Section 3.5, Figure 3(a), reports higher DIR accuracy in 128 of 130 eligible entries against Prior vote, which pools each gene's responses across perturbations in the four source cell lines, independently of the query drug. The additional comparator in Appendix F.4 votes over source responses matched to both the query drug and gene. The 88 negative differences among 125 entries refer to this drug-specific comparator; its available-query subset also differs from that used for Prior vote.

The complete-test-set evaluation summarizes overall and macro-per-gene performance, while the Hallmark analyses characterize performance within specified biological processes. Both cross-cell-line comparisons evaluate DIR accuracy, but use different source responses for voting and different query subsets. They therefore describe performance relative to distinct response-based references, rather than opposite outcomes against the same baseline on identical queries.

We have clarified the evaluation scopes and comparators of these process-level analyses in the revised manuscript (Appendix F.4).

AuthorsOpenReview ↗

Response to Reviewer rp1M

Thank you again for your valuable feedback and for taking the time to review our work. We wanted to kindly follow up on our previous response to ensure we have addressed your main concerns. Please let us know if you have any further questions or need additional clarifications. We remain committed to improving the paper.

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

This paper proposes MMCA-CT, a framework for predicting whether genes respond to drug treatment and whether their expression increases or decreases. It uses a factorized memory to separate general gene-response tendencies from perturbation-specific effects, then learns to weight and select response estimates according to their expected predictive value. Selected estimates and observed responses are combined into a numerical prediction, while a frozen LLM uses the evidence and biological knowledge to provide explanations. Experiments across five cancer cell lines show improvements in cross-perturbation response detection and cross-cell-line direction prediction, with ablations supporting the contributions of the main components.

Critical strengths

The paper separates general gene-response tendencies from perturbation-specific effects and learns which response estimates are useful for each query. Gene-prior baselines and macro-per-gene evaluation help assess predictive value beyond general gene tendencies. Strong cross-cell-line direction prediction and component ablations further support the effectiveness of the framework.

Critical weaknesses

Some baselines are ranked using binary scores to calculate the sorting indicators, while in this paper, continuous probabilities are employed. The experiments do not clearly separate the contribution of the numerical prediction model from that of the LLM. It remains unclear whether evidence selection improves LLM reasoning or mainly improves the numerical predictor. A comparison with the same pipeline without the LLM would help clarify its role.

Questions for authors
  1. Could you also report prediction accuracy using the numerical predictor alone, without the LLM?
  2. The proposed method has similar performance as the prior vote baseline. The advantage of using the complex LLM based pipeline beyond simple voting needs to be justified
6 comments
AuthorsOpenReview ↗

Response to Reviewer BXqb [part 1]

Thank you for recognizing the value of our gene-prior comparisons and macro-per-gene evaluation. Below, we provide binary-score results, assess the LLM's predictive contribution, and clarify the improvements over Prior vote.

Q1. Score representation and the contribution of the LLM

MMCA-CT retains its overall AUROC/AUPRC advantage with binary scores (Table R3), and the complete agent achieves higher mean accuracy than both MMCA-CT without LLM and LLM + Base evidence (Table R1). These supplementary comparisons address score representation and the final LLM-based decision stage separately.

Classification and ranking metrics

  • Common decision metrics. Accuracy, precision, recall, and F1 use final binary DE/DIR decisions for every method under Appendix D. Both overall and macro-per-gene results are reported for every method.
  • Ranking-score inputs. Expression-based baselines output expression profiles rather than task probabilities. Following the task conversion in VCWorld (ICLR 2026, Section 4.2), our adapter compares these profiles with matched controls to obtain DE calls and uses effect signs for DIR. These binary outputs also serve as the expression-based baselines' AUROC/AUPRC scores. MMCA-CT uses fusion probabilities or confidence-derived positive-class scores; Prior vote uses training positive-label frequencies, as specified in Appendix D.
  • Reason for retaining continuous scores. They preserve distinctions among queries receiving the same binary decision and assess ordering beyond a single threshold. CORE (Sections 2.2 and D.3) uses continuous ranking scores, with macro-per-gene evaluation assessing discrimination across perturbations within each gene.
  • Direct supplementary comparison. Table R3 uses MMCA-CT's final binary predictions as AUROC/AUPRC scores, with query and gene eligibility unchanged, to address the difference in score representation.

Table R3. Binary-score AUROC and AUPRC for MMCA-CT (QWEN3.5-9B).

SettingTaskOverall AUROCOverall AUPRCMacro-per-gene AUROCMacro-per-gene AUPRC
Cross-perturbationDE0.951 ± 0.0110.947 ± 0.0170.568 ± 0.0010.722 ± 0.049
DIR0.849 ± 0.0160.822 ± 0.0720.560 ± 0.0140.559 ± 0.080
Cross-cell-lineDE0.765 ± 0.0400.787 ± 0.0340.574 ± 0.0050.639 ± 0.041
DIR0.862 ± 0.0480.849 ± 0.0550.861 ± 0.0360.845 ± 0.065

Overall ranking metrics pool answered queries within each cell line. Macro-per-gene ranking metrics are computed across perturbations for each eligible gene and averaged equally over genes. Values are unweighted means and sample standard deviations across five cell lines in each setting.

With binary scores, overall AUROC and AUPRC remain above the expression-based baselines for both tasks and settings (manuscript Tables 1 and 2). Cross-cell-line DIR macro-per-gene AUROC/AUPRC are 0.861/0.845, versus CRISP's 0.763/0.772. These gains persist when MMCA-CT is evaluated with binary scores.

We have added this evaluation, its protocol, and analysis to the revised manuscript (Appendix F.2, Table S10; Sections 3.2 and 3.3).

AuthorsOpenReview ↗

Response to Reviewer BXqb [part 2]

Response fusion and the LLM

Utility-guided evidence selection directly targets reductions in the fused prediction's log loss. The numerical predictor corresponds to the existing response-fusion component, which combines transfer-weighted route corrections with the gene-prior reference into a task probability, integrating potentially conflicting directions rather than selecting the largest route probability. The comparison below separately evaluates the additional predictive contribution of the final LLM-based decision stage.

Table R1 compares three configurations using frozen QWEN3.5-9B.

  • LLM + Base evidence predicts from retrieved KG facts and training response cases.
  • MMCA-CT without LLM retains factorized response memory, conditional transfer estimation, utility-guided evidence selection, and response fusion, thresholding the fused task probability at 0.5.
  • MMCA-CT is the complete method. Supported fused predictions determine the class; otherwise, the LLM assesses the available evidence (Section 2.4; Appendix B.2).

Table R1. Overall and macro-per-gene accuracy of three configurations with QWEN3.5-9B.

SettingTaskConfigurationOverall AccuracyMacro-per-gene Accuracy
Cross-perturbationDELLM + Base evidence0.686 ± 0.0100.608 ± 0.015
MMCA-CT without LLM0.921 ± 0.0110.788 ± 0.013
MMCA-CT0.933 ± 0.0110.809 ± 0.013
DIRLLM + Base evidence0.537 ± 0.0180.523 ± 0.017
MMCA-CT without LLM0.758 ± 0.0260.718 ± 0.018
MMCA-CT0.784 ± 0.0260.739 ± 0.010
Cross-cell-lineDELLM + Base evidence0.558 ± 0.0130.483 ± 0.012
MMCA-CT without LLM0.709 ± 0.0410.612 ± 0.042
MMCA-CT0.731 ± 0.0400.633 ± 0.039
DIRLLM + Base evidence0.562 ± 0.0140.518 ± 0.012
MMCA-CT without LLM0.726 ± 0.0500.711 ± 0.045
MMCA-CT0.792 ± 0.0480.787 ± 0.036

Overall accuracy covers all test queries within each cell line; macro-per-gene accuracy averages eligible genes equally. Values are unweighted means and sample standard deviations across five cell lines, following Appendix D.

Results. MMCA-CT without LLM improves both mean accuracies over LLM + Base evidence in every setting-task combination. Adding the final LLM-based decision stage further improves overall accuracy by 1.2-6.6 percentage points and macro-per-gene accuracy by 2.1-7.6 percentage points. For cross-cell-line DIR, these increase from 0.726 to 0.792 and from 0.711 to 0.787, respectively. The comparison supports complementary predictive contributions from the learned response components and the final LLM-based decision stage.

The LLM also provides biological rationales linked to response records and KG relations, with confidence and evidence citations (Appendix B.4). The accuracy comparison quantifies the predictive contribution of the final LLM-based decision stage; the recorded examples show how retrieved evidence supports the accompanying interpretations.

We have added this comparison, its protocol, and analysis to the revised manuscript (Appendix F.5, Table S13; Section 3.6).

Q2. Prediction accuracy without the LLM

The requested numerical-only results are reported as MMCA-CT without LLM in Table R1 above. This configuration retains the learned response components and obtains its binary decision from the fused probability at threshold 0.5, without invoking the LLM.

  • Cross-perturbation. Mean overall DE/DIR accuracies are 0.921/0.758; macro-per-gene accuracies are 0.788/0.718.
  • Cross-cell-line. Mean overall DE/DIR accuracies are 0.709/0.726; macro-per-gene accuracies are 0.612/0.711.

The complete agent improves both mean accuracy measures in all four setting-task combinations, supporting an additional predictive contribution from the final LLM-based decision stage. Table R1 includes standard deviations and the corresponding complete-agent results. These results are included in the revised manuscript (Appendix F.5, Table S13).

AuthorsOpenReview ↗

Response to Reviewer BXqb [part 3]

Q3. Added value over Prior vote

Prior vote captures general gene-response tendencies; MMCA-CT additionally distinguishes responses to different perturbations for the same gene while improving overall prediction. The two aggregation schemes assess these complementary aspects.

  • Why Prior vote performs well overall. It assigns each cell-gene pair its training positive-label frequency, independently of the test perturbation. Overall performance can therefore benefit from distinguishing broadly responsive genes from rarely responsive genes and predicting each gene's predominant response.
  • What macro-per-gene evaluation reveals. Metrics are computed across perturbations within each eligible gene and averaged equally over genes. Prior vote's score is constant within a cell-gene pair, so all perturbations are tied and its macro-per-gene AUROC is 0.500. CORE (Sections 2.2 and 3.2-3.3) motivates this distinction between aggregate performance and within-gene perturbation discrimination.
  • Quantitative evidence. With QWEN3.5-9B, manuscript Table 1 reports DE macro-per-gene AUROC/AUPRC of 0.577/0.739 for MMCA-CT versus 0.500/0.628 for Prior vote; DIR values are 0.562/0.568 versus 0.500/0.476. Overall accuracy also improves from 0.903 to 0.933 for DE and from 0.747 to 0.784 for DIR.

These results support predictive information beyond general gene-response tendencies, particularly for distinguishing which perturbations affect a given gene and how its response direction varies. In Table R1, the complete agent also exceeds MMCA-CT without LLM in both mean accuracy measures across all four setting-task combinations, supporting the final LLM-based decision stage's predictive contribution. Biological rationales with evidence citations make the supporting response records and KG context inspectable alongside the decision (Appendix B.4).

Reviewer_ac1NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce an agent for perturbation prediction. The agent uses a factorized response memory component to separate perturbation-specific effects from stable genes' behavior. It learns to weight predictions based on evidence: expected gain, reliability, and incremental predictive value. The proposed approach is evaluated on differential expression (DE) and direction (DIR) for one dataset, GeneTAK. Experiments show that the proposed approach outperforms baselines, and ablations demonstrate the improvement of each component.

Critical strengths

Evaluated the proposed approach (MMCA-CT) against STATE and CRISP, which are state-of-the-art perturbation prediction models, and several other older baselines.

Evaluated on both DE and DIR, which provide complementary insights about the model, and performed biological analyses on DIR to better understand model strengths/limitations.

Benchmarked on both cross-perturbation and cross-cell-line to better understand generalizability.

Evaluated MMCA-CT with three backbones: Llama3.1-8B, QWEN3.5-9B, and QWEN3.5-27B.

Conducted ablation analyses in which components of MMCA-CT are successively added to understand which component(s) yielded the largest gains.

Performance is measured via standard metrics (Accuracy, AUROC, AUPRC) for overall and macro-per-gene, with standard deviation across cell lines.

Critical weaknesses

Regarding retrieval from public KG, which contain information about baseline expression and co-expression: To what extent are the perturbation effects already observed in the KG? How do you ensure that there is no leakage?

There are lots of components beyond the ones ablated in Table 3. For example, route probabilities, transfer reliability/weights, memory and gene co-expression routing, and transfer fusion. Could you clarify the contributions of these components?

Standard metrics for perturbation response prediction are not included: RMSE, MAE, R^2, and MMD. Could you please include them?

Questions for authors

See weaknesses.

6 comments
AuthorsOpenReview ↗

Response to Reviewer ac1N [Part 1]

Thank you for your careful review and for recognizing our evaluation across perturbations and cell lines. Below, we clarify KG sources and component contributions, explain the evaluation metrics, and provide additional validation.

Q1. KG information and data leakage

The two KGs are constructed without held-out drug-treated responses and contain no corresponding DE/DIR labels. The public KG supplies pre-existing biological annotations; the cell-line context KG describes the untreated control state. Observed perturbation responses are handled separately under the training and test splits.

  • Public KG. Drug-target, gene, protein, and pathway relations can describe known molecular actions of a test drug. These annotations are distinct from the measured response of a queried gene in a particular cell line. They support retrieval, memory readouts, and mechanistic context for LLM reasoning (Appendix A.3).
  • Cell-line context KG. Baseline expression, detection rates, and CS-CORE coexpression estimates are computed exclusively from DMSO controls. The edges describe control-state associations; drug-treated profiles and DE/DIR labels are not used to estimate them (Appendix A.4).
  • Cross-perturbation separation. Canonical active-moiety groups are disjoint between training and test. Held-out groups' responses are excluded from response-memory fitting and response retrieval.
  • Cross-cell-line separation. All response-dependent components are refitted within each outer fold using the four source cell lines. Coexpression routing also uses neighboring genes' drug responses from source lines. Target-cell inputs are limited to untreated controls, assay metadata, and pre-existing annotations; its drug-treated responses and labels are reserved for evaluation (Appendices C.2-C.3).
AuthorsOpenReview ↗

Response to Reviewer ac1N [Part 2]

Q2. Roles and contributions of the listed components

The listed items are response routes, intermediate outputs, and computations within the three modules, rather than each constituting an additional standalone methodological contribution. Table 3 evaluates cumulative module-level contributions in cross-perturbation prediction.

  • Factorized response memory (A) and route probabilities. Three memory routes use molecular structure, targets and action types, or mechanism descriptions to estimate unseen-perturbation responses from the factorized response memory. For held-out cell lines, a memory-transfer route combines latent cell–perturbation states and broad perturbation effects from source lines. Task-specific logistic regression maps each route's features to DE or DIR positive-class probabilities for transfer estimation and fusion (Section 2.2).
  • Conditional transfer estimation and fusion (B). Gain estimates the expected log-loss reduction relative to the gene-prior reference; transfer reliability estimates the probability that this reduction is positive. Query-specific transfer weights combine these estimates with adjustments for response support and extrapolation to scale each route's logit correction. Fusion combines the weighted corrections with the gene-prior reference logit using learned coefficients shared across queries (Sections 2.3–2.4).
  • Utility-guided evidence selection (C). Utility is the reduction in fused log loss from adding a memory or coexpression route to the observed-response route set, evaluated with fusion parameters and calibration fixed. Predicted utility guides route selection under the evidence budget (Section 2.4).
  • Coexpression routing. The cell-line context KG supplies positively coexpressed neighbors and coexpression edge weights. Their responses to the query drug in source cell lines are aggregated to estimate the target gene's response. This cross-cell-line response route uses the same probability-estimation, transfer-weighting, fusion, and selection framework; it is distinct from retrieving KG facts for LLM reasoning (Appendix A.4).

Empirical contributions. Under shared training partitions, backbone, and base evidence, Table 3 reports overall DE accuracy of 0.686, 0.904, 0.921, and 0.933 from Base evidence through A, B, and C; DIR accuracy is 0.537, 0.747, 0.765, and 0.784. Appendix F.3 directly evaluates whether learned reliability identifies positive-gain route predictions. For DE, its positive-gain discrimination AUROC is 0.86 versus 0.54 for retrieval/similarity scores in cross-perturbation evaluation, and 0.84 versus 0.47 in cross-cell-line evaluation. These results support module-level gains and the value of learned reliability beyond biological similarity.

Additional configuration comparison. Table R1 compares three configurations using frozen QWEN3.5-9B.

  • LLM + Base evidence predicts from retrieved KG facts and training response cases.
  • MMCA-CT without LLM retains factorized response memory, conditional transfer estimation, utility-guided evidence selection, and response fusion, thresholding the fused task probability at 0.5.
  • MMCA-CT is the complete method, including the final LLM-based decision stage (Section 2.4; Appendix B.2).

Table R1. Overall and macro-per-gene accuracy of three configurations with QWEN3.5-9B.

SettingTaskConfigurationOverall AccuracyMacro-per-gene Accuracy
Cross-perturbationDELLM + Base evidence0.686 ± 0.0100.608 ± 0.015
MMCA-CT without LLM0.921 ± 0.0110.788 ± 0.013
MMCA-CT0.933 ± 0.0110.809 ± 0.013
DIRLLM + Base evidence0.537 ± 0.0180.523 ± 0.017
MMCA-CT without LLM0.758 ± 0.0260.718 ± 0.018
MMCA-CT0.784 ± 0.0260.739 ± 0.010
Cross-cell-lineDELLM + Base evidence0.558 ± 0.0130.483 ± 0.012
MMCA-CT without LLM0.709 ± 0.0410.612 ± 0.042
MMCA-CT0.731 ± 0.0400.633 ± 0.039
DIRLLM + Base evidence0.562 ± 0.0140.518 ± 0.012
MMCA-CT without LLM0.726 ± 0.0500.711 ± 0.045
MMCA-CT0.792 ± 0.0480.787 ± 0.036

Overall accuracy covers all test queries within each cell line; macro-per-gene accuracy averages eligible genes equally. Values are unweighted means and sample standard deviations across five cell lines, following Appendix D.

MMCA-CT without LLM improves both mean accuracies over LLM + Base evidence in every setting-task combination, supporting the joint contribution of the learned response components. The complete agent further improves both accuracies in all four combinations.

We have added this comparison, its protocol, and analysis to the revised manuscript (Appendix F.5, Table S13; Section 3.6).

AuthorsOpenReview ↗

Response to Reviewer ac1N [Part 3]

Q3. RMSE, MAE, R2R^2, and MMD

We evaluate the agent's DE/DIR decisions with classification and ranking metrics and additionally report R2R^2 for the memory's continuous reconstruction. For expression-based perturbation models, RMSE, MAE, and R2R^2 assess agreement with expression values or changes, while MMD compares expression distributions. MMCA-CT predicts whether a gene is differentially expressed and, for DE-positive queries, its direction of change (Section 2.1), rather than generating complete expression profiles or cell populations.

This task formulation follows VCWorld (ICLR 2026, Section 4.2), which evaluates perturbation-response decisions using classification metrics and converts expression-based baselines to these tasks, and SUMMER (ICLR 2025, Sections 3 and 5.2), which distinguishes expression regression from discrete biological outcomes and evaluates DE/direction using per-gene AUROC.

  • Overall evaluation summarizes performance across perturbation-gene queries within each cell line. Accuracy, precision, recall, and F1 assess the correctness of final decisions and the balance between detecting positive responses and avoiding false positives; AUROC/AUPRC assess discrimination between positive and negative responses.
  • Macro-per-gene evaluation measures performance across perturbations within each eligible gene, then averages genes equally. CORE (Section 2.2) motivates this distinction because strong aggregate performance can reflect general gene-response tendencies without demonstrating discrimination among perturbations for the same gene. We report both evaluation levels for DE and DIR in both generalization settings.

Component-level validation. The factorized response memory reconstructs continuous expression effects as intermediate features for DE/DIR probability estimation. We evaluate it using masked-response reconstruction with five random holdout splits within the response-memory fitting partition, covering all 2,000 genes and recorded perturbations in each of the five cell lines. In each repetition, we randomly withhold 30% of the entries in each perturbation column of the gene-by-perturbation response matrix and refit the memory on the remaining 70%.

  • Cell–gene tendency consistency (ρ\rho). For each gene, we estimate its median expression effect separately from visible and held-out responses. We pair the estimates by gene, compute Spearman correlation across genes, and average the correlation over repetitions; higher values indicate greater rank consistency.
  • Held-out gene-by-perturbation recovery (R2R^2). We double-center the observed response matrix and the fitted structured response, subtracting row and column means and adding the grand mean, to remove additive gene and perturbation effects. Predicted and observed variation are compared at the same held-out gene–perturbation entries, pooling squared errors across repetitions. Predicting no such variation gives R2=0R^2=0; exact recovery gives R2=1R^2=1, with higher values indicating more accurate recovery.

Table R2. Component-level validation of the factorized response memory.

Cell lineCell–gene tendency consistency (ρ\rho)Held-out gene-by-perturbation recovery (R2R^2)
C320.9010.643
HepG2/C3A0.8690.654
HOP620.8540.683
Hs 766T0.8090.561
PANC-10.8500.593

The results show consistent cell–gene tendency rankings across disjoint response subsets. The structured response recovers gene-by-perturbation variation with less than half the held-out squared error of predicting no such variation. This component-level validation complements the complete agent's DE/DIR evaluation.

We have added this validation, its protocol, and analysis to the revised manuscript (Appendix F.5, Table S14; Section 3.6).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.