acceptodds
Under review as a conference paper at ICLR 2027

Looped Virtual Cell

Abstract

Virtual cell perturbation models often collapse a dynamic cellular response into a static endpoint expression profile. We introduce Looped Virtual Cell, instantiated as Looped X-Cell, which turns endpoint prediction into a learned recurrent rollout over genes. Rather than predict all genes in one pass, Looped X-Cell repeatedly ranks unresolved genes, commits selected expression values, carries the partial endpoint state forward, and uses lookahead supervision so stages preview genes that will be committed shortly afterward. A scorer-latent outer loop can then revisit the staged rollout through recurrent latent state without directly feeding predicted expression values back into the input. Across four Replogle-Nadig cellular contexts, Looped X-Cell improves endpoint prediction over a fine-tuned X-Cell model, scGPT, and a linear baseline under matched held-out perturbation splits. The learned ranks also expose biological structure that is never used as supervision: they recover Reactome pathways across Th17, Replogle-Nadig, and Marson, recover CORUM protein-complex relationships in rank-feature perturbation clusters, and reveal GO-enriched RNA/protein rank gaps in a multimodal Th17 model. Component ablations identify staged global ranking and lookahead supervision as the main drivers of HepG2 endpoint gains, while compute sweeps show that inner-loop resolution improves prediction more consistently than outer-loop recurrence. Together, these results show that learned gene rollout is a structured compute axis for virtual-cell endpoint models and an auditable internal trace for perturbation biology.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Cell, single-cell perturbation prediction, recurrent computation, iterative inference, interpretable representation learning

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_2RVaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper extends X-Cell with a staged procedure for predicting perturbation-induced expression profiles. Each inner stage ranks unresolved genes, commits selected predictions and applies lookahead supervision; optional outer passes reuse scorer-latent state. The method improves several endpoint metrics over its parent model across four Replogle-Nadig contexts and five training seeds. Biological analyses associate the learned ranks with pathways and protein complexes and compare RNA and antibody-derived tag (ADT) ranks. Ablations attribute the strongest gains to global ranking and lookahead, while a limited compute sweep favors additional inner stages over outer recurrence. The learned order describes model computation rather than measured biological time.

Critical strengths
  1. The consistent improvement over the parent X-Cell model is the strongest result. With shared pretrained weights and inherited objectives, Looped X-Cell reduces mean MAE by approximately 7.6–10.4% across the four contexts, with results reported over five training seeds. Pearson Δ, DE direction and DE LFC also improve in each context. This provides useful evidence for staged prediction in the evaluated setting on a useful, broad sweep of metrics.

  2. The model exposes internal ranks and stages that can be inspected and tested. The biological probes and recurrence controls are useful attempts to characterize these representations. In particular, the HepG2 experiment in Table A3 shows that scorer-latent recurrence preserves accuracy over additional passes more reliably than direct value feedback. This is a useful engineering observation.

Critical weaknesses
  1. The benchmark and ablations do not establish the predictive contribution of the complete architecture for transcriptomic perturbation prediction.

The improvement over X-Cell is promising, but the benchmark needs stronger simple controls and contemporary comparisons. Table 1 includes a linear baseline whose specification is insufficient to judge its strength. Recent benchmarking work [1–3] motivates including control/no-change and training-perturbation-mean predictions, alongside comparisons with STATE, STACK, the public-data version of TxPert, and LPM in the settings they support [6–9]. Licensing restrictions may legitimately prevent some comparisons. Have necessary permissions be requested, and what specific barriers remain? More broadly, this reviewer encourages developers including Xaira and Arc to enable reproducible head-to-head evaluations and share evaluation code, splits, predictions, and results. TxPert’s public-data implementation remains subject to non-commercial terms, whereas LPM’s Apache-2.0 implementation provides Replogle training configurations. For each omitted model, please identify the restriction or task mismatch and limit comparative claims accordingly.

Metric calibration is also needed to determine whether the improvements reflect perturbation-specific prediction gains rather than changes shared across perturbations. Systema [4] and Miller, Mejia et al. [5] motivate this distinction and the use of informative positive and negative controls. The reported centroid accuracy deserves credit, but I would also like calibration using an appropriate split-half or technical-replicate positive reference, uninformative negative controls, and an explicit definition of Pearson Δ, including how predicted and observed expression changes are calculated and which control expression profile is subtracted (the one in Appendix A6 erroneously copies the MAE definition). The generalization protocol should distinguish unseen perturbation identities from unseen identity–context pairs, establish whether test identities are excluded across all fine-tuning contexts, and quantify overlap with the pretraining data. The paper should state exactly which data were used for training, validation, and testing; which pretrained models they started from and what those models had already seen; and how much training and tuning each method received. This is needed to judge whether the comparisons are fair.

The ablations raise a more direct concern about the full architecture. In Table 2, global ranking plus lookahead has better reported means than the complete model on all six metrics, including MAE of 0.0611 versus 0.0618. The three-seed versus five-seed comparison does not establish superiority, but it provides no positive evidence that the additional components improve endpoint prediction. Please compare the simpler and complete variants under matched seeds and budgets, including fixed/random staged ordering, and establish the executor’s contribution or present it as an implementation choice. Likewise, the limited compute sweep and additional-pass stability experiments do not establish a general compute-scaling benefit; claims should distinguish stage-count sensitivity, recurrence stability, and improved accuracy at matched compute.

  1. The biological analyses do not establish the added value of the learned ordering, and the framing exceeds the demonstrated capability.

Recovering pathway or complex annotations without using them as explicit labels does not establish independence from biological knowledge, especially because the model uses external priors. CORUM rank permutations show sensitivity to gene correspondence, but do not establish an advantage over predicted expression/deltas or inherited/prior-only representations. Table A4 provides control-expression comparisons for selected Reactome terms only. The paper should check whether the learned gene ordering reveals more biological structure than the predicted expression levels, predicted expression changes, or information already present in the original model from the prior knowledge it was fed. They should use the same tests for each comparison and state what external biological information each model was given. Otherwise, the results support annotation-associated representations rather than independent biological discovery.

The RNA/ADT analysis also warrants narrower interpretation. The GO analysis should compare the selected genes against all genes whose proteins were measured and matched to RNA. Instead, it uses a broader background, which may make features of the antibody panel appear to be findings about the learned ordering. The Gene Ontology Consortium recommends using only genes that could have entered the selected list as the background (https://geneontology.org/docs/go-enrichment-analysis/). The paper acknowledges this limitation in B.3, but the main text’s claim of greater biological coherence remains stronger than the analysis supports. The protein-earlier and protein-later sets contain 6 and 56 genes, respectively, so differences in enrichment may partly reflect list size and statistical power. The weak abundance–rank-gap association does not exclude abundance, normalization, or initialization effects. The permutation statistic and dependence assumptions also require clarification.

Finally, the demonstrated endpoint prediction capability covers only a limited subset of the virtual cell vision described by Bunne et al. [10]. The paper should present its contribution as progress toward that vision. Applying the virtual cell label to a narrowly evaluated perturbation prediction model and a comprehensive model of cellular behavior obscures differences in scope and makes progress harder to assess. Similarly, the explicit acknowledgement that computational ranks establish neither biological time nor causal response paths should consistently govern the motivation and interpretation of the results throughout the paper. Although the paper states this in the limitations, the motivation and interpretation elsewhere sometimes imply a progression of biological processes that the analysis does not establish.

To make these results reproducible, please release the training and evaluation code, specify the data and pretrained models used in each experiment, and describe how the biological analyses were performed, including the genes used as the enrichment background.

References

[1] Ahlmann-Eltze, Huber, and Anders. Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines. Nature Methods (2025). https://doi.org/10.1038/s41592-025-02772-6

[2] Wenteler et al. PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction. ICML (2025). https://proceedings.mlr.press/v267/wenteler25a.html

[3] Wu et al. PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis. NeurIPS (2025). https://proceedings.neurips.cc/paper_files/paper/2025/hash/8aee537279a66ced96319dfca3c00002-Abstract-Datasets_and_Benchmarks_Track.html

[4] Viñas Torné et al. Systema: a framework for evaluating genetic perturbation response prediction beyond systematic variation. Nature Biotechnology (2025). https://doi.org/10.1038/s41587-025-02777-8. Implementation: https://github.com/mlbio-epfl/systema

[5] Miller, Mejia, et al. Deep Learning-Based Genetic Perturbation Models Do Outperform Uninformative Baselines on Well-Calibrated Metrics. bioRxiv (2025). https://doi.org/10.1101/2025.10.20.683304. Implementation: https://github.com/shiftbioscience/Perturbation-Models-Outperform-Baselines

[6] Wenkel et al. TxPert: using multiple knowledge graphs for prediction of transcriptomic perturbation effects. Nature Biotechnology (2026). https://doi.org/10.1038/s41587-026-03113-4. Public-data implementation: https://github.com/valence-labs/TxPert

[7] Adduri et al. Predicting cellular responses to perturbation across diverse contexts with State. Cell (2026). https://doi.org/10.1016/j.cell.2026.07.052. Implementation: https://github.com/ArcInstitute/state

[8] Dong et al. Stack: In-context learning of single-cell biology. bioRxiv (2026). https://doi.org/10.64898/2026.01.09.698608. Implementation: https://github.com/ArcInstitute/stack

[9] Miladinovic et al. In silico biological discovery with large perturbation models. Nature Computational Science 5, 1029–1040 (2025). https://doi.org/10.1038/s43588-025-00870-1

[10] Bunne et al. How to Build the Virtual Cell with Artificial Intelligence: Priorities and Opportunities. Cell (2024). Public author preprint: https://arxiv.org/abs/2409.11654

Questions for authors
  1. Please provide the missing contemporary and strong simple baseline comparisons, together with metric calibration against positive/negative controls? For any omitted model, what exact version, restriction or task mismatch prevents a valid comparison? These results would determine whether the reported gains represent a meaningful advance beyond just the parent model.

  2. Are each context's test perturbation identities excluded from every fine-tuning context, and what overlap remains with the actual pretraining checkpoint? Please provide the split/checkpoint information needed to distinguish identity holdout from identity–context-pair holdout.

  3. What evidence favors the full architecture over global ranking plus lookahead under matched seeds and budgets, and how does it compare with fixed/random staged schedules? Please distinguish stability under additional inference passes from improvements in prediction accuracy.

  4. Do learned ranks improve biological recovery over predicted expression/deltas and inherited/prior-only representations under the same analysis? If those comparisons are unavailable, which narrower biological conclusions does the paper intend to retain?

  5. How robust are the RNA/ADT enrichment results to using the eligible measured-and-matched protein panel as background and accounting for the unequal gene-set sizes? Please also specify the statistic used in the rank-gap permutation test and how the shuffling preserves the relevant dependence structure.

5 comments

Thank you very much for your exceptionally detailed and thoughtful review! Two of your suggestions were especially helpful. We were able to run LPM on all four contexts, and using the measured panel as the GO background changed our interpretation of the RNA/protein result.

Q1. Baselines and metric calibration

Following your suggestion, we ran a separate pseudobulk comparison with LPM and report the results here. For each perturbation, we averaged Looped X-Cell’s predictions across cells so that both models could be evaluated at the same resolution. Across all four Replogle–Nadig contexts, LPM achieved lower mean MAE, while Looped X-Cell achieved lower mean weighted MSE (wMSE) and higher mean Pearson Δ\Delta and centroid accuracy.

ContextModelMAE ↓\downarrowwMSE ↓\downarrowPearson Δ\Delta ↑\uparrowCentroid acc. (L2L_2) ↑\uparrow
HepG2LPM0.0589±0.0003\mathbf{0.0589 \pm 0.0003}0.03363±0.000660.03363 \pm 0.000660.6646±0.00300.6646 \pm 0.00300.7859±0.01000.7859 \pm 0.0100
Looped X-Cell0.0618±0.00050.0618 \pm 0.00050.02154±0.00069\mathbf{0.02154 \pm 0.00069}0.6838±0.0033\mathbf{0.6838 \pm 0.0033}0.8512±0.0150\mathbf{0.8512 \pm 0.0150}
JurkatLPM0.0712±0.0002\mathbf{0.0712 \pm 0.0002}0.05859±0.001480.05859 \pm 0.001480.6490±0.00170.6490 \pm 0.00170.7448±0.00770.7448 \pm 0.0077
Looped X-Cell0.0772±0.00190.0772 \pm 0.00190.04178±0.00168\mathbf{0.04178 \pm 0.00168}0.6553±0.0047\mathbf{0.6553 \pm 0.0047}0.8021±0.0187\mathbf{0.8021 \pm 0.0187}
K562LPM0.0578±0.0002\mathbf{0.0578 \pm 0.0002}0.02503±0.000350.02503 \pm 0.000350.6721±0.00300.6721 \pm 0.00300.7687±0.00450.7687 \pm 0.0045
Looped X-Cell0.0631±0.00320.0631 \pm 0.00320.01982±0.00151\mathbf{0.01982 \pm 0.00151}0.6771±0.0069\mathbf{0.6771 \pm 0.0069}0.8347±0.0333\mathbf{0.8347 \pm 0.0333}
RPE1LPM0.0701±0.0003\mathbf{0.0701 \pm 0.0003}0.05794±0.001510.05794 \pm 0.001510.7452±0.00230.7452 \pm 0.00230.7588±0.00570.7588 \pm 0.0057
Looped X-Cell0.0706±0.00220.0706 \pm 0.00220.04286±0.00284\mathbf{0.04286 \pm 0.00284}0.7656±0.0064\mathbf{0.7656 \pm 0.0064}0.8291±0.0176\mathbf{0.8291 \pm 0.0176}

Values are mean ±\pm sample s.d. across five seeds (n=5n=5). The LPM learning rate was 0.005. Bold indicates the better mean for each context and metric. Following Miller, Mejia et al. [5], we included wMSE because it gives greater weight to genes that distinguish each perturbation from the other perturbations in the same context. Across all four contexts, LPM achieved lower MAE, while Looped X-Cell achieved 21–36% lower wMSE, indicating smaller weighted prediction errors on these perturbation-specific expression profiles.

In the manuscript, we added Control and scDFM to Table 1. Control returns the same reference-control cells for every perturbation. Its centroid accuracy was close to chance in all four contexts (0.5013–0.5018), even though its MAE was lower than Linear’s in three. Looped X-Cell achieved lower mean MAE and higher mean Pearson Δ\Delta than scDFM in all four contexts. For Pearson Δ\Delta, we subtract the same target-context control mean from the predicted and observed pseudobulk profiles, then correlate the two changes across genes (Appendix B.4).

Appendix B.3 now describes Linear, a mean-based baseline that transfers each perturbation’s average effect from the source contexts to the target context. Table A2 summarizes the training and inference settings for the models in Table 1 and the public release terms for STATE, STACK, and TxPert.

[5] Miller, Mejia et al. Deep Learning-Based Genetic Perturbation Models Do Outperform Uninformative Baselines on Well-Calibrated Metrics. https://doi.org/10.1101/2025.10.20.683304

Q2. Which perturbations were held out?

Table A1 now lists the fine-tuning and evaluation data for each setting. In Replogle–Nadig, every test perturbation was held out in the evaluated context and appeared during fine-tuning in at least one other context. Table 1 therefore measures transfer to held-out perturbation–context pairs. Appendix B.2 identifies the processed data release and the notebook defining the HepG2 split. Appendix B.5 describes the checkpoint shared by X-Cell and Looped X-Cell.

Q3. What does the full architecture add?

We agree that the simpler model is a strong comparison. Table 2 now uses the same three HepG2 seeds throughout. With four stages and lookahead disabled, all four orders reduced mean MAE relative to fine-tuned X-Cell (0.0668). The means were 0.0632 for learned order, 0.0648 for random order, 0.0645 for control mean, and 0.0627 for control variance. Learned order gave the highest mean Pearson Δ\Delta and Overlap@N among the four. Adding lookahead to learned global ranking reduced MAE to 0.0611, compared with 0.0616 for the full model. We therefore attribute the main HepG2 MAE gains to staged prediction and lookahead.

Local ranking gives us a within-stage order we can inspect. In Marson, it recovered 34 CORUM modules and 38 same-complex target pairs missed by the other two rollout views (Figure A5). The scorer-latent model retained an MAE of 0.0609 to four decimal places at T=2,4,8,16T=2,4,8,16 outer passes (Table A5). This shows stability across additional passes. Figure 4 varies the stage and pass counts without matching total compute.

Q4. What do the ranks recover beyond expression and the supplied priors?

In Marson, we applied the same clustering and CORUM scoring procedure to each representation. Stage assignments, global ranks, and local ranks recovered 135, 167, and 133 modules, respectively, compared with 82 for predicted expression and 71 for measured expression. Combining the three rollout views recovered 273 modules, and adding the two expression views brought the total to 413 (Figure A6).

We also repeated CORUM scoring on HepG2 after removing the five external prior tokens at inference and retaining the learned perturbation-gene embedding. Recovery decreased, although each rollout view still recovered about three times as many modules and five to six times as many same-complex pairs as its shuffled control (Table A8). The checkpoint was trained with the priors, so this experiment measures the effect of supplying them at inference. For Reactome, the learned stages recovered 17 of 19 curated context–pathway pairs, compared with one for control mean and eight for control variance. Ten were unique to the learned stages (Table A6). Our conclusion is that the learned ranks recover known annotations missed by the expression baselines tested here. A comparison with expression deltas or the pretrained backbone alone remains open.

Reviewer_3qHAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This article addresses the problem of predicting expression profiles after genetic perturbations, and proposes a learnable ranking and staged prediction method based on X-Cell, called Looped X-Cell, which improves prediction accuracy in four cellular contexts and finds that the learned rankings have certain biological relevance.

Critical strengths

1.Clear research motivation. Treating the order of genetic perturbations as a learnable computational strategy—and leveraging the context from staged predictions to improve the prediction of final expression states—offers significant methodological value. 2.Consistent performance improvements. Across four Replogle–Nadig cellular contexts and five different training random seeds, the model demonstrates improvements in multiple predictive metrics compared to "X-Cell" (which utilizes the same pre-trained backbone).

Critical weaknesses

1.The ablation study fails to isolate the sources of gain for the core methodology. Table 2 does not decouple the contributions of staged computation from those of learning-to-rank; it lacks a baseline comparison—using either fixed or random ranking—within the same staged framework. Consequently, it is impossible to determine whether performance improvements stem from the staged computation or the learning-to-rank component. 2.Furthermore, the necessity of the additional modules remains unproven. In the HepG2 dataset, the configuration comprising "global ranking + staged prediction + lookahead" yields an average MAE of 0.0611, which is lower (better) than that of the full model (0.0618). This indicates that the added modules do not provide any clear performance benefit.

Questions for authors
  1. Holding the staged framework, supervision strategy, and training budget constant, does "learned ordering" consistently outperform fixed or random ordering? Please provide a direct comparison of training and evaluation results for these respective strategies.

  2. What specific, verifiable benefits does the full model offer over the simpler configuration? How do the current results justify the retention of these additional modules and the associated increase in complexity?

2 comments

Thank you very much for your thoughtful suggestions! Your insightful comments prompted a substantial revision of the HepG2 ablations in Section 3.2 and Table 2. We now compare learned ordering with random ordering and two fixed orders (control-cell mean expression and variance) within the same four-stage rollout. In Section 3.3, we further compare Reactome enrichment in the learned stages with enrichment in gene sets defined by fixed rankings based on the same control-cell mean expression or variance. Of the 19 curated context-relevant pathways in Table A6, the learned stages recovered 17, compared with one for the control-mean ranking and eight for the control-variance ranking. Ten pairs were recovered only by the learned stages.

Q1. Does learned ordering consistently outperform fixed or random ordering?

On HepG2, learned ordering gave the highest mean Pearson Δ and Overlap@N across the four orders, in the staged framework. Ordering by control variance gave the lowest mean MAE with a slight margin. We used four stages and the same three training seeds in each case, with the auxiliary lookahead loss disabled. The table below reports held-out means, with standard deviations in Table 2.

Model / gene orderMAE ↓Pearson Δ ↑Overlap@N ↑
Fine-tuned X-Cell0.06680.65980.3618
Looped X-Cell (Learned)0.06320.68520.3642
Looped X-Cell (Random)0.06480.67810.3331
Looped X-Cell (Fixed control mean)0.06450.67060.3518
Looped X-Cell (Fixed control variance)0.06270.67640.3574

Finetuned X-Cell achieved a mean MAE of 0.0668 and a mean Pearson Δ of 0.6598 on the same three seeds. Looped X-Cell with all four orderings improved over finetuned X-Cell, suggesting that staged prediction accounts for much of the endpoint gain. Learned ordering gave the highest mean Pearson Δ and Overlap@N, with MAE close to that of the best fixed order (0.0632 vs. 0.0627 for control variance).

Q2. What measurable benefit does the full model offer over global ranking and lookahead?

On HepG2, global ranking, staged commitment, and lookahead were sufficient to achieve the lowest mean MAE in Table 2 (0.0611, compared with 0.0616 for full Looped X-Cell across the same three seeds). To isolate the contribution of local ranking, we removed the within-stage ranking step while retaining the Mamba executor and the same training losses (Global-rank Looped X-Cell in Table 2), with MAE from 0.0616 to 0.0624, and Pearson Δ from 0.6848 to 0.6803, showcasing modest performance fluctuation. Local ranks also let us compare perturbations by the order of genes within each stage. In Marson, local rank recovered 34 CORUM modules and 38 same-complex target pairs that the stage and global-rank views missed (Figure A5).

We explored outer-loop recurrence as a way for the model to revisit its staged predictions. The endpoint gains in Table 1 were obtained with a single outer pass. On HepG2, the model trained with two scorer-latent passes had the same reported MAE (0.0609) at 2, 4, 8, and 16 inference passes (Table A3). In Th17, the second outer pass preserved the broad pathway ordering while resolving more specific Reactome terms, including interferon signaling in early stages and spindle-checkpoint pathways in late stages (Figure A4). We therefore treat outer recurrence as an optional extension for refining inspectable internal traces, rather than a prerequisite for the reported endpoint gains.

Thank you again for suggesting the ordering controls. They helped us clarify how much of the endpoint gain comes from staged prediction and where learning the order helps.

We also substantially expanded Appendix B to document the benchmark splits, model initialization, and evaluation protocol, with training and inference settings for each model in Table A2. The Reactome comparison cited in our earlier reply now appears in Table A6, and the outer-loop stability results are in Table A5.

Reviewer_P1xhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Looped X-Cell, a staged and recurrent extension of X-Cell for predicting gene-expression responses to perturbations. It scores unresolved genes, commits predictions in stages, adds lookahead supervision, and explores an outer loop that carries scorer state between passes. On four cellular contexts, the paper reports improvements over X-Cell on several endpoint metrics. It further examines the learned ordering using pathway enrichment, CORUM complex recovery, and RNA–protein rank differences. The main claim is that staged prediction can improve endpoint accuracy while exposing a biologically interpretable internal order.

Critical strengths

An explicit gene ordering could, in principle, make perturbation predictions easier to inspect than a single endpoint output. The evaluation covers four cellular contexts and reports several complementary endpoint metrics. The authors also attempt to relate the learned order to external pathway and protein-complex annotations, rather than relying only on qualitative examples. These are useful directions, but the current experiments do not establish the central claims strongly enough.

Critical weaknesses
  1. The method is not self-contained, and its novelty is not isolated. The paper repeatedly refers back to X-Cell in the method section, while much of the architecture, tokenization, perturbation-prior pathway, decoder, and losses are inherited. For readers unfamiliar with X-Cell, the method is difficult to follow.

  2. It is not clear which parts are necessary, how global/local ranking, staged commitment, lookahead supervision, and outer-loop recurrence fit together, or what is technically new beyond an extension of an existing predictor. The experiments do not establish that the added components are effective or that the interpretation supports the method.

  3. The pathway/GO/CORUM analyses show associations with known annotations, but do not directly validate the proposed ranking mechanism or show that its learned order is more meaningful than expression-based or shuffled alternatives. The interpretability experiments therefore feel weakly connected to the method's main novelty.

Questions for authors
  1. The method relies heavily on X-Cell, making it difficult to follow independently. Please clarify the novel components, explain how the method differs from X-Cell's iterative remasking/refinement, and isolate the contribution of the learned gene order.

  2. The baselines are limited to Linear, scGPT, and X-Cell. More competitive baselines and challenging evaluations, such as Norman double-gene perturbations and cross-cell-type generalization without fine-tuning, are needed.

  3. The experiments mainly focus on perturbations seen during training. It remains unclear whether the method generalizes to unseen perturbations or provides improvements beyond the pretrained X-Cell backbone.

  4. In Table 2, several simpler variants outperform the full model on MAE, Pearson Δ, or Overlap@N. The authors should clarify the contribution and necessity of the additional components.

  5. Pathway, GO, and CORUM enrichment alone does not establish the biological relevance of the learned gene order. Comparisons with random, expression-based, and variance-based rankings would provide stronger evidence.

  6. It is unclear whether the test perturbations or their response profiles were included in X-Cell pretraining. The authors should clearly distinguish what was held out during pretraining from what was held out during fine-tuning.

4 comments

Thank you very much for your insightful comments! We have substantially revised the manuscript in response. Section 2 and the new Appendix A explain what Looped X-Cell inherits from X-Cell and how learning which genes to commit changes inference. We also expanded the ordering comparisons in Table 2, added cross-donor endpoint results in Table A5, and revised Section 3.3 to show which biological signals the learned ranks recover beyond expression-based and shuffled controls.

We appreciate your feedback on the clarity of the writing. We revised the sections above for clarity and highlighted the changes in blue.

Q1. What does Looped X-Cell add to X-Cell’s iterative prediction, and how much does learning the order help?

Inspired by diffusion language models, X-Cell uses varying masking ratios during training and supports iterative prediction at inference time. At each step, X-Cell rolls out predicted expression values for a randomly selected subset of the remaining genes until all genes have been revealed. The rollout schedule is therefore independent of the perturbation and not learned during training. Looped X-Cell learns a perturbation- and control-conditioned order for committing gene-expression predictions, while carrying recurrent computational state across stages during training. Within each stage, it orders genes for sequential processing by the interleaved Mamba layers. During training, lookahead also asks the model to predict values one stage before they are committed. Both models use the same encoders, transformer backbone, perturbation conditioning, and endpoint decoder (Section 2 and Appendix A).

To test how much the choice of order matters, we kept the four-stage HepG2 rollout and disabled the auxiliary lookahead loss for all four orderings (Table 2). With a random order or a fixed order based on control mean or variance, the staged rollout improved mean MAE and Pearson Δ over fine-tuned X-Cell. Learning the order gave the highest mean Pearson Δ (0.6852, compared with 0.6781 for random order, 0.6706 for control mean, and 0.6764 for control variance) and Overlap@N. The variance-based order gave slightly lower MAE than the learned order (0.0627 vs. 0.0632). This suggests that staging itself accounts for much of the gain, with the clearest additional benefit of learned ordering in Pearson Δ and Overlap@N. Beyond endpoint prediction, the full model provides three views of the learned rollout that we can inspect: which genes are committed at each stage, how genes rank globally, and how they are ordered within each stage. The learned ranks also reflect known biological organization, with genes ranked together enriched for Reactome pathways and perturbations with similar rollouts tending to target members of the same CORUM protein complexes (Section 3.3).

Q2. Can you include stronger baselines and more challenging transfer settings?

We added cross-donor comparisons on Th17 and Marson (Section 3.3 and Table A5). The linear baseline takes the mean expression change for each perturbation in the source donor and adds it to the target donor’s control profile. Looped X-Cell achieved higher Pearson Δ in both Th17 (0.3075 vs. 0.1303) and Marson D2-to-D3 transfer after 8 hours of stimulation (0.2009 vs. 0.0202). The linear baseline achieved lower MAE in both settings (0.0539 vs. 0.0608 for Looped X-Cell in Th17 and 0.0769 vs. 0.0914 in Marson). The donor experiments showcase zero-shot transfer across individuals within a cellular context, without fine-tuning in the target donor. Cross-cell-type transfer without target-context fine-tuning and Norman double perturbations remain open comparisons for Looped X-Cell. The original X-Cell study evaluated transfer from resting to stimulated Jurkat cells against a linear baseline (Wang et al., 2026). The same study reported higher Pearson Δ and lower MAE for X-Cell in a separate comparison with STATE (Adduri et al., 2026) on Replogle-Nadig HepG2 split and Tahoe. We are also evaluating scDFM (Yu et al., 2026) on the Replogle–Nadig benchmark and will report the results if they are available before the rebuttal deadline.

Yu, Chenglei, et al. (2026). scDFM: Distributional Flow Matching Model for Robust Single-Cell Perturbation Prediction. arXiv:2602.07103.

Wang, Chloe, et al. (2026). X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models. bioRxiv.

Q3. Which perturbations are held out, and what improvement comes from Looped X-Cell beyond the pretrained backbone?

Q6. Were the test perturbations or their response profiles present during pretraining?

In Replogle–Nadig fine-tuning, across the four available cellular contexts, we follow the few-shot perturbation prediction setting where training includes perturbations from three contexts and a fraction of the perturbations from the remaining test context. Up to 380 perturbations were held out within each test context. Each evaluated held-out perturbation appeared during fine-tuning in at least one other context (Section 3.1 and Appendix B.2). Table 1 therefore evaluates generalization to held-out perturbation–context pairs.

X-Cell and Looped X-Cell start from the same checkpoint, pretrained on X-Atlas/Pisces perturbation screens in HCT116, HEK293T, iPSCs, and HepG2 (Appendix B.5). We have not verified whether the test targets or their response profiles appeared during pretraining (Appendix B.3). In the HepG2 seed-42 fine-tuning run with Looped X-Cell, MMD loss on the 5% validation set went from 4.06 to 3.81 over 20 epochs of fine-tuning, and validation centroid accuracy went from 0.798 to 0.825. Across five training seeds, Looped X-Cell achieved lower mean MAE and higher mean Pearson Δ than fine-tuned X-Cell in all four contexts (Table 1).

Q4. What do the additional components contribute when a simpler model already performs well?

The ablations suggest that the simpler model with global ranking and lookahead gives the lowest mean HepG2 MAE. Across three matched seeds, global ranking and staged commitment reduced mean MAE from 0.0668 to 0.0632. Adding lookahead reduced it further to 0.0611, compared with 0.0616 for the full model (Table 2). We attribute the main HepG2 gains to staged prediction and lookahead.

We retain local ranking because it specifies an order within each stage that we can inspect. In Marson, local ranks recovered 34 CORUM modules and 38 same-complex target pairs missed by both the stage and global-rank views (Figure A5). Outer recurrence remains optional, and all four contexts in Table 1 use a single outer pass.

Q5. Do the learned ranks recover biological structure beyond expression-based or shuffled rankings?

The learned stages recovered Reactome pathways missed by both control-expression rankings, and perturbation clusters based on learned ranks recovered more CORUM modules than their shuffled-rank controls. Across the four contexts in the Replogle–Nadig dataset, we repeated the same Reactome enrichment analysis after ranking genes by mean expression or variance in control cells. Of the 19 curated context–relevant pathways in Table A6, the learned stages recovered 17, compared with one for the control-mean ranking and eight for the control-variance ranking. Ten pairs were recovered only by the learned stages, including all eight Jurkat immune terms. For CORUM, we shuffled gene ranks within each perturbation and repeated the clustering. In Marson cross-donor predictions, global rank recovered 6.9 times as many modules and 8.8 times as many same-complex target pairs as its shuffled control (Figure 2B).

Thank you again for suggesting a stronger baseline. We have completed the scDFM comparison mentioned in our earlier response and added the results to Table 1. Across five training seeds, Looped X-Cell achieved lower mean MAE and higher mean Pearson Δ than scDFM in all four Replogle–Nadig contexts.

Following your suggestions, we also substantially expanded Appendix B to clarify which perturbation–context pairs were held out, how the models were initialized, and how predictions were evaluated. Table A2 summarizes the training and inference settings for each model. The cross-donor results cited in our earlier reply now appear in Table A7.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.