acceptodds
Under review as a conference paper at ICLR 2027

DeFo: Decomposed Dynamics for Forecasting Temporal Single-Cell Perturbation Responses

Abstract

Forecasting future cellular trajectories for perturbations unseen during training from limited early-response observations is a fundamental challenge in single-cell perturbation modeling. Unlike conventional perturbation prediction that focuses on observed conditions or static endpoints, this task requires generalizing dynamic response patterns across perturbations while preserving shared cellular progression. We propose DeFo, a decomposed dynamics framework for held-out perturbation forecasting in temporal single-cell systems. DeFo introduces a control-referenced formulation that decomposes cellular evolution into a shared base field capturing perturbation-independent temporal progression and a perturbation-specific response component modeling deviations from the control trajectory. To infer future responses for unseen perturbations, we introduce Temporal Response Memory Attention (TRMA), a trajectory-valued memory retrieval mechanism that retrieves and softly composes temporal response patterns from known perturbations according to their early effects. Experiments on time-resolved chemical and genetic perturbation datasets demonstrate competitive recovery of population-level expression, control-referenced perturbation effects, and differential-expression directions. These results establish DeFo as a principled framework for learning transferable cellular dynamics and forecasting responses to unseen perturbations from limited temporal observations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Single-cell perturbation modeling, Time-series scRNA-seq, Flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_M2ETOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the problem of forecasting future single-cell transcriptomic trajectories for unseen (held-out) perturbations given only their earliest observed post-perturbation response and time-matched control populations across the forecast horizon. The proposed framework, DeFo, uses optimal transport to align perturbed cells with control trajectories and decomposes the velocity field into a control-driven base field (vbasev_{\text{base}}) and a perturbation-specific response field (vrespv_{\text{resp}}) inferred via Temporal Response Memory Attention (TRMA), which retrieves and linearly combines finite-difference mean effect trajectories from training perturbations based on t0t_0 response similarity. Experiments on a 2D synthetic system and two real time-resolved perturbation datasets (HSPC-Mk and RENGE) evaluate pseudobulk expression recovery and differential-expression direction concordance.

Critical strengths
  1. Formulating held-out temporal perturbation forecasting conditioned on early-response observations and time-matched controls addresses a practical experimental design scenario, and separating shared control progression from perturbation-specific deviations provides a clean control-referenced coordinate system (Lotfollahi et al., 2019).

  2. The manuscript is clearly organized, evaluates five condition-level holdout splits with reported standard deviations, and achieves strong pseudobulk differential-expression direction agreement on the RENGE genetic perturbation dataset.

Critical weaknesses
  1. Despite being framed as a continuous flow-matching dynamical system, the forecasting ODE in Equation (14), dxipred(t)/dt=vbase(xictrl(t),t)+∑cαi,cm~c(t)\mathrm{d}x_i^{\text{pred}}(t)/\mathrm{d}t = v_{\text{base}}(x_i^{\text{ctrl}}(t), t) + \sum_c \alpha_{i,c}\tilde{m}^c(t), is entirely open-loop and does not depend on the evolving perturbed cell state xipred(t)x_i^{\text{pred}}(t) on its right-hand side. Because future control populations xctrl(tℓ)x^{\text{ctrl}}(t_\ell) are already provided at inference time and m~c(t)\tilde{m}^c(t) is simply a piecewise-linear interpolation of finite differences of population means in Equation (6), integrating Equation (14) via fourth-order Runge–Kutta is mathematically nearly equivalent to directly adding a static convex combination of known perturbations' mean displacements ∑cαi,c(Δx‾c(tℓ)−Δx‾c(t0))\sum_c \alpha_{i,c}(\overline{\Delta x}^c(t_\ell) - \overline{\Delta x}^c(t_0)) to the observed target-time control state, without learning state-dependent cellular dynamics (Tong et al., 2020; Zhang et al., 2024a).

  2. Both the TRMA memory construction and the quantitative evaluation collapse single-cell distributions into population averages (pseudobulk), discarding single-cell heterogeneity. Specifically, Equations (4)–(6) average Δxi,tℓc\Delta x_{i,t_\ell}^c across all cells within each known perturbation before storing memory velocities, and all three metrics in Appendix A.4 (MSE, Pearson-Δ, and DE Dir.) evaluate only pseudobulk mean vectors or marginal gene-wise rank tests, with no single-cell distributional metrics (such as Wasserstein distance W2W_2, Energy Distance, or MMD) to verify whether subpopulation structure and cell-level variance are preserved (Yeo et al., 2021).

  3. The empirical results in Tables 1 and 2 do not convincingly justify the proposed modules across datasets. On HSPC-Mk (Table 1), the naive unconditional baseline Pert—which completely ignores the held-out perturbation's early response and simply pools known training perturbations—achieves a higher Pearson-Δ (0.5890±0.09070.5890 \pm 0.0907) than DeFo (0.5627±0.08000.5627 \pm 0.0800) and virtually identical DE Dir. (0.80080.8008 vs. 0.80210.8021), indicating that performance on this dataset is driven by shared homogeneous drug shifts rather than TRMA retrieval; meanwhile, on RENGE (Table 2), removing the core decomposition (w/o Deco), temporal velocity (w/o Temp), or soft composition (w/o Comp) yields nearly identical scores (e.g., MSE of 0.00100.0010 for both DeFo and w/o Deco, and Pearson-Δ of 0.46670.4667 vs. 0.45520.4552, well within one standard deviation).

  4. The pipeline relies heavily on chained deterministic optimal-transport couplings across both time points and control-versus-perturbation conditions in a fixed 50-dimensional PCA space, which can introduce pairing artifacts when perturbations alter proliferation/apoptosis rates or shift subpopulation proportions (Zhang et al., 2024b). Furthermore, computing time-invariant attention weights αi,c\alpha_{i,c} solely from the earliest snapshot t0t_0 assumes that initial transient deviations uniquely determine long-term trajectories, which fails when distinct genetic perturbations share generic early stress responses or exhibit delayed transcriptional cascades.

Questions for authors

see weaknesses

2 comments

We thank the reviewer for the comments on DeFo’s dynamical form, single-cell heterogeneity, module effectiveness, and the assumptions underlying OT matching and early-response retrieval. We address these points below. Major revisions in the updated manuscript are shown in blue.

W1: Open-loop dynamics and anchor-time conditioning

We agree that Equation (14) does not depend on the evolving prediction x^(t)\hat{\mathbf{x}}(t), so DeFo is not a state-feedback dynamical system. However, prediction remains cell-specific because each TRMA query is the cell-level control-referenced effect Δxi(t0)\Delta\mathbf{x}_i(t_0) at the anchor time. Different cells can therefore obtain different response combinations, although these weights remain fixed during forecasting.

Fixed weights do not imply a single static displacement because the memory values are time-varying response-velocity trajectories that are combined with the corresponding control progression. Continuous-time integration accumulates these time-varying components but does not introduce dependence on the generated state. We have therefore revised the model positioning to a continuous-time, control-guided trajectory model rather than a general state-dependent flow model.

W2: Single-cell heterogeneity and distributional evaluation

TRMA uses condition-average response keys and values to reduce sampling noise and avoid artificial cell-to-cell correspondence across perturbations. Cell specificity is retained through the query, attention weights, and cell-specific control reference, allowing different cells to retrieve different response combinations.

To evaluate population-level distribution recovery directly, we added Energy Distance (ED) in PCA-50 space.

MethodHSPC-Mk ED ↓RENGE ED ↓
Ctrl0.7969 ± 0.11670.5463 ± 0.1741
Pert0.6786 ± 0.07870.5554 ± 0.1497
scNODE1.2302 ± 0.06790.9444 ± 0.1352
scGPT1.9823 ± 0.04621.5933 ± 0.0613
CellFlow0.6537 ± 0.05780.7488 ± 0.1802
scDFM2.2877 ± 0.64011.5231 ± 0.2785
Departures1.2890 ± 0.23930.8434 ± 0.0582
DeFo0.7814 ± 0.07560.5144 ± 0.1177

DeFo achieves the lowest ED on RENGE and outperforms most learned methods on HSPC-Mk, although it is weaker than CellFlow and Pert there. These results indicate that cell-specific retrieval with control references preserves meaningful population-level structure, while condition-level memory remains limited for subgroup-specific effects.

W4: OT pseudo-trajectories and early-response retrieval

OT-reference robustness

DeFo propagates a single anchor-time perturbation–control OT coupling along the aligned control trajectory. To test whether the results depend on a particular coupling, we added Random fixed, which uses a random correspondence fixed across time, and Control mean, which removes cell-level matching and uses the control-population mean.

DatasetControl referenceMSE ↓Pearson-Δ\Delta ↑DE Dir. ↑
HSPC-MkAnchor-time OT0.0017 ± 0.00090.5627 ± 0.08000.8021 ± 0.0290
HSPC-MkRandom fixed0.0022 ± 0.00080.5167 ± 0.11780.7739 ± 0.0459
HSPC-MkControl mean0.0018 ± 0.00070.5215 ± 0.11400.7725 ± 0.0408
RENGEAnchor-time OT0.0010 ± 0.00070.4667 ± 0.05740.9465 ± 0.0110
RENGERandom fixed0.0010 ± 0.00070.4448 ± 0.05850.9366 ± 0.0323
RENGEControl mean0.0010 ± 0.00070.4603 ± 0.05360.9447 ± 0.0196

Both alternatives remain functional, while anchor-time OT gives the strongest response-level results. This indicates that the conclusions are not driven by one accidental coupling and that OT provides an informative control reference. This sensitivity analysis does not establish biological lineage correspondence, but it shows robustness to the construction of the control reference.

Early retrieval and delayed responses

TRMA determines its composition weights from the anchor-time effect, but the retrieved memory values are time-varying response-velocity trajectories rather than static displacements. Thus, early effects select reference response programs whose later memory values can contain delayed responses.

We compared condition-pair similarity matrices between the anchor time and each future time point. Their Spearman correlations were 0.62, 0.65, and 0.81 for HSPC-Mk and 0.69, 0.60, and 0.51 for RENGE. These positive correlations indicate that the relative response structure among perturbations remains partly preserved over time, supporting early-effect-based retrieval.

The limitation is that perturbations with indistinguishable early effects but different unobserved late mechanisms cannot be reliably separated. DeFo can transfer delayed patterns represented in memory, but cannot infer entirely novel late mechanisms unsupported by the early observation.

W3: Module effectiveness and dataset dependence

Pearson-Δ\Delta and DE Direction assess response patterns and directions but not whether predictions preserve held-out-condition identity. We therefore added PDS, whose random expectation is 0.5.

MethodHSPC-Mk PDS ↑RENGE PDS ↑
Ctrl0.5000 ± 0.00000.5600 ± 0.0435
Pert0.5056 ± 0.01240.5633 ± 0.0545
scNODE0.4333 ± 0.06390.5600 ± 0.0303
scGPT0.7667 ± 0.11550.5667 ± 0.0717
CellFlow0.6833 ± 0.11550.5667 ± 0.0645
scDFM0.5056 ± 0.06630.6167 ± 0.0264
Departures0.6889 ± 0.04560.5700 ± 0.0908
DeFo0.7833 ± 0.07190.6333 ± 0.0486

Pert has near-random PDS on HSPC-Mk despite strong directional scores, indicating that averaging reference perturbations preserves shared directions without reliably identifying the held-out condition. DeFo achieves the highest PDS on both datasets.

The original w/o Deco is a conventional condition-aware field v(x,c,t)v(\mathbf{x},c,t) and does not match DeFo’s inputs. We therefore added information-matched w/o Add, which receives the same matched control state, TRMA response, and time as DeFo but predicts total velocity with a single field.

ModelHSPC MSE ↓Pearson-Δ\Delta ↑DE Dir. ↑RENGE MSE ↓Pearson-Δ\Delta ↑DE Dir. ↑
w/o Add0.0034 ± 0.00070.4308 ± 0.11640.7148 ± 0.04100.0019 ± 0.00100.2753 ± 0.11760.8242 ± 0.1282
DeFo0.0017 ± 0.00090.5627 ± 0.08000.8021 ± 0.02900.0010 ± 0.00070.4667 ± 0.05740.9465 ± 0.0110

Under matched conditioning information, replacing the full decomposed formulation with a unified field reduces performance on both datasets, supporting the utility of the full decomposed formulation.

To explain the smaller effects of w/o Temp and w/o Comp on RENGE, we computed median across-condition/across-time effect correlations: 0.475/0.291 for HSPC-Mk and 0.211/0.457 for RENGE. HSPC-Mk’s lower temporal consistency makes temporal transfer more useful, whereas RENGE’s higher consistency makes static responses and top-1 retrieval more competitive. Thus, the benefits of temporal memory and soft composition depend on response structure.

Finally, the reported standard deviations are computed across five condition splits after within-split averaging and therefore reflect held-out-condition composition, not initialization uncertainty or significance intervals. To assess initialization sensitivity separately, we repeated DeFo with seeds 43 and 44 on the same splits:

DatasetSeedMSE ↓Pearson-Δ\Delta ↑DE Dir. ↑
HSPC-Mk420.0017 ± 0.00090.5627 ± 0.08000.8021 ± 0.0290
HSPC-Mk430.0018 ± 0.00070.5422 ± 0.08210.7843 ± 0.0338
HSPC-Mk440.0018 ± 0.00060.5817 ± 0.11290.8069 ± 0.0351
RENGE420.0010 ± 0.00070.4667 ± 0.05740.9465 ± 0.0110
RENGE430.0010 ± 0.00070.4768 ± 0.04480.9569 ± 0.0265
RENGE440.0010 ± 0.00070.4515 ± 0.05340.9435 ± 0.0382

These three runs suggest limited sensitivity to initialization on the evaluated splits. Cross-split standard-deviation overlap alone therefore does not establish equivalence between models.

Overall, we have clarified DeFo’s model scope and added targeted evaluations of distribution recovery, component effectiveness, initialization stability, and OT robustness. We have also stated the limitations of early-response retrieval and condition-level response memory.

Reviewer_FTDyOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new problem formulation of perturbation prediction: completing temporal perturbation trajectories for which only the initial part of the trajectory is available. The authors then propose DeFo, a modeling strategy for addressing that problem. DeFo learns a convex combination of observed complete perturbation trajectories. The methods outperforms established baselines on single cell perturbation prediction that are not designed for this particular problem formulation.

Critical strengths

The problem formulation is interesting. Partial observations from perturbation is common and could speed up experimental campaigns. The method shows better performance at this problem than existing methods (that are not designed for this task).

Critical weaknesses

The model dynamics are much simpler than suggested in the introduction of the paper. The response model does not really learn cell-state-dependent dynamics. According to the methods, the velocity field is a convex combination of average perturbation response from the training set. As such, it is not really a dynamical system that responds to changes in cellular state. Given this, a very simple baseline would be to just compute weights as a cosine similarity between the observed part of the target trajectory and the other perturbations.

Transferability is also restricted to convex combinations of previously observed responses, which may be too limiting in many cases.

The paper compares the approach against other perturbation methods that are not designed for this problem. While this is fine in principle because the problem formulation is new, I think the paper would benefit from a deeper analysis of different baselines dedicated to this problem (see my first point for instance). If no other method is truly fitting this task, the bar is a bit too low and you should provide more obvious ways of solving this problem and show why yours is better among these.

Questions for authors

Could the authors add other baselines that are adapted to the problem definition ? In particular, one where one computes weights as a cosine similarity between the observed part of the target trajectory and the other perturbations.

Given the simple model, the authors could compute the optimal performance achievable by the model. Given the full test trajectory, what is the convex combination that is closest ? This would be interesting to report as a ceiling performance of your method.

3 comments

We thank the reviewer for the comments on task-adapted baselines, dynamical structure, and transferability. We added Raw-state cosine, Effect-space cosine, and Oracle-L2 analyses and clarified DeFo’s cell-state conditioning and transfer scope. Major revisions in the updated manuscript are shown in blue.

1. Task-adapted baselines

Following the reviewer’s suggestion, we constructed two cosine-retrieval baselines. Raw-state cosine combines future trajectories according to anchor-time cosine similarities between the held-out condition and all known conditions, including control. Effect-space cosine instead combines future responses according to similarities between control-referenced early effects. We also performed Oracle-L2, which uses the true future trajectory to find the minimum-L2 convex combination within the fixed response memory.

In addition to MSE, Pearson-Δ\Delta, and DE Direction, we report W2 for distributional discrepancy and PDS for condition identity. PDS measures whether the predicted control-referenced effect is closer to the correct held-out condition than to other held-out conditions; its random expectation is 0.5.

DatasetMethodMSE ↓PDS ↑W2 ↓Pearson-Δ\Delta ↑DE Dir. ↑
HSPC-MkRaw-state cosine0.0019 ± 0.00050.7611 ± 0.089171.3472 ± 2.12050.5076 ± 0.14140.7570 ± 0.0511
HSPC-MkEffect-space cosine0.0029 ± 0.00090.6833 ± 0.138372.3289 ± 2.24500.4400 ± 0.14220.7311 ± 0.0724
HSPC-MkDeFo0.0017 ± 0.00090.7833 ± 0.071966.3029 ± 1.17450.5627 ± 0.08000.8021 ± 0.0290
HSPC-MkOracle-L20.0009 ± 0.00050.9111 ± 0.066370.4080 ± 2.10400.7117 ± 0.04320.8648 ± 0.0062
RENGERaw-state cosine0.0010 ± 0.00070.6233 ± 0.036541.5698 ± 0.90680.4815 ± 0.04750.9435 ± 0.0376
RENGEEffect-space cosine0.0009 ± 0.00070.6367 ± 0.029841.5497 ± 0.87490.4880 ± 0.04910.9461 ± 0.0354
RENGEDeFo0.0010 ± 0.00070.6333 ± 0.048637.9210 ± 0.89330.4667 ± 0.05740.9465 ± 0.0110
RENGEOracle-L20.0007 ± 0.00060.7400 ± 0.059641.3199 ± 0.81350.5487 ± 0.03780.9518 ± 0.0104

Bold indicates the best non-oracle result.

DeFo outperforms both cosine baselines on all reported metrics for HSPC-Mk. This pattern is consistent with the lower within-condition temporal consistency of HSPC-Mk (0.291 versus 0.457 for RENGE), where a fixed anchor-time similarity may be less predictive of later responses. On RENGE, Effect-space cosine is competitive and achieves slightly better MSE, PDS, and Pearson-Δ, whereas DeFo obtains lower W2 and slightly higher DE Direction. Thus, the results suggest that learned response retrieval is most beneficial when early and later responses are less directly aligned, while simple effect-space similarity can be a strong baseline when response patterns are more temporally consistent.

Oracle-L2 is a retrospective reference rather than a deployable baseline or a theoretical upper bound. Because it selects convex-combination weights using the true future trajectory, it indicates the extent to which the fixed response memory can represent held-out trajectories under an oracle L2 criterion. Its higher W2 further shows that this oracle is aligned with expression-level L2 recovery rather than distributional similarity.

2. Dynamical structure and transfer scope

TRMA separates condition-level response estimation from cell-level retrieval. Keys and values summarize each training condition, whereas each query is the cell-specific control-referenced effect at the anchor time. Different cells can therefore retrieve different response combinations. However, the resulting weights remain fixed during forecasting; DeFo is conditioned on the initial cell state but is not a fully state-feedback dynamical system.

Unlike fixed cosine retrieval, DeFo learns the query–key mapping in a control-referenced effect space and combines time-varying response trajectories. Its improvements on HSPC-Mk are consistent with a benefit beyond static anchor-time similarity.

DeFo is intended for two-stage drug or CRISPR screens in which reference perturbations have complete trajectories and candidate perturbations have only early measurements. Its transfer remains restricted to response patterns represented in the reference memory and does not provide unconstrained extrapolation to entirely novel mechanisms. We have clarified this positioning, application setting, and limitation in the revised manuscript.

Overall, these analyses distinguish DeFo from fixed cosine retrieval, support its compositional response memory, and clarify its initial-response conditioning and transfer scope.

Reviewer_FTDyOpenReview ↗

Thank you for your response

I thank authors for their answers and for providing the results of the requested experiments.

However, I disagree with the conclusions of the authors about the new results. The simple cosine-based baselines are highly competitive (most often likely within non-significance range). Given the results of the oracle, MSE and PDS are the most reliable metrics (given W2 is worse for the oracle). On these metrics, raw-state cosine is highly competitive.

In the light of these new findings, the methodological details should be reconsidered and/or simplified.

Although I believe the evaluation framework is meaningful, the new results confirm my initial concerns about the methodological contribution and the lack of in-depth benchmarking and I thus unfortunately need to recommend rejection at this stage.

We thank the reviewer for suggesting this informative baseline. Unlike the conventional Ctrl and Pert baselines, cosine-based retrieval directly examines how much future response can be explained by similarity at the anchor time, allowing us to evaluate DeFo from a complementary perspective. Simple empirical baselines can be highly competitive on certain datasets and metrics in single-cell prediction tasks; however, the results also show that they do not consistently replace DeFo across datasets and evaluation dimensions. Relative to the existing baselines, DeFo remains competitive overall.

Raw-state cosine is the most direct implementation of this idea, combining future trajectories solely according to absolute-state similarity at the anchor time. On HSPC-Mk, DeFo outperforms Raw-state cosine on all reported metrics, whereas the two methods perform comparably on RENGE. Effect-space cosine is more closely aligned with the task setting because it retrieves responses in the control-referenced perturbation-effect space, which is also a central design choice of DeFo. DeFo continues to show clearer gains on HSPC-Mk while performing comparably to this baseline on RENGE, indicating that the additional benefit of learned attention is partly dataset-dependent.

Oracle-L2 is a retrospective diagnostic that uses ground-truth future trajectories to select L2-optimal convex-combination weights and is therefore an oracle for L2, which explains its favorable MSE. Its worse W2 does not make W2 less meaningful; rather, it indicates that L2-optimal condition-level composition can still miss population geometry. DeFo’s lower W2 is consistent with the benefit of its cell-specific learned retrieval for this distributional aspect.

Overall, these results support the rationale for representing and composing perturbation effects in the Δ\Delta space, while showing that DeFo’s gain over fixed cosine retrieval depends on the temporal response structure of the dataset. The reviewer’s suggestion has therefore helped us more precisely understand and delineate the effective scope of the method.

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

This paper introduces a new variant of perturbation effect prediction that involves forecasting cellular temporal trajectories for perturbations not seen in training given the earliest timepoint under perturbation, alongside time-matched controls for the entire timecourse. To tackle this problem the author(s) propose DeFo, which decomposes the velocity field for inferring single-cell trajectories into two components:

  • A base field, trained using flow matching on OT-chained pseudo-trajectories in the control conditions.
  • A response component, which models deviations from the control trajectory and is produced by Temporal Response Memory Attention (TRMA). TRMA compares a cell’s effect (its difference from a matched control cell at the earliest timepoint) with each known condition’s mean effect and outputs an attention-weighted average of their mean velocity curves.

The author(s) run experiments on the timecourse perturbation datasets HSPC-Mk (21 chemical perturbations across 5 timepoints) and RENGE (23 genetic perturbations across 4 timepoints) and report that DeFo performs the best on five of six metric-dataset pairs. The paper also include ablations, a synthetic study, and a biological validation case study that measures the ability of DeFo to recapitulate known biology in held-out perturbations.

Critical strengths
  1. The task is well-formulated and it is novel compared to previous tasks in this space such as endpoint prediction, static perturbation prediction, and in-distribution trajectory modeling.
  2. The data boundary is clean. I confirmed in the code that the held-out perturbations are loaded only at anchor time and that the PCA and response memory key/value computations are computed only on samples from conditions in the train set. (The fact that code is provided, and it matches the reported hyperparameters in Section A.5, is also a strength.)
Critical weaknesses
  1. The central claim of the paper, that the decomposition into a base field and a response component provides an improvement over Pert baselines and other approaches, is inconsistently supported by the data:
    • On RENGE, “w/o Deco” is within noise of DeFo on all metrics.
    • On HSPC-Mk, the untrained Pert baseline beats DeFo on Pearson-delta and ties on DE direction.
    • Each split holds out only 4–5 perturbations; there are five splits, one training seed per split, and no significance tests. Many deltas between models in Tables 1 and 2 fall within one standard deviation.
  2. There are several key issues in the DeFo framework, its implementation, and how it is explained in the paper:
    • No leave-one-condition-out masking: the response memory contains every training condition, and nothing excludes a condition’s own entry when its cells are queried (model.py, lines 232–269). Since the mean of the condition population that the cell came from is probably closer to the cell’s state than the mean of any other population, during training the TRMA might just learn to map a cell to its own population mean. Meanwhile at inference time the held-out condition’s own entry isn’t in the response memory, so the predicted weightings may be arbitrary.
    • Section 3.1 says perturbed and control cells are matched at each timepoint, but the code matches them only at the anchor time and follows the control trajectory forward.
    • The first line of the abstract says “Forecasting future cellular trajectories for perturbations unseen during training from limited early-response observations is a fundamental challenge in single-cell perturbation modeling,” but to my knowledge this specific task isn’t actually a “fundamental” problem in the field. In fact, the novelty of this task is a key element of the paper. The paper could also better motivate this task. In what experimental settings does one have timecourse data for a set of conditions but only the first timepoint for another set of conditions?
Questions for authors
  1. Per weakness #2 above, have the author(s) tried examining the TRMA mechanism to see if it mostly learns a one-hot encoding for class assignment during training or if it learns something more sophisticated?
  2. In the code there are two ablations (--ablate_decomposition and --fair_ablate_decomposition). Which one was used for the “w/o Deco” ablation results in the paper? For HSPC-Mk the improvement of DeFo over the ablation is quite large (Pearson-delta falls from 0.563 to 0.294, and DE direction from 0.802 to 0.654), but whether this is impressive depends on which ablation was used. If the condition-ID variant (--ablate_decomposition) were used, it would receive no information about the held-out perturbation’s early response, which alone could explain the large HSPC-Mk improvements.
  3. Where are the “w/o Temp” and “w/o Comp” implementations, the evaluation code, and the baseline adaptations?
3 comments

We thank the reviewer for the detailed comments on leave-one-condition-out masking in TRMA, perturbation–control matching, task motivation, decomposition ablations, and evaluation stability. In response, we examined whether TRMA learns a one-hot class assignment, evaluated strict leave-one-condition-out masking, clarified the matching procedure and practical setting, and added an information-matched ablation, PDS evaluation, and multi-seed validation. Major revisions in the updated manuscript are shown in blue.

Q1 & W2: TRMA and Method Clarifications

1. Leave-one-condition-out masking in TRMA

The reviewer raised the concern that access to a condition’s own memory entry might reduce TRMA to one-hot class assignment. However, the query is a cell-level anchor-time effect relative to its OT-matched control, whereas each key is the mean anchor-time effect of a training condition. They therefore differ in source and granularity.

To test whether TRMA relies on the query condition’s own entry, we trained a strictly masked variant on the five predefined splits. Each training query was restricted to keys and values from other conditions; at inference, the held-out condition was absent from memory by construction. We compared both attention distributions and forecasting performance.

We report the maximum attention weight and the effective number of attended entries, Neff=exp⁡(H(w))N_{\mathrm{eff}}=\exp(H(w)). Here, Neff=1N_{\mathrm{eff}}=1 corresponds to one-hot attention, whereas larger values indicate contributions from multiple entries.

DatasetStageModelMax. weightNeffN_{\mathrm{eff}}Own-entry weight
HSPC-MkTrainOriginal0.503 ± 0.0784.700 ± 0.8500.168 ± 0.010
HSPC-MkTrainMasked0.525 ± 0.0264.350 ± 0.3300
HSPC-MkInferenceOriginal0.508 ± 0.0844.560 ± 0.840N/A
HSPC-MkInferenceMasked0.502 ± 0.0254.540 ± 0.310N/A
RENGETrainOriginal0.512 ± 0.0304.940 ± 0.4900.142 ± 0.012
RENGETrainMasked0.460 ± 0.1005.850 ± 2.2200
RENGEInferenceOriginal0.497 ± 0.0355.100 ± 0.550N/A
RENGEInferenceMasked0.446 ± 0.0966.130 ± 2.280N/A

Both models show similar attention distributions: maximum weights are approximately 0.45–0.53, and each query uses approximately 4–6 effective entries. In the original model, only 0.14–0.17 of the attention is assigned to the condition’s own entry, showing that it does not dominate retrieval.

We also compared forecasting performance:

DatasetModelMSE ↓Pearson-Δ\Delta ↑DE Dir. ↑
HSPC-MkOriginal0.0017 ± 0.00090.5627 ± 0.08000.8021 ± 0.0290
HSPC-MkMasked0.0020 ± 0.00080.5444 ± 0.09310.7957 ± 0.0374
RENGEOriginal0.0010 ± 0.00070.4667 ± 0.05740.9465 ± 0.0110
RENGEMasked0.0010 ± 0.00070.4662 ± 0.04070.9443 ± 0.0376

Strict masking leaves RENGE nearly unchanged and causes only modest decreases on HSPC-Mk. Together, these results suggest that TRMA is not primarily driven by one-hot self-retrieval under these splits: retrieval remains distributed across multiple response entries and remains effective when the query condition’s own entry is excluded.

2. Control-trajectory construction

We thank the reviewer for identifying this ambiguity. For reference perturbations, OT first aligns within-condition trajectories; an anchor-time perturbation–control coupling is then fixed and applied to later aligned control states. For held-out inference, the anchor-time coupling is retained as a fixed reference operator and applied to each observed time-matched control population, without future re-matching. We have clarified these procedures in Section 3.1 and Appendix A.2.

3. Task motivation

The task is motivated by the complementary limitations of two experimental settings: high-throughput perturbation screens cover many conditions but usually have sparse temporal measurements, whereas time-resolved single-cell experiments capture response dynamics but cover fewer conditions because measurements are destructive and costs scale with both conditions and time points.

We therefore consider a two-stage design in which complete trajectories are collected for a limited set of reference perturbations, while a larger set of candidates is measured only at an early time point. For example, a drug screen could collect complete single-cell measurements at 0, 24, 48, and 72 hours for several representative compounds, while measuring tens or hundreds of new candidates only at an early time point. DeFo can then forecast candidate-specific states at 48 or 72 hours to support candidate prioritization and follow-up experimental design, rather than replace experimental validation.

We have added this two-stage experimental setting and the corresponding drug-screening example to the Introduction.

Q2 & W1 & Q3: Decomposition Ablations and Experimental Reliability

1. Decomposition ablations

The original w/o Deco variant (--ablate_decomposition) predicts the complete dynamics with a condition-aware field v(x,c,t)v(\mathbf{x},c,t) using cell state, condition-ID embedding, and time. It therefore compares DeFo with a conventional single-field model rather than isolating additive decomposition under matched information. We have renamed it Single Field.

We further added an information-matched w/o Add variant (--fair_ablate_decomposition). It receives the same matched control state, anchor-time effect, TRMA response, and time as DeFo, but concatenates them into a single field instead of using the additive form vctrl+vpertv_{\mathrm{ctrl}}+v_{\mathrm{pert}}.

DatasetModelMSE ↓Pearson-Δ\Delta ↑DE Dir. ↑
HSPC-MkSingle Field0.0053 ± 0.00150.2939 ± 0.12760.6540 ± 0.0534
HSPC-Mkw/o Add0.0034 ± 0.00070.4308 ± 0.11640.7148 ± 0.0410
HSPC-MkDeFo0.0017 ± 0.00090.5627 ± 0.08000.8021 ± 0.0290
RENGESingle Field0.0010 ± 0.00060.4552 ± 0.07950.9385 ± 0.0343
RENGEw/o Add0.0019 ± 0.00100.2753 ± 0.11760.8242 ± 0.1282
RENGEDeFo0.0010 ± 0.00070.4667 ± 0.05740.9465 ± 0.0110

w/o Add performs worse than DeFo on both datasets, supporting the utility of the full decomposed formulation under matched conditioning information. To examine the dataset-dependent behavior of Single Field, we also computed median effect correlations across conditions and across time.

ComparisonHSPC-MkRENGE
Across conditions0.4750.211
Across time0.2910.457

HSPC-Mk shows stronger cross-condition similarity but greater temporal variation, favoring temporal response transfer. RENGE has higher within-condition temporal consistency, helping Single Field remain competitive.

2. Directional performance of the Pert baseline

MSE measures absolute error and response magnitude, whereas Pearson-Δ\Delta and DE Direction emphasize effect patterns and directions. Pert averages known perturbations, which removes condition-specific differences but can retain shared response directions. This explains its strong directional metrics on HSPC-Mk, where cross-condition responses are more similar.

We additionally report the Perturbation Discrimination Score (PDS), which evaluates whether a predicted effect is more similar to the correct held-out condition than to other held-out conditions; its random expectation is 0.5.

DatasetPertDeFo
HSPC-Mk0.5056 ± 0.01240.7833 ± 0.0719
RENGE0.5633 ± 0.05450.6333 ± 0.0486

Pert is close to random on HSPC-Mk despite its strong directional scores, whereas DeFo achieves higher PDS on both datasets. We now report PDS for all methods in Table 1 and define it in the Appendix.

3. Splits, random seeds, and result stability

Each HSPC-Mk split contains four held-out perturbations, and each RENGE split contains five. Although an individual split is small, the five predefined splits cover 15 of 21 HSPC-Mk conditions and 19 of 23 RENGE conditions while retaining approximately 80% of the conditions for training.

We repeated DeFo with seeds 43 and 44 on the same five splits:

DatasetSeedMSE ↓Pearson-Δ\Delta ↑DE Dir. ↑
HSPC-Mk420.0017 ± 0.00090.5627 ± 0.08000.8021 ± 0.0290
HSPC-Mk430.0018 ± 0.00070.5422 ± 0.08210.7843 ± 0.0338
HSPC-Mk440.0018 ± 0.00060.5817 ± 0.11290.8069 ± 0.0351
RENGE420.0010 ± 0.00070.4667 ± 0.05740.9465 ± 0.0110
RENGE430.0010 ± 0.00070.4768 ± 0.04480.9569 ± 0.0265
RENGE440.0010 ± 0.00070.4515 ± 0.05340.9435 ± 0.0382

These three runs suggest limited sensitivity to initialization on the evaluated splits. In Table 1, each metric is first averaged over held-out perturbations and forecast time points within a split, and the reported mean and standard deviation are then calculated across the five splits. The reported standard deviation therefore reflects variation in held-out-condition composition, not a significance interval; a mean difference smaller than one model’s standard deviation does not establish equivalence. We have clarified this distinction in the revised manuscript.

4. Code availability

We have added the missing w/o Temp and w/o Comp entry points, together with a CellFlow adaptation example, to the revised codebase. For standardized evaluation, we use the publicly available VirtualCell evaluation implementation.

Overall, these additions clarify the decomposition comparisons, explain the complementary behavior of the evaluation metrics, and suggest limited sensitivity of the reported results to initialization on the evaluated splits.

Reviewer_vUuXOpenReview ↗

Response to rebuttal

I thank the authors for their detailed rebuttal and for the revisions to their paper. In particular I appreciate:

  • the cell matching procedure in Section 3.1 is now clarified
  • the renaming of “w/o Deco” to “Single Field” to clarify the actual ablation being performed

However, I still feel that my main concerns regarding the central claim of the paper—that DeFo is superior to baselines at forecasting future perturbation cellular trajectories from early timepoints in unseen conditions—have not been addressed:

  • For the RENGE dataset, the Single Field ablation, which has no trained representation of the held-out condition, matches DeFo on MSE, Pearson-Δ and DE direction, and exceeds it on the new PDS metric (0.677 vs 0.633). This suggests the response memory adds little measurable benefit on this dataset.
  • For HSPC-Mk: Pert still exceeds DeFo on Pearson-Δ across all three DeFo seeds. PDS does not address this, since Pert's prediction is condition-agnostic by construction.
  • While I agree the cross-split standard deviation is not a significance interval, the splits are paired here; a paired per-split or per-perturbation comparison would help settle the ambiguity around the performance of DeFo vs. baselines, yet this comparison was not included in the analysis.
  • Reproducibility: The masking, attention, and multi-seed analyses are not in the revised paper, and the evaluation code and most baseline adaptations are still missing.

While I appreciate the author’s effort, and I still think the framework is interesting, I remain skeptical of the claim that DeFo is actually superior at this task compared to the baselines. I am sticking with my original score of 2: Weak Rejection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.