acceptodds
Under review as a conference paper at ICLR 2027

RADF: Reference-Anchored Dynamic Flow for Observed-Perturbation Spatial Completion

Abstract

Spatial perturbation profiling reveals how genetic interventions alter transcription across tissue locations, but destructive assays leave many perturbation-by-location response populations unmeasured. We study observed-perturbation spatial completion: given a reported response population for one intervention and a query coordinate scaffold, estimate the held-out response population. Predictions are unpaired and evaluated as populations; we do not claim prediction for unseen interventions or cell-to-location matching. RADF is fitted separately for each observed intervention. It first constructs a population-valued anchor using equal reference contribution and then applies a bounded coordinate-conditioned residual flow. On matched Perturb-map tasks, RADF reduces macro E-distance, an unordered-population distance, from 138.05 to 40.60 relative to CONCERT (70.6%) and from 62.98 for scGen to 40.60. Validation-gated residual refinement improves E-distance over the learned anchor by only 0.24% overall, indicating that reference anchoring provides the main gain while the flow supplies conservative refinement. Code will be made publicly available upon acceptance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Spatial transcriptomics, spatial perturbation modeling, observed-perturbation completion, population-level prediction, reference-based learning, flow matching, optimal transport

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_mtx5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies observed-perturbation spatial completion: estimating an unmeasured response population from a same-perturbation reference expression set and a query coordinate scaffold. A separate model is fitted for each intervention, and predictions are unpaired with held-out observations and primarily evaluated as unordered populations. RADF first learns a Sinkhorn-balanced convex allocation of reference profiles, ensuring equal total reference contribution and preservation of the reference population mean. A state-dependent spatial attention network then predicts residual velocities, with projected Euler updates restricting changes to a fixed neighborhood of the anchor. Training uses a reference-derived OT endpoint with flow-matching and reference-distribution losses. Validation selects the radius and checkpoint and can return the anchor exactly.

Critical strengths
  1. The task formulation and constrained predictor provide consistent, auditable information and behavioral boundaries. Section 2 separates reference, validation, and evaluation expression and restricts the task to same-perturbation-supported population completion. Candidates are frozen before validation expression is accessed, and evaluation expression is reserved for scoring.

  2. The empirical gains have practical value, with useful transparency about their source and the method’s failures. The matched results in Table 1, cross-slide evaluation, and SPAC-seq evaluation with settings frozen from Perturb-map validation support reference-supported completion over the compared learned predictors in multiple settings.

Critical weaknesses
  1. The independent contributions of the core algorithmic components over simpler alternatives remain unestablished, limiting novelty and significance.

  2. On spatial blocks, the anchor scores 58.209 and full RADF 58.550, a degradation of approximately 0.58%. Thus, the large gain in Table 1 mainly supports same-perturbation reference reuse rather than dynamic spatial relations.

  3. Section 3.3’s explanation that general predictors lag because they lack joint query geometry is insufficiently supported when an anchor without query geometry already achieves nearly the full result.

  4. The spatial evidence does not yet establish reliable regional response completion beyond coordinate dependence, and the selection rule’s conservatism is not established on evaluation data. Section 2 primarily supplies an unordered reference expression set and query geometry, while the endpoint in Equations 4-5 is derived entirely from reference expression without native-location response supervision.

Questions for authors
  1. On the matched benchmark and one main external track, what are the per-unit gains of the learned anchor over balanced resampling, empirical-reference prediction, and a simple coreset when output count, expression processing, and validation budget are matched?

  2. What are the task-paired intervals for full RADF versus the anchor and self-only flow, particularly on spatial-block and regional metrics? Is there evidence of a stable benefit from dynamic spatial relations beyond near-identity expression correction?

  3. Does any modeling stage access the association between reference expression and native reference coordinates? If not, what assumption or supervision links regional responses to query regions?

1 comment

Q1. What does learned allocation add over simple reference reuse on individual units? We evaluated the original ten matched units and five cross-slide Perturb-map units. Within each unit, the fixed balanced resampler, greedy coreset, and learned anchor use the same reference expression, preprocessing, gene order, and 31 outputs. All three use zero validation selections. The original “empirical reference” comparator is the balanced resampler itself. Scoring the full empirical measure uses a different output count, so we report that result separately.

The learned anchor beats balanced resampling in 4/10 matched units and 0/5 cross-slide units. It beats the coreset in 2/10 matched units and 2/5 cross-slide units. The size of the difference also matters. On the GSM5808056/Ifngr2 spatial block, balanced resampling scores 105.581, the anchor improves to 89.011, and the coreset scores 88.520. That unit contributes substantially to the anchor’s mean improvement over resampling, yet the simple coreset recovers the same gain. Tables S15–S16 report the E-distances for every unit.

Full RADF has a separate selection budget. On matched tasks it evaluates eight nonzero refinements plus the anchor. Giving balanced resampling nine validation candidates yields mean E-distance 42.806, compared with 42.941 for fixed resampling, 40.108 for the fixed coreset, and 40.595 for full RADF. This matches the number of validation scores queried. Training costs and model capacities still differ. The cross-slide track uses fixed settings without local validation selection (Tables S13 and S17).

Q2. Do task-paired intervals support a spatial-flow benefit? Let Δ=DE(comparator)−DE(RADF)\Delta=D_E(\mathrm{comparator})-D_E(\mathrm{RADF}), so positive values favor RADF. Across the five spatial-block tasks, mean Δ\Delta is −0.340 against the anchor, with a descriptive 95% paired interval of [−0.841,0.054][-0.841, 0.054]. Against self-only flow it is 0.022, with interval [−0.506,0.668][-0.506, 0.668]. Across all matched units the corresponding intervals are [−0.156,0.352][-0.156, 0.352] and [−0.236,0.299][-0.236, 0.299] (Table S14). These comparisons do not establish a stable gain from dynamic spatial relations. The validation fallback also has a measurable limit. In three of the four spatial-block units where validation selected a nonzero refinement, evaluation favored the anchor.

For the regional pathway endpoint, only three random-split Tgfbr2 units satisfy the common-support rule, and no spatial-block unit qualifies. Mean contrast error is 0.199 for RADF and 0.201 for the anchor, with both recovering two of three directions (Table S10). We report these as descriptive results. We have no supported spatial-block regional interval or regional comparison against self-only flow. A separate whole-lesion diagnostic reinforces this distinction. Flow improved population E-distance from 92.45 to 90.26 while worsening held-out stromal-module RMSE from 0.713 to 0.769.

Q3. What links a reference response to its native position and then to a query region? Native coordinates enter upstream partitioning and scaffold construction. During fitting, the anchor uses reference expression, the OT cost uses expression-space distances, and the flow receives query-to-query geometry and summaries of reference expression. None of these stages uses paired reference expression and native position as supervision. The modeling assumption is that nearby query slots may benefit from related refinements. It supplies no observation identifying which reference response belongs in a particular query region. Permuting the predicted profiles across query slots leaves population E-distance unchanged.

We therefore agree with your distinction between coordinate dependence and regional accuracy. We have removed the attribution of the headline performance gap to joint query geometry. The revised claim is that same-perturbation references support population-level response recovery. Whether learned spatial allocation improves region-specific predictions remains open.

Reviewer_pGw3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studied observed-perturbation spatial completion. Given a reference expression, and query coordinates, the model outputs predicted population of expression that are evaluated on the held out population through E-distance. The paper explicitly said they do not claim unseen perturbation or cell-to-location matching.

Evaluating on five perturb-map tasks, RADF reports a lower E-distance compared to CONCERT and scGen.

Critical strengths
  1. The problem is well scoped and no over claiming. The paper repeatedly say the method is not for predicting unseen perturbation and cell to location matching. So readers are clear what was been tested.

  2. The report is honest. The 0.24% marginal improvement is reported honestly in the abstract.

Critical weaknesses
  1. The spatial component does not support the main claim that paper is making. Not enough measurable effects are shown. The 0.24% marginal gain on the learned anchor, and is worse than on the spatial block splits (-0.58%). The method setting was designed to reward spatial component but failed.

  2. The motivation and evaluation are not well-aligned. The authors stated the usage is to profile unmeasured response population and then compare gene or pathway summaries across spatial regions. But outputs are evaluated as unordered sets without row-to-location claim, so they cannot be used for "comparing gene or pathway summaries across spatial regions". The introduction has one sentence of biological motivation but never addressed same perturbation completion is the problem worth solving.

Questions for authors
  1. Was CONCERT given R_p as perturbed input?

  2. Whats the case that spatial flow help? On spatial-block splits, the anchor beats the full model. Is there a regime spatial flow provides measurable gain?

  3. Whats the intended downstream biological application with the method?

1 comment

Q1. Did CONCERT receive the perturbed reference population RpR_p? Yes, during fitting. For each matched task, CONCERT’s training file contained the same target-reference expression rows RpR_p available to RADF, together with the non-target training rows. We verified that the target-validation and target-evaluation rows were excluded. The methods differ in how they use those observations. CONCERT learns from them during fitting, whereas RADF also uses RpR_p explicitly to construct its prediction anchor. We have clarified this distinction in the experimental setup.

Q2. When does spatial flow help? The clearest measurable improvement over the anchor occurs on the five random-split tasks. Mean held-out E-distance falls from 23.176 to 22.641, a 2.31% reduction. The descriptive 95% task-paired interval for anchor minus RADF is [0.158,0.913][0.158, 0.913]. However, the corresponding comparison against self-only flow gives [−0.118,0.221][-0.118, 0.221]. Thus we can identify a setting where refinement helps relative to the anchor, but the comparison does not establish that spatial messages supply that gain. On the five spatial-block tasks, the anchor performs better, 58.209 versus 58.550. Four of nine refinements selected on validation also reverse direction on evaluation. The revised paper reports both outcomes (Table S14 and the main-text flow analysis).

Q3. What biological question can the output answer? Consider a perturbation already measured in one tissue section. Given those reference profiles, one may want to estimate the population-wide response in another section with incomplete measurements. For example, which genes show the largest average expression changes, and does a predefined pathway’s average response have the same direction in the held-out population? These questions can be checked against held-out observations without pairing individual spots.

In our retrospective audit, the gene-wise response ranking has Spearman correlation 0.690 with the held-out ranking for RADF, 0.688 for the anchor, and 0.423 for CONCERT. This illustrates both the utility and its limit. Reference reuse preserves much of the measured population response. The evaluation does not tell us which tissue region contains that response. We agree that our original motivation implied this stronger application and have removed the claim about comparing pathway responses across query regions.

If the clarification of CONCERT’s reference access and the revised biological scope address those concerns, we would be grateful if you would consider whether they change your overall assessment.

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

RADF completes an observed perturbation's response population using same-perturbation reference cells and a query coordinate scaffold. It learns a balanced convex allocation of reference profiles, then applies a bounded spatial residual flow selected on validation data. Matched and transfer evaluations show large improvements over several adapted predictors, with nearly all of the matched-task gain supplied by the reference anchor.

Critical strengths
  1. The paper demonstrates the importance of retaining observed response support. The anchored predictor substantially improves population-level error over the evaluated spatial and adapted single-cell baselines, including transfer settings. This is a useful finding for the stated completion task.
  2. The evaluation scope is explicitly defined. The paper separates fitting, validation, and evaluation expression, and correctly distinguishes unordered population completion from recovering cell-to-location assignments.
Critical weaknesses
  1. The distinctive dynamic-flow component contributes almost no demonstrated benefit. Overall E-distance changes from 40.693 for the anchor to 40.595 for RADF, while spatial-block error worsens from 58.209 to 58.550 (§3.4). Downstream biological outcomes are also essentially tied (§3.6). These results do not justify the added relational flow as a consequential methodological contribution.
  2. The value beyond straightforward reference reuse remains unclear. The anchor is fitted separately for each perturbation to reproduce that same observed reference population. The supplied PDF does not establish how much this learned allocation improves over simply resampling those references. Coordinate-intervention results demonstrate dependence on geometry, but do not establish that geometry explains the headline population gains. Thus the central empirical result may primarily concern the strength of reference reuse rather than a new spatial modeling capability.
Questions for authors
  1. Assessment-critical comparison: How do direct reference resampling, the balanced anchor, and full RADF compare under identical reference access and output sizes, especially on spatial blocks and cross-slide/replicate transfer? Please identify any supporting results in the separately referenced supplement.
  2. Assessment-critical evidence: On which held-out population or regional endpoint does the dynamic spatial flow materially outperform the anchor and self-only refinement? Coordinate sensitivity alone does not establish this advantage.
1 comment

Thank you for distinguishing two questions that our original presentation left too close together. How much does the model gain from having observed responses to the same perturbation, and how much does it gain from learning what to do with those responses?

Q1. Direct reference reuse, the learned anchor, and full RADF.

We compared the three methods using the same reference expression, gene order, preprocessing, and 31 output profiles within each Perturb-map unit. The empirical reference in the original supplement already performs balanced resampling. It permutes the reference rows and cycles through them until it has produced 31 outputs. We also evaluated a simple coreset that selects 31 reference profiles to approximate the full reference distribution. It uses neither coordinates nor validation scores.

In matched Perturb-map overall (10 units), E-distance is 42.941 for balanced resampling, 40.693 for the learned anchor, 40.595 for full RADF, and 40.108 for the simple coreset. On the spatial-block subset (5 units), the respective values are 64.166, 58.209, 58.550, and 57.286. On cross-slide Perturb-map (5 units), they are 69.883, 71.710, 71.199, and 70.775.

The spatial-block comparison illustrates your point particularly well. The learned anchor improves on balanced resampling, yet the simpler coreset improves further. On cross-slide transfer, balanced resampling has the lowest mean error of the four. These are retrospective comparisons on previously evaluated units. Tables S13, S15, and S16 of the revised supplement give the protocol, averages, and individual results.

For cross-replicate SPAC-seq, Table S7 reports the existing nine-unit comparison. Full RADF and empirical reference reuse both score 2.203 in population E-distance at the reported precision. RADF has lower spatial contrast error, 0.300 versus 0.376 for empirical reuse (Table S8). That comparison does not isolate the flow because it lacks a separate learned-anchor and self-only ablation on this track.

Q2. On which endpoint does dynamic flow materially outperform both controls?

We have not identified a held-out endpoint on which it does so reliably. Across the ten matched units, the mean E-distance gain is 0.098 over the anchor and 0.029 over self-only flow. The corresponding descriptive 95% task-cluster intervals are [−0.156, 0.352] and [−0.236, 0.299] (Table S14). Random splits show a small improvement over the anchor, while spatial-block error increases from 58.209 to 8.550. On the three units supporting a regional pathway comparison, RADF and the anchor recover the same two of three contrast directions (Table S10).

We have revised the abstract and conclusion accordingly. The finding that survives the stronger comparisons is that retaining the observed response population accounts for most of the advantage over the adapted learned predictors. The present results leave the independent value of dynamic spatial refinement unresolved.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.