acceptodds
Under review as a conference paper at ICLR 2027

DRIFT: Directed Developmental Geometry for Spatiotemporal Transcriptomic Alignment

Abstract

Aligning unpaired spatial transcriptomic snapshots across developmental time is essential for reconstructing tissue dynamics, yet existing approaches primarily rely on transcriptomic similarity and spatial organization, which may yield developmentally implausible correspondences. We introduce DRIFT, a framework that incorporates directed developmental geometry into spatiotemporal transcriptomic alignment. DRIFT infers cell-type developmental reachability from velocity-informed single-cell dynamics and lifts this geometry to heterogeneous spatial spots through composition-aware optimal transport. The resulting developmental distance is integrated with transcriptomic and spatial information in a unified transport objective for cross-time alignment. Across four human spatiotemporal transcriptomics datasets, DRIFT produces correspondences that are more consistent with reference lineage structures and exhibit lower developmental deviation than representative transport-based baselines. These improved alignments also support more accurate held-out stage reconstruction under both linear transport and flow matching, with strong performance in both interpolation and extrapolation settings. Our code is publicly available at [https://github.com/xxx/DRIFT](https://anonymous.4open.science/r/DRIFT-review-3F38/README.md).

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Optimal Transport, Spatiotemporal Alignment, Developmental Trajectory Inference

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_4nEROpenReview ↗
Rating
1
Confidence
4(+1)
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors present DRIFT, a framework for aligning unpaired spatial transcriptomic snapshots across developmental time by incorporating directed developmental geometry. Using a paired single-cell reference, DRIFT builds a velocity-informed absorbing Markov chain to estimate cell-type developmental reachability, converts it into a cell-type cost, and lifts it to heterogeneous spatial spots through composition-aware entropic optimal transport, giving a spot-to-spot developmental distance. This distance is added to the expression cost of an entropic fused Gromov-Wasserstein objective that also preserves within-slice spatial geometry, yielding cross-time spot couplings. Empirical results on four human datasets (fetal pancreas, intestine, embryonic limb, and skin wound healing) suggest correspondences more consistent with reference lineage graphs and lower developmental deviation than OT, FGW-OT, PASTE, PASTE2, and DeST-OT, as well as improved held-out stage reconstruction with linear transport and flow matching in interpolation and extrapolation settings.

Critical strengths

The visuals are truly curated and intuitive. I really appreciated the overall polish, both in the figures and the text. In general, I think the paper's quality is satisfying. I also commend the authors for presenting applications across multiple datasets to showcase its versatility. Reported results are promising, overall.

Critical weaknesses
  1. L152: Minor, define nmn_m.
  2. A lot of the paper is based on spots, but the future of spatial transcriptomics technology seems to be leading towards higher resolution, single-cell measurements. I feel this slightly undermines the relevance of the method.
  3. I feel the first part is quite similar to CellRank (Lange et al., 2022), if not in all the technical details, at least in spirit. I currently do not see a lot of novelty in the methodology. The velocity-cosine kernel on a kNN graph (Eqs. 5–6) mirrors CellRank's velocity kernel, and the absorption probabilities (I−Q)−1R(I-Q)^{-1}R (Eqs. 7–8) are CellRank's fate probabilities. The main differences are the Gaussian distance factor, the pseudotime-based endpoint selection (App. B.2.2), and the "stay" probability (App. B.2.1). I wonder why one should use this method rather than CellRank. The latter is also quite flexible, as it identifies terminal states by itself, without for them to be supplied from the outside.
  4. Connected to the above, I think the paper does make some unique choices, like the choice of the cost, but it does not seem like a significant-enough contribution for ICLR. As said, part 1 is very similar to CellRank, part 3 (4.3) is a standard FGW OT framework (and I think it would be more transparent for the paper to address it this way). The novelty boils down to lifting the cost at a cell type level, but this feels a bit A+B as a contribution and dies out as soon as one moves to more granular datasets, unannotated or unreliably annotated settings. Or even settings where lineage mapping is not the core question (e.g., spatiotemporal brain datasets).
  5. I am a bit critical of the interpolation experiment overall. Why would a model with no intermediate supervision be able to predict biological meaningful interpolations, and why would a different coupling help? This is even more significant for extrapolation.
  6. I know space is slim, but metrics are not introduced. It would be great to at least categorize the metrics to understand what they measure. For now they are mostly acronyms. Moreover, ωsp\omega_{sp} is not defined before introducing it in the main.
  7. Limitations should be added to the discussion.
Questions for authors
  1. Is the neighborhood in Eq. 5 symmetrized?
  2. Why isn't MOSCOT compared with the other methods?
  3. Why are all the shapes of Fig. 5 the same for different methods? Also, what makes DRIFT better performing than Ordinary OT qualitatively? Or how exactly are intermediate overlays obtained?
  4. Since I am skeptical regarding qualitative extrapolations, I wonder if you could provide some overlays for it as well, similar to interpolation. It would be useful to see whether such a difficult task can still preserve target morphology and composition.
  5. Looking at Fig. 3, I feel the biggest advantage of method is that it is directly cell type informed, while the counterparts are not. Am I missing something? Are there other advantage beyond that? I believe it would be more fair to compare to standard OT with some sort of mask for inadmissible transitions. That would be a much more convincing counterpart.
5 comments

We thank the reviewer for the detailed feedback. Your advices help us a lot to revise our work. Below we clarify the methodological contribution, address limitations, and provide additional controlled comparisons and qualitative results.

(1) Weakness 1: We will define nmn_m as the number of spatial transcriptomic spots at stage tmt_m in Section 3.

(2) Weakness 2: We agree that spatial transcriptomics is moving toward single-cell resolution. Importantly, DRIFT can naturally extend to single-cell-resolution data without modifying its formulation, as single-cell observations are a special case of our composition-aware framework.

The developmental distance is defined as

+\tau\mathrm{KL}\left(\gamma \Big\|\ \theta_i\theta_j^\top\right)\}.$$ For single-cell observations, the compositions become one-hot vectors, $\theta_i=e_{a_i}$ and $\theta_j=e_{a_j}$, and the distance **naturally reduces to $D_{ij}^{\mathrm{single}}=C_{a_i a_j}^{\mathrm{dev}}$**. Thus, the same developmental alignment framework applies at both spot and single-cell resolutions, with no additional modeling components required. **To empirically examine this simplified representation**, we evaluate the Dominant variant, which approximates each mixed spot by a single cell type and therefore uses the same one-hot developmental cost as the single-cell case. Although this is not a direct single-cell-resolution benchmark, it assesses whether DRIFT remains effective when full cell-type mixtures are replaced by single-type representations. **Table 1. Comparison of Full DRIFT and Dominant across four spatiotemporal datasets.** Higher LCR and LSC and lower MLDD and CWLD indicate better alignment performance. | Dataset | Method | LCR ↑ | MLDD ↓ | CWLD ↓ | LSC ↑ | |---|---|---|---|---|---| | Pancreas | Full DRIFT | 0.845 | 0.152 | 0.284 | 0.485 | | | Dominant | 0.8643 | 0.1348 | 0.2783 | 0.4181 | | Intestine | Full DRIFT | 0.573 | 0.384 | 0.599 | 0.433 | | | Dominant | 0.6194 | 0.3527 | 0.5874 | 0.4488 | | Limb | Full DRIFT | 0.888 | 0.092 | 0.161 | 0.076 | | | Dominant | 0.8393 | 0.1304 | 0.2425 | 0.0825 | | Skin Wound | Full DRIFT | 0.682 | 0.316 | 0.357 | 0.237 | | | Dominant | 0.6438 | 0.3551 | 0.4361 | 0.1779 | **DRIFT provides a unified formulation across spatial resolutions, naturally reducing to cell-type-level developmental alignment at single-cell resolution**. We will clarifiy this applicability in the revised manuscript. **(3) Weakness 3:** We agree that velocity-informed Markov kernels and absorbing probabilities are established techniques, and **we do not claim them as independent innovations**. Our aim is to use biologically directed reachability as a criterion for **cross-time spatial alignment**, rather than terminal-fate inference. CellRank typically estimates the probability of eventually reaching terminal populations, which may cover only a subset of annotated types. By contrast, DRIFT needs a stage-dependent $K\times K$ reachability matrix over the cell-type vocabulary used for spatial alignmen. We therefore specify absorbing populations according to that context. It is true that CellRank can also implement this with a **user-defined absorbing state**, as we also explored in developing DRIFT. Under matched references, graphs, annotations, and absorbing states, the two kernels produce comparable alignment performance (Table 2), without a consistent winner. Table 2. Comparison of CellRank-based and DRIFT-based developmental transition modeling across four datasets. All variants use the same single-cell references, neighborhood graphs, cell-type annotations, absorbing endpoints. Higher LCR and LSC and lower MLDD and CWLD indicate better performance. | Dataset | Transition modeling | LCR ↑ | MLDD ↓ | CWLD ↓ | LSC ↑ | |---|---|---:|---:|---:|---:| | Pancreas | CellRank kernel | 0.8472 | 0.1503 | 0.3057 | 0.4821 | | | Drift kernel | 0.845 | 0.152 | 0.284 | 0.485 | | Intestine | CellRank kernel | 0.5589 | 0.3981 | 0.6016 | 0.4237 | | | Drift kernel | 0.573 | 0.384 | 0.599 | 0.433 | | Limb | CellRank kernel | 0.9048 | 0.0840 | 0.1594 | 0.0767 | | | Drift kernel | 0.888 | 0.092 | 0.161 | 0.076 | | Skin Wound | CellRank kernel | 0.6847 | 0.3135 | 0.3553 | 0.2410 | | | Drift kernel | 0.682 | 0.316 | 0.357 | 0.237 | We chose our own implementation mainly for **flexibility and computational simplicity**, not to replace CellRank; its distance-aware kernel is a task-specific choice, **not our central contribution**. DRIFT instead aggregates cell-level hitting probabilities into directed cell-type reachability and lifts this geometry to heterogeneous spatial spots via composition-aware OT, directly informing cross-time spatial correspondence.

(4) Weakness 4: We acknowledge that DRIFT uses established Markov modeling and FGW-OT. Our motivation is a concrete gap in all existing methods: expression-similar spatial observations can still represent biologically implausible developmental transitions. Existing methods model expression, geometry, population dynamics, or differentiation potential, but do not explicitly place cell-type developmental reachability into the cross-time spatial alignment cost (Table 3).

Table 3. Comparison of representative spatial and spatiotemporal alignment methods.

MethodMain Modeling StrategyDirected ReachabilityComposition-Aware Cost
Ordinary OTTranscriptomic similarityNoNo
FGW-OTExpression and spatial geometryNoNo
PASTE / PASTE2Spatial and expression alignmentNoNo
MOSCOTMultimodal and spatiotemporal OTNoNo
DeST-OTPopulation growth and developmental transportNoNo
STORIESSpatially informed differentiation potentialNoNo
stVCRDifferentiation, proliferation, and migrationNoNo
SOCSSpatially coherent trajectory inferenceNoNo
DRIFT (Ours)Velocity-guided developmental geometryYesYes

Our contribution is therefore not a new Markov-chain method or FGW solver: we construct directed cell-type reachability, translate it into a developmental cost between heterogeneous spatial observations, and integrate it with expression and geometry. Consequently, similarly expressed spots can receive different alignment costs according to their underlying developmental compatibility—something expression or spatial similarity alone cannot capture. As noted above, the cost naturally reduces to a cell-type transition cost at single-cell resolution. We have also clarified the FGW connection in Section 4.3 and refined our positioning in the Introduction and Related Work.

(5) Weakness 5: Thank you for this question. About what can be reconstructed without intermediate supervision and why coupling matters. First, observed spatial snapshots constrain the endpoints, while the single-cell reference informs plausible cell-type transitions. Multiple paths may fit these observations, so we evaluate population-level agreement with a measured held-out stage, rather than claiming to recover individual cell trajectories. Held-out interpolation is also used in CellBRIDGE (Estévez et al., 2026), Waddington-OT (Schiebinger et al., 2019), TrajectoryNet (Tong et al., 2020), PRESCIENT (Yeo et al., 2021), and ContextFlow (Rathod et al., 2025); stVCR and STORIES also evaluate interpolation/extrapolation on spatial omics data.

Second, coupling controls which source–target pairs drive the prediction. Under linear interpolation, x^(u)=(1−u)xi+uxj\hat x(u)=(1-u)x_i+ux_j for a coupled pair (i,j)(i,j), where xx contains expression and coordinates; different couplings therefore produce different intermediate distributions. In flow matching, these pairs also determine the displacements used to train the velocity field.

And we isolate this effect by keeping the predictor fixed across methods and evaluating predictions against measured held-out slices using expression, spatial, and composition metrics. The new figures in response to Questions 3–4 display predicted slices by each methods, hope they can give you a more clear view.

(6) Weakness 6: Thank you for your advice. We will revise Section 5.1 to group the metrics into alignment (LCR, MLDD, CWLD, LSC) and reconstruction/prediction (MMD, cell-type JS divergence, Moran's I error), and defined ωsp\omega_{\mathrm{sp}} at first use as the spatial-geometry weight.

(7) Weakness 7: Thank you for your advice. We will expand the Discussion to acknowledge reliance on single-cell reference availability and quality, RNA-velocity estimates, cell-type annotations, and deconvolution; inferred correspondences are probabilistic associations rather than experimentally validated cell trajectories, especially for extrapolation.

(8) Question 1: They are not symmetrized. N(i)\mathcal N(i) contains the kk nearest neighbors of cell ii in joint PCA space, without symmetrization; j∈N(i)j\in\mathcal N(i) need not imply i∈N(j)i\in\mathcal N(j).

(9) Question 2: We have added moscot to the alignment benchmarks (Table 4).

Table 4. Lineage alignment quality across developmental spatial datasets.

MethodLCR ↑MLDD ↓CWLD ↓LSC ↑LCR ↑MLDD ↓CWLD ↓LSC ↑
PancreasIntestine
Ordi. OT0.5390.4490.5570.2700.3540.5930.6110.267
FGW OT0.4320.5550.6570.2170.3580.5830.6420.263
PASTE0.2390.7480.7490.1120.2200.7110.7080.168
PASTE20.2680.7190.7220.0780.2430.6870.6750.101
DeST-OT0.5860.4060.4580.3000.3860.5600.5750.295
DRIFT0.8450.1520.2840.4850.5730.3840.5990.433
moscot0.4540.5350.6070.2220.3380.6060.6220.254
MethodLCR ↑MLDD ↓CWLD ↓LSC ↑LCR ↑MLDD ↓CWLD ↓LSC ↑
LimbSkin Wound
Ordi. OT0.8770.1160.2020.0700.6230.3750.4460.175
FGW OT0.8430.1470.2760.0670.5610.4350.5320.158
PASTE0.4250.5720.5730.0150.3320.6660.6710.089
PASTE20.3930.6030.5930.0090.2680.7290.7270.047
DeST-OT0.8390.1550.1580.0630.6530.3430.3870.185
DRIFT0.8880.0920.1610.0760.6820.3160.3570.237
moscot0.8290.1680.2160.0630.6800.3190.3550.173

(10) Question 3: Dear reviewer you may refer to Fig.4. We apologize that Figure 4 made morphology difficult to compare: the original panels shared the measured slice's outline because we displayed expression similarity scores on that same reference slice. Each method first predicted the intermediate slice using its own coupling; we then used FGW-OT to match its predicted spots to the held-out slice and visualized matched expression similarities on the reference outline. The held-out slice was used only after prediction, for visualization.

The new figure instead displays each method's actual predicted slice. Ordinary OT fills much of the gap between the two measured tissue regions, producing an almost continuous band, whereas DRIFT retains a clearer separation.

Interpolation prediction results: https://hackmd.io/_uploads/rkmz-sHAfe.png

(11) Question 4: We agree that extrapolation is difficult. Here it serves as a downstream test of alignment quality under fixed observed stages and predictor, assessed against a measured future slice; similar evaluations appear in ContextFlow (Rathod et al., 2025), scNODE (Zhang et al., 2024), PI-SDE (Jiang and Wan, 2024), stVCR (Peng et al., 2026), and STORIES (Huizing et al., 2026).

The new figure below shows the actual forecasts slices with the held-out slice. The measured tissue has two separated regions; Ordinary OT produces a largely continuous band, while DRIFT better preserves the observed separation and spatial organization.

Extrapolation prediction results: https://hackmd.io/_uploads/Sy6iHjrAze.png

(12) Question 5: We agree that leveraging biological information is an intended contribution of DRIFT—not an accidental advantage.

Following the reviewer's suggestion, we add a standard OT baseline with a binary mask defined by biological lineage rules: a source–target spot pair is allowed only if at least one cell-type pair present in the two spots has an allowed developmental transition; otherwise we impose Πij=0\Pi_{ij}=0. Table 5 compares this baseline with DRIFT.

Table 5. DRIFT versus standard OT with a developmental mask. Higher LCR and LSC and lower MLDD and CWLD indicate better alignment performance.

DatasetMethodLCR ↑MLDD ↓CWLD ↓LSC ↑
PancreasOrdinary OT0.5390.4490.5570.270
OT + developmental mask0.5150.4280.5180.258
Full DRIFT0.8450.1520.2840.485
IntestineOrdinary OT0.3540.5930.6110.267
OT + developmental mask0.3450.6000.6110.259
Full DRIFT0.5730.3840.5990.433
LimbOrdinary OT0.8770.1160.2020.070
OT + developmental mask0.8760.1180.1740.069
Full DRIFT0.8880.0920.1610.076
Skin WoundOrdinary OT0.6230.3750.4460.175
OT + developmental mask0.6720.3250.3660.212
Full DRIFT0.6820.3160.3570.237

From Table 5, DRIFT outperforms masked OT on all four metrics in Pancreas, Intestine, Limb and Skin Wound. Masked OT performs better in MLDD, and CWLD comparing to standard OT. Thus, a binary lineage constraint can be effective, but does not consistently reproduce the benefits.

Reviewer_s7aKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the alignment of unpaired spatial transcriptomic snapshots across time. DRIFT converts multi-step cell-type reachability inferred from RNA velocity into spot-level developmental costs through composition-aware optimal transport, then combines these costs with expression and spatial structure. Evaluation covers lineage consistency across four human datasets and held-out stage reconstruction and future extrapolation.

Critical strengths
  1. The paper is clearly structured, and the pipeline from local cellular dynamics to spot alignment is easy to follow.

  2. Across four datasets, the method improves multiple lineage-alignment metrics over the baselines.

Critical weaknesses
  1. The benchmarks omit complete dynamical frameworks such as STORIES and stVCR, and the shared-predictor extrapolation experiments do not establish DRIFT’s relative advantage in dynamical inference.

  2. Using full cell-type mixtures does not consistently help, as Dominant outperforms Full on all four alignment metrics in the intestine dataset (Table S6). DRIFT also uses cell-type and RNA-velocity information beyond that used by the main baselines.

Questions for authors
  1. How do the native STORIES and stVCR frameworks perform on reconstruction and extrapolation under comparable data and evaluation protocols? If DRIFT can enhance these frameworks, that would be even more helpful.

  2. Why does Dominant, which retains only the dominant cell type, perform better on the intestine dataset? Could you provide your interpretation of this result?

  3. Could you clarify how many cell types are used to construct the cell-type reachability matrix and how the cell-type annotations are obtained?

3 comments

We thank the reviewer for the constructive comments. We address the dynamical benchmarks, DRIFT-enhanced STORIES & stVCR, composition-aware modeling, and annotation details below.

(1) Weakness 1: Thank you for the question. We mainly use shared-predictor experiments to evaluate the quality of the inferred cross-time alignments, rather than claiming superiority over complete dynamical models. And this evaluation has been commonly used in previous studies (ContextFlow, scNODE, PI-SDE, STVCR, STORIES) . We also add native STORIES/stVCR comparisons under matched evaluation protocols (Shown in Question 1).

(2) Critical Weakness 2: We agree that full mixtures do not always help: Dominant improves all four Intestine metrics. We investigate when detailed compositions and optimized inner OT are beneficial, and independently assess the contributions of velocity and cell-type information.

We measure dominant-type concentration as ri=max⁡kθikr_i=\max_k\theta_{ik} and compare its mean and median with the original Dominant ablation (Table 1; Figure 1).

Table 1. Dominant-type concentration and ablation. Δ=Dominant−Full\Delta=\mathrm{Dominant}-\mathrm{Full} (Table S6); arrows show favorable directions.

DatasetMean Dominant Proportion ↑Median Dominant Proportion ↑ΔLCR ↑ΔMLDD ↓ΔCWLD ↓ΔLSC ↑
Pancreas0.6570.639+0.019−0.017−0.006−0.067
Intestine0.6900.721+0.046−0.031−0.012+0.016
Limb0.6400.672−0.049+0.038+0.082+0.007
Skin Wound0.5410.499−0.038+0.039+0.079−0.059

Figure 1. Distribution of dominant cell-type proportions. https://hackmd.io/_uploads/ry3QTEXAMe.png

Intestine has the highest dominant-type concentration (mean/median: 0.6902/0.7208), consistent with Dominant preserving much of the relevant composition information and improving all four metrics. Skin Wound has the lowest (0.5408/0.4993), where Dominant worsens all four metrics, suggesting greater importance of non-dominant types in heterogeneous spots.

Separating information contributions. Because DRIFT uses more biological information than some baselines, we add controlled ablations for velocity and composition (Tables 3–4), holding all other components fixed.

Table 3. RNA velocity ablation. Same reference cells, neighborhood graphs, cell-type labels, absorbing endpoints, spot compositions, and optimization settings.

DatasetVelocity SettingLCR ↑MLDD ↓CWLD ↓LSC ↑
Pancreasw/o Velocity0.7820.2040.3230.439
Shuffled Velocity0.7510.2380.3350.423
Reverse Velocity0.6500.3220.4260.364
Full DRIFT0.8450.1520.2840.485
Intestinew/o Velocity0.5430.4140.6290.362
Shuffled Velocity0.5300.4280.6320.358
Reverse Velocity0.5380.4190.6300.359
Full DRIFT0.5730.3840.5990.433
Limbw/o Velocity0.8490.1260.1950.075
Shuffled Velocity0.8440.1270.1820.075
Reverse Velocity0.7760.1910.2220.075
Full DRIFT0.8880.0920.1610.076
Skin Woundw/o Velocity0.5380.4570.4170.169
Shuffled Velocity0.5190.4790.4270.172
Reverse Velocity0.5240.4730.4310.168
Full DRIFT0.6820.3160.3570.237

Table 4. Cell-type composition ablation. Same velocity-guided transitions and reachability, reference cells, neighborhood graphs, absorbing endpoints, and optimization settings.

DatasetComposition SettingLCR ↑MLDD ↓CWLD ↓LSC ↑
PancreasUniform Composition0.4360.5010.6300.220
Shuffled Composition0.3840.5630.6320.188
Full DRIFT0.8450.1520.2840.485
IntestineUniform Composition0.3480.5950.6420.254
Shuffled Composition0.3370.6000.6440.253
Full DRIFT0.5730.3840.5990.433
LimbUniform Composition0.8840.1150.2480.069
Shuffled Composition0.8500.1470.2500.064
Full DRIFT0.8880.0920.1610.076
Skin WoundUniform Composition0.7140.2810.3980.181
Shuffled Composition0.6710.3240.4050.169
Full DRIFT0.6820.3160.3570.237

Full DRIFT outperforms all velocity-ablated variants across all four datasets and metrics (Table 3), demonstrating the distinct value of velocity-derived directionality beyond annotations and reference geometry. Uniform or shuffled compositions degrade most metrics (Table 4), showing an additional contribution from informative spot mixtures. Together these controls distinguish the benefits of directed development from merely adding cell-type information. We will clarify these findings and the limitations of composition-aware OT in the revision.

(3) Question 1: We are so grateful to you for giving such a very constructive advice. We just realized that our method can also be integrated in stVCR and STORIES. We compare native STORIES & stVCR and DRIFT-enhanced STORIES & stVCR under matched observed/held-out stages. The results are shown in Table 5&6. And the details of how integrating DRIFT into these two methods are also provided behind. We will add this new improvements in our revised paper.

Table 5. Interpolation benchmark. Reconstruction of withheld stages using MMD, Moran’s I error, and cell-type JS divergence. Lower values are better.

MethodPancreasIntestineLimbSkin wound
MMDMoran ICell JSMMDMoran ICell JSMMDMoran ICell JSMMDMoran ICell JS
Linear Methods
Ordi. OT0.7120.1860.0230.6210.4300.1280.4270.3770.0500.3110.4220.071
FGW OT0.6410.1970.0240.6580.3930.1280.4420.3630.0440.3250.4710.071
PASTE0.7500.1430.0210.4840.2220.1290.3870.1910.0330.3290.2000.071
PASTE20.6460.2820.0220.4840.3590.1370.4280.2380.0390.3140.3320.060
DeST-OT0.5440.1490.0240.6650.3580.1270.3830.3080.0330.3110.4120.065
DRIFT0.5430.1540.0210.4690.3390.1220.3930.2890.0260.3030.3130.055
moscot0.5510.1700.0390.4710.3610.0690.3910.3610.0280.3110.4230.049
Flow matching Methods
Ordi. OT0.5630.3450.0220.4540.2280.1280.3990.1940.0500.3380.3600.057
FGW OT0.6000.3220.0230.4760.2480.1280.3770.1900.0560.3280.3480.057
PASTE0.6670.2120.0220.4580.2390.1280.4140.1900.0500.3360.2730.065
PASTE20.5660.4600.0250.5060.3460.1360.4210.1880.0390.3260.3030.063
DeST-OT0.5560.2860.0220.5390.2560.1280.3680.1930.0450.3310.3640.057
DRIFT0.5500.1220.0210.4430.1980.1200.3680.1820.0270.3070.2330.056
moscot0.5550.1740.0380.4760.4300.0670.3920.4000.0260.3080.5020.056
Dynamic prediction methods
STORIES0.5820.1250.0430.4650.2590.0650.3740.1490.0260.3080.2290.059
stVCR0.6670.1380.0470.4500.2910.0730.3900.2360.0380.3930.3750.069
STORIES + DRIFT0.5730.1180.0390.4550.2570.0610.3670.1450.0250.3050.2480.041
stVCR + DRIFT0.6800.1320.0420.3570.3460.0470.2940.2340.0370.3940.3710.047

Table 6. Extrapolation benchmark. Prediction beyond the observed time range measured by MMD, Moran’s I error, and cell-type JS divergence. Lower is better.

MethodPancreasIntestineLimbSkin wound
MMDMoran ICell JSMMDMoran ICell JSMMDMoran ICell JSMMDMoran ICell JS
Linear Methods
Ordi. OT0.7550.2370.0280.6900.3770.3910.7320.2490.1720.6150.3900.086
FGW OT0.7170.1990.0270.6880.3930.3790.7070.2430.1590.6260.3770.092
PASTE0.6930.1780.0220.7080.4140.3740.7100.3980.1420.6680.1860.086
PASTE20.6950.2810.0370.7110.5240.3880.7150.3350.1620.6700.2230.127
DeST-OT0.7210.2200.0280.6800.3290.4250.7110.1810.1680.6170.3710.089
DRIFT0.7050.1740.0130.6640.3610.3800.7020.1890.1570.6210.3390.088
moscot0.7040.1920.0440.7020.3660.4250.7290.2240.1610.6100.3460.069
Flow Matching Methods
Ordi. OT0.7560.2690.0220.7320.3480.3940.7480.1790.1720.6320.4960.085
FGW OT0.7240.2540.0230.7310.3560.3810.7370.2130.1600.6310.4740.086
PASTE0.6980.2660.0240.7310.4720.3720.7400.3820.1410.6530.3550.086
PASTE20.7240.3010.0320.7320.5370.3910.7440.6100.1570.6740.4360.127
DeST-OT0.7160.2540.0240.7310.2570.4220.7380.2040.1690.6280.4500.091
DRIFT0.7090.1820.0130.7230.2180.3770.7270.1650.1520.6050.1700.085
moscot0.7580.2780.0540.7520.2930.4970.7760.3070.1830.6390.3830.071
Dynamic prediction methods
STORIES0.8760.2380.1410.6710.2350.4460.8070.4360.1830.7240.3000.177
stVCR0.4570.2730.1610.7670.3910.3210.8430.2910.2130.7600.3390.092
STORIES + DRIFT0.6930.1980.0170.6720.2330.4480.7470.4010.1270.7260.3010.176
stVCR + DRIFT0.7010.1700.0421.1050.4010.0920.8360.2470.2010.5970.2350.071

As shown in Tables 5 and 6, integrating DRIFT consistently improves STORIES (19/24 cases) and benefits stVCR in most cases (18/24). These results demonstrate that DRIFT provides a valuable developmental prior that can effectively enhance existing dynamical prediction methods, highlighting its flexibility and broader applicability. DRIFT with shared flow matching also achieves lower extrapolation Moran's I error than native STORIES and stVCR on all four datasets, while results vary across other metrics.

Integration of DRIFT. We add our developmental cost to the native cross-time matching objectives of stVCR and STORIES. Let DijD_{ij} denote the developmental cost, with its scale absorbed into λ\lambda.

For stVCR, the augmented OT matching objective is

LmatchstVCR+DRIFT=min⁡π∈Π(a,b)∑i,jπij[κ∣x^i−xj∣22+(1−κ)∣r^i−rj∣22+λDij].\mathcal{L}_{\mathrm{match}}^{\mathrm{stVCR+DRIFT}} = \min_{\pi \in \Pi(a,b)} \sum_{i,j} \pi_{ij} \left[ \kappa \\|\widehat{x}_i-x_j\\|_2^2 + (1-\kappa) \\|\widehat{r}_i-r_j\\|_2^2 + \lambda D_{ij} \right].

For STORIES, the augmented FGW cross-matching objective is

\mathcal{F}_{\lambda}(A,B) = \min_{\pi \in \Pi(a,b)} \left\\{ (1-\alpha) \sum_{i,j} \pi_{ij} \\|\widehat{x}_i-x_j\\|_2^2 + \alpha \sum_{i,k,j,l} (S^A_{ik}-S^B_{jl})^2 \pi_{ij}\pi_{kl} + \lambda \sum_{i,j} \pi_{ij} D_{ij} + \varepsilon \mathrm{KL}(\pi \parallel ab^\top) \right\\}.

STORIES retains its native self-matching correction:

LmatchSTORIES+DRIFT=Fλ(A,B)−12F0(A,A)−12F0(B,B).\mathcal{L}_{\mathrm{match}}^{\mathrm{STORIES+DRIFT}} = \mathcal{F}_{\lambda}(A,B) - \frac{1}{2}\mathcal{F}_0(A,A) - \frac{1}{2}\mathcal{F}_0(B,B).

Here, AA and BB are the predicted and observed distributions, xx denotes expression features, rr denotes spatial coordinates, and SA,SBS^A,S^B contain within-distribution squared spatial distances. The coupling π\pi has the native marginals a,ba,b. All other native model components and regularization terms are retained.

(4) Question 2: Why does Dominant outperform Full on Intestine? As shown under Weakness 2 (Table 1; Figure 1), Intestine has the highest dominant-type concentration, so retaining only its leading type may preserve most relevant information while reducing uncertainty from minor components.

(5) Question 3: Cell-type numbers and annotation sources. The annotation sources and type vocabularies are listed below.

Table 7. Cell-type counts and annotation sources.

DatasetCell Types (KK)Single-cell Type AnnotationSpatial Composition Source
Pancreas10Annotated by us using Seurat clustering and cell-type markers (GSE197064)Provided in the original dataset (GSE197317)
Intestine7Already annotated by the authors in the original data (GSE158702)Computed by us using RCTD (spacexr) with full mode and normalize weights
Limb10Author-provided labels from CELLxGENEAuthor provided from CELLxGENE (E-MTAB-10367)
Skin Wound10Already annotated in the original dataset (GSE241132)Author provided from Xu Landen WebAtlas

Table 8. Cell-type vocabularies.

DatasetTypes used in TT and spatial compositions
PancreasMesenchymal; Immune; Endothelial; Ductal; Schwann; Endocrine progenitors; Acinar; Alpha; Delta; Beta
IntestineEpithelium; Fibroblasts; Myofibroblasts; Neural; Pericytes; Endothelium; Muscularis
LimbMesenchymal; Chondrogenic; Osteogenic; Tendon; Myogenic; Endothelial; Epithelial; Neural; Hematopoietic; Melanocyte
Skin WoundKeratinocyte; Fibroblast; Myeloid; Lymphoid; Endothelial; Mast; Pericyte and smooth muscle; Melanocyte; Schwann; Hair follicle
Reviewer_1UoqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DRIFT, a framework for aligning unpaired spatial transcriptomic snapshots across developmental time by incorporating developmental reachability, rather than relying solely on transcriptomic similarity and spatial organization. DRIFT constructs a velocity-informed Markov process from a corresponding single-cell reference and uses first-hitting probabilities to estimate directed reachability between cell types. It then converts this information into a spot-level developmental cost through optimal transport between cell-type compositions. The final alignment objective combines normalized transcriptomic, developmental, and spatial costs with entropy regularization. The method is evaluated on four human datasets covering pancreas development, intestinal development, limb morphogenesis, and skin wound healing. The evaluation considers cross-time spatial alignment and held-out stage reconstruction and prediction, with shared linear and flow-matching predictors used for interpolation and extrapolation. The reported results show improved dataset-level lineage consistency and competitive held-out reconstruction, although the gains vary across tissues, predictors, and evaluation metrics.

Critical strengths
  1. The paper addresses an important limitation of similarity-based alignment: two states can be close in expression space without representing a plausible developmental transition. Using directed reachability to inform alignment provides an interpretable way to address this ambiguity. Furthermore, translating cell-type reachability into a cost between mixed-cell spatial spots offers a natural connection between single-cell dynamics and spatial measurements. This gives the method a coherent biological motivation and strengthens the significance of the proposed alignment framework.

  2. The benchmark is not restricted to a single cross-modality matching setting. For pancreas, the modalities come from the same study and developmental stages, but donor-level correspondence is not established. For intestine, the single-cell reference has denser temporal coverage than the spatial measurements, and the modalities are treated as population-level rather than pointwise matched observations. These settings broaden the empirical evaluation and provide evidence of the method’s applicability when strict cross-modality pairing is unavailable. Although they do not constitute a controlled study of reference mismatch, they make the evaluation more informative than testing only a closely matched reference–spatial pair.

Critical weaknesses
  1. The “w/o development” ablation removes the entire developmental cost. It therefore removes both the velocity-derived directional structure and the cell-type composition information incorporated through that term. While the resulting degradation in alignment metrics supports the usefulness of the combined developmental cost, it does not establish how much of the improvement comes specifically from RNA-velocity directionality rather than additional cell-type information or the geometry of the single-cell reference. A matched control that removes or shuffles velocity information in the transition weights would more directly isolate this contribution. Such a comparison should retain the same reference cells, neighborhood graph, cell-type labels, absorbing endpoints, spot compositions, and optimization settings while recomputing the transition and reachability matrices. This would test the incremental value of velocity-guided transitions conditional on the same endpoint selection.

  2. Table S6 shows that the Dominant variant, which replaces each spot’s mixture with its dominant cell type, improves all four intestinal alignment metrics relative to the full model. In skin wound, the Independent variant, which replaces the inner OT problem with an independent-composition cost, improves LCR, MLDD, and CWLD, although it reduces LSC. These results do not negate the benefits of composition-aware OT in other settings, but they leave unclear when this component is preferable to simpler alternatives. Since lifting reachability through full cell-type compositions is presented as a central contribution, a clearer characterization of the conditions under which it improves alignment would strengthen the paper. For example, relating the observed differences to spot heterogeneity or composition-estimation quality could help explain the tissue-dependent results. Without such an explanation, the empirical justification for the additional modeling complexity remains incomplete.

Questions for authors

Clarification requests:

  1. Appendix B.1 describes the compositions as precomputed inputs but does not clearly identify the deconvolution procedure. Could the authors specify the method, reference data, and principal settings used for each dataset, including whether these compositions were recomputed or obtained from the original studies?

  2. The manuscript reports shared settings of (α,ωsp,λ)=(0.5,1.0,1.0)(\alpha,\omega_{\mathrm{sp}},\lambda)=(0.5,1.0,1.0) across datasets. Were these values fixed in advance or selected using validation data, and what criterion guided their selection? In particular, were held-out prediction results or evaluation lineage scores used in that process? Clarifying the selection procedure would help assess how dependent the reported improvements are on configuration choices and whether the baseline comparisons use comparable tuning protocols.

2 comments

We thank the reviewer for the constructive feedback. We address the ablations, methodological details, and hyperparameter settings below.

(1) Weakness 1: We are very grateful for this important and constructive advice. Removing the entire developmental cost conflates velocity directionality and cell-type information. We therefore isolate these contributions with matched ablations (Tables 1–2).

Table 1. RNA velocity ablation. Reference cells, neighborhood graphs, cell-type labels, absorbing endpoints, spot compositions, and optimization settings are identical across variants.

DatasetVelocity settingLCR ↑MLDD ↓CWLD ↓LSC ↑
Pancreasw/o Velocity0.7820.2040.3230.439
Shuffled Velocity0.7510.2380.3350.423
Reverse Velocity0.6500.3220.4260.364
Full DRIFT0.8450.1520.2840.485
Intestinew/o Velocity0.5430.4140.6290.362
Shuffled Velocity0.5300.4280.6320.358
Reverse Velocity0.5380.4190.6300.359
Full DRIFT0.5730.3840.5990.433
Limbw/o Velocity0.8490.1260.1950.075
Shuffled Velocity0.8440.1270.1820.075
Reverse Velocity0.7760.1910.2220.075
Full DRIFT0.8880.0920.1610.076
Skin Woundw/o Velocity0.5380.4570.4170.169
Shuffled Velocity0.5190.4790.4270.172
Reverse Velocity0.5240.4730.4310.168
Full DRIFT0.6820.3160.3570.237

Table 2. Cell-type composition ablation. Velocity-guided transitions and reachability, reference cells, neighborhood graphs, absorbing endpoints, and optimization settings are fixed.

DatasetComposition settingLCR ↑MLDD ↓CWLD ↓LSC ↑
PancreasUniform Composition0.4360.5010.6300.220
Shuffled Composition0.3840.5630.6320.188
Full DRIFT0.8450.1520.2840.485
IntestineUniform Composition0.3480.5950.6420.254
Shuffled Composition0.3370.6000.6440.253
Full DRIFT0.5730.3840.5990.433
LimbUniform Composition0.8840.1150.2480.069
Shuffled Composition0.8500.1470.2500.064
Full DRIFT0.8880.0920.1610.076
Skin WoundUniform Composition0.7140.2810.3980.181
Shuffled Composition0.6710.3240.4050.169
Full DRIFT0.6820.3160.3570.237

Full DRIFT outperforms all three velocity-ablated variants on all metrics across all four datasets (Table 1), demonstrating that velocity-derived directionality adds value beyond cell-type annotations and reference geometry. Uniform or shuffled compositions worsen most metrics (Table 2), confirming a separate contribution from informative spot mixtures. These matched controls distinguish the two sources of improvement.

(2) Weakness 2 part1, about Dominant variant: We agree that Full does not always outperform simpler variants. We investigate whether spot heterogeneity helps explain this tissue dependence, while acknowledging that composition-estimation quality may also matter.

We measure dominant-type concentration by ri=max⁡kθikr_i=\max_k\theta_{ik} and report its dataset-wise mean and median alongside the submitted Dominant ablation (Table 3); Figure 1 shows the distributions.

Table 3. Dominant-type proportions and submitted ablation. Δ=Dominant−Full\Delta=\mathrm{Dominant}-\mathrm{Full} ; ↑\uparrow/↓\downarrow indicate favorable directions.

DatasetMean Dominant Proportion ↑Median Dominant Proportion ↑ΔLCR ↑ΔMLDD ↓ΔCWLD ↓ΔLSC ↑
Pancreas0.65660.6389+0.0193−0.0172−0.0057−0.0669
Intestine0.69020.7208+0.0464−0.0313−0.0116+0.0158
Limb0.64020.6723−0.0487+0.0384+0.0815+0.0065
Skin Wound0.54080.4993−0.0382+0.0391+0.0791−0.0591

Figure 1. Boxplot of dominant cell-type proportions in each dataset: https://hackmd.io/_uploads/ry3QTEXAMe.png

Intestine has the highest dominant-type concentration (mean/median: 0.6902/0.7208), consistent with Dominant improving all four metrics because the leading type may retain much of the relevant information. Skin Wound has the lowest (0.5408/0.4993), where Dominant worsens all four metrics, suggesting that non-dominant types matter more in heterogeneous spots.

(2) Weakness 2 part2, about Independent variant: Independent retains full compositions but replaces optimized inner OT with the independent coupling Γind=pq⊤\Gamma^{\mathrm{ind}}=pq^\top, giving Dind(p,q)=p⊤GqD_{\mathrm{ind}}(p,q)=p^\top Gq. We compare it with Full across low-, medium-, and high-heterogeneity spot pairs under identical settings (Table 4).

Table 4. Full versus Independent by spot heterogeneity. Δ=Independent−Full\Delta=\mathrm{Independent}-\mathrm{Full}.

DatasetHeterogeneityΔLCR ↑ΔMLDD ↓ΔCWLD ↓ΔLSC ↑
PancreasLow+0.0227-0.0213-0.0209+0.0094
Medium-0.0892+0.0711+0.0313-0.0236
High-0.1188+0.1023+0.0825-0.0118
IntestineLow+0.0517-0.0412-0.0056+0.0293
Medium-0.0959+0.1005+0.0097-0.0212
High-0.1360+0.0992+0.0401-0.0585
LimbLow+0.0475-0.0428-0.0120-0.0001
Medium+0.0073+0.0070+0.0206+0.0019
High-0.1038+0.1083+0.0267-0.0046
Skin WoundLow+0.0303-0.0300-0.0400-0.0105
Medium-0.0980+0.0915+0.0602-0.0199
High-0.0024+0.0024+0.0899-0.0002

Independent has higher LCR and lower MLDD across all low-heterogeneity strata, whereas Full has both advantages across all high-heterogeneity strata (Table 4). Thus, optimized cell-type transport is most valuable when multiple cell-type transitions matter, while Independent may suffice for more homogeneous spots. We will discuss these in the revision.

About how to divide heterogeneity groups. We average the normalized Shannon entropies of each source–target spot pair and split all candidate pairs within each stage pair into low, medium, and high groups using the 1/3 and 2/3 quantiles.

(3) Question 1: Deconvolution details. The composition sources and processing settings are summarized below.

DatasetComposition Source / MethodReference DataProcessing
PancreasProvided by the original authorsGSE197317-
IntestineComputed by us using RCTD (spacexr)GSE158702 (scRNA-seq); GSE158328 (spatial)Full mode; normalize weights
LimbProvided by the original authorsE-MTAB-10367-
Skin WoundProvided by the original authors through WebAtlasXu Landen WebAtlas; GSE241124-

(4) Question 2: Hyperparameter selection. Following CellBRIDGE (Estévez et al., 2026) and ContextFlow (Rathod et al., 2025), we fixed (α,ωsp,λ)=(0.5,1.0,1.0)(\alpha,\omega_{sp},\lambda)=(0.5,1.0,1.0) for all datasets, without dataset-specific tuning or using held-out predictions or lineage evaluation scores for selection. We additionally vary each parameter independently while holding the others fixed (Table 5).

Table 5. Hyperparameter sensitivity. Equal averages across four datasets; arrows indicate favorable directions.

ParameterValueLCR ↑MLDD ↓CWLD ↓LSC ↑
α\alpha0.30.73990.24050.33500.3013
0.5 (Default)0.7470.2360.3500.308
0.70.75140.23280.36910.3118
0.90.73060.24800.40350.3046
ωsp\omega_{sp}0.50.73190.24750.33580.3008
1.0 (Default)0.7470.2360.3500.308
2.00.75790.22730.37370.3114
5.00.72980.24870.44050.3009
λ\lambda0.50.75440.23020.36520.3033
1.0 (Default)0.7470.2360.3500.308
2.00.72920.25090.34380.3000
5.00.68090.29320.35930.2831

DRIFT remains relatively stable under moderate parameter changes; the default is competitive across metrics, while improvements under alternative settings generally involve trade-offs.

Comments on the paper

1 comment
Area_Chair_UnCYOpenReview ↗

Please Join the Discussion

Dear Reviewers,

The authors have submitted their responses. Please read them and provide your feedback. Thank you!

Best regards, Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.