Organizations: Center for Quantitative Biology, Peking University · Center for Machine Learning Research, Peking University · LMAM and School of Mathematical Sciences, Peking University · National Engineering Laboratory for Big Data Analysis and Applications, Beijing · AI for Science Institute, Beijing
Learning the spatiotemporal dynamics of cells from spatial transcriptomics requires connecting snapshots of different cells acquired in different coordinate frames. A rotation or translation of a section can appear as cell movement and distort the inferred correspondence. We introduce horizontal flow matching, which learns changes in tissue organization, molecular state, and cell abundance up to shared rigid motion. The dynamics follow a continuity--reaction equation whose action assigns no cost to shared rigid motion, and the resulting horizontal projector removes this component from spatial prediction errors during learning and from predicted velocities during generation. An equivalent endpoint problem jointly estimates correspondence and relative pose in Wasserstein--Fisher--Rao geometry and yields analytic paths whose velocity and growth serve as regression targets. Regressing these targets recovers the horizontal dynamics under ideal population conditions. Our model, HORIZON, uses an equivariant transformer to predict velocity and growth from the current tissue, so that a tissue can be generated continuously from a single initial snapshot. On a simulated tissue observed in perturbed coordinate frames, alignment improves reconstruction and projection reduces rigid drift. On four spatial transcriptomics time courses, HORIZON reconstructs tissue geometry, gene expression, and cell abundance more accurately than existing spatial and unbalanced trajectory methods, and interpolates held-out developmental stages.
Figures & tables
Figure 1: Horizontal flow matching with Horizon . (a) Snapshots contain different cells in unrelated frames. (b) Quotient optimal entropy transport (qOET) jointly estimates a rigid pose and an unbalanced correspondence between cells. (c) Closed-form WFR bridges give intermediate states, masses, velocities, and growth rates. (d) An equivariant transformer learns velocity and growth from these states, and the horizontal projector is applied both in the spatial loss and to the velocity used for integration. Panels (a–c) use the 18–24 hpf zebrafish interval.
Direct spatial SW
Severe regime
Method
n
Canonical
Severe
Quotient sp. SW
Quotient joint SW
Gene SW
Mass RE (%)
Horizon (ours)
9/9
0.111±0.004
0.115±0.003
0.042±0.002
0.037±0.001
0.070±0.005
0.02±0.01
registered, no projector
9/9
–
0.127±0.004
0.039±0.002
0.034±0.001
0.073±0.007
0.02±0.01
no quotient (raw, no projector)
9/9
–
1.246±0.093
0.113±0.017
0.121±0.039
0.195±0.065
0.23±0.04
no growth (reaction set to zero)
9/9
–
0.117±0.003
0.045±0.003
0.039±0.002
0.085±0.004
9.32±0.00
WFR-FM
9/9
0.223±0.165
1.213±0.027
0.110±0.008
0.142±0.016
0.298±0.028
0.14±0.03
Table 1: Simulated tissue under hidden changes of coordinate frame. Mean ± standard error over completed fits ( n = completed/attempted, severe regime). Direct spatial SW is scored against intrinsic ground truth without alignment, measuring sensitivity to the frames used for training, whereas quotient metrics are computed after SE(2) alignment. The indented rows use the same network and training settings as Horizon . Registered, no projector aligns the snapshots in advance with the qOET poses and then trains and integrates without the projector, no quotient does the same on the raw snapshots, and no growth sets the reaction field to zero.
Dataset
Method
Protocol
Joint SW
Joint MMD
Spatial SW
Chamfer
Gene SW
Gene MMD
Mass RE (%)
Axolotl
Horizon
first
0.063
0.003
0.240
0.367
0.122
0.018
0.09
WFR-FM
first
1.441
0.167
5.907
2.860
1.171
0.214
146.3
stVCR
first
1.040
0.218
4.294
1.008
0.248
0.044
1.30
STORIES
first
–
–
–
–
1.087
0.207
–
Horizon
restarted
0.054
0.001
0.224
0.323
0.046
0.006
–
NicheFlow raw
restarted
0.201
0.025
0.843
0.360
0.279
0.060
–
Table 2: Reconstruction of real spatial series. We evaluate full-trajectory generation from the initial snapshot ( first ) and interval-wise generation restarted from each observed source snapshot ( restarted ). All scores average targets after the initial snapshot under proper SE(2) evaluation alignment, where lower values are better and boldface marks the best entry under each protocol. Horizon (first) and WFR-FM average three seeds, other rows use one seed, and Horizon (restarted) evaluates one trained model (seed 17). Dashes mark outputs not predicted or scored (Appendix D ).
Figure 2: Reconstructing zebrafish development from the initial snapshot. Observations are shown above and a representative Horizon trajectory below, with time in hours post fertilization. Predictions use no restarts and are aligned for evaluation. Points carry equal mass so point count reflects abundance, while colors show observed annotations above and annotations decoded from predicted expression below (Appendix J ).
Projection
Axolotl
Chicken
Drosophila
Zebrafish
Variant
loss
gen.
Gene SW
Direct SW
Gene SW
Direct SW
Gene SW
Direct SW
Gene SW
Direct SW
Horizon
✓
✓
0.123
0.239
0.581
0.215
0.103
0.041
0.283
0.192
registered, no projector
0.126
0.228
0.641
0.260
0.107
0.043
0.289
0.208
projection in loss only
✓
0.123
0.549
0.581
0.243
0.103
0.281
0.283
0.741
projection in generation only
✓
0.126
0.231
0.641
0.213
0.107
0.041
0.289
0.201
Table 3: Registration and projection on real data. All rows use the same qOET bridges (written in the frame of the first snapshot for the ambient loss) and the same network and optimizer, averaged over seeds 17, 29, and 43. Gene SW is frame-free, while direct spatial SW scores the generated trajectory against targets registered to the first frame by the composed qOET poses and thus also counts rigid drift. Appendix K gives all metrics and standard deviations.
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Times ∗
Full cells per snapshot
B
λq
Reference seed
Time input
Axolotl
2, 5, 10, 15, 20
7,668, 8,106, 9,440, 9,676, 11,319
1,024
0.15
17
piecewise Fourier
Chicken
4, 7, 10, 14
147, 528, 908, 1,967
768
0.05
17
piecewise Fourier
Drosophila
5.6, 6.3, 7, 9.1, 10.5
1,263, 1,248, 1,039, 1,078, 1,110
768
0.05
17
piecewise Fourier
Zebrafish
5.25, 10, 12, 18, 24
563, 1,036, 2,081, 3,048, 5,271
1,024
0.05
17
piecewise Fourier
Appendix
Table 4: Shared Horizon configuration. One architecture and optimizer are used on every real dataset (9,000 updates, width 384, five blocks, eight heads, feed-forward width 768, initial learning rate 1.5×10−3 with cosine decay to 2×10−5 , EMA 0.995, gradient clip 100, δ=10 , λx=1 , and projected RK4 maximum normalized step 0.25). B is the number of bridge states per update. Figures use seed 17, and Table 8 also includes seeds 29 and 43. ∗ For Drosophila, section depth.
Figure 3: Spatial comparison at the final snapshot. All methods use the same inputs and display canvas, and predictions are integrated from the first snapshot and aligned by the evaluation pose. Each displayed point carries equal mass so point count shows total mass, and colour shows the shared molecular PCs.
Method
Regime
n
Direct Wx
Direct Wx,g
WxQ
Wx,gQ
Horizon (ours)
Canonical
3/3
0.111±0.004
0.079±0.005
0.040±9.84×10−4
0.035±9.90×10−4
Mild
9/9
0.114±0.004
0.082±0.003
0.039±0.002
0.037±0.002
Severe
9/9
0.115±0.003
0.081±0.002
0.042±0.002
0.037±0.001
WFR-FM
Canonical
2/3
0.223±0.165
0.295±0.222
0.219±0.162
0.293±0.221
Mild
9/9
0.337±0.007
0.227±0.004
0.075±0.003
0.087±0.005
Severe
9/9
1.213±0.027
0.786±0.016
0.110±0.008
0.142±0.016
Appendix
Table 5: Direct and quotient sliced-Wasserstein metrics on the frame-perturbation benchmark (mean ± standard error over successful runs, where n reports successes/attempts and lower is better).
Method
Regime
n
Direct Chamfer
Direct joint MMD
Quotient Chamfer
Quotient joint MMD
Horizon (ours)
Canonical
3/3
0.116±0.005
0.023±0.002
0.096±1.91×10−4
0.006±2.58×10−4
Mild
9/9
0.117±0.004
0.024±0.001
0.095±6.91×10−4
0.006±5.75×10−4
Severe
9/9
0.119±0.004
0.023±9.84×10−4
0.096±5.36×10−4
0.006±3.90×10−4
WFR-FM
Canonical
2/3
0.273±0.164
0.020±1.37×10−5
0.272±0.163
0.019±3.48×10−4
Mild
9/9
0.227±0.009
0.085±0.002
0.127±0.003
0.021±0.001
Severe
9/9
0.721±0.026
0.311±0.007
0.148±0.009
0.033±0.006
Appendix
Table 6: Direct and quotient Chamfer and joint-MMD metrics on the frame-perturbation benchmark (mean ± standard error over successful runs, where n reports successes/attempts and lower is better).
Method
Regime
n
Wg
Gene MMD
Mass RE
Horizon (ours)
Canonical
3/3
0.076±0.008
0.011±0.002
2.68×10−4±1.06×10−4
Mild
9/9
0.084±0.008
0.012±0.002
2.03×10−4±4.93×10−5
Severe
9/9
0.070±0.005
0.010±9.51×10−4
1.92×10−4±5.57×10−5
WFR-FM
Canonical
2/3
0.484±0.298
0.078±0.007
7.10×10−4±9.89×10−5
Mild
9/9
0.204±0.011
0.068±0.004
6.23×10−4±4.72×10−5
Severe
9/9
0.298±0.028
0.115±0.009
0.001±3.00×10−4
Appendix
Table 7: Gene-space and finite-mass metrics on the frame-perturbation benchmark (mean ± standard error over successful runs, where n reports successes/attempts and lower is better).
Figure 4: Trajectories under severe frame perturbations. Observed snapshots and predictions of Horizon , WFR-FM, and stVCR at three times for the first severe variant. Colors show the three molecular programs, and all panels share the canvas of the observations after evaluation alignment.
Dataset
Method
Joint SW
Joint MMD
Spatial SW
Chamfer
Gene SW
Gene MMD
Mass RE (%)
Axolotl
Horizon
0.0629 ± 0.0049
0.0034 ± 0.0008
0.2400 ± 0.0223
0.3667 ± 0.0103
0.1218 ± 0.0037
0.0178 ± 0.0005
0.0912 ± 0.0173
WFR-FM
1.4411 ± 0.1103
0.1669 ± 0.0294
5.9074 ± 0.4223
2.8603 ± 0.4860
1.1715 ± 0.2544
0.2137 ± 0.0466
146.2522 ± 67.1410
stVCR
1.0405
0.2180
4.2944
1.0078
0.2482
0.0436
1.2956
Chicken
Horizon
0.1341 ± 0.0071
0.0367 ± 0.0011
0.1994 ± 0.0269
0.3916 ± 0.0099
0.5971 ± 0.0067
0.0982 ± 0.0079
0.4523 ± 0.1870
WFR-FM
0.1873 ± 0.0343
0.0637 ± 0.0178
0.5397 ± 0.1989
0.5761 ± 0.1281
0.7513 ± 0.0141
0.1352 ± 0.0049
1.5590 ± 0.6506
stVCR
0.5765
0.2048
1.8059
2.0790
1.5178
0.2617
99.1742
Appendix
Table 8: Free-generation results of Table 2 with seed variation. Lower is better. Horizon and WFR-FM report mean ± sample standard deviation over three seeds, and stVCR uses one seed.
Horizon
stVCR
stVCR/ours
Dataset
qOET
NN
fit
NN
fit speedup
Axolotl
794.1
298.3
1,092.4
12,424.4
11.4 ×
Chicken
465.5
281.2
746.7
1,950.3
2.6 ×
Drosophila
361.4
440.3
801.7
5,704.7
7.1 ×
Zebrafish
929.3
288.8
1,218.2
4,752.6
3.9 ×
Appendix
Table 9: Fitting time on one GPU, in seconds. HORIZON fit includes the one-time sparse qOET solve and simulation-free neural regression. stVCR uses its published training procedure. Speedup compares complete HORIZON fitting with stVCR.
Dataset
Method
Joint SW
Joint MMD
Spatial SW
Chamfer
Gene SW
Gene MMD
Mass RE (%)
Chicken
Horizon
0.134
0.037
0.199
0.392
0.597
0.098
0.45
no quotient
0.242
0.070
0.448
0.527
1.234
0.205
2.06
Drosophila
Horizon
0.035
0.013
0.043
0.125
0.107
0.023
0.10
no quotient
0.050
0.021
0.049
0.127
0.111
0.023
0.14
Appendix
Table 10: Removing the quotient on real data. Lower is better. no quotient fixes the pose to the identity and removes the projector, keeping the architecture, optimizer, step size, and evaluator of Horizon , and both rows generate from the first snapshot. The Horizon rows repeat Table 2 .
STORIES evaluator
Shared evaluator
Dataset
Method
Sinkhorn ϵ=.1
Gene SW
Gene MMD
Gene SW
Gene MMD
Axolotl
Horizon (first)
0.2132
0.1248
0.0184
0.1178
0.0173
STORIES (first)
0.2709
1.1093
0.2101
1.0867
0.2071
STORIES (restarted)
0.1847
0.6373
0.1332
0.6525
0.1376
Chicken
Horizon (first)
0.4407
0.6750
0.1089
0.6019
0.0969
STORIES (first)
0.2654
1.4526
0.2705
1.4761
0.2706
Appendix
Table 11: STORIES under its own evaluator and under the shared evaluator. Lower is better (seed 17 throughout). STORIES’ evaluator uses its normalization and Sinkhorn objective, whereas the shared evaluator is the one used in Table 2 , where the free-generation STORIES row appears. Bold marks the best entry per column and dataset.
Dataset
Method
Protocol
Joint SW
Joint MMD
Spatial SW
Chamfer
Gene SW
Gene MMD
Fit (h)
Axolotl
Horizon
first
0.0605
0.0029
0.2292
0.3616
0.1178
0.0173
0.30
NicheFlow raw
restarted
0.201±0.005
0.025±0.001
0.843±0.023
0.360±0.003
0.279±0.008
0.060±0.001
9.36
NicheFlow aligned
restarted
0.270±0.001
0.038±0.000
1.055±0.007
0.383±0.004
0.345±0.004
0.076±0.001
9.82
Chicken
Horizon
first
0.1328
0.0356
0.1855
0.3898
0.6019
0.0969
0.21
NicheFlow raw
restarted
0.064±0.002
0.025±0.001
0.198±0.005
0.150±0.002
0.217±0.007
0.042±0.001
6.91
NicheFlow aligned
restarted
0.079±0.001
0.035±0.001
0.268±0.004
0.149±0.001
0.231±0.007
0.047±0.001
6.80
Appendix
Table 12: NicheFlow with generation-seed variation and fitting time. Lower is better, and all rows use the shared evaluator. Horizon generates from the first snapshot (seed 17), and NicheFlow restarts every interval from its observed source (Table 2 compares both methods under this protocol). NicheFlow is trained for 100,000 steps, and its values are mean ± sample standard deviation over five generation seeds. Its aligned input uses the proper SE(2) qOET poses.
Support
Dense
Sparse
Speedup
Agreement
sec
GiB
sec
GiB
ΔJ/J / marg.
1,0242
10.3
0.14
11.2
0.08
0.92 ×
1.1 e–8 / 2.8 e–6
2,0482
25.7
0.55
17.8
0.16
1.44 ×
9.7 e–7 / 1.0 e–4
4,0962
110.2
2.18
30.1
0.33
3.66 ×
1.6 e–5 / 2.0 e–4
8,1922
649.1
8.69
62.0
0.77
10.48 ×
7.7 e–6 / 1.5 e–4
3,048×5,271
290.8
2.18
173.0
1.35
1.68 ×
3.7 e–5 / 1.2 e–4
Appendix
Table 13: Sparse qOET scaling at a fixed pose. Dense and column-generated solves use the same 10−4 full-gap tolerance. Memory is peak allocated GPU GiB, and ΔJ/J and marginal error compare the two converged solutions.
Configuration
n
Direct sp. SW
Quotient sp. SW
Quotient joint SW
Gene SW
Mass RE (%)
δ=10 , λq=0.25 (reference)
3
0.120±0.003
0.044±0.004
0.038±0.004
0.086±0.019
0.03±0.01
δ=5
3
0.081±0.010
0.038±0.001
0.034±0.002
0.069±0.005
0.06±0.01
δ=20
3
0.126±0.009
0.052±0.004
0.043±0.003
0.066±0.006
0.00±0.00
λq=0.1
3
0.107±0.011
0.038±0.001
0.041±0.005
0.107±0.021
0.02±0.01
λq=0.5
3
0.123±0.002
0.043±0.002
0.036±0.001
0.072±0.004
0.01±0.00
Appendix
Table 14: Sensitivity to the WFR scale δ and the molecular weight λq on the severe regime of the frame-perturbation benchmark (mean ± standard error over three variants, one per intrinsic seed, with n completed runs).
Dataset
Held out
Method
Spatial SW
Chamfer
Gene SW
Gene MMD
Joint SW
Joint MMD
Mass RE (%)
Axolotl
5 DPI
Horizon
1.136
0.476
0.816
0.220
0.363
0.069
0.2
WFR-FM
2.025
0.892
0.899
0.223
0.542
0.080
18.7
stVCR
5.390
0.750
0.918
0.222
1.251
0.300
4.4
10 DPI
Horizon
1.201
0.803
2.182
0.436
0.574
0.106
10.2
WFR-FM
3.394
3.057
1.825
0.381
0.934
0.121
24.4
stVCR
5.162
1.317
2.168
0.488
1.367
0.268
17.0
Appendix
Table 15: Timepoint holdout with out-of-fold preprocessing. Lower is better. In every fold the library-size target, gene filter and ranking, gene standardization, PCA and spatial scale are fitted without the held-out snapshot, and every interval’s qOET is solved from scratch. Horizon averages seeds 17, 29 and 43, WFR-FM averages seeds 11, 29 and 61, and stVCR uses seed 17. Bold marks the best entry per fold.
Figure 5: Population growth across all real time courses. Observed cell count ratios and Horizon predictions are normalized to the first snapshot. Curves show the predictive mean over three seeds with sample standard deviation. The model follows both expansion and contraction, with mean relative error over target snapshots ranging from 0.091% to 0.506%.
Dataset
Held out
Method
Spatial SW
Chamfer
Gene SW
Gene MMD
Joint SW
Joint MMD
Axolotl
5 DPI
Horizon
1.1144
0.4706
1.1985
0.2801
0.4030
0.07918
WFR-FM
2.0499
0.8889
1.1741
0.2824
0.5615
0.09006
stVCR
5.2877
0.7662
1.2062
0.2793
1.2498
0.2958
10 DPI
Horizon
1.1916
0.7883
2.2785
0.4573
0.6121
0.1125
WFR-FM
3.2498
2.6579
1.8388
0.3835
0.8631
0.1261
stVCR
5.0112
1.3247
2.3118
0.4996
1.3296
0.2670
Appendix
Table 16: Results for individual holdout folds with full-series preprocessing. Lower is better. Each row is an independently trained fold. Preprocessing was fitted once on the full series, whereas Table 15 refits it within each fold.
Figure 6: Biological program dynamics and coupling structure. For Drosophila the horizontal axis is section depth. The four panels on the left compare observed and predicted program trajectories. The panel on the right summarizes label affinity weighted by transported mass, averaged over intervals and relative to independent marginals.
Figure 7: Additional generated time courses. Zebrafish is shown in Figure 2 . Observations are on top and Horizon predictions on the bottom. Each displayed point carries equal mass, so point count shows abundance. Colors show observed annotations on top and annotations decoded from the predicted molecular PCs on the bottom.
Dataset
Variant
Joint SW
Spatial SW
Chamfer
Gene SW
Mass RE (%)
Direct sp. SW
Direct Chamfer
Rotation ( ∘ )
Rigid share (%)
Axolotl
Horizon
0.063 ± 0.002
0.243 ± 0.010
0.377 ± 0.001
0.123 ± 0.005
0.10 ± 0.02
0.239 ± 0.001
0.378 ± 0.000
0.00 ± 0.00
12.3 ± 1.3
registered, no projector
0.061 ± 0.006
0.234 ± 0.025
0.365 ± 0.020
0.126 ± 0.003
0.12 ± 0.02
0.228 ± 0.019
0.366 ± 0.020
0.64 ± 0.02
0.2 ± 0.0
projection in loss only
0.063 ± 0.002
0.243 ± 0.010
0.377 ± 0.001
0.123 ± 0.005
0.10 ± 0.02
0.549 ± 0.041
0.398 ± 0.004
0.82 ± 0.35
12.3 ± 1.3
projection in generation only
0.061 ± 0.006
0.234 ± 0.025
0.365 ± 0.020
0.126 ± 0.003
0.12 ± 0.02
0.231 ± 0.020
0.366 ± 0.020
0.00 ± 0.00
0.2 ± 0.0
Chicken
Horizon
0.128 ± 0.007
0.212 ± 0.006
0.401 ± 0.003
0.581 ± 0.036
0.45 ± 0.07
0.215 ± 0.007
0.403 ± 0.001
0.00 ± 0.00
2.3 ± 1.5
registered, no projector
0.138 ± 0.006
0.211 ± 0.020
0.401 ± 0.011
0.641 ± 0.011
0.45 ± 0.19
0.260 ± 0.039
0.405 ± 0.012
0.94 ± 1.08
0.9 ± 0.5
Appendix
Table 17: Registration and projection on real data, all metrics. Mean ± standard deviation over seeds 17, 29 and 43. Quotient metrics (first five columns) use the shared evaluator. Direct metrics compare with targets registered to the first frame by the composed qOET poses. Rotation is the accumulated angle of the rigid part of the integrated velocity at the last snapshot, and rigid share is the fraction of the network’s raw spatial kinetic energy carried by a shared rigid motion, averaged along the generated trajectory. Rows with projection in generation integrate no rigid motion, since the projector removes it.
Dataset
Interval
Creation+decay (%)
Targets/source
From split sources (%)
Rigid share, qOET (%)
Rigid share, identity (%)
Axolotl
2 → 5
0.000
2.13
91.6
0.030
–
5 → 10
0.021
2.28
95.6
0.025
–
10 → 15
0.021
2.14
83.8
0.069
–
15 → 20
0.018
2.28
90.9
0.420
–
Chicken
4 → 7
0.000
5.80
100.0
0.005
98.09
7 → 10
0.000
2.97
100.0
0.025
97.48
Appendix
Table 18: Properties of the fitted qOET couplings. For each interval we report creation and decay mass as a fraction of target mass, the mean number of targets reached from one transported source cell, the fraction of target mass reached from sources that split, and the share of the bridges’ spatial kinetic energy carried by a shared rigid motion under the qOET pose and (where converged) under the identity pose.
Dataset
K
σ
Carriers
Spatial SW
Chamfer
Gene SW
Gene MMD
Joint SW
Mass RE (%)
Chicken
1
0
147
0.184
0.389
0.613
0.098
0.135
0.27
2
0.25
294
0.165
0.316
0.592
0.096
0.124
0.26
2
0.5
294
0.189
0.332
0.590
0.099
0.125
0.23
2
1
294
0.273
0.406
0.619
0.104
0.143
0.19
4
0.25
588
0.163
0.267
0.564
0.092
0.117
0.27
4
0.5
588
0.165
0.273
0.556
0.090
0.115
0.23
Appendix
Table 19: Initial carrier support. Each initial cell is replaced by K carriers with its molecular state and weight /K , at positions jittered by σ times the median nearest-neighbour spacing, keeping the trained model and evaluator unchanged. Values average three jitter seeds, where K=1 is the single-carrier baseline. Drosophila joint SW is omitted because the Chamfer-based evaluation pose of its nearly half-turn-symmetric embryo flips between runs at 10.5.
Dataset
Interval
Width ( ∘ )
Other (%)
Other ( ∘ )
Replicas
ΔJ/J (%)
Δθ ( ∘ )
Shift
Axolotl
2 → 5
0.0
24.77
180
9/9
+0.037
1.07
0.024
5 → 10
0.0
23.07
180
9/9
+0.001
0.19
0.009
10 → 15
0.0
13.88
180
9/9
-0.005
0.78
0.022
15 → 20
0.0
16.55
180
9/9
-0.184
1.96
0.039
Chicken
4 → 7
109.3
0.01
20
9/9
+0.000
4.06
0.058
7 → 10
10.1
0.45
22
9/9
-0.004
1.93
0.040
Appendix
Table 20: Initialization of the pose search. Per interval, 120 coreset starts (72 rotations at 5∘ spacing with centroid-matched translation and 24 rotations with each of two translation offsets of 0.3 tissue radii). Width is the angular spread of starts within 0.1% of the best coreset objective, Other is the relative objective excess and angle of the best start more than 20∘ away after full-support refinement, and Replicas counts how many of the nine rotated versions of the 8-start schedule reach the best orientation. The last three columns compare the best refined start with the 8-start reference coupling in objective, angle, and mean shift of a source cell’s barycentric target (in tissue radii).
Coupling (s)
Training step (ms)
Generation (s)
Dataset
qOET
identity pose (dense)
horizontal
ambient
projected
ambient
Axolotl
794
not converged
28.0
25.2
69.2
68.9
Chicken
466
681
21.6
20.4
0.99
0.85
Drosophila
361
1,452
23.9
22.8
0.91
0.84
Zebrafish
929
not converged
23.9
22.4
1.87
1.64
Appendix
Table 21: Cost of the quotient. Training steps and trajectory integrations are medians of alternating repeats on a single dedicated GPU, and the coupling is the wall-clock time of the qOET solves. Identity-pose OET is the coupling of the no-quotient ablation (Table 10 ), solved densely with the pose fixed. It did not converge within tolerance on axolotl and zebrafish. Training steps use the batch sizes of Table 4 , and integrations propagate all cells of the first snapshot. One call of the projector takes 0.8 ms for 103 – 6.6×104 cells.
Symbol or term
Meaning
z=(x,q)
State of a cell, consisting of spatial position x and molecular representation q (32 PCs).
μ,ν,ρt
Tissues as weighted sets of cells, where the weight of a cell is the abundance it represents.
ux,uq,g
Spatial velocity, molecular velocity and relative growth rate.
δ,λx,λq
Cost of mass change, and weights of spatial and molecular displacement.
WFRδ
Wasserstein–Fisher–Rao distance, the least action to move cells and change their mass (Eq. 1 ).
G=SE(dx) , h=(R,b)
Proper rotations R and translations b , applied to all positions at once.
Appendix
Table 22: Notation and terms used in the main text.
Understanding cellular trajectories via time-resolved single-cell transcriptomics is vital for studying development, regeneration, and disease. A key challenge is inferring continuous trajectories from discrete snapshots. Biological complexity stems from stochastic cell fate decisions, temporal proliferation changes, and spatial environmental influences. Current methods often use deterministic interpolations treating cells in isolation, failing to capture the probabilistic branching, population shifts, and niche-dependent signaling driving real biological processes. We introduce Manifold Interpolating Optimal-Transport Flow (MIOFlow) 2.0. This framework learns biologically informed cellular trajectories by integrating manifold learning, optimal transport, and neural differential equations. It models three core processes: (1) stochasticity and branching via Neural Stochastic Differential Equations; (2) non-conservative population changes using a learned growth-rate model initialized with unbalanced optimal transport; and (3) environmental influence through a joint latent space unifying gene expression with spatial features like local cell type composition and signaling. By operating in a PHATE-distance matching autoencoder latent space, MIOFlow 2.0 ensures trajectories respect the data's intrinsic geometry. Empirical comparisons show expressive trajectory learning via neural differential equations outperforms existing generative models, including simulation-free flow matching. Validated on synthetic datasets, embryoid body differentiation, and spatially resolved axolotl brain regeneration, MIOFlow 2.0 improves trajectory accuracy and reveals hidden drivers of cellular transitions, like specific signaling niches. MIOFlow 2.0 thus bridges single-cell and spatial transcriptomics to uncover tissue-scale trajectories.
Xingzhi Sun, João Felipe Rocha, Brett Phelan +11
Yale University, USA · University of Wisconsin-Madison, USA · Université de Montréal; Mila - Quebec AI Institute, Canada +2
Single-cell RNA sequencing (scRNA-seq) provides high-dimensional profiles of cellular states, enabling data-driven modeling of cellular dynamics over time. In practice, time-resolved scRNA-seq is collected at only a few discrete time points as unpaired snapshot populations, leaving substantial temporal gaps. This motivates trajectory inference at unmeasured time points. Existing methods mainly follow two directions, optimal-transport (OT) alignment provides distribution-level matching between observed snapshots, while continuous-time generative models support forecasting via learned dynamics. However, two challenges remain: (i) unpaired snapshots render local transitions between adjacent time points ambiguous, leading to unstable supervision; and (ii) long-horizon prediction relies on repeated integration, where small modeling errors compound and cause distribution drift. To address these challenges, we propose single-cell Flow Matching (scFM), a latent generative framework based on coupling-conditioned flow matching. First, we compute entropically regularized OT couplings between adjacent snapshots and use them to construct soft, weighted flow-matching targets for learning time-dependent velocity fields. Second, we learn bidirectional velocity fields and leverage their consistency to refine couplings and improve temporal coherence under sparse supervision. Third, we introduce distribution-level alignment and latent dynamic regularization to anchor long rollouts and mitigate drift. Experiments on real-world time-series scRNA-seq datasets show that scFM consistently improves distributional prediction performance for both temporal interpolation and extrapolation. Moreover, scFM yields more accurate trajectory reconstruction and temporally coherent visualizations where intermediate time points are absent, indicating a more faithful recovery of underlying temporal gene expression dynamics.
Siyu Pu, Qingqing Long, Xiaohan Huang +7
Computer Network Information Center, Chinese Academy of Sciences · University of California, Los Angeles
Spatial transcriptomics (ST) prediction from histology aims to infer spatially resolved gene expression profiles directly from tissue images. This is a challenging problem that is currently often addressed with complicated model architectures, additional inputs, and complicated training procedures, while remaining limited to fixed, small gene panels. This raises a natural question: Is there an efficient method for large-panel ST prediction? Here, we present GeneFlow, a simpler yet effective framework for large-scale spatial gene expression prediction that unifies flow-matching generative modeling with structured multimodal fusion. Specifically, GeneFlow first encodes both histological morphology and cross-gene interactions in spatial spots through histology and gene tokenizers. The model directly predicts spatial gene expression, rather than an intermediate velocity field, while being trained with a velocity-based flow-matching objective. We show that, even without relying on heavy generative pipelines, GeneFlow achieves state-of-the-art performance across multiple ST benchmarks, with up to 56.8% gains in 1,000-gene prediction. Extensive experiments further show that GeneFlow produces biologically meaningful spot profiles while reducing computational cost by orders of magnitude, thereby improving efficiency and scalability, particularly in clinical settings. We believe GeneFlow provides an opportunity for leveraging widely available histology images for downstream biological discovery. Code is available at: https://github.com/ZZhangsm/GeneFlow
Songming Zhang, Yuxiao Luo, Shifu Luo +3
School of Computer Science, Nanjing University · Faculty of Computer Engineering and Artificial Intelligence, Shenzhen University of Advanced Technology · Department of Land Surveying and Geospatial Science (LSGS), The Hong Kong Polytechnic University +2