A cell observed after a gene knockdown has an uncertain future, both in how its appearance will evolve and when it will divide. Existing approaches predict endpoints or population-level changes without jointly modeling individual trajectories and division timing. In this work, we present MitoWorld, a stochastic world model that generates these future appearances as continuous evolution in a Gaussian reference latent, terminated by a division hazard and converted into images by a conditional diffusion renderer. The ordinary transition provably preserves its Gaussian reference, tying its noise to its dynamics; paired observations identify its discrete transition where stationary marginals cannot. The hazard turns division timing into a survival likelihood that can be tested against real cycle lengths. On 325 held-out MitoCheck tracks spanning 17 knockdowns and a control, MitoWorld lowers FID by 34.6% and the video metric JEDi by 45.6% relative to the strongest baselines. Its latent-space prediction intervals achieve near-nominal 90% coverage, are 35% narrower than those of a model with freely fitted noise. The hazard also yields cell-cycle length distributions with KS distance close to the model-based finite-sample reference. Testable single-cell stochastic dynamics thus offer a generative tool for time-lapse perturbation screens.
Figures & tables
Figure 1: Cell-cycle prediction. (a) One-step endpoint predictors learn to map control populations to perturbed populations, while video world models synthesize population-level cell dynamics. Neither explicitly models how an individual cell grows or when it divides. (b) From one cell xt under knockdown g , MitoWorld rolls out stochastic latent trajectories, renders every step (faded box), and stops each at a division (red point) sampled from p(Tdiv) .
Figure 2: The pipeline of MitoWorld. (a) trains the encoder to place cell states in a Gaussian reference space ( Sec. 3.2 ). (b) freezes the encoder and moves each state with a reference-preserving transition whose operator splits into a contraction Sg and a rotation Kg ( Sec. 3.3 ), and a diffusion renderer converts every state into an image ( Sec. 3.5 ). (c) uses a division hazard h to determine the survival probability Sk and the division-time distribution. A hazard trigger initiates the terminal transitions, which end the cycle with a two-daughter composite ( Sec. 3.4 ).
Method
FID +3h ↓
FID +10h ↓
FID div ↓
FID all ↓
R@1 ↑
JEDi ↓
Size ↓
Intensity ↓
Intensity dist. ↓
Texture ↓
Shape ↓
Generative
CellFlux [ICML25] [ 78 ]
74.72
99.39
97.96
85.74
37.33%
7.14
0.3056
0.0789
0.4073
0.1343
0.1476
IMPA [Nat. Commun.25] [ 51 ]
68.61
64.65
55.10
54.38
5.67%
6.50
0.1626
4.0374
0.5267
2.4320
0.1812
PhenDiff [MICCAI24] [ 9 ]
306.09
300.45
279.16
289.75
3.59%
9.61
1.3789
3.6650
0.8561
3.2474
4.7492
MorphoDiff [ICLR25] [ 46 ]
104.21
123.30
110.95
112.68
6.46%
9.78
0.5144
3.8396
0.2490
2.9829
0.1356
World Model
DIAMOND [NeurIPS24] [ 2 ]
89.36
95.69
73.22
80.09
5.64%
8.31
0.3126
3.4623
0.7152
3.7944
0.3780
DreamerV3 [Nature25] [ 27 ]
176.59
183.83
176.49
177.08
18.92%
3.75
0.2734
0.0789
0.7617
0.1191
0.4845
Table 1: Image prediction and biological features on 325 test MitoCheck [ 48 ] tracks. FID measures frame fidelity at +3h, +10h, pre-division ( div ), and across future frames ( all ); R@1 measures cell-specific retrieval, and JEDi measures sequence realism. The five biological categories average marginal W1 distances. BEST and SECOND results are highlighted.
Figure 3: Single-cell growth, division, and latent trajectories. (a) Image-level cell trajectory. Observed trajectory and predictions from strong baselines IMPA and HI, and our MitoWorld. (b) Latent space cell trajectory. Top to bottom: the observed encoded trajectory, HI native-latent trajectory, and MitoWorld coupling rollout, shown on the same frozen UMAP. Thin lines show full state trajectories; thick arrows connect phase-averaged states. Background points are colored by cell area (pixels 2 ), reflecting variation in cell size. (c) Cycle-length distributions. Observed and generated cycle lengths for three conditions.
Model
Valid. ↑
Temp. ↑
Agr. ↑
Div. ↑
Overall ↑
IMPA
2.27 ±0.73
2.63 ±0.69
1.87 ±0.57
1.03 ±0.08
2.67 ±0.72
CellFlux
3.57 ±0.45
2.93 ±1.02
2.20 ±0.79
1.13 ±0.18
3.50 ±0.74
DIAMOND
2.47 ±0.19
1.30 ±0.32
1.60 ±0.25
1.03 ±0.08
1.60 ±0.37
OC-STORM
3.77 ±0.68
1.80 ±0.46
2.63 ±0.64
1.00 ±0.00
2.10 ±0.09
HI
4.30 ±0.51
2.43 ±0.63
2.00 ±0.72
1.03 ±0.08
2.97 ±0.94
MitoWorld
4.03 ±0.36
4.23 ±0.36
4.93 ±0.15
4.93 ±0.09
4.33 ±0.26
Table 2: Human expert evaluation on a five-point Likert scale (1: strongly disagree; 5: strongly agree).
Feature
Age + gene
Latent z
Area
0.483
0.804
[0.425,0.542]
[0.773,0.832]
Eccentricity
0.050
0.802
[0.019,0.074]
[0.779,0.822]
Mean intensity
0.270
0.885
[0.208,0.326]
[0.869,0.897]
Table 3: Phenotypic information in the frozen encoder. Equal-track test R2 for linear readouts from age and gene or from z . Brackets give 95% lineage-root bootstrap intervals.
Variant
Err ↓
NLL ↓
R@1 z ↑
ρI
0.9093
1.3760
36.62%
Diagonal
0.9072
1.3740
36.24%
w/o Kg
0.8996
1.3575
39.66%
Shuffled pairs
0.9185
1.3343
31.18%
No gene cond.
0.8932
1.3596
40.55%
Uniform clock
0.9002
1.4229
41.16%
Table 4: Effects of coupling components on latent-state prediction . Red and blue mark the BEST and SECOND BEST values.
Noise
Horizon
CRPS ↓
ES ↓
Cov90
Width ↓
None
+3h
0.62
6.07
–
–
Isotropic
0.48
4.72
98.02%
4.03
Free
0.47
4.69
97.88%
3.96
MitoWorld
0.44
4.36
89.72%
2.58
None
div
0.79
7.55
–
–
Isotropic
0.59
5.84
99.28%
5.15
Table 5: Calibration results of latent futures.
Figure 4: Long-horizon latent dynamics .
State z
Age ak
Gene eg
Brier ↓
KS ↓
FID ↓
✓
✓
0.0211
0.2263
30.61
✓
✓
0.0220
0.4743
32.95
✓
✓
0.0202
0.2414
30.65
✓
✓
✓
0.0201
0.2229
30.50
Table 6: Effects of the factors on division timing prediction.
Figure 5: Underestimated nuclear growth. Both methods generate trajectories from the first observed frame.
Predicting how a cell will change its transcriptional state under a developmental signal or a genetic perturbation is the computational core of in-silico biology and the AI Virtual Cell program. Existing approaches either fit static control-to-treated maps that discard time, or solve multi-step ODE / Schrödinger-bridge problems on each dataset independently. We introduce Chreode, a one-step cell world model that predicts action-conditioned cell-state transitions through a structured residual transition operator. It shifts distributional evolution from inference time to training time, enabling single-pass generation while preserving a Waddington-inspired decomposition into downhill landscape flow, rotational in-tangent dynamics, and stochastic spread. The model is pretrained with a shared scVI encoder and a DiT-based dynamics backbone on a 2.4M-cell mouse embryonic atlas spanning 7 datasets. As a fine-tuning initialization, Chreode improves per-target Sinkhorn distance on Weinreb hematopoiesis and Veres islet differentiation over matched scratch models, PI-SDE, and PRESCIENT. As a transferable gene-state embedding for GEARS, the pretrained dynamics representation reduces shared-vocabulary DE20 mean squared error on Norman Perturb-seq from 0.2121 to 0.1858, a 12.4% relative improvement, without changing the GEARS training procedure. We interpret this transfer to perturbation prediction as evidence that pretrained developmental-trajectory dynamics encode differentiation primitives transferable to CRISPR-induced state shifts, since both involve cell-state transitions in a shared latent geometry. The pretrained backbone additionally produces zero-shot clonal fate scores on Weinreb that are competitive with strong dynamic-OT baselines.
Mufan Qiu, Genhui Zheng, Yinuo Xu +4
University of North Carolina at Chapel Hill · The University of Texas at Austin · University of Pennsylvania
Building Virtual Cells that can accurately simulate cellular responses to perturbations is a long-standing goal in systems biology. A fundamental challenge is that high-throughput single-cell sequencing is destructive: the same cell cannot be observed both before and after a perturbation. Thus, perturbation prediction requires mapping unpaired control and perturbed populations. Existing models address this by learning maps between distributions, but typically assume a single fixed response distribution when conditioned on observed cellular context (e.g., cell type) and the perturbation type. In reality, responses vary systematically due to unobservable latent factors such as microenvironmental fluctuations and complex batch effects, forming a manifold of possible distributions for the same observed conditions. To account for this variability, we introduce PerturbDiff, which shifts modeling from individual cells to entire distributions. By embedding distributions as points in a Hilbert space, we define a diffusion-based generative process operating directly over probability distributions. This allows PerturbDiff to capture population-level response shifts across hidden factors. Benchmarks on established datasets show that PerturbDiff achieves state-of-the-art performance in single-cell response prediction and generalizes substantially better to unseen perturbations. See our project page (https://katarinayuan.github.io/PerturbDiff-ProjectPage/), where code and data will be made publicly available (https://github.com/DeepGraphLearning/PerturbDiff).
Xinyu Yuan, Xixian Liu, Ya Shi Zhang +3
Mila - Québec AI Institute · University of Montréal · University of Ottawa +3
Although conventional world models - auto-regressive or diffusion models based on transformers or convolutional networks - may learn surface statistics of world dynamics, can they learn the exact world dynamics from its observed history? Leveraging cellular automata as a simple testbed, we find the answer to be no in many cases. Conventional architectures predict most pixels correctly yet rarely complete a rollout: a CNN predicts 96.3% of cells but completes 18.9% of rollouts; a joint diffusion model completes none. We trace the gap to three failure modes of these world models - namely, they fail to exactly capture spatial locality, temporal locality or temporal stability. Simple changes repair each: (1) for spatial locality, two-dimensional rotary positions lift a transformer from 39.1% to 100% on the Game of Life; (2) for temporal locality, handing each token its cell's previous-frame neighbourhood lifts the same transformer from 25.8% to 99.9% on unseen rules; (3) for temporal stability, causal freezing lifts the same diffusion weights from 42.2% to 99.9%. None of the three changes touches the architectural backbone; each only modifies the information flow within it. We also compare joint and ordered sampling on billiards and, in an exploratory study, on a simulated Burgers equation.
Shaoyang Guo, Ziming Liu
Meta Circle(元环智能) · Peking University · Tsinghua University