TL;DR
MitoWorld is a stochastic world model designed to predict the future of individual cells following genetic perturbation, jointly modeling their morphological evolution and the exact timing of their division. By combining a reference-preserving latent transition with a division hazard model and a diffusion renderer, it significantly outperforms existing population-level predictors in both visual fidelity and temporal accuracy.
What Problem Does This Paper Address?
In biological research, "knocking down" a gene allows scientists to observe how its absence affects cellular behavior. However, cells are inherently stochastic: even cells that look identical at the start of an experiment can follow wildly different growth paths and divide at different times.
Existing computational methods generally fall into two camps:
- One-step endpoint predictors: These map a starting population to an end state but ignore the journey (the trajectory) in between.
- Video world models: These generate realistic-looking videos of cell populations but don't explicitly track or model the life cycle of individual cells.
The gap is the lack of a model that can take a single observed cell and predict its specific, uncertain future—both how it will change and when it will stop growing and divide.
Methods
MitoWorld treats the single-cell cycle as a "world model" problem, consisting of four main components:
- Latent Space Encoding: A nonlinear encoder maps microscopy frames into a Gaussian-regularized latent space. This regularization ensures that the "map" of cell states remains well-behaved for long-term predictions.
- Stochastic Transition: To move through time, the model uses a gene-conditioned Ornstein–Uhlenbeck (OU) transition. Crucially, this transition is "reference-preserving," meaning it keeps the latent distribution Gaussian over any number of steps. The transition operator is structured into contraction (representing natural decay/stability) and rotation (representing directional temporal evolution, like growth).
- Biological Clock: A learned monotone clock adjusts the speed of evolution based on the cell's chronological age, allowing the model to capture how progression rates vary throughout the cycle.
- Division Hazard: Instead of a simple "stop" flag, MitoWorld uses a hazard function. This calculates the probability of division at each step based on the cell's current state, age, and genetic condition. When triggered, it initiates terminal transitions that simulate chromatin condensation and the formation of two daughter cells.
- Conditional Diffusion Renderer: Finally, a diffusion model converts these abstract latent trajectories back into high-fidelity microscopy images.
Datasets
The authors evaluated MitoWorld using a subset of the MitoCheck dataset, a massive genome-wide RNAi screen of HeLa cells.
- Scope: 17 different gene knockdown conditions plus one negative control.
- Data: 1,508 complete single-cell cycles with recorded birth and division events.
- Split: 1,183 trajectories for training and 325 for testing.
Key Insights and Ideas
- Paired Supervision is Essential: The authors prove that you cannot uniquely identify a cell's transition dynamics just by looking at population snapshots (marginals). You need "same-cell" pairs over time to learn the true cross-time dependence.
- Tying Noise to Dynamics: By using a reference-preserving OU process, the model's uncertainty (noise) is mathematically tied to its dynamics. This prevents "drift" where the model might otherwise generate increasingly unrealistic images during long-horizon rollouts.
- Rotation Captures Life: A simple contractive (decaying) model can't capture the "loop" of a cell cycle. Adding a rotational component to the latent transition was key to modeling the directional progression of a cell toward division.
- Hazard as a Survival Likelihood: Modeling division as a hazard allows the model to produce a full distribution of possible cycle lengths, which can be directly tested against real-world biological data.
Key Results
MitoWorld demonstrates a step-change in performance over both generative and world-model baselines:
- Visual Fidelity: It reduced the Fréchet Inception Distance (FID) by 34.6% and improved the video realism metric (JEDi) by 45.6% compared to the strongest baselines.
- Identity Preservation: In a retrieval task (matching a predicted future back to the correct real cell), MitoWorld achieved 41.87% accuracy, outperforming the best baseline (37.33%).
- Uncertainty Calibration: Its latent-space prediction intervals achieved near-perfect 90% coverage while being 35% narrower than models with unconstrained noise, indicating much "sharper" and more confident predictions.
- Biological Accuracy: The generated cell-cycle length distributions closely matched observed data, capturing how different gene knockdowns shift the timing of division.
Limitations and Future Work
- Generalization: The model was trained and tested on known knockdowns within a single dataset. A major next step is improving generalization to entirely unseen genetic conditions.
- Rare Phenotypes: The model occasionally underestimates growth in cells with "unusually large" nuclei, likely because it defaults to the more common growth patterns it saw during training.
- Unpaired Data: While this work relies on high-quality time-lapse tracking, many biological datasets are just "snapshots." Adapting these stochastic dynamics to learn from unpaired snapshots remains an open challenge.
Summary updated