August 19, 2026
Kinema
A generative model of cells in time, trained on real time-lapse microscopy.
A culture you can experiment on.
Kinema is a generative model of cells in time, trained on real time-lapse microscopy. The goal is a culture you can experiment on before anyone does it at the bench: start from a real frame, roll the population forward in microns and minutes as its cells move, grow, divide and die, then change one rule of their behaviour and see what follows. Cells are individually noisy, so Kinema is built to get populations right rather than to say where any one cell will be: how many divide, how neighbours move together, how a culture responds when something changes. One model spans six cell types, learns from less data than a conventional video model, and switching off its proliferation rule changes the simulation the way the matching parameter does in a physics simulator. Every capability is a numbered claim with the test that decides it, and the claims that failed stay on the record.
Results.
Measured by Essarion on held-out microscopy, failures included.
Learning from less data
Accuracy at predicting what happens when two cells meet, as the training data grows; three seeds, compute-matched. No model yet beats predicting that nothing changes (0.511).
Predicting a cell’s next division
AUC for a division within four frames, on 73,690 single-cell crops, leave-one-corpus-out. 0.5 is chance.
Cells that move together
How alike neighbouring cells move, one cell diameter apart; three training seeds per arm, paired t = +21.2.
An intervention does what physics does
Agreement (cosine) between switching off one of Kinema’s operators and changing the matching parameter in a physical simulator. It holds for proliferation and fails for adhesion.
One model across cell types
Error on a cell type held out entirely, against a model trained on it; 1.0× means no penalty. Medians over six cell types. Lower is better.
What holds.
- Holds
Interventions reach mechanism
+0.97Switch off one operator and the model moves as the simulator does when the matching physical parameter changes. Re-verified on deterministic rollouts.
- Holds, post-hoc
Learns from less data
+0.312At 11% of the data it beats the video model trained on 31% and on all of it, 3 of 3 seeds.
- Holds at short horizon
One library across cell types
1.02×Error on a cell type it never saw, against a model trained on it, one frame ahead.
- Holds, once trained for it
Collective motion
+0.120Neighbouring cells move alike, against +0.169 in real cells: 71% of the real correlation.
- Holds
The inside predicts behaviour
0.704AUC for a cell’s next division from the learned code, against 0.596 from its outline.
- Holds
A population that evolves
r = +0.963Response tracks selection, and collapses to +0.603 when inheritance is scrambled.
What fails.
Kept on the record, beside what holds.
Individual paths
About 1.9× above the floor a perfect sampler could reach. Single-cell motion here is close to unpredictable.
Physically possible fields
Simulated cells overlap more than real ones, on every corpus.
Cell shape
Predicted outlines are 4.7× as rough by 32 frames.
Seeing the cell
The generator largely ignores which cell it is shown: its output correlates 0.036 with it.
New doses
Loses to a straight line through four training doses.
Open-ended evolution
The trait moves about 0.05 and stops.
Four earlier results were withdrawn or weakened after their own controls caught them: a reproducibility bug in rollouts, now fixed; a headline for a separate size channel, withdrawn; speed calibration, 5 of 7 cell types rather than 6; and long-horizon transfer, a median of 1.04× rather than the mean first reported.
Kinema is private, built for Essarion Science.
Ask harder questions.
If you are working on cells in time, we would like to hear from you.