HeLa cells treated with the same dose of staurosporine do not all die at once. From one image of a treated cell, how early can its fate be read? And given that first image and a fate, can a generative model draw the cell's future?
Data: one plate of the Way lab's ChromaLIVE time-lapse (S-BIAD2515, CC0), ten doses, 13 frames over 6 h, then an AnnexinV stain. The project tracks 6,676 cells from the first frame to the stain, labels their fate, checks that label, and trains a flow-matching video model written from scratch.
One field per dose at 0 h and 6 h, and the AnnexinV image taken afterwards. Blue Hoechst, green ChromaLIVE 488, red ChromaLIVE 561. Circles: tracked nuclei by fate group.
| Step | Result | Figure | |
|---|---|---|---|
| 1 | Data. Segment, track and crop single cells in 119 fields | 23,902 nuclei at 0 h, 6,676 followed to the endpoint with a fate label | 1, gallery |
| 2 | Label, and its audit. AnnexinV above the control 90th percentile, or a condensed nucleus; checked against two markers read from the movie alone | The AnnexinV clause has a 10% background by construction: 44% of the "dies, AnnexinV only" cells at the model doses are expected from the threshold alone, and the stain misses a third of cells with a condensed nucleus. Both outcomes are common at one dose only, 39.06 nM | 2 |
| 3 | A new live marker. The cytoplasm turns from red to green | Of the cells whose nucleus condenses, 95 to 98% turn green, most of them first, by a median of half an hour | 4 |
| 4 | How early is fate readable from one frame? | Not before the cell starts to change. The rise in AUROC under the fate label (0.50 at 0 h to 0.84 at 6 h) also appears in untreated wells (0.41 to 0.79). Frozen DINOv2 and V-JEPA 2 do not beat eleven nucleus measurements | 5 |
| 5 | Drawing the future. A latent flow-matching video model given the 0 h frame and the fate group | Right look, colour and motion for each group, and texture close to its autoencoder's limit. But its futures are closer to the cell's real future than the 0 h frame only for condensing cells, and they are not recognisable as that cell | 6, 7, 9 |
The full account with numbers and confidence intervals is in docs/report.md.
The same 0 h cell drawn at 6 h under each of the three fate classes, from the same noise.
- One plate, three wells per dose, dose not randomised across the plate; held-out wells are one row.
- The fate label is the weakest part (step 2). The three groups are stages reached by 6 h more than three fates.
- The model is told the fate. Nothing here predicts it, and counterfactual futures cannot be checked cell by cell.
- 28% of nuclei end with a label, and almost no dividing cell.
- One training run per model.
| docs/report.md | the direct path: data, label audit, two questions, limits, next steps |
| docs/gallery.md | many more images of real and generated cells |
| docs/code_tour.md | the code file by file |
| docs/log.md | all 25 experiments in the order they were run, dead ends included |
| notes/notebook.md | dated notes and corrections |
src/static2fate/ |
library: fields, cells, audit, flow, video, autoencoder, evaluate, features, display |
scripts/ |
the pipeline, in order (scripts/README.md); scripts/archive/ holds superseded experiments |
kaggle/ |
the GPU jobs for the autoencoder and the latent models |
results/figures/ |
the report's figures; supporting/ holds the log's |
results/tables/ |
every number as CSV |
One well is about 1.5 GB and runs on a laptop; all 30 wells are 45 GB.
uv sync
scripts/00_download_subset.sh C-08 # one well of S-BIAD2515, 19.53 nM; no argument = all 30 wells
scripts/run_all.sh # segmentation, tracking, fate labels, crops, baseline
uv run scripts/46_label_audit.py # label audit and readability (needs all wells)
uv run scripts/47_gallery.py real # image galleries
uv run pytestThe models were trained on Kaggle T4 GPUs: uv run scripts/export_kaggle.py <user> writes the
crops as a dataset, and kaggle/ae and kaggle/fate3 train the autoencoder and the latent
video model. Scoring and figures then run on a laptop (see docs/report.md).
@software{gerami2026static2fate,
author = {Gerami, Matin},
title = {static2fate: single-cell fate from live-cell imaging},
year = {2026},
version = {0.1.0},
url = {https://github.com/MMatinGerami/static2fate}
}Data: S-BIAD2515, Lippincott et al., Way lab, CC0 (BioImage Archive).
MIT