Can AI Predict Whether a Cancer Cell Will Divide or Die?

Chemotherapy rarely kills every cancer cell. Cells with identical DNA, in the same dish and at the same dose, can still split into two groups: some die, and others keep dividing and may seed a relapse. Telling which cell will do what, and how early, is a basic question about drug response. A new study trained a deep-learning model to answer it from microscope video alone.¹

The gap

Single signals are known to matter. One is p53, a protein that responds to DNA damage. A 2012 study found that cells showing repeated p53 pulses recovered from damage, while cells held in a sustained p53 response frequently entered senescence, a permanent stop in dividing.⁵

What the authors say was missing is a systematic test of how predictable fate is from several moving signals at once, and when during a cell's life that information shows up. They also say earlier deep-learning work mostly classified cell-cycle stage or modeled tissues, and that Transformer models had not been applied to this chemotherapy setting. We did not independently test that last claim.¹

How the model works

The data. The team used a public set of 274 fluorescence time-lapse videos of cancer cells. Frames came every 30 minutes for 93 hours. In 118 videos the cells received cisplatin, a DNA-damaging chemotherapy drug, 48 hours in. Three glowing reporters tracked the cell cycle, a circadian-clock protein, and p53 activity.^{1, 3} The dataset comes from the same group's earlier cell-tracking preprint.⁴

Cells were tracked and labeled by hand. A division was marked when a parent cell separated into two daughters, and a death at the first frame showing clear collapse. Table 1 lists 21,898 labeled trajectories: 17,442 ending in division and 4,456 in death. Each cell was cropped around its center from its first appearance up to the frame before the event, so the split or collapse itself was never shown to the model.¹

The model. A network called ResNet-50 turns each cropped frame into a list of 2,048 numbers, a compact description of what the image looks like. These lists form a sequence, which goes into a Transformer, a model type built around self-attention. In self-attention, every time point can weigh every other time point, so the model can link a clue early in a video to one near the end.⁶ A final attention layer gives each frame a weight, which later lets the authors ask which frames mattered.¹

This design suits sequences of varying length. A reasonable alternative is hand-measured features such as cell area or shape, and the authors tested a version of that: the same Transformer fed those features instead of image embeddings.¹

They split trajectories randomly into 80% training, 10% validation, and 10% test sets. Because dying cells are rarer, they balanced training samples and judged results with balanced accuracy, the average of the share correct within each class, so rare dying cells count as much as common dividing ones.¹

What the paper reports

The test set held 2,089 trajectories: 439 deaths and 1,650 divisions. It was drawn at random from the combined labeled dataset. In the full dataset, Table 1 lists 693 untreated controls, 4,722 trajectories from three known cisplatin doses, and 16,483 with no dose information. The paper does not break the test set down by dose, so the scores below are not a treated-only result. The paper restricts its interpretation analyses to cisplatin-treated sequences. Balanced accuracy was 0.94 and the F1 score, another summary of accuracy, was 0.93. The model correctly flagged 401 of 439 deaths (91.3%) and wrongly labeled 3.5% of dividing cells as dying. The hand-feature version scored 0.90 balanced accuracy and 0.877 F1, so images helped, though hand features also did well.¹

The authors trimmed frames off the end of each trajectory. With the last 15 frames removed, which is 7.5 hours at 30-minute spacing, balanced accuracy was still 0.903. At 20 or more frames removed it fell steeply, reaching 0.565 at 40. Using only the final frame gave 0.713.¹

The two fates behaved differently. Division could be recognized from very few late frames: recall, the share of true divisions caught, reached 0.918 using the last four frames. Death needed far more context, reaching a recall of 0.907 only with about 80 frames.¹

To check that attention reflects real reliance, the authors masked frames. Hiding half the frames with the highest weights cut balanced accuracy to 0.836. Hiding random frames left 0.912, and hiding the lowest-weight frames left 0.931.¹

The weights themselves told a similar story. For dividing cells, attention piled up at the end: normalized weight was 0.62 at the second-to-last frame and 0.22 ten frames earlier. For dying cells it was spread out: 0.88 at the last frame, about 0.42 thirty frames earlier, and 0.07 at 45 frames.¹

Comparing high- and low-attention frames, p53 intensity showed the largest effect for both fates, measured with Cliff's delta, a statistic of how often values in one group exceed the other. Frames before division showed larger, rounder cells, while frames before death showed larger cells that did not round. The cell-cycle reporter showed no significant difference.¹

Caveats. The paper describes a random split of trajectories and does not say whether cells from one video were kept together. Cells in a video share a dish and imaging conditions, so performance on a new experiment is unknown. This is our reading, not a flaw the authors report. The untreated control had only 41 deaths, so the authors did not interpret those estimates.¹

Attention is also a debated explanation. Work on language models found that attention weights often do not line up with other importance measures.⁷ The masking test supports the authors' use here, but it shows what the model relies on, not what the cell is doing. The authors suggest testing features by removing them from a model.¹

Future directions

These are proposals anchored to the paper's limits. As of October 19, 2026, each idea was searched in PubMed, bioRxiv, arXiv, and the web, then given a separate adversarial check. Anyone picking up one of these ideas should repeat a fresh literature search before committing to a project.

1. Test on whole held-out experiments. Scoring the model on complete videos, a different dose, or a different drug or cell line would show whether it generalizes. The searches did not surface such a study for a division-versus-death classifier. Related work probes cell-fate predictors by virtually changing their inputs.⁸

2. Compare attention with other explanation methods. The authors validate attention by masking frames. Gradient-based attribution could be run on the same model and compared. The searches did not surface such a comparison for time-lapse cell-fate models. The closest adjacent work is the virtual perturbation study.⁸

3. Link predicted fate to gene activity. The authors propose profiling cells with single-cell transcriptomics at the time points when fate is decided. This is partly explored. A 2025 study predicted survival or death after heat stress from images, then picked cells for single-cell RNA sequencing and compared predicted-dead and predicted-survivor cells.⁹ The searches did not surface this for division versus death after chemotherapy.

Why it matters

Time-lapse imaging is usually descriptive. This work shows it can also be predictive, and a model's focus can suggest when fate takes shape. The open question is whether it holds on new experiments.