A Transformer reads raw microscopy video and calls cell death up to 8 hours early

Cell fate can be predicted directly from raw time-lapse image sequences, without predefined morphological or molecular features. Earlier cell-fate work relied on manually selected and quantified features, or focused on cell-cycle staging rather than directly predicting divergent outcomes under treatment. On a held-out test set of 2,089 trajectories the model reached 0.94 balanced accuracy and an F1-score of 0.9326, correctly identifying 91.3% of apoptotic samples while misclassifying only 3.5% of mitotic samples as apoptotic.

The predictive signal is not confined to the final frames of a cell trajectory; for apoptosis it is spread over a wider time window. When only the last few frames are kept, mitosis reaches 0.641 recall from the final frame alone, whereas apoptosis reaches only 0.699 recall from the last five frames. Frame truncation showed balanced accuracy falling to 0.919 after removing the last 5 frames and to 0.565 after removing 40; on a length-matched subset apoptosis recall dropped only from 0.91 to 0.81 while mitosis declined more steeply.

Attention weights identify which frames matter most, and those frames differ statistically in morphology and fluorescence intensity. Comparing high- and low-attention frames had not been done systematically under a Transformer's attention mechanism. Masking high-attention frames degraded performance more than masking random or low-attention frames, with balanced accuracy dropping to 0.836 at a masking fraction of 0.5 while random and low-attention masking retained 0.912 and 0.931.

The report covers single-cell fate prediction for cancer cell populations under cisplatin treatment, with long-term live-cell image sequences as input and a binary apoptosis-versus-mitosis output plus attention-based explanation. The explainability analyses focus on cisplatin-treated sequences because the control condition contains only 41 apoptotic sequences, too few for a robust characterization of apoptotic dynamics. It suits readers interested in single-cell imaging, drug-response heterogeneity and interpretable deep learning.

In the three-class setting the model reached 0.8434 balanced accuracy and 0.7946 macro F1, with most errors coming from confusion between the neither class and apoptosis, a label that is inherently ambiguous. The control condition holds only 41 apoptotic sequences, so those estimates are noisy and the authors do not interpret them quantitatively. Dose-stratified and condition comparisons are noisier still, and the authors refrain from drawing strong conclusions from them.