Fig 1.
Overview of microfluidic platform.
(A) Design layout of the microfluidic device. An inlet channel distributes fresh medium to cell traps, which immobilize single mother cells; dashed box highlights the trapping region. (B) Representative phase-contrast frame showing yeast mothers trapped at the bottom ends of individual traps while daughters are flushed away by the medium flow. Each mother cell is continuously tracked throughout its replicative lifespan.
Fig 2.
BudFinder framework for automated division detection.
Raw time-lapse movies are used in two distinct training stages: [1] Masked Auto-Encoder (MAE) pretraining: single-frame crops from movies are used to pretrain an MAE in a self-supervised manner, learning visual representations without the need for division annotation data. [2] Division detection: temporally ordered sequences of MAE embeddings are passed to a lightweight temporal Transformer trained with annotated division events to predict whether a budding event occurs.
Fig 3.
Learning single cell image features in MAE via reconstruction for integrating with temporal transformer for division detection.
(A) Input image reconstruction process. Images are patchified and randomly masked, then positional embeddings are added and patches are encoded. Following encoding, mask tokens are introduced and the latent along with the mask tokens are decoded. Decoded patches are unshuffled using stored indices and the original image is then reconstructed. (B) Example of reconstructions produced by the pretrained MAE. Despite heavy masking during training, the MAE accurately reconstructs key cell morphology, indicating successful learning of yeast cell shape and internal structure. (C) After the reconstruction portion of training, weights are carried over, and the encoder is used on a frame-by-frame basis for each frame in an 11-frame tiff stack input. A temporal transformer takes the 11 latent representations as input and performs classification.
Fig 4.
Automated division-detection model accurately reproduces ground-truth replicative lifespan and outperforms an untrained baseline.
(A) Scatter plot of predicted versus manually annotated RLS for wild-type mother cells (n = 121). Points cluster tightly around the unity line (dashed red), indicating high agreement between model predictions and ground truth. (B) Bar graph comparing the division-event detection performance of the trained transformer model to a naïve heuristic using a ± 1-frame tolerance. Trained-versus-untrained comparisons are presented only as diagnostic controls, confirming that model accuracy depends on learned visual and temporal representations rather than intrinsic architectural priors.
Fig 5.
Automated division-detection accurately reproduces budding events irrespective of aging mode and daughter cell morphology.
(A) Representative phase-contrast images from the mother-machine device illustrating the two budding patterns observed in wild-type cells. Mode 1 shows elongated daughter cells (left), whereas Mode 2 cells exhibit small round daughter cells (right). Green arrowheads mark daughter cells; white dashed outlines delineate the mother cell boundary. (B) Mean ± SD replicative lifespan for Mode 1 and Mode 2 mothers measured by manual annotation (black bars) versus BudFinder model (grey bars). Model predictions closely match ground truth scores for both modes (two-tailed paired t-test, n.s. for Mode 1 and Mode 2). (C) Cell-cycle length trajectories over the course of life, expressed as a percentage of each cell’s total lifespan. Scatter points represent individual cycles; solid lines are Savitzky–Golay–smoothed trajectories (left, manual ground-truth annotations; right, model predictions). The model recapitulates the rapid budding characteristic of young cells, followed by progressive late-life cell-cycle lengthening, with Mode 2 cycles consistently longer than Mode 1 across both ground-truth and BudFinder model outputs.
Table 1.
Frame-level agreement between BudFinder model and manually annotated budding events across three genetic backgrounds. Percentages indicate the fraction of predicted division ticks that coincide with the ground-truth annotation in the same frame (Exact match) or within a tolerance window of ±1 or ±2 frames. Even under the rapid cycling of the oscillator strain and the shortened lifespan of sir2Δ, the model captures > 78% of events within ±1 frame and > 89% within ±2 frames, demonstrating robust temporal precision comparable to wild-type performance.
Fig 6.
BudFinder reproduces manual budding timelines, detects divisions in genetically perturbed strains, and accurately predicts replicative lifespan across diverse genetic backgrounds.
(A) Raster plots show ground-truth (blue) and model-predicted (orange) budding events for at least n = 15 representative mother cells per strain. The near-perfect superposition of orange and blue ticks across all three panels demonstrates that BudFinder preserves frame-level accuracy despite the distinct division frequencies and lifespan trajectories characteristic of each genotype, confirming its robustness to genetic perturbations. Prediction accuracy was quantified via the Hungarian algorithm. In the WT dataset, 54.5% of predictions matched ground truth exactly, with a mean error of +0.27 frames indicating a mild late bias; 30.3% of events were late, 8.1% early, and 7.1% unmatched. The sir2Δ dataset exhibited late bias, with 32.4% exact matches, 58.5% late predictions, 3.1% early, and 6.0% unmatched. The oscillator dataset showed a moderate early bias, with 40.8% exact matches, 35.5% early predictions, 15.6% late, and 8.1% unmatched. (B) Bar plots show F1 scores calculated with a 1-frame tolerance for the trained division-detection model (left bar in each panel) versus a naïve baseline (right bar) in sir2Δ cells (left) and an engineered long-lived oscillator strain (right). The trained model achieves F1 ≈ 0.75 in sir2Δ and F1 ≈ 0.68 in the oscillator, more than four-fold higher than the baseline in both cases. (C) Scatter plots compare predicted versus ground-truth replicative lifespan for sir2Δ (left, n = 172) and the engineered oscillator strain (right, n = 90). Each point represents one mother cell; the red dashed line denotes the 1:1 correspondence.
Fig 7.
Benchmarking division detection performance across different imaging platforms.
(A) Scatter plot comparing predicted versus ground-truth replicative lifespan for sir2 overexpression (n = 58). (B) Comparison of predicted and ground-truth division events for representative single cells.