Fig 1.
Overview of our deep learning pipeline for analyzing cellular autophagy.
Panel A: Biological experiments generate raw imaging data of Drosophila melanogaster S2 cells under different conditions. A theoretical model of autophagy illustrates key cellular stages, guiding the analysis. Processed data, derived from computational methods, provides insights into autophagy dynamics, enabling further interpretation. Panel B: Our computational pipeline follows five steps: (1) Detection to identify cells, (2) Segmentation to isolate structures, (3) Classification to categorize autophagy states, (4) Tracking to monitor cell dynamics, and (5) Explainable AI to enhance interpretability and biological relevance.
Fig 2.
Time-lapse imaging from the CELLULAR dataset.
The top panel shows progressive changes in a cell population over five time points. The middle panel (Fed Cells Sequence) depicts an individual fed cell remaining in a basal autophagy state throughout the time course. The bottom panel (Starved Cells Sequence) shows a starved cell transitioning from a basal to an activated autophagy state.
Table 1.
Comparison of segmentation outcomes between pre-trained and fine-tuned MedSAM, DeepLabV3+, and U-Net++ models applied to cropped cell images using bounding boxes.
Fig 3.
Starvation induces an autophagic response in S2 cell.
A: Western blot showing free mRFP levels from mRFP-EGFP-Atg8a S2 cells that were kept fed or starved for the indicated times. β-actin/Act5C was used as a loading control. B and C: Fed (yellow) and starved (orange) mRFP-EGFP-Atg8a S2 cells that were subjected to flow cytometry to detect cells with green (B) or red fluorescence (C). S2 cells with no reporter expression were used as negative control.
Fig 4.
Traditional image analysis of the CELLULAR dataset.
A: Spot area and spot count of mRFP-EGFP-Atg8a spots segmented based on the RFP channel in the annotated CELLULAR data set. B: PCA plots of the individual cells based on the ground truth annotations. C: Metrics for classification based on the activated autophagy threshold calculated from the 95th percentile of RFP spot area from fed cells at timepoint 1. D: Confusion matrix for the classification as in C.
Table 2.
Comparative classification performance using VGG, ResNet, and ViT models, with each model applied in two different training approaches on segmented cells: trained from scratch and fine-tuned.
Fig 5.
Temporal progression of a starved cell with explainable AI heatmaps.
The figure shows a five-time-point sequence from the test dataset, accompanied by GradCAM, EigenCAM, and AblationCAM heatmaps. These highlight key image regions influencing the ViT model’s predictions, with the predicted autophagy status shown above each frame. Notably, an error occurs at time point 3, where basal autophagy is misclassified as activated.
Fig 6.
t-SNE plot of processed test data cells.
The visualization reflects how cells cluster by class and time. Different colors indicate distinct cell classes, while varying shades within each color represent different time points.
Fig 7.
ViT model classification performance over time.
Classification results are shown for fed (a) and starved (b) cell sequences across multiple time points in the test dataset. Performance metrics include Accuracy, F1 Score, Precision, Recall, and MCC (Matthews Correlation Coefficient).
Fig 8.
Temporal progression with expert-annotated bounding boxes.
The figure illustrates annotation inconsistencies across five time points. While a single cell is identified at time points 2 and 5, multiple cells are annotated at the other time points, indicating variation in expert interpretation.