Explainable AI to unveil cellular autophagy dynamics

Abstract

Autophagy is a fundamental intracellular renovation process vital for maintaining cellular homeostasis through the degradation and recycling of damaged components. It is implicated in numerous pathological conditions, including cancer and neurodegenerative diseases. However, its dynamic nature and complexity pose challenges for manual analysis. In this study, we present a computational pipeline that leverages advanced deep learning models to automate the analysis of autophagic processes in 6,240 fluorescence microscopy images from the CELLULAR dataset. Our framework integrates object detection, cell segmentation, classification by autophagic state, cellular tracking, and explainability methods for interpretability. We achieved optimal results using YOLOv8 for object detection with a mAP50 of 0.80, U-Net++ for segmentation with an IoU of 0.82, and a vision transformer for classification with an accuracy of 0.86. To track cells, we developed a custom algorithm capable of handling complex scenarios such as cell division and morphological changes, all without requiring annotated tracking data. To enhance transparency, we employed explainability techniques based on class activation mappings to analyze model decision-making processes and validate classification outcomes, complemented by t-SNE visualizations for deeper insights into the data. Collaboration with biology experts validated our findings, highlighting the pipeline’s potential to advance autophagy research. This study demonstrates the potential of deep learning and explainable AI to streamline biomedical research, reduce manual effort, and uncover key autophagy dynamics.