A Deep Learning Pipeline for Genome-Wide Imaging Screen Uncovering Cell Death Regulators

In numerous image-based studies, such as cell studies, a plethora of information is embedded in cell objects, which must first be accurately extracted before it can be utilized for further downstream analysis and linked to biological processes explored in the original experiment. One such technique for this purpose is instance segmentation. In our study, which cell death and associated pathways and regulatory networks are investigated through a genome-wide microscopic screen, the dataset is extensive and challenging to manually curate or annotate. Therefore, we opted to train distinct U-shape nuclei segmentation models using only a small but meticulously annotated subset of the high-content screen data to ensure high accuracy where the superior U-Net model achieved an F1-score of 89%. Subsequently, we have deployed the best U-Net model to segment whole objects in the nuclei channel of the original genome-wide dataset, comprising nearly 5 million image frames across 213 plates. The extracted features from the nuclei channel served as biological metrics to explore cell death. The results identified several well-known genes involved in established cell death pathways, such as apoptosis. Additionally, other genes were discovered that merit further investigation for their effects on altering the visual features of cell nuclei.