Deep-Learning-Assisted Assessment of DNA Damage Based on Foci Images and Its Application in High-Content Screening of Lead Compounds.

DNA damage is one of major culprits in many complex diseases; thus, there is great interest in the discovery of novel lead compounds regulating DNA damage. However, there remain plenty of challenges to evaluate DNA damage through counting the amount of intranuclear foci. Herein, a deep-learning-based open-source pipeline, FociNet, was developed to automatically segment full-field fluorescent images and dissect DNA damage of each cell. We annotated 6000 single-nucleus images to train the classification ability of the proposed computational pipeline. Results showed that FociNet achieved satisfying performance in classifying a single cell into a normal, damaged, or nonsignaling (no fusion-protein expression) state and exhibited excellent compatibility in the assessment of DNA damage based on fluorescent foci images from various imaging platforms. Furthermore, FociNet was employed to analyze a data set of over 5000 foci images from a high-content screening of 315 natural compounds from traditional Chinese medicine. It was successfully applied to identify several novel active compounds including evodiamine, isoliquiritigenin, and herbacetin, which were found to reduce 53BP1 foci for the first time. Among them, isoliquiritigenin from Glycyrrhiza uralensis Fisch. exerts a significant effect on attenuating double strand breaks as indicated by the comet assay. In conclusion, this work provides an artificial intelligence tool to evaluate DNA damage on the basis of microscopy images as well as a potential strategy for high-content screening of active compounds.