Generative Modeling for Synthesis of Cellular Imaging Data for Low-Cost Drug Repurposing Application

Advances in high-content high-throughput fluorescence microscopy have emerged as a powerful tool for several stages of drug discovery process, leading to the identification of a drug candidate with the potential for becoming a marketed drug. This high-content screening (HCS) technology has recently involved the application of machine learning methods for automated analysis of large amount of data generated from screening of large compound libraries to identify drug induced perturbations. However, high costs associated with large-scale HCS drug assays and the limitations of producing abundant high-quality data required to train machine learning models, pose major challenges. In this work, we have developed a computational framework based on deep convolutional generative adversarial network (DCGAN), for the generation of synthetic high-content imaging data to augment the limited real data. The proposed framework was applied on cell-based drug screening image data to derive phenotypic profiles of drug induced effects on the cells and to compute phenotypic similarities between different drugs. Such analysis can provide important insights into repurposing of previously approved drugs for different conditions. Moreover, a generative modeling-based approach of creating augmented datasets can allow to screen more drug compounds within the same imaging assay, thus reducing experimental costs.