Discovering effective drugs and demonstrating their safety are significant challenges facing the pharmaceutical industry, due to the high costs of development, long lead times, and low success rates of late stage clinical trials. There is a need for new tools and technologies to help identify safe and effective drugs during the early stages of development. Over the last decade, there has been significant progress in using human induced pluripotent stem cells (hiPSCs) for modeling of human disease, drug screening, and toxicity testing. Numerous studies have demonstrated that these cells have physiologically relevant characteristics and can be used for preclinical testing of new drugs using high-throughput assays. In such assays, image and signal analysis algorithms are used to generate quantitative measurements that relate to cell degradation, death, or changes in function. Such approaches may be missing subtle changes that are not easily visualized, are too complex to measure with traditional data analysis methods, and/or suffer from lack of consistent quality control metrics on the input data.
Artificial intelligence (AI) techniques, and specifically deep convolutional neural networks, are perfectly suited to address the challenges of these high-throughput assays by analyzing large amounts of imaging data robustly and with a level of sensitivity that has not been previously possible. We present case studies for using AI for high-throughput image-based phenotypic screening, toxicity testing, and quality control. First, we present data from a drug discovery program for dilated cardiomyopathy using high-throughput imaging of sarcomere structure in stem cell-derived cardiomyocytes. We were able to build disease models with high accuracy, which were then deployed to identify small molecules that showed to reverse the disease phenotype. The identified small molecules were further validated with functional assays and preclinical mouse studies. Second, we present data from a pilot toxicity testing study using stem cell-derived cardiomyocytes. Our novel image-based AI method was successful in capturing dose-dependent structural changes on a panel of drugs with known cardiotoxicity profiles, while no change was detected for the negative control. The detected structural changes correlated strongly with contractility. Finally, we present data from a pilot quality control study using current-trace signals from a patch clamp instrument. We successfully built an AI model that can accurately classify signals as good versus poor quality, which enables automated and consistent filtering of data during high-throughput experiments.