Deep Learning-Based HCS Image Analysis for theEnterprise

摘要

Drug discovery programs are moving increasingly toward phenotypic imaging assaysto model disease-relevant pathways and phenotypes in vitro. These assays offerricher information than target-optimized assays by investigating multiplecellular pathways simultaneously and producing multiplexed readouts. However,extracting the desired information from complex image data poses significantchallenges, preventing broad adoption of more sophisticated phenotypic assays.Deep learning-based image analysis can address these challenges by reducing theeffort required to analyze large volumes of complex image data at a quality andspeed adequate for routine phenotypic screening in pharmaceutical research.However, while general purpose deep learning frameworks are readily available,they are not readily applicable to images from automated microscopy. During thepast 3 years, we have optimized deep learning networks for this type of data andvalidated the approach across diverse assays with several industry partners.From this work, we have extracted five essential design principles that webelieve should guide deep learning-based analysis of high-content images andmultiparameter data: (1) insightful data representation, (2) automation oftraining, (3) multilevel quality control, (4) knowledge embedding and transferto new assays, and (5) enterprise integration. We report a new deeplearning-based software that embodies these principles, Genedata Imagence, whichallows screening scientists to reliably detect stable endpoints for primary drugresponse, assess toxicity and safety-relevant effects, and discover newphenotypes and compound classes. Furthermore, we show how the software retainsexpert knowledge from its training on a particular assay and successfullyreapplies it to different, novel assays in an automated fashion.