Deep learning of virus infections reveals mechanics of lytic cells

Abstract

Imaging across scales gives insight into disease mechanisms in organisms, tissues and cells. Yet, rare infection phenotypes, such as virus-induced cell lysis have remained difficult to study. Here, we developed fixed and live cell imaging modalities and a deep learning approach to identify herpesvirus and adenovirus infections in the absence of virus-specific stainings. Procedures comprises staining of infected nuclei with DNA-dyes, fluorescence microscopy, and validation by virus-specific live-cell imaging. Deep learning of multi-round infection phenotypes identified hallmarks of adenovirus-infected cell nuclei. At an accuracy of >95%, the procedure predicts two distinct infection outcomes 20 hours prior to lysis, nonlytic (nonspreading) and lytic (spreading) infections. Phenotypic prediction and live-cell imaging revealed a faster enrichment of GFP-tagged virion proteins in lytic compared to nonlytic infected nuclei, and distinct mechanics of lytic and nonlytic nuclei upon laser-induced ruptures. The results unleash the power of deep learning based prediction in unraveling rare infection phenotypes.