Non-invasive quantification of viability in liver spheroids using deep learning

Introduction: viability assays are essential in drug discovery, development, and pharmacovigilance. However, traditional methods for evaluating cell viability rely on destructive processes that render cultures non-viable, limiting them to single endpoint measurements and precluding further analyses. Methods: We present Neural Viability Regression (NViR), a deep learning-based method that enables real-time, non-invasive quantification of culture viability from microscopy images. Although developed and validated on liver spheroids, the framework includes a retrainable pipeline adaptable to other spheroid types. To demonstrate its applicability, we exposed human liver spheroids to 108 FDA-approved drugs and captured microscopy images over time, using NViR's viability estimates to predict Drug-Induced Liver Injury (DILI). Results: NViR's viability assessments accurately predicted whether a drug induces DILI in humans. Its non-invasive nature enabled frequent viability evaluations throughout experiments, capturing subtle temporal changes while preserving the structural integrity of the cultures and substantially reducing both culture and labor costs. Discussion: The cost-effectiveness and non-destructive characteristics of NViR enable high-frequency, high-throughput viability assessments, positioning it as a tool to enhance liver safety protocols and reduce both the costs and failure rates in drug discovery and development.