Introduction: Obtaining large quantities of functionally mature hepatocytes remains a significant challenge for clinical applications, including treatment of liver failure via hepatocyte transplantation. Differentiating human pluripotent stem cells (hPSCs) into hepatocyte-like cells (HLCs) offers a promising solution, with numerous protocols developed for this purpose. However, these methods typically yield immature cells lacking the full repertoire of functionalities characteristic of primary human hepatocytes.

We developed machine learning (ML) algorithms capable of predicting hepatocyte maturation directly from live-cell, label-free phase contrast (PC) images. These algorithms were successfully deployed to analyze phenotypic screening data, enabling the optimization of differentiation protocols to produce HLCs with improved functional maturation.

Methods: hPSC differentiation into HLCs was performed following an 18-day protocol described in Ang et al. 2018. ML models were trained on paired PC and immunofluorescence (IF) images of markers involved in metabolism, detoxification, and protein synthesis, as well as in vitro assays of hepatocyte function. Models took solely PC images as input and produced a hepatocyte maturation score as output. 

Validated ML models were used to rank and select hits in phenotypic screening experiments where cells were exposed to small molecules and biologics at varying concentrations, durations, and combinations. After iterative cycles of screening, the top-ranking hits were compared to controls in a follow-up validation experiment using gold standard in vitro assays of hepatocyte function – CYP3A4 enzymatic activity, albumin secretion, and urea secretion. A mixed-effect model was used to assess statistical significance.

Results: ML-predicted hepatocyte maturation scores showed strong correlations (Pearson r > 0.9) across a panel of IF markers and functional assays. Leveraging these ML predictions in live monitoring of cultures undergoing differentiation, we rapidly and cost-effectively screened for combinations of molecules that significantly enhanced hepatocyte maturation, with median fold-change of 3.3x albumin secretion, 5.8x CYP3A4 enzymatic activity, and 10.6x urea secretion compared to baseline.

Conclusion: Our approach highlights a novel, ML-driven approach to optimizing stem cell differentiation protocols for improved functional maturation. Not only are the resulting HLCs attractive as a more potent candidate for hepatocyte transplantation, but also we believe the same strategy may be used to further improve differentiation protocols for other high-value cell types.