Biliary epithelial cell-specific RAGE controls ductular reaction-mediated fibrosis during cholestasis

Transcriptomic datasets were featurized using a state-of-the-art archetypal analysis based framework.Using each of our transcriptomic and imaging datasets separately, and viewing pHSCs as ground truth, we quantified three quality metrics for iHSCs: (i) predictiveness of TGF-beta response in an ML classifier trained on pHSCs; (ii) separation between TGF-beta and DMSO in an unsupervised ML model; (iii) similarity of the overall cell state distribution between iHSCs and pHSCs.Results: The established ML framework provided informative featurizations of our cellular assays and allowed for the correction of experimental artifacts.We showed robustness of TGF-betainduced ML phenotype in pHSCs both in imaging and transcriptomics (mean AUC 0.96 for out-of-line assessment in imaging, Figure 1A; mean AUC 0.98 in transcriptomics, data not shown).Interpretation of learned ML phenotypes revealed insights into the TGF-beta responsiveness of pHSCs.This allowed us to utilize multi-parametric criteria to evaluate the cohort of iHSCs.We observed a strong correlation between transcriptomic and imaging characterizations used for the ranking of iHSC lines (spearman = 0.66, p = 8.7 * 10 -5 , Figure 1B). Conclusion:We developed a state-of-the-art approach for the characterization of HSC morphological and transcriptional phenotypes.Using the developed framework, we successfully modeled TGF-beta-induced activation signatures in HSCs.These image-based phenotypic assays paired with high-performance ML models present unique opportunities for genetic and chemical screens and the discovery of novel fibrosis targets.