Single-cell genomic methods now provide unprecedented resolution for characterizing the component cell types and states of tissues such as the epithelial subsets of the gastrointestinal tract. Nevertheless, functional studies of these subsets at scale require faithful in vitro models of identified in vivo biology. While intestinal organoids have been invaluable in providing mechanistic insights in vitro, the extent to which organoid-derived cell types recapitulate their in vivo counterparts remains formally untested, with no systematic approach for improving model fidelity.
Harvard Medical School and MIT researchers present a generally applicable framework that utilizes massively parallel single-cell RNA-seq to compare cell types and states found in vivo to those of in vitro models such as organoids. Furthermore, they leverage identified discrepancies to improve model fidelity. Using the Paneth cell (PC), which supports the stem cell niche and produces the largest diversity of antimicrobials in the small intestine, as an exemplar, the researchers uncover fundamental gene expression differences in lineage-defining genes between in vivo PCs and those of the current in vitro organoid model. With this information, they nominate a molecular intervention to rationally improve the physiological fidelity of our in vitro PCs. They then perform transcriptomic, cytometric, morphologic and proteomic characterization, and demonstrate functional (antimicrobial activity, niche support) improvements in PC physiology.
Transcriptional benchmarking of in vitro Paneth cells (PCs) to in vivo
a Schematic of intestinal epithelial cell isolation from terminal ileum for unbiased identification of in vivo PC signature genes, and system for intestinal stem cell (ISC) enrichment to characterize in vitro PCs, via high-throughput scRNA-seq. b Marker gene overlay for binned count-based expression level (log(scaled UMI + 1)) of Lyz1, a canonical PC marker gene, on a tSNE (t-stochastic neighbor embedding) plot of 7667 small intestinal epithelial cells isolated from the terminal ileum; receiver operating characteristic (ROC)-test area under the curve (AUC) = 0.995, n = 2 mice, independent experiments. c Violin plot for the count-based expression level (log(scaled UMI + 1)) of Lyz1 across clusters identified through shared nearest neighbor (SNN) analysis over small intestinal epithelial cells; n = 196 cells in cluster 11, 7667 cells in total. d A tSNE plot of 2513 cells, with clusters identified through SNN (Additional file 1: Table S1 for full gene lists with ROC-test AUC > 0.60) from conventional ENR organoids; n = 6 wells of ENR organoids. e Marker gene overlay for binned count-based expression level (log(scaled UMI + 1)) of Lyz1 on a tSNE plot from; ROC-test AUC = 0.856. f Violin plot of expression contribution to a cell’s transcriptome of PC genes across ENR organoid clusters from (d) (In vivo PC gene list AUC > 0.65); effect size 0.721, ENR-4 vs. all ENR, *t test p < 2.2 × 10−16. g Row-normalized heatmap of top differentially expressed genes using bimodal test over single-cells from the top 200 PC-like cells from ENR-4 and the 196 in vivo PCs (cluster 11, from (c)); *bimodal test, all displayed genes p < 1.89 × 10−16 or less with Bonferroni correction. h Violin plots for the count-based expression level (log(scaled UMI + 1)) of Lyz1, Ang4, and Defa3 in ENR and in vivo PCs; *bimodal test, all p < 2.92 × 10−37 or less with Bonferroni correction. i Violin plot of expression contribution to a cell’s transcriptome of PC genes (effect size 1.25, InVivo vs. ENR, *t test p < 2.2 × 10−16), Wnt pathway (effect size 0.559, InVivo vs. ENR, *t test p < 2.035 × 10−8) and Notch pathway (effect size −0.500, InVivo vs. ENR, *t test p < 5.25 × 10−7) genes
This systematic approach provides a simple workflow for identifying the limitations of in vitro models and enhancing their physiological fidelity. Using adult stem cell-derived PCs within intestinal organoids as a model system, the researchers successfully benchmark organoid representation, relative to that in vivo, of a specialized cell type and use this comparison to generate a functionally improved in vitro PC population. They predict that the generation of rationally improved cellular models will facilitate mechanistic exploration of specific disease-associated genes in their respective cell types.