A 3D system to model human pancreas development and its reference single-cell transcriptome atlas identify signaling pathways required for progenitor expansion

Human organogenesis remains relatively unexplored for ethical and practical reasons. Here, we report the establishment of a single-cell transcriptome atlas of the human fetal pancreas between 7 and 10 post-conceptional weeks of development. To interrogate cell–cell interactions, we describe InterCom, an R-Package we developed for identifying receptor–ligand pairs and their downstream effects. We further report the establishment of a human pancreas culture system starting from fetal tissue or human pluripotent stem cells, enabling the long-term maintenance of pancreas progenitors in a minimal, defined medium in three-dimensions. Benchmarking the cells produced in 2-dimensions and those expanded in 3-dimensions to fetal tissue identifies that progenitors expanded in 3-dimensions are transcriptionally closer to the fetal pancreas. We further demonstrate the potential of this system as a screening platform and identify the importance of the EGF and FGF pathways controlling human pancreas progenitor expansion.


GATA6
Flow cytometry data for INS, GCG and PDX1 showing the low differentiation capacity at passage 1 after the last thawing both at day 10 (N=6) and day 17 (N=2). This can be compared to the increased differentiation more than 1 passage after thawing shown in Fig. 5c. c Summary of qPCR analyses after exocrine differentiation. Expression levels were normalized to PP-spheroids, N=2. Data shown as mean ± SEM. *P < 0.05, **P < 0.01, ***P < 0.001. P values were determined by two-sided Mann-Whitney test.

Network Inference Methods
In order to demonstrate the advantages of Intercom, we set out to compare the cell-cell communication networks reconstructed by Intercom with those obtained by state-of-the-art tools.
In particular, we employed three different tools, CellChat 1 , SingleCellSignalR 2 and ICELLNET 3 , which solely rely on a scRNA-seq dataset for network reconstruction. Although other tools, such as CellTalker 4 and iTALK 5 , have been developed, we excluded them from this assessment as they only consider differentially expressed ligands and receptors.
To guarantee an unbiased analysis, most parameters have been set to their default values. However, we set the significance parameter of SingleCellSignalR to 0.9, which is the same cutoff employed in Intercom, and employed CellChat on projected data as recommended for sparse scRNA-seq samples. As a result, we were able to obtain reconstructed cell-cell communication networks from CellChat and SingleCellSignalR for the 7 and 9 wpc (weeks post conception) samples. However, ICELLNET failed to return networks of significant interactions even under varying parameter settings. Therefore, we excluded ICELLNET from the subsequent network comparison.
We first compared the reconstructed network quantitatively with respect to the number and type of interactions as well as the participating ligands and receptors. As a result, we observed vast differences between the networks reconstructed by Intercom, CellChat and SingleCellSignalR. In total, at most four ligands and five receptors participate in interactions inferred by all tools whereas up to 59% of molecules are unique to a single method ( Supplementary Fig. 7a, b). As expected from these results, all methods predict vastly different interactions even without considering the sending and receiving cell populations ( Supplementary Fig. 7c). To gain insight into the different types of interactions that are predicted by each method, we annotated each interaction as one of "Cell-Cell Contact", if it is an interaction between two membrane proteins, "ECM-Receptor", if the ligand is an extracellular matrix protein, and "Secreted Signaling", if the ligand is secreted and not an extracellular matrix protein. While CellChat provides this information as a result of the network reconstruction, we annotated the results of Intercom and SingleCellSignalR based on the Uniprot 6 location annotation of the ligand. In case it is annotated as "Secreted" and not as "Extracellular Matrix" the interaction is considered to be "Secreted Signaling". If the ligand is annotated as "Extracellular matrix" the interaction is considered to be "ECM-Receptor" interaction. Finally, in case both proteins are receptors that are not annotated to as "Secreted" the interaction is considered to be "Cell-Cell Contact". As a result, we observed that each method has a unique proportion of interaction types that is similar in both the 7 and 9 wpc networks ( Supplementary  Fig. 7d). For instance, while more than 50% of interactions in the SingleCellSignalR networks are related to cell-cell contact, the interaction types in the CellChat networks are more balanced. In contrast, between 50% and 60% of interactions in the Intercom networks belong to "Secreted Signaling" whereas the remaining interactions are "ECM-Receptor" interactions. However, we observed that 10 interactions in the 9+6 wpc network are classified as "Cell-Cell Contact" although only secreted ligands are included in the interaction scaffold. Nevertheless, these interactions all involve Ephrin A1 (EFNA1), a membrane protein that can be secreted as well. Finally, we assessed the interaction frequency between different cell populations in the networks reconstructed by each method (Supplementary Fig. 7e). To account for the different number of interactions inferred by each method, we normalized the number interactions between a pair of cell populations by the total number of interactions. We observed that, depending on the interacting cell populations, the frequency of interactions in the Intercom networks is more similar either CellChat or SingleCellSignalR.
Despite the quantitative comparison, we qualitatively assessed the cell-cell communication networks of both CellChat and SingleCellSignalR with respect to the one identified by Intercom (Supplementary Data 3). For the 7 wpc, the communication network by SingleCellSignalR contained a total of 95 interactions out of which only six interactions matched with those of Intercom with respect to the ligand, the receptor as well as the sending and receiving cell populations. When examining the remaining 89 interactions, we found 27 false positives. More specifically, 25 of these interactions involve UBA52 as a ligand whereas the remaining two interactions are predicted to be between HSP90AA1 and FGFR3. However, UBA52 is neither secreted nor a membrane protein and HSP90AA1 is not a native ligand of FGFR3 (Supplementary Data 3). The remaining cell-cell interactions unique to SingleCellSignalR are related to Notch signaling (Notch1, Notch2, Notch3), Laminin signaling (RPSA) and ITGB1 signaling through midkine. Similarly, the 9 wpc communication network by SingleCellSignalR contains a total of 70 interactions out of which 8 interactions can also be found in the Intercom network. Overall, the interactions of the 7 and 9 wpc network are largely similar with 35 unique interactions related to Notch signaling (Supplementary dataset 1). Thus, the unique interactions predicted by SingleCellSignalR are largely related to cell-cell contact, which is expected given the design considerations of Intercom. However, the tool fails to predict important cell-cell interactions, such as Bmp and Fgf signaling.
Following the same rationale, we compared the cell-cell communication networks reconstructed by CellChat and Intercom. Since CellChat also contains receptor complexes, we considered an interaction to be common if Intercom predicts an interaction with at least one of their subunits. As a result, we observed that CellChat predicts 449 and 226 interactions for the 7 and 9 wpc sample, respectively, of which 10% and 3% are also predicted by Intercom (Supplementary Data 3). Similar to the SingleCellSignalR networks, the networks reconstructed by CellChat contain interactions related to cell-cell contact. In particular, CellChat predicts CD99, CDH1 and Notch signaling in both cases, which are not captured by Intercom. However, CellChat also predicts more than 45% of interactions whose participating ligand or receptor (subunit) is only expressed in few cells (below 5%). Notably, in some cases the ligand or receptor is not expressed in any cell (Supplementary Data 3). We speculate that this is due to CellChat's data transformation strategy to cope with sparse scRNA-seq data. The remaining interactions unique to the CellChat networks are related to NCL, NTRK2, Plexin and syndecan-2 signaling.
In summary, both CellChat and SingleCellSignalR are predicting many interactions that are unlikely to take place. Nevertheless, the cell-cell contact interactions these methods predict are complementary to the cell-cell communication network reconstructed by Intercom. Nevertheless, the absence of a ground truth network prevents a quantitative evaluation of the accuracy of each method. Especially CellChat potentially shows an improved performance when presented with a more dense expression dataset, for instance obtained from SmartSeq-based technologies.