Fig 1.
(A) Cell type annotation is used to extract cell type fractions in each sample. Next cell type fraction is discretized by learning Gaussian Mixture Model (GMM) for this type, respectively. (B) A Bayesian network (BN) is learned using the discretized cell abundance information. Bootstrapping is performed to identify high scoring differential interactions between cell types. (C) For pairs identified in the directed bootstrap BN analysis, a ligand-target regression (LTR) model is learned. In this model we use the expression change of ligands in the cell type with the outgoing edge to predict the expression of targets genes in the cell type with incoming edge. (D) Finally, LTR is used to select key ligands that underlie the cell-cell interactions identified in the BN. cell interaction.
Fig 2.
Bayesian Networks (BN) learned for lung cell types in healthy and IPF individual.
(A) BN for controls (healthy individuals). (B) BN for IPF patients. Nodes represent specific cell types and are colored accordingly, edges represent directed interactions between the cell types. Edge width corresponds to its bootstrap score.
Table 1.
Top differential cell type interactions identified by CINS for the IPF dataset.
The IPF-Control column lists the difference in the number of times the edge between the two cells was identified in 100 bootstrap runs for each of the two datasets. Negative values indicate that it was identified more for the Control whereas positive numbers mean that the interaction is more prevalent in IPF. For all listed edges the interaction was only identified in one of the two datasets (score of 100 or -100).
Fig 3.
Interactions learned by the BN are more significant than interactions between cells of the same type.
Comparison between the ability of the LTR model to predict target expression change when learning the model using cell pairs identified by the BN (A) and the same cell type (B). The x axis represents the bootstrapped edge count (score) of the interaction in the BN for a cell type pair, and the y axis represents the LTR model performance (higher is better) for the same cell pair.
Fig 4.
(A) BN for young mice. B) BN for adult mice. Nodes and edges notations and colorings are similar to those used in Fig 2.
Fig 5.
LTR comparison for the aging data.
Comparison between the ability of the LTR model to predict target expression change when learning the model using cell pairs identified by the BN (A) and the same cell type (B).
Fig 6.
Permutation analysis highlights the agreement between the two aging networks.
(A) Leftmost–learning using the Angelidis (15 samples) dataset. (B) Top–Learning combined networks using both Angelidis and real new data. Bottom–Learning combined networks using both Angelidis and permutation of cell type fractions in the new data. (C) Overlap in bootstrapped edges between the original and combined model when using the real data (red dashed line) and the permutation data (blue distribution).