Fig 1.
Schematic of CoCoNet for identifying trait relevant tissue/cell type using information from gene co-expression networks.
From left to middle: we calculate gene level effect measurements from GWAS summary statistics for the trait of interest and treat them as an m dimensional vector of outcomes. From right to middle: we infer the tissue specific gene co-expression network for each tissue/cell type using the PANDA software, which requires the tissue-specific gene expression matrix, existing protein-protein interaction network information, as well as existing transcription factor and gene binding information as input. For the trait of interest, we examine one tissue at a time and we model the gene-level effect measurements for the trait as a function of the gene co-expression matrix using a covariance regression network model. We infer parameters in the model through composite likelihood. We calculate the maximum composite likelihood for each tissue and eventually rank tissues by the corresponding log likelihoods.
Fig 2.
Power of CoCoNet to identify trait relevant tissues in simulations.
(A): In scenario I, we selected one of the ten tissues as the trait-relevant tissue (second row) and simulated gene-level effect measurements based on the adjacency matrix (first row). We treated the adjacency matrices all as observed (third row) and fit the model using each of the ten matrices. (B): In scenario II, the data are generated in the same way as in scenario I but we fit the model using tissue-specific matrices that are noisy versions of the truth (third row). In particular, we assumed that each connected gene pair in the true adjacency matrices has a probability of p being un-observed, and each unconnected gene pair has a probability q to be falsely assigned as connected. (C): In scenario III, the data are again generated in the same way as in scenario I but we fit the model using tissue-specific matrices that are noisy versions of the truth (third row). In particular, we randomly converted a proportion q of unconnected gene pairs in the true adjacency matrix to be connected. (D): The power of CoCoNet for identifying the correct tissue (y-axis) increases with increasing signal strength measured by (x-axis) in scenario I. (E): The power of CoCoNet for identifying the correct tissue (y-axis) gradually decreases with increasing noise level characterized by p (x-axis) in scenario II. (F): The power of CoCoNet for identifying the correct tissue (y-axis) gradually decreases with increasing noise level characterized by q (x-axis) in scenario III.
Fig 3.
Comparison of tissue rankings by CoCoNet, LDSC-SEG, and RolyPoly for four neurological traits.
(A): For each of the four neurological diseases, we plotted the rank (y-axis) of three brain tissues (colored blue; including brain cerebellum, brain basal ganglia, and brain other) and the rank of the remaining 35 tissues (colored green) in separate boxplots. The rank of brain tissues obtained using CoCoNet is often lower than that of the non-brain tissues (first column in each panel). The rank difference between the two types of tissues obtained by CoCoNet (first column in each panel) is often more pronounced than that obtained by either LDSC-SEG (second column) or RolyPoly (third column) for the four neurological diseases. (B): We compared the ranking of tissues obtained by CoCoNet with that obtained by PubMed search for 4 neurological traits, in the bulk RNAseq data. The CoCoNet and PubMed search results are positively correlated with each other for three traits. (C): We compared the ranking of cell types obtained by CoCoNet with that obtained by PubMed search for 4 neurological traits, in the single cell RNAseq data. The CoCoNet and PubMed search results are positively correlated with each other for two traits.
Fig 4.
Comparison of tissue rankings by CoCoNet, LDSC-SEG, and RolyPoly for four autoimmune traits.
(A): For each of the four autoimmune diseases, we plotted the rank (y-axis) of three colon tissues (colored orange; including colon sigmoid, colon transverse, and intestine terminal ileum) and the rank of the remaining 35 tissues (colored green) in separate boxplots. The rank of colon tissues obtained using CoCoNet is often lower than that of the non-colon tissues (first column in each panel). The rank difference between the two types of tissues obtained by CoCoNet (first column in each panel) is often more pronounced than that obtained by either LDSC-SEG (second column) or RolyPoly (third column) for the four autoimmune diseases. (B): We compared the ranking of tissues obtained by CoCoNet with that obtained by PubMed search for 4 autoimmune traits, in the bulk RNAseq data. The CoCoNet and PubMed search results are positively correlated with each other for two traits.
Table 1.
Top five tissue types identified among the 38 tissues in the GTEx bulk RNAseq data for each of the eight GWAS traits.
Table 2.
Top five cell types identified among 10 cell types in the single cell RNAseq data for each of the four neurological disorders.