A functional genomics predictive network model identifies regulators of inflammatory bowel disease

Abstract

A major challenge in inflammatory bowel disease (IBD) is the integration of diverse IBD data sets to construct predictive models of IBD. We present a predictive model of the immune component of IBD that informs causal relationships among loci previously linked to IBD through genome-wide association studies (GWAS) using functional and regulatory annotations that relate to the cells, tissues, and pathophysiology of IBD. Our model consists of individual networks constructed using molecular data generated from intestinal samples isolated from three populations of patients with IBD at different stages of disease. We performed key driver analysis to identify genes predicted to modulate network regulatory states associated with IBD, prioritizing and prospectively validating 12 of the top key drivers experimentally. This validated key driver set not only introduces new regulators of processes central to IBD but also provides the integrated circuits of genetic, molecular, and clinical traits that can be directly queried to interrogate and refine the regulatory framework defining IBD.

Figure 1

An integrative approach for constructing a predictive network model of IBD, and identifying and validating master regulators of these networks. (a) Identification of causal IBD genes. We identified IBD-associated DNA variants in immune cells and digestive-tissue-derived CRE regions, some of which also corresponded to eQTLs derived from patients with IBD. Identification of a core IAM. Three different populations representing distinct states of disease were profiled, and the resulting data were integrated to build predictive molecular networks of the intestine. The core IAM was derived from coexpression modules from pediatric and adult patients with moderate and advanced IBD who were screened to link the immune network to active IBD in a disease-relevant context. The core IAM from all three populations was identified as the most highly enriched for genes in the immune network but then also highly enriched for causal IBD GWAS genes and macrophage expression. (b) Identification and annotation of IBD networks. Subnetworks for each IBD cohort representing the core IAM were identified as different instances of the CIC IBD network model and annotated using a diverse set of data. Identification and ranking of KDGs. KDGs were identified from the CIC IBD networks and prioritized by the size of their effect on the network as well as the degree of disease subnetwork support. (c) KDG experimental validation. To validate the utility of the CIC IBD model as a regulatory framework for modeling IBD, the top KDGs were experimentally validated in IBD mouse models and human macrophage cell systems. (d) Molecular validation of the CIC IBD model. Signatures from mouse intestine and human macrophages were projected onto the CIC IBD model to evaluate the degree to which the model predicted KDG perturbation signatures. (e) Phenotype validations. In vitro and in vivo phenotypes of inflammation were measured to evaluate the functional role of the KDGs in the context of IBD. CD, Crohn's disease; UC, ulcerative colitis; KD, knockdown.

Figure 2

Ranking KDGs of the CIC IBD networks. One hundred and thirty-three KDGs were identified and scored with respect to genetic association to IBD, involvement in inflammatory processes associated with IBD, and association with IBD-related clinical traits. The KDGs are listed clockwise around the disk in rank order, starting with the top ranked KDG at the top of the disc. Each of the internal rings represents either a summary statistic used in the ranking or a type of disease support used to form the summary statistics for the ranking. The first two tracks are normalized rank and rank by disease trait enrichment, respectively, and the remaining tracks 3–10 are KDG rankings per trait (Supplementary Table 16): track 3, IBD GWAS; track 4, original immune network; track 5, Crohn's disease versus control (CL)_ileum; track 6, Crohn's disease versus CL_colon; track 7, Crohn's disease versus ulcerative colitis colon; track 8, CRP; track 9, disease duration; track 10, VEO IBD genes. The color depicted in each element signifies the degree of support provided by the trait rank for the corresponding KDG, with cooler colors reflecting weaker support and hotter colors reflecting stronger support. Black asterisks denote KDGs that have already been linked to IBD experimentally. Orange asterisks indicate prioritized KDGs for experimental validation.

Figure 3

Transcriptional responses in stimulated macrophages to perturbations in macrophage KDGs are predicted by the IBD networks. (a) Table values represent fold change in expression of the indicated KDG in stimulated versus unstimulated macrophages. The network image illustrates the LPS-induced gene expression changes (red increased expression; blue decreased expression; intensity indicates magnitude of fold change) in the CERTIFI KDG (diamond nodes) network neighborhoods. (b) The heat map represents the −log₁₀ P value for the enrichment of genes whose expression levels change in response to siRNA-mediated knockdown of the KDGs. The CERTIFI IBD network adjacent to the heat map is a representative example of genes that were upregulated (blue nodes) and downregulated (red nodes) in response to knockdown of the macrophage KDG TNFAIP3. The network-predicted TNFAIP3 signature was 1.84-fold enriched for genes in the TNFAIP3 macrophage knockdown signature (Fisher's exact test (FET) P = 0.003). (c) Subnetwork of cytokines whose protein levels change in response to siRNA knockdown of TNFAIP3 siRNA in macrophage. This subnetwork of differential protein cytokine expression contains TNFAIP3 as well as other KDGs including LAPTM5, DOCK2, and GBP5. The red node represents the macrophage KDG TNFAIP3; blue nodes represent cytokines significantly differentially expressed in response to siRNA knockdown of TNFAIP3; purple nodes represent KDGs.

Figure 4

Enrichment analysis of KDG subnetworks. (a) Predicted transcriptional signature of 5 KDGs in the adult IBD networks. (b) The KDG (diamond) GPSM3 subnetwork is 2.76-fold enriched (Fisher's exact test, P < 0.008) for monocyte and macrophage IBD CRESNPs (light green) and 3.58-fold enriched (Fisher's exact test, P < 1.89 × 10⁻⁹) for differentially expressed nodes in the Gpsm3 DSS knockout signature (forest green) or both (bright green). Nodes present in the CRP, calprotectin, and lactoferrin trait signatures (blue border) are represented. The GPSM3 subnetwork reflects genes involved in macrophage function. C1orf228 (encoding p40, the molecular target of ustekinumab) is also present in this subnetwork (red). (c) The KDG (diamond) DOCK2 subnetwork is 2.2-fold enriched (Fisher's exact test, P = 0.02) for T cell IBD CRESNPs (light green), 5.13-fold enriched in T cell expression (Fisher's exact test, P = 1.84 × 10⁻⁷), and 4.59-fold enriched for genes upregulated in the DOCK2 perturbation signature (Fisher's exact test, P = 7.91 × 10⁻¹⁸) (forest green) or both (bright green). The DOCK2 subnetwork contains many genes represented in the CRP, calprotectin, and lactoferrin trait signatures from the CERTIFI cohort (blue border). Also represented is the IL-12Rβ1 receptor chain (in red) that comprises a chain in the IL-12 and IL-23 receptor and binds p40, the ligand to ustekinumab. The DOCK2 subnetwork is also 2.2-fold enriched (Fisher's exact test, P = 0.005) in a ROR-γT knockout differential expression signature (triangle). Each circular node represents an expressed gene, and the directed edges connecting genes represent causal or correlative relationships among the genes in the populations from which the network was built.

Figure 5

FACS analysis of immune cells in the colonic lamina propria of KDG-knockout mice as compared to wild-type littermate controls. (a) Events are electronically gated on CD45⁺CD3⁺CD4⁺ cells, and cells within colored contour plots show staining for IFN-γ and IL-17A. (b) Box plots show percentages of CD4⁺ T cells producing IL-17, IL-17 and IFN-γ, or IFN-γ in the KDG-knockout mice. (c) Colonic lamina propria cells were isolated from the indicated knockout strains and wild-type controls and stained with anti-CD45, anti-CD11c, anti-CD11b, anti-CD103, anti-CD64, and anti–MHC II antibodies. Cells were electronically gated on CD45⁺CD11⁺MHC II⁺ cells and further subdivided by staining for CD103, CD11b, and CD64. Box plots show the percentages of CD103⁺ DCs, CD103⁺CD11b⁺ DCs, CD103⁻CD11b⁺ DCs, and CD64⁺ macrophages. Data shown are representative of four independent experiments. One-way ANOVA with Bonferroni's multiple comparison was performed. Box limits, first and third quartiles; line, median; whiskers, minimum and maximum values. Statistical significance is indicated as follows: *P < 0.05, **P < 0.01, ***P < 0.001.

Figure 6

Differential weight loss and intestinal inflammation of KDG-knockout models as compared to sex-matched wild-type littermate controls. (a,b) Weight-loss curves for the KDG models. Nckap1l^−/−, Gpsm3^−/−, Dock2^−/−, Aif-1^−/−, and Dok3^−/− mice under DSS colitis conditions (a) and Dock2^−/− mice under TNBS conditions (b) relative to wild-type littermate controls. Comparisons were performed using an autoregressive model to maximize use of the time-series data. (c) Colonoscopy severity score on day 7 or 12. Pairwise comparison of endoscopy results was performed using the Mann–Whitney test. (d) Images shown are representative of endoscopy scoring performed with blinding to mouse group. (e) Histology scores. (f–k) Images of sections stained with hematoxylin and eosin. (f) A DSS-treated wild-type littermate shows mucosal inflammation, submucosal inflammation, and focal erosion (black arrowhead). (g) A DSS-treated Dok3^−/− mouse shows marked mucosal, submucosal, and muscularis inflammatory infiltrate and ulceration (black arrowhead). (h) A DSS-treated Gpsm3^−/− mouse shows more pronounced inflammation with involvement of the mucosa, submucosa, and muscularis propria and erosion (black arrowhead) as compared to the DSS-treated wild-type mouse. (i) Mild mucosal and submucosal inflammation in a DSS-treated Aif1^−/− mouse. (j) A representative TNBS-treated wild-type mouse displaying only focal mucosal inflammation. (k) A TNBS-treated Dock2^−/− mouse shows additional submucosal edema and inflammation as compared to the TNBS-treated wild-type control. *, mucosal inflammation; #, submucosal inflammation; &, muscularis inflammation. Images taken at 10× magnification have a scale bar representing 100 μm, while those taken at 20× magnification have a scale bar representing 50 μm. Data shown represent pooled results from males and females with two independent experiments for each KDG unless otherwise stated. Data are expressed as mean ± s.e.m. Box limits, first and third quartiles; line, median; whiskers, minimum and maximum values; asterisks, significant difference between mice homozygous null for the KDG and wild-type littermate controls treated with DSS: *P < 0.05; **P < 0.01; ***P < 0.001, ****P < 0.0001.

Figure 7

Schematic of crosstalk of KDG molecular and network pathways. The icons represent multiscale datastreams from populations of patients with IBD, including DNA, RNA, and protein collected from blood and intestine of patients across different disease stages. KDG network nodes (diamonds) regulate other network nodes in subnetworks defined by the orange edges, depicting causal regulatory relationships among the network nodes. For example, the NF-κB pathway, RAC, and its actin cytoskeleton rearrangement, RAS, the NLRP3 inflammasome pathways, and TLR and chemokine receptors are modulated by the KDGs we identified, and all have been reported as associated with IBD.