Transcriptional Behavior of Regulatory T Cells Predicts IBD Patient Responses to Vedolizumab Therapy

Abstract

Background

Inflammatory bowel disease (IBD) involves chronic T cell–mediated inflammatory responses. Vedolizumab (VDZ), a monoclonal antibody against α4β7 integrin, inhibits lymphocyte extravasation into intestinal mucosae and is effective in ulcerative colitis (UC) and Crohn’s disease (CD).

Aim

We sought to identify immune cell phenotypic and gene expression signatures that related to response to VDZ.

Methods

Peripheral blood (PBMC) and lamina propria mononuclear cells (LPMCs) were analyzed by flow cytometry and Cytofkit. Sorted CD4⁺memory (Tmem) or regulatory T (Treg) cells from PBMC and LPMC were analyzed by RNA sequencing (RNA-seq). Clinical response (≥2-point drop in partial Mayo scores [UC] or Harvey-Bradshaw index [CD]) was assessed 14 to 22 weeks after VDZ initiation. Machine-learning models were used to infer combinatorial traits that predicted response to VDZ.

Results

Seventy-one patients were enrolled: 37 received VDZ and 21 patients remained on VDZ >2 years. Fourteen of 37 patients (38%; 8 UC, 6 CD) responded to VDZ. Immune cell phenotypes and CD4⁺Tmem and Treg transcriptional behaviors were most divergent between the ileum and colon, irrespective of IBD subtype or inflammation status. Vedolizumab treatment had the greatest impact on Treg metabolic pathways, and response was associated with increased expression of genes involved in oxidative phosphorylation. The strongest clinical predictor of VDZ efficacy was concurrent use of thiopurines. Mucosal tissues offered the greatest number of response-predictive biomarkers, whereas PBMC Treg-expressed genes were the best predictors in combinatorial models of response.

Conclusions

Mucosal and peripheral blood immune cell phenotypes and transcriptional profiles can inform VDZ efficacy and inform new opportunities for combination therapies.

Key Messages.

• What is already known? Vedolizumab therapy in IBD is thought to function through inhibition of effector T cell recruitment to the intestine, but no biomarkers exist to predict response to this therapy.

• What is new here? The metabolic state of regulatory T cells in peripheral blood and mucosa and the use of thiopurines at baseline had the greatest predictive value for response to vedolizumab.

• How can this study help patient care? Our data inform potential biomarkers of response to vedolizumab using gene expression of peripheral blood Tregs and highlights the opportunity for drug development to enhance Treg function that may be rationally combined with vedolizumab.

Introduction

Treatment for inflammatory bowel disease (IBD) has improved dramatically over the past 3 decades. The mainstay of IBD therapy until recently has been antitumor necrosis factor (anti-TNF) agents. One new class of medicines—that includes vedolizumab (VDZ)—has been rationally designed to interfere with intestinal homing of leukocytes.^1,2 Vedolizumab has proven effective for the treatment of both ulcerative colitis (UC) and Crohn’s disease (CD). Approximately 30% of patients enter remission in clinical trials, although these numbers are lower in patients previously exposed to anti-TNFs.^3,4 A head-to-head study of VDZ and adalimumab in UC patients showed superiority of VDZ in inducing endoscopic improvement and clinical remission after 1 year.⁵ Patients with CD continue to improve over the course of 2 years, demonstrating a need to predict who will ultimately respond before giving up on a potentially effective therapy.⁶ Thus, further understanding of disease pathogenesis and predictive biomarkers are needed to tailor specific interventions to relevant patient subsets.

Vedolizumab is thought to act primarily by inhibiting α4β7⁺T effector/memory (Tmem) cell extravasation into intestinal mucosae. However, saturation of α4β7 receptors on circulating and mucosal Tmem cells does not directly relate to VDZ response,⁷ suggesting that the mechanism of VDZ action may not be fully explained by α4β7 blockade alone.^7–10 A recently completed study demonstrated that using therapeutic drug levels to augment the treatment response to VDZ was not effective—again suggesting a distinct mechanism of action.¹¹ Thus, we sought to develop data sets and computational models useful for predicting VDZ responses among IBD patients, while also informing therapeutic mechanism(s) of VDZ action.

Given that α4β7 antibodies inhibit lymphocyte trafficking from blood into intestinal mucosae, we reasoned that a combined analysis of immunophenotypic and transcriptomic signatures in both blood and mucosal immune cells may enhance our ability to discriminate VDZ responders from nonresponders. We further hypothesized that these same cellular and transcriptional signatures could provide novel insight into VDZ’s mechanism of action and identify opportunities to complement or enhance its effect. Here, we present IBD patient data on clinical parameters, immunophenotypic patterns, and transcriptomic alterations that are different between VDZ responders and nonresponders in peripheral blood and mucosal biopsies. To maximize biomarker discovery, we used unsupervised approaches to analyze immunophenotypic and transcriptomic (RNA-seq) data. The results of our study highlight the compartmentalized nature of mucosal immune-regulatory mechanisms that underlie IBD and clinical responses to VDZ therapy. Our data highlight Tregs as both predictors and pharmacodynamic markers of response to VDZ. Similar approaches could be used to improve efficacy rates of other IBD medicines and to rationally inform new opportunities for combination therapy.

Materials and Methods

Patient Population

Patients with CD or UC were enrolled if they were considered for VDZ therapy as part of their clinical care at the University of Miami Crohn’s and Colitis Center. Seventy-one patients provided consent to participate and provided detailed demographic information, blood, and tissue biopsies collected at baseline. Of these 71 patients, only 37 ultimately received VDZ (see Supplementary Table 1 for list of reasons patients did not receive VDZ). Of the 37, a first cohort (n = 29) met all the criteria used in the pivotal clinical trials of VDZ.^3,4 A parallel, second cohort (n = 8) received VDZ for active CD or UC at the discretion of the treating physician but did not strictly meet the criteria used for the VDZ pivotal trials (Supplementary Table 2). Because the same data were collected for both cohorts, we combined the 2 cohorts to simplify the presentation of the data. Patients received VDZ at weeks 0, 2, and 6, followed by maintenance therapy every 8 weeks (Supplementary Figure 1). After the first maintenance dose at 14 to 22 weeks, we collected a second blood sample and assessed the clinical response to VDZ (Supplementary Figure 1). Long-term follow up clinical data were also collected up to 62 months after the start of VDZ. The University of Miami Human Subject Research Office Institutional Review Board (UM HSRO IRB) approved this project (IRB ID 20150750, 20081100), and all participants provided written informed consent to conduct this study.

Assessment of Response and Inflammation Status

Responders to VDZ were defined as patients who had >2-point decrease in disease activity score, partial Mayo (UC), or Harvey-Bradshaw (CD) index between pre- (baseline) and post- (week 14-22) VDZ treatment. We also considered patients as responders who had a 2-point decrease in disease activity by week 14 to 22 and who remained on VDZ for >2 years (Figure 1). Biopsies from terminal ileum, ascending colon, and sigmoid colon were collected and scored as noninflamed, mild, or moderate/severe inflammation based on endoscopy and pathologic evaluation.

Figure 1.

Flow chart and Kaplan-Meier curve of vedolizumab persistence in IBD patients over time. A, Distribution of patients’ biospecimens (blood and biopsy) and responsiveness to VDZ throughout the course of clinical study (> 2 years). B, A total of 37 patients initiated VDZ, with discontinuation over time. Most discontinued within 6 months to a year after initiation. Twenty-one patients remained on VDZ for 2 years. At 5 years, no patients remained on VDZ.

Cell Isolation From Blood and Mucosal Biopsies

Tissue biopsies collected during colonoscopy were kept in HypoThermosol FRS Preservation Solution (Millipore Sigma) at 4°C and processed within 24 hours after collection. To isolate lamina propria mononuclear cells (LPMCs), biopsies were depleted of mucus and intestinal epithelial cells by incubation in 10 mM dithiothreitol followed by chelation in 0.5 mM ethylenediaminetetraacetic acid solution prepared in Dulbecco’s Modified Eagle Medium (DMEM) containing penicillin/streptomycin. Tissue was then digested in a mixture consisting of 250 µg/mL of Liberase (Millipore Sigma) and 10 µg/mL of DNase I (Epicenter) in DMEM at 37°C for 20 minutes, followed by mechanical dissociation by pipetting and straining through 70 µm of nylon filters Becton, Dickinson and Company (BD). Single cell suspensions of LPMCs were subsequently stained for fluorescence-activated cell sorting (FACS) analysis. Blood samples were collected in BD Vacutainer Plastic Blood Collection Tubes with Lithium Heparin (Fisher Scientific) and processed within 24 hours of collection. To isolate PBMCs, the remaining blood sample was diluted 1:1 with phosphate-buffered saline (PBS) and layered over a Histopaque-1077 (Millipore Sigma) gradient. After low-speed centrifugation, the buffy coat containing mononuclear cells was collected, washed with PBS, and centrifuged at 350x g for 5 minutes. Each fraction was stained and analyzed using 2 separate antibody panels with overlapping lymphocyte markers in order to characterize a broad array of immune cell subsets analyzed.

Flow Cytometry and FACS Sorting

Surface antigen antibodies with specific fluorochrome conjugates (Supplementary Table 3) were used to segregate immune cell subpopulations. Details of staining and CytofKit analysis are online in Supplementary Methods.

RNA Sequencing and Bioinformatics

Total RNA from Tmem and Treg T cell subsets sorted from PBMCs and biopsies were isolated by means of the miRNAeasy Micro Kit (catalog #217084; Qiagen, Germantown, MD) according to the manufacturer’s instructions. RNA concentration and integrity were determined with the Agilent Bioanalyzer RNA Pico LabChip analysis (PN 5067-1513; Agilent, Santa Clara, CA); RNA-seq was done using the SMART-Seq HT Kit (Catalog #634455, Takara Bio USA, Inc.). Libraries were individually indexed and pooled in groups of 20. Each pool was run on a NextSeq500 flow cell (9 flow cells total) generating 75-base single reads average 20 million reads per sample (3.5 billion reads total).

Reads from T cell RNA-seq were mapped to the human genome (GRCh38) using STAR (ver.2.5.0) aligner.¹⁰ Raw counts were generated based on Ensembl genes (Ensembl release 91) with HTSeq (ver.0.11).¹² To normalize different numbers of reads in the samples, we used DESeq2 median of ratios method.¹³ Venn diagrams were constructed by matching ENSEMBLE gene IDs from different comparison groups where padj values <0.05 were used. Venn Diagram JMP Add-In in JMP (Version 13.2.1 SAS Institute Inc., Cary, NC, 1989-2007) was used to construct the Venn diagram. The ingenuity pathway analysis (IPA) is achieved by the knowledgebase, which serves as a repository of biologically important molecules that are grouped based on biological interactions and functional relationships curated from published literature and scientific databases. The list of significantly regulated genes (padj <0.05) for each comparison were uploaded to IPA and the molecules that were mapped by IPA were analyzed using Core Analysis. When available, z scores were used to establish the direction of change. All RNA sequencing data discussed in this publication have been deposited in NCBI’s Gene Expression Omnibus¹⁴ and are publicly accessible through GEO Series accession number GSE184593 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE184593) on September 22, 2024. A secure token will grant access to reviewers. All other data supporting the findings of this publication are available on request from the corresponding author.

Predictive Modeling of Clinical, Transcriptomic, and Cellular Data

Four models were chosen for comparison due to their ability to prioritize relevant variables without prior specification. Modeling was restricted to those patients with experimental data in either the Cytofkit or RNAseq data sets (n = 33). The R package Glmnet version 4.0-2 was used for creating Lasso penalized linear regression models.¹⁵ The lambda value used was calculated using the cv.glmnet method, with the minimum plus 1 standard error (“min.1se”) option for prediction. The R package rpart version 4.1-15¹⁶ was employed for recursive partitioning analysis using the control options xval = 10, cp = 0, minbucket = 2, minsplit = 5. The R package randomGLM version 1.02-1¹⁷ was used for ensemble voting models based on linear regression with forward selection with the settings nBags = 500 and nFeaturesInBag = 40. The randomForest R package¹⁸ version 4.6-14 was used for ensemble voting based on recursive partitioning with the settings nTREE = 500 and mtry = 40. Model performance was compared using 10-fold cross-validation studies.

All authors had access to the study data and reviewed and approved the final manuscript.

Results

Response to Vedolizumab in a Cohort of Anti-TNF-Refractory CD and UC Patients

Patients initiating VDZ therapy between 2014 and 2018 were enrolled prospectively to explore clinical, cellular, and transcriptional biomarkers that predict clinical response. Seventy-one patients (35 UC, 36 CD; Table 1) provided consent for blood and tissue collection at baseline. A second follow-up blood collection was performed in patients who received VDZ (n = 37). Unlike the pivotal trials of VDZ in UC and CD,^3,4 we examined clinical response to VDZ between 14 and 22 weeks after the first maintenance dose. This follow-up time point was chosen to give more time for patients to respond.^19,20 Of the 37 patients who received VDZ, a total of 14 patients were considered responders (Figure 1a; Table 2). Thirty-four patients never started VDZ therapy due to one of several reasons (Supplementary Table 1). More than 40% were Hispanic, and most had previously received at least 1 anti-TNF agent prior to VDZ induction (Table 1).

Table 1.

Demographics and Baseline Characteristics

Total Participants Enrolled at Baseline (n = 71)		Total Participants Who Completed the Study (n = 37) (subset of total patients enrolled)
Gender (n)		Gender (n)
Male	37 (52%)	Male	18 (49%)
Female	34 (47%)	Female	19 (51%)
Ethnicity (n)		Ethnicity (n)
Hispanic	30 (42%)	Hispanic	15 (40%)
Non-Hispanic	41 (57%)	Non-Hispanic	22 (60%)
Race (n)		Race (n)
White	67 (94%)	White	35 (94%)
African American	3 (4.2 %)	African American	2 (6%)
Other	1 (1.4 %)
Smoking status		Smoking status
Non-smokers (n)	53 (75%)	Non-smokers (n)	30 (81%)
Ex-smokers (n)	15 (21%)	Ex-smokers (n)	7 (19 %)
Smokers (n)	3 (4.2%)	Smokers (n)	0
Age (median)	44 (20-79)	Age (median)	46 (20-79)
Medications (n)		Medications (n)
Current use of aminosalicylates	25 (35%)	Current use of aminosalicylates	10 (27%)
Previous use of aminosalicylates	40 (56%)	Previous use of aminosalicylates	18 (48%)
Current use of steroids	21 (30%)	Current use of steroids	11(30%)
Previous use of steroids	38 (54%)	Previous use of steroids	18 (48%)
Current use of immunomodulators	29 (41%)	Current use of immunomodulators	20 (54%)
Previous use of immunomodulators	39 (55%)	Previous use of immunomodulators	19 (51%)
Current use of biologics	29 (40%)	Current use of biologics	18 (48%)
Previous use of anti-TNF agents	42 (59%)	Previous use of anti-TNF agents	31 (83%)
• 1 anti-TNF agent	27 (38%)	• 1 anti-TNF agent	14 (45%)
• 2 anti-TNF agents	16 (23%)	• 2 anti-TNF agents	12 (39%)
• 3 anti-TNF agents	6 (9%)	• 3 anti-TNF agents	4 (13%)
• > than 3 anti-TNF agents	4 (5.6%)	• > than 3 anti-TNF agents	1 (3.2%)
UC (n = 35)	49%	UC (n = 19)	51%
Male	19 (54%)	Male	9 (47%)
Female	16 (45%)	Female	10 (52%)
Age at diagnosis (mean)	29 (7-63)	Age at diagnosis (median)	28 (7-63)
Duration of UC (mean)	15 (1-65)	Duration of UC (mean)	15.27 (1-21)
Location of UC (n)		Location of UC (n)
Pancolitis	26 (72%)	Pancolitis	13 (68%)
Left sided	5 (13%)	Left sided	3 (16%)
Proctosigmoiditis	5 (21%)	Proctosigmoiditis	3 (16%)
History of G.I surgeries	2 (2.8%)	History of G.I surgeries	0 (0%)
Partial Mayo at baseline (mean ± SD)	4.6 ± 3.2	Partial Mayo at baseline (mean ± SD)	3.2 ± 2.1
CD (n = 36)	51%	CD (n = 18)	49%
Male	18 (50%)	Male	9 (50%)
Female	18 (50%)	Female	9 (50%)
Age at diagnosis (mean)	28 (6-51)	Age at diagnosis (mean)	28 (12-62)
Duration of CD (mean)	15 (2-44)	Duration of CD (mean)	16 (1-40)
Location of CD		Location of CD	27 (12-62)
Ileum	11 (33%)	Ileum	7 (39%)
Ileum + colon	17 (51%)	Ileum + colon	10 (56%)
Colon	5 (15%)	Colon	1 (5 %)
History of G.I surgeries	24 (72%)	History of G.I surgeries	24 (72%)
Harvey Bradshaw Score (mean ± SD)	4.5 ± 3.3	Harvey-Bradshaw Score (mean ± SD)	5.6 ± 4

Table 2.

Demographic Characteristics of Short and Long-Term Responders

Responders week 14-22 (n = 14)		Responders > 2 years VDZ treatment (n = 21)
Gender (n)		Gender (n)
Male	9 (64%)	Male	11 (52%)
Female	5 (35%)	Female	10 (48%)
Ethnicity (n)		Ethnicity (n)
Hispanic	3 (21%)	Hispanic	6 (29%)
Non-Hispanic	11 (78%)	Non-Hispanic	15 (71%)
Race (n)		Race (n)
White	12 (86%)	White	20 (95%)
African American	2 (14%)	African American	1 (5%)
Smoking status		Smoking status
Non-smokers (n)	9 (64%)	Non-smokers (n)	15 (71%)
Ex-smokers (n)	5 (35%)	Ex-smokers (n)	6 (29%)
Age (median)	42 (27-69)	Age (median)	43 (22-69)
Medications (n)		Medications (n)
Current use of aminosalicylates	6 (42%)	Current use of aminosalicylates	8 (38%)
Previous use of aminosalicylates	6 (42%)	Previous use of aminosalicylates	10 (48%)
Current use of steroids	3 (21%)	Current use of steroids	6 (29%)
Previous use of steroids	7 (50%)	Previous use of steroids	9 (43%)
Current use of immunomodulators	11(78%)	Current use of immunomodulators	9 (43%)
Previous use of immunomodulators	11(78%)	Previous use of immunomodulators	17 (81%)
Current use of biologics	9 (64%)	Current use of biologics	10 (48%)
Previous use of biologics	9 (64%)	Previous use of biologics	15 (71%)
•1 biologic agent	7 (50%)	•1 biologic agent	9 (43%)
•2 biologic agents	3 (21%)	•2 biologic agents	7 (33%)
•3 biologic agents	1 (7%)	•3 biologic agents	1 (5%)
•> than 3 biologic agents	2 (14%)	•> than 3 biologic agents	2 (9%)
UC (n = 8)	57%	UC (n = 11)	52%
Male	4 (50%)	Male	4 (36%)
Female	4 (50%)	Female	7 (63%)
Location of UC		Location of UC
Pancolitis	4 (50%)	Pancolitis	7 (64%)
Left sided	2 (25%)	Left sided	2 (18%)
Proctosigmoiditis	2 (25%)	Proctosigmoiditis	3 (18%)
History of G.I surgeries	0 (0%)	History of G.I urgeries	0 (0%)
CD (n = 6)	43%	CD (n = 10)	48%
Male	5 (83%)	Male	7 (70%)
Female	1 (16%)	Female	3 (30%)
Location of CD		Location of CD
Ileum	3 (50%)	Ileum	5 (50%)
Ileum + colon	3 (50%)	Ileum + colon	5 (50%)
History of G.I surgeries	5 (83%)	History of G.I surgeries	7 (70%)

We were able to obtain more than 5 years of follow-up data on many of the enrolled patients (Figure 1b). Early discontinuation of VDZ occurred after the first follow-up visit either because of a lack of efficacy or, in fewer instances, because of onset of adverse events including ankylosing spondylitis, uveitis, or new or worsening joint pain (Supplementary Table 4). Importantly, 21 patients remained on VDZ for more than 2 years (Table 2, Figure 1b). Eleven of these 21 were considered responders by our definition, but an additional 9 had little to no response by week 22 yet remained on VDZ for > 2years. Thus, VDZ was largely effective in this cohort of anti-TNF-refractory IBD patients.

Immunological Differences Between the Ileum and Colon Exceed Those Between CD and UC

We first leveraged this patient cohort to examine if core immunological differences exist between circulating and/or mucosal immune cells from CD and UC patients. Blood and biopsy samples collected from all 71 patients, regardless of whether they ultimately received VDZ, were analyzed by flow cytometry; and immune cell clusters were compared between CD and UC patients (Figure 2). To maximize biological insights and limit bias, we performed unsupervised clustering of the flow cytometry data using the Phenograph package within Cytofkit and discriminated 57 unique cell clusters in PBMC and 29 in LPMC (Figure 2A). Phenotypic determinants of each cell cluster are shown in Figure 2B, and absolute frequencies of the clusters were enumerated and compared between either CD and UC patients or between biopsy site (ie, ileum, colon). Only 1 of 57 immune cell clusters in peripheral blood and none of the 29 mucosal clusters were differentially abundant between CD and UC patients. The sole differentially abundant cluster observed in peripheral blood, cluster 44, was enriched in UC patient blood and corresponded to a small population of activated CD3^loCD4⁺CD8⁺T cells that express high levels of both α4β7 and CCR9, as well as CD25, CD69, and CX₃CR1 (ie, fractalkine receptor; Figure 2B-C). This phenotype could be consistent with a subset of intraepithelial lymphocyte (IELs) that have either re-entered circulation or have recently emerged from the thymus and are preprogrammed to home to intestinal mucosae.²¹

Figure 2.

Immunophenotypic and gene expression differences between tissue locations are more pronounced than differences between disease subtype or inflammation state. A, PhenoGraph t-SNE clustering maps of PBMC of all available samples (n = 52), PBMC of only CD patients (n = 29), and PBMC of only UC patients (n = 23; top) or all ileum + colon (biopsy, bottom) combined by disease state (CD, n = 21 and UC, n = 10). B, Heatmap of expression levels of all markers (x-axis) across all defined clusters (y-axis) in PBMC (left) or biopsy (right) of all patients. C, Volcano plots of fold-change/fold-change of PhenoGraph cell percentages for each cluster comparing disease type (CD/UC) in PBMC (top left), disease type (CD/UC) in ileum and colon biopsies (top), tissue types (ileum/colon; bottom left), or inflammation in ileum and colon biopsies (noninflamed/inflamed; bottom). P < .05 (grey dashed horizontal line). D, Comparison of the PhenoGraph cluster cell percentages, from the biopsy volcano plot shown in (C, bottom left), with P < .05. Clusters with lowest p-values chosen: 3, 17, 28. E, Venn diagram of RNA-seq DEG’s in Treg (top) and Tmem (bottom) cells comparing disease (CD/UC), sample type (blood or biopsy) and tissue location (colon vs ileum).

In contrast to these minor differences between immune phenotypes in CD and UC patients, we found more marked differences in the abundance of immune cell clusters between ileum and colon; nearly one-third (9 of 29) of all mucosal cell clusters identified in this analysis were differentially abundant in ileum or colon (Figure 2C) regardless of IBD subtype or inflammation status (Figure 2D, data not shown). Exemplar clusters that appear to reflect core differences between immune cell lineages in ileum (cluster 28) or colon (clusters 3, 17) are highlighted in Figure 2D. Cluster 28, for example, was consistently enriched in ileum vs colon of all IBD patients, whereas clusters 3 and 17 were higher in the colon (Figure 2D). These differences would have been missed through conventional (ie, targeted) FACS analysis, illustrating the utility of unsupervised clustering methods for agnostic biomarker discovery. As predicted by a prior study,²² neither pregating on viable mononuclear cells nor application of manual compensation affected Phenograph’s ability to discern these core immunological differences between ileal and colonic immune compartments (data not shown). Further, direct (ie, manual) gating of clusters 3, 17, and 28—based on expression matrices inferred by Phenograph—corroborated that these phenotypes were significantly different between ileal and colonic mucosa (Supplementary Figure 2).

Consistent with our flow cytometry results, many more differentially expressed genes (DEGs) were detected in T cells (Tmems or Tregs) from colon vs ileum compared with these same cells stratified by IBD diagnosis (eg, UC vs CD; Figure 2E). Given the central roles that CD4⁺T cells are thought to play in the pathogenesis of both CD and UC, we elected to focus our gene expression studies on both pro-inflammatory Tmem and tissue-protective Treg cell subsets. Across all comparisons, Treg cells displayed greater transcriptional differences than Tmem cells, suggesting that Treg cells, while smaller in number, more dynamically reflect differences in immunological microenvironments. Using ingenuity pathway analysis, we inferred the identity of molecular pathways underlying discrepant transcriptional behaviors of Treg cells and found dozens of pathways predicted to be selectively activated in either the colon or ileum (Supplementary Figure 3A). Notable pathways that most clearly segregated between colon and ileum Treg cells included (1) the phospholipase C pathway, which is activated in response to T cell antigen recognition and was elevated in colonic Treg cells (Supplementary Figure 3B); and (2) the protein kinase A (PKA) signaling pathway, an immunosuppressive pathway activated during states of energy (ATP)-deficiency that showed increased activity in Treg cells from ileum (data not shown). Taken together, these results suggest that differences in the immunological makeup of the ileum and colon supersede those between either the ileum or colon of CD vs UC patients. Functional attributes of Tregs may inform new approaches for tailored IBD therapy.

Pharmacodynamic Effects of Vedolizumab in PBMCs Demonstrate a Greater Change in Gene Expression Among Treg vs Tmem Cells

Immunological processes other than intestinal homing might contribute to therapeutic effects of VDZ.^7,23 To assess pharmacodynamic effects of VDZ in IBD patients, we analyzed matched flow cytometric and RNA-seq data from PBMCs of patients before and after VDZ therapy. Although mucosal biopsies were not collected post-VDZ for this study, unsupervised clustering of PBMCs using CytofKit (Figure 3A and 3B) showed a decrease in several agnostically defined clusters, especially in UC patients (Figure 3C). Clusters that decreased most following VDZ treatment (clusters 3, 16, 26) were not canonical lymphocyte subsets stained by the flow panel and may represent monocyte populations (Figure 3B-C).

Figure 3.

Pharmacokinetic effects of VDZ in PBMC demonstrate the greatest change in Treg gene expression. A, PhenoGraph t-SNE clustering maps of PBMC of all samples (n = 25), only PBMC of pre-VDZ samples (n = 25), and only PBMC of matched post-VDZ samples (n = 25). B, Heatmap of expression levels of all markers (x-axis) across all found clusters (y-axis) in PBMC of patients who received vedolizumab. C, Volcano plots of fold-change/fold-change of PhenoGraph cell percentages of each cluster between patient samples before or after VDZ treatment (pre/post). P < .05 (grey dashed horizontal line). D, Comparison of the PhenoGraph cluster cell percentages in patient samples pre VDZ treatment (red) and samples post-VDZ treatment (blue), using clusters from PBMC volcano plots having P < .05. Pre- and postanalysis furthered with data on patient responsiveness to VDZ to highlight correlations (or lack of) between differentially expressed clusters pre/post-VDZ and response to VDZ. E, Simple linear regression analysis between pharmacodynamic changes (post/pre) of cluster 9 in UC patients (y-axis) and Mayo score (UC) changes post/pre-VDZ treatment (x-axis). F, Volcano plot of PBMC Treg DEGs between post-VDZ treatment and pre-VDZ treatment (post/pre. G, Heatmap showing significantly enriched pathways of regulatory T cells in post/pre-VDZ samples. Values indicate -log₁₀ (P) for each pathway, with only those having P < .05 (log₁₀P ≥ 1.3) included. Red pathways signal pathways higher post-VDZ. Blue pathways illustrate pathways higher pre-VDZ. IPA could not predict directionality in black pathways. H, Upregulator pathway network analysis for oxidative phosphorylation in PBMC Treg cells between post/pre-VDZ treatment. Pathway had P < .05 and z score >1 (higher post-VDZ). The figure legend highlights different properties/functions of the proteins encoded by genes representing the regulator network pathway (noted by different symbols). Colored genes highlight up/down regulation, and colored lines denote forms of interactions between participating genes.

One cluster (cluster 9) consistently accumulated in peripheral blood of all VDZ-treated patients; these cells appeared to reflect a subset of activated (CD69⁺) CD4⁺T cells that also displayed the highest expression among all clusters of α4 (CD49d) and β7 integrin, as well as the gut-tropic chemokine receptor, CCR9 (Figure 3B-D). This finding is consistent with the idea that VDZ blocks the homing of these cells to the inflamed intestine. Although cluster 9 increased most in the periphery of UC responders, the magnitude of this increase did not significantly correlate with response in UC (or CD) patients (Figure 3E). Attempts to manually gate clusters 14 and 9 revealed consistent trends but not the significant differences, discerned by unsupervised clustering (Supplementary Figure 4).

By RNA-seq, Treg cells again showed the greatest number of DEGs before and after VDZ treatment compared with Tmem cells (Figure 3F). Using IPA, we found 18 molecular pathways that differed significantly (P < .05) different in Tregs before vs after VDZ treatment (Figure 3G). The oxidative phosphorylation pathway, which is directly linked with efficient mitochondrial respiration and Treg cell bioenergetic fitness in vivo was elevated in PBMC Tregs following VDZ treatment (Figure 3H).^24,25 Reciprocally, the smaller number of Treg genes whose expression decreased post-VDZ therapy were enriched for those associated with antioxidant signaling via the Sirtuin family of NAD+-dependent class III histone deacetylases (Figure 3G). These data suggest that VDZ treatment may reduce activation-induced oxidative stress in Treg cells while increasing genes linked to oxidative phosphorylation, which collectively may increase the persistence and suppressive function of Tregs.²⁶ These findings in peripheral blood Tregs may indirectly reflect mucosal changes related to VDZ-mediated improvement in mucosal inflammation.

Immune Features Predicting Response to Vedolizumab Treatment Are More Abundant in Mucosal Biopsies Than Peripheral Blood

Patients in this cohort were mostly refractory or previously exposed to anti-TNF agents which has been previously associated with decreased response to VDZ.²⁷ We next asked which peripheral or mucosal immunophenotypic and/or transcriptomic features at baseline could predict which patients will go on to benefit clinically from VDZ therapy. Using the same unsupervised clustering approaches and data sets as above, we compared the baseline (pre-VDZ treatment) immune phenotypes in PBMC or LPMC (biopsy) between patients that went on to show response or no response on follow-up visit. No baseline PBMC clusters were sufficient to predict VDZ response in this cohort (Figure 3A-B), whereas 9 immunophenotypic clusters from mucosal biopsies did (Figure 4A-C). Within the gut, 3 clusters were consistently elevated at baseline in patients who went on to show clinical response to VDZ—21, 22, and 24 (Figure 4B-C). Based on inferred expression matrices, each of these clusters appeared to reflect subsets of CD14⁺CD15⁺macrophages (Figure 4B-C). Manual gating of these clusters again corroborated the trends, and is some cases statistical differences (eg, cluster 15), discerned by unsupervised clustering amongst all patients (Supplementary Figure 5).

Figure 4.

Immune features that predict response to vedolizumab treatment are more prevalent in tissue. A, PhenoGraph t-SNE clustering maps of ileum and colon matched biopsy samples (n = 25), only biopsies of patients with response to VDZ (n = 14), and only biopsies of patients with no response to VDZ (n = 11). B, Heatmap of expression levels of all markers (x-axis) across all found clusters (y-axis) in biopsies of patients who received VDZ. C, Volcano plots of fold-change/fold-change of PhenoGraph cell percentages of each cluster, in colon + ileum combined (biopsy) between patients responsive to VDZ and those nonresponsive to VDZ (no response/response). P = .05 (grey dashed horizontal line). D, Comparison of the PhenoGraph cluster cell percentages in nonresponsive patients (blue) and responsive patients (red) from the biopsy volcano plot shown in C (bottom left). Clusters having a P < .05 were chosen for deeper analysis. E, Venn diagram of RNA-seq DEGs in Tmem cells (top) and Treg cells (bottom) comparing tissue location-dependent response to VDZ. F, Heatmap showing significantly different enriched pathways of regulatory T cells in nonresponders/Responders to VDZ therapy. Values indicate -log₁₀(p-value) for each pathway, with only those having a P < .05 (log₁₀P ≥ 1.3) included. Directionality for pathways in black text is not defined in IPA. Blue pathways illustrate pathways higher pre-VDZ.

As before, RNA-seq analyses supported and extended our immunophenotypic results. For both Treg and Tmem cells, relatively few DEGs were found between VDZ-responders and nonresponders in peripheral blood, whereas many more DEGs were observed between these T cell subsets from mucosal biopsies (Figure 4E). Treg cells, especially from the ileum, showed the most transcriptional differences at baseline in responders vs nonresponders to VDZ regardless of CD or UC diagnosis (Figure 4E). Treg pathways that differed significantly in baseline biopsies from those who did or did not respond to VDZ therapy were enriched for cell metabolic pathways, including branched chain amino acid metabolism, glycolysis, and cAMP-mediated signaling (Figure 4F). These results suggest that the metabolic state of mucosal Tregs at the time of VDZ initiation could modulate VDZ biological activity.

We also examined clinical factors correlated with response. Use of nonbiologic medications was most correlated with response to VDZ (Supplementary Figure 6A). This finding was driven by thiopurine use at VDZ initiation, which was higher in responders than nonresponders (Supplementary Figure 6B-D). Finally, we examined patients who discontinued VDZ therapy due to development of adverse events. We were not able to discern either discrete immune phenotypes or Tmem or Treg cell transcriptomic features that predicted patients who discontinued VDZ therapy due to development of adverse events (Supplementary Table 4). Thus, transcriptomic markers in ileal Tregs and use of thiopurine were the best predictors of response to VDZ.

Machine-learning Approaches to Multiplex Clinical, Cellular, and Gene Expression Data Increases the Predictive Accuracy of VDZ Response

We next used our data set to compare the performance of several multiparametric computational models to determine which model and which combination of clinical parameters and biomarkers provided the most power for predicting patient responses to VDZ therapy. As expected, receiver operating characteristic (ROC) curve analyses of individual (continuous) clinical or immunophenotypic variables showed only minimal predictive power (Figure 5A). We employed models that incorporate multiple (baseline) clinical, flow cytometric, and gene expression variables. We included patients for whom not all data were empirically available due to either low cell recovery or poor RNA-seq library quality. Missing data were thus inferred using a k-nearest neighbor algorithm.²⁸ We excluded transcriptomic data from mucosal T cells in these analyses because the percentage of inferred values for these samples exceeded the recommended threshold of 50% for modeling.²⁹ Clinical, immunophenotypic, and PBMC DEGs from 33 patients were ultimately combined and used to train 4 binomial predictive models (glmnet, rpart, randomGLM, randomForest).^15–18 The predictive power of each was then computed and compared.

Figure 5.

Machine-learning approaches to multiplex data in IBD patients increases the predictive accuracy of VDZ response. A, Individual ROC curves for continuous clinical demographics (chosen randomly to provide examples; top left) or clusters from Cytofkit Colon (top right), Ileum (bottom left) and PBMC (bottom right). Comparisons made and statistical significances calculated between responders vs nonresponders to VDZ. Top 6 clusters for Cytofkit Colon, Cytofkit Ileum, and Cytofkit PBMC shown as ranked by t test P value, without a specific cutoff used. All 6 continuous clinical variables included, without exclusion based on P value. B, Model performance comparison: penalized regression (glmnet = 67%), recursive partitioning (rpart = 70%), regression with forward selection random sampling and bootstrap aggregation (randomglm = 82%), and recursive partitioning with random sampling and bootstrap aggregation (randomForest = 82%). Predicted probability of response, at a cutoff of 0.5, for each patient (shown to the left of each model output). Red and black coloring corresponds to actual classification as responders (red) or nonresponders (black) in the original data. Predicted probability of response used to measure prediction accuracy. C, Ranked variable importance to determine which input is most capable of driving the predictive power of this model. Relative variable importance graphed for glmnet, randomGLM, rpart, and randomForest computational methods. “Model inputs” plot highlights percentage of total variables for each individual variable: clinical, cytofkit colon, cytofkit ileum, cytofkit PBMC, PBMC Tmem, and PBMC Treg. “Model Outputs top 20 variable importance union” highlights representation of the 6 variable inputs in a union list of the top 20 variables ranked by their relative importance to each model. D, Top 10 PBMC Treg genes most capable of predicting responsiveness in patients with VDZ, organized by predictive rank where the size of the circle represents the fold change. Predictive score highlights the robustness of each gene in predicting patient responsiveness to VDZ. Predictive rank highlights the ranking of each gene in comparison to the 10 genes noted and is directly linked to predictive score (highest predictive score = highest rank). E, PBMC Treg expression (TPM counts), in responders (green) and nonresponders (pink), for top 10 PBMC Treg genes with greatest predictive power.

The models chosen represented a mix of both regression (glmnet, randomGLM) and recursive partitioning (rpart, randomForest), as well as simple (glment, rpart) and ensemble voting (randomGLM, randomForest). For this data set, the ensemble predictors achieved equal predictive accuracy in cross-validation studies (82%) and substantially outperformed the simpler models (Figure 5B). Despite a generally equivalent distribution of input variables (eg, clinical, immunophenotypic, and DEGs), those that contributed the most substantially to the predictive power of the models came from DEGs among PBMC Treg cells (Figure 5C). Expression of BLOC1S1, TLCD1, and TMEM223, in particular, were all increased in PBMC Treg cells at baseline from patients that went on to respond to VDZ therapy (Figure 5D-E). Conversely, PHLDA1, OSBPL11, and CXCL3 expression were each elevated in PBMC Treg cells at baseline in nonresponders (Figure 5D-E). These results reveal a constellation of biomarkers and clinical parameters that could increase the therapeutic utility of VDZ in IBD patients; they also inform computational best practices for inferring VDZ responses using multiple, agnostically defined biomarkers.

Discussion

Medical therapy for IBD has improved dramatically since the advent of anti-TNF therapy. At present, only VDZ works by targeting lymphocyte homing to the intestine and is used in both CD and UC. Studies demonstrate that VDZ is superior to the anti-TNF agent adalimumab in treatment of UC⁵; an ongoing study is now examining combinations of VDZ with adalimumab in CD (ClinicalTrials.gov Identifier: NCT02764762). Yet very few studies to date have delved into the mechanistic effects of a therapeutic agent in IBD, either to inform who will respond or to predict rational new approaches for combination therapy. Given its unique mechanism of action, studies of VDZ can lead to significant new insights in the interface between peripheral and mucosal immune compartments.

Our study included detailed clinical data and samples collected prospectively from patients. As can be seen from the demographics of our patient population, the patients studied here had a longer duration of disease (on average 15 years), and many had failed multiple anti-TNF agents. This cohort therefore represented a mostly refractory group, which includes the patients most in need of new therapeutic options and which may also offer a window into the mechanism of “refractoriness.” We also included a large percentage of Hispanic patients that are typically not represented in translational studies. We performed a comprehensive analysis of the peripheral and mucosal immune profiles using unsupervised analyses of both flow cytometry and bulk RNA-seq data on sorted peripheral blood and mucosal Tmem and Tregs to define the frequency and functional characteristics of these cells. Several themes emerged during our analyses. First, the largest differences in both the abundance and transcriptional behaviors of immune cells were not between UC and CD patients, or even inflamed vs uninflamed tissues. Rather, quantitatively larger and more fundamental differences appear to exist between mucosal immunophenotypes and T cell transcriptional characteristics in the ileum compared with the colon. Fundamental differences between ileal and colonic immune compartments are consistent with emerging evidence that ileal CD is a unique genetic entity and is more refractory to current therapies compared with either UC or colonic CD.^29–31 In a study of VDZ, ileal CD had decreased endoscopic healing compared with colonic inflammation.³² Importantly, the observed differences between ileal and colonic Tregs appear to lie primarily in metabolic and cell signaling pathways, including phospholipase C and protein kinase A signaling, which are therapeutically tractable targets and could suggest new concepts for combination therapies with VDZ. These studies thus highlight how Treg function, rather than Treg number, could be an opportunity to enhance treatment in refractory IBD patients.^31,33

The second theme that emerges from this work is the relationship between baseline Treg characteristics and VDZ response. When examining predictors of response to VDZ, we found that while numerically less abundant than memory T cells, Treg-derived gene expression signals generally—and genes associated with Treg cell metabolism specifically—were the most effective at predicting response to VDZ therapy (Figure 4-5). This result is both striking and timely as a growing body of recent literature has highlighted the role of Treg cell metabolism in dictating their pro- and anti-inflammatory functions during infection, inflammation, and cancer.²⁴ For example, FOXP3 directly shifts Treg metabolism from aerobic glycolysis towards mitochondrial respiration (eg, oxidative phosphorylation).^24,26 In the mouse T cell transfer colitis model, Tregs deficient in mitochondrial complex 3 display reduced suppressive activity and failed to protect against colitis.²⁶ Our identification of Treg genes involved in lipid metabolism and signaling (eg, BLOC1S1, TLCD1, and TMEM223) as the strongest predictive features of IBD patients that go on to respond to VDZ therapy is potentially consistent with the notion that stable, highly suppressive Treg cells use fatty acid β-oxidation to fuel oxidative phosphorylation.³⁴ Increased CXCL3 expression by Treg cells in VDZ nonresponders is also notable, as this is a neutrophil chemoattractant and could suggest that the presence of Tregs in an inflammatory state is detrimental to VDZ response. Our studies suggest the metabolic status of Treg cells may be a key determinant in dictating which patients ultimately respond to VDZ therapy.

An important implication of our study is to target the Treg metabolic pathways inferred here with currently existing drugs to augment the therapeutic efficacy of VDZ. For example, our findings suggest that pharmacologic approaches to enhance cAMP and/or PKA pathway activation in Tregs may increase patient responses to VDZ. Apremilast is an oral phosphodiesterase4 (PDE4) inhibitor approved for psoriasis³⁵ that results in increased intracellular cAMP, an intracellular second messenger that activates protein kinase A and controls a network of pro-inflammatory and anti-inflammatory mediators.^36,37 In the context of Tregs, cAMP has been shown to maintain peripherally induced (i)Tregs and inhibit pro-inflammatory function.³⁸ Although a clinical trial had been planned for apremilast in UC, the study has yet to be carried out. Pentoxyphylline is another PDE inhibitor that raises intracellular cAMP, activates PKA, and decreases TNF production.³⁹ Our group has recently completed a pilot study of pentoxyphylline combined with VDZ. We found that CD patients on the combination of pentoxyphylline with VDZ responded more quickly than VDZ alone.⁴⁰ Considering that both apremilast and pentoxyphylline are safe, FDA-approved medications, combining these agents with VDZ could be one strategy to bolster VDZ response rates among IBD patients. Such investigational studies could provide important new insights into the best and safest approaches to target Treg metabolic pathways for therapeutic benefit in IBD.

The best computational and machine-learning models for prediction of response to VDZ applied ensemble voting, randomForest, and randomGLM, with equal 82% accuracy. Interestingly, we found that as a stand-alone clinical parameter, use of thiopurines at the time of treatment with VDZ correlated with response. The US clinical trials of VDZ required discontinuation of thiopurines, which does not permit studying the combined effect. Thiopurine use has decreased because of toxicity and is typically not combined with VDZ, which is less immunogenic than anti-TNF agents. Our data suggest that SONIC-type studies may reveal improved efficacy of VDZ when combined with thiopurines.⁴¹

Limitations of our study include the size and scope of our flow cytometry panels that prioritized T cell markers compared with larger CyTOF or spectral flow panels, for example, as well as our use of bulk rather than single-cell RNA-seq.^42,43 When we initiated this prospective study, the main mechanism of action of VDZ was thought to occur through blockade of peripheral blood CD4 T cell homing to the intestine. The original logic was that T cells represent most circulating cells in human peripheral blood, whereas mucosal biopsies, like most nonlymphoid human tissues, can contain many more discrete immune cell lineages. Thus, we sought to simultaneously achieve “deep profiling” of circulating T cell phenotypes in IBD patients pre- and post-VDZ treatment, while also capturing broader mucosal immune features (including B cells and myeloid subsets) that may be relevant to VDZ-mediated immune modulation. We also did not specifically stain for CD8+ T cells in our data set, which may also play a role. We did, however, use unsupervised clustering to enhance discovery power and mitigate bias.

With flow and mass cytometry becoming increasingly multiparametric, automated (ie, unsupervised) clustering algorithms are becoming essential tools for downstream analyses and interpretation. Phenograph is but one of the now widely adopted automated clustering platforms, along with viSNE, SPADE, X-shift, and Citrus³¹ but has the added benefit of quantifying discrete cell clusters. Other groups are working to establish best practices and guidelines for preparing input flow cytometry data for unsupervised clustering analyses; a growing consensus from these studies is that less manual processing up front (ie, prior gating on select cell subsets, application of manual compensation, etc.) results in improved performance and more information.²² For our study, we analyzed flow cytometry data sets without prior gating or compensation, although we undertook efforts to confirm that key findings were not influenced by prior gating or compensation variables (data not shown). We also validated that most but not all differentially abundant cell phenotypes inferred by unsupervised clustering could be reproduced by manual gating (Supplementary Figure 2, 4-5). A degree of discordance between unsupervised clustering and direct gating is to be expected; one is quantitative and precise, the other qualitative and imprecise and reinforces the message that unsupervised clustering algorithms can extract biologically meaningful information from dense flow cytometry data sets that are not perceivable to the human eye. Still, it is reasonable to estimate that some of the immune cell clusters we identified and highlighted here could be impacted by alternative clustering methods or platforms. In addition, we used distinct sets of antibodies for profiling peripherally circulating and mucosal immune cell compartments. Thus, further unbiased efforts to analyze all cell types that interact to predict or reflect VDZ response is necessary to translate initial discovery into clinically actionable information.

We performed bulk RNA-seq rather than single cell but used sorted Tregs and Tmems from blood or tissue. We also used mucosal biopsies rather than surgical resections, the latter of which reflect skewed, end-stage perspectives of IBD immunopathogenesis. In addition, it is likely that the best chance to translate our findings to the bedside is to devise relatively simple flow cytometric and/or gene expression panels that could be analyzed and interpreted quickly in precision medicine approaches.

Finally, our work establishes hypothesis-generating data sets, in which we have performed a comprehensive exploration of the relationship of peripheral blood and tissue immunophenotypes in IBD. At the end of our study, we saw several patients who persisted on VDZ for years. We hope in the future to examine factors associated with persistence on VDZ as we look for improved ways to maintain durable remission in patients with IBD.

Funding

This study was funded by Takeda Pharmaceuticals, U.S.A. [grant number IISR-2014-1000892], Inc. and the National Institute of Diabetes and Digestive and Kidney Diseases (grant number R01DK09907). The small acquisition cohort #2 was funded by The Micky & Madeleine Arison Family Foundation Crohn's & Colitis Discovery Laboratory and The Martin Kalser Chair in Gastroenterology, University of Miami. Work in the Sundrud laboratory was supported by a National Institute of Health grant R01AI118931, and a Senior Research Award (422515) from the Crohn’s & Colitis Foundation (CCF).

Conflicts of Interest

M.T.A. has served as a consultant and scientific advisory board member for Abbvie Inc., Arena Pharmaceuticals, Bristol Myers Squibb, Eli Lilly Pharmaceuticals, Gilead, Janssen Biotech, LLC, and Prometheus Biosciences; UCB, a speaker for Alimentiv and has had projects funded by Pfizer, Prometheus Laboratories and Takeda Pharmaceuticals. A.R. has served as a consultant to GenapSys Inc. A.R.D. has served as a consultant for GI Health Foundation and the American Board of Internal Medicine and received research funding from Takeda; D.H.K. has served as a scientific advisory board member for AbbVie Inc., a consultant for Cleveland Clinic, and a trainer or lecturer for PRIME Continuing Medical Education and The Academy for Continued Healthcare Learning. O.M.D. received honoraria from Pfizer and has a funded grant from Pfizer. J.P.K. serves as a consultant to Global Urgent and Advanced Research and Development (GUARD). M.S.S. serves as a consultant to Sigilon Therapeutics and Sage Therapeutics. These activities do not alter the authors’ adherence to the journal’s policies on data and material sharing. Other authors declare no conflicts.