Metabolism and response to stress gene signatures reveal ulcerative colitis heterogeneity and identify patients with increased response to therapy

Abstract

Background and Aims

Ulcerative colitis (UC) therapies lead to variable remission and response rates in patients participating in clinical trials, likely due to interindividual target variability, differences in active biological pathways, feedback, and/or resistance mechanisms. Here, we stratified patients into subtypes by characterizing heterogeneity using mucosal biopsy transcriptomics data.

Methods

Transcriptomics data from an andecaliximab phase 2/3 study in patients with UC were scored for gene signature enrichment. Eleven Reactome gene sets, moderately correlated with histological disease activity using Robarts Histopathology Index with low correlation to each other, were selected and evaluated in baseline gene expression data of ustekinumab, infliximab, and adalimumab clinical trials in patients with UC.

Results

Of 11 gene sets, referred to as “Metabolism and Response to Stress” (MARS) signatures, 5 correlated with “non-disease” mucosa and 6 with “disease-related” mucosa. Clustering baseline andecaliximab samples scored with MARS revealed 3 clusters with low non-disease/high disease-related, high non-disease/low disease-related, or a mixture. Importantly, these clusters did not correlate with patient demographics, clinical characteristics, or disease activity metrics. Clustering baseline data from other clinical trials (anti-interleukin-12/23 and anti-tumor necrosis factor) in patients with UC scored with MARS showed that patients in low non-disease/high disease-related baseline score clusters less likely to achieve treatment response.

Conclusions

We identified and evaluated a novel, multi-dimensional signature gene set to characterize previously undefined heterogeneity in patients with UC and identify patients less likely to respond to therapy. This approach offers potential utility to define clinical trial populations, enrich for clinical responders, and identify difficult-to-treat populations for therapeutic development.

1. Introduction

Ulcerative colitis (UC) is a chronic relapsing-remitting inflammatory bowel disease (IBD) that is characterized by inflammation of the mucosa and submucosa, a loss of epithelial barrier integrity, and dysregulated immune responses in the large intestine,^1,2 which can cause frequent bloody stools and bowel movements, fatigue, weight loss, mucus discharge, and abdominal discomfort.^3,4 Management of these symptoms and disease progression in patients with moderate to severe UC can be achieved using a wide range of drug classes that target various pathways mediated by tumor necrosis factor (TNF) (infliximab, adalimumab, and golimumab), integrin α₄β₇ (vedolizumab), Janus kinases (tofacitinib, filgotinib, and upadacitinib), interleukin 12/23 (IL-12/23; ustekinumab), interleukin 23 (IL-23; risankizumab, mirikizumab, and guselkumab), and sphingosine-1-phosphate (S1P; ozanimod and etrasimod).^3–7 Despite the array of available therapies and advancements in drug development aiming to overcome primary non-response, secondary loss of response, and immunogenicity,⁸ the response to treatment and remission of patients with UC in clinical trials is heterogeneous, and molecular determinants are not well understood.⁹

Since the development of high-throughput molecular profiling techniques, including the RNA technologies of microarray and RNA-sequencing, molecular markers have been used to investigate and potentially characterize patient heterogeneity.^10–14 The ability of these markers to predict response or characterize disease heterogeneity varies widely, and these markers have not been commonly used in IBD clinical trials or clinical practice to date. The use of molecular data to elucidate interindividual variability and biological pathways involved in disease severity, pathogenesis, and response to therapy may shed light on the molecular basis behind the response to therapy, may be used to refine patient populations, and may aid in the development of targeted therapies.¹⁵

Gene expression analysis of mucosal biopsies from patients with UC has been used to define the molecular processes and pathways associated with UC as well as those associated with treatment responses.^9,10,16,17 However, many of these signatures, by design, are defined by the largest component of disease, which is mucosal inflammation. For example, the biopsy molecular inflammation score for patients with UC was developed using the Mount Sinai Crohn’s and Colitis registry using genes that correlated with inflammation status and was shown to be associated with early response outcomes, better than endoscopy.¹⁰ Although inflammation is a major component of this disease, other aspects of pathophysiology, such as epithelial barrier function, metabolic dysfunction, and microbial defense, would not be measured by such a score and may be critical processes that contribute to treatment response.

Here, we approach the goal of stratifying patients into treatment agnostic subtypes by first characterizing heterogeneity using RNA signature scores and identifying those that represent multiple orthogonal aspects of disease. Next, we evaluate whether the subtypes identified are novel compared to those that could be identified using available signatures, clinical data, or demographic data. Lastly, we evaluate whether the differences observed are relevant to treatment response in external clinical trial datasets.

2. Materials and methods

2.1. Ethical considerations

All datasets analyzed were obtained from studies that received approval from an institutional review board/independent ethics committee and that were conducted in accordance with the Declaration of Helsinki. Patients consented to the use of their data for research purposes.

2.2. Datasets of patients with UC

Gilead Sciences Inc. provided RNAseq gene expression data (GS-US-326-1100) derived from mucosal biopsies obtained from patients with moderately to severely active UC in a combined phase 2/3, double-blind, randomized, placebo-controlled, induction and maintenance study evaluating the safety and efficacy of andecaliximab (GS-5745) (ClinicalTrials.gov identifier: NCT02520284).¹⁸ Two colonic biopsies were obtained from adjacent regions of the colon at baseline (pretreatment, n = 145) and at week 8 (n = 141). One biopsy was preserved in RNAlater at baseline and Week 8 for RNAseq analysis, and the other biopsy was formalin fixed paraffin embedded for histological assessment of mucosal healing using Geboes scores. Robarts Histopathology Index (RHI)¹⁹ scores were calculated from the requisite set of component items of the Geboes score.

Publicly available microarray gene expression datasets used to predict response to ustekinumab (GSE206285) and anti-TNF therapies (GSE16879, GSE23597, GSE73661, and GSE92415) were obtained from the National Center for Biotechnology Information Gene Expression Omnibus (GEO) data repository. Details of each dataset and corresponding trial information are summarized in Table 1.

Table 1.

Summary of public microarray transcriptome datasets.

GEO series accession number	ClinicalTrials.gov identifier	Official trial title	Drug	Response definition	Organizational name of the dataset provider	PMID (publication year)	Sample number of UC-treated patients, baseline (n)
GSE206285	NCT02407236	A Phase 3, Randomized, Double-blind, Placebo-controlled, Parallel-group, Multicenter Protocol to Evaluate the Safety and Efficacy of Ustekinumab Induction and Maintenance Therapy in Subjects With Moderately to Severely Active Ulcerative Colitis	Ustekinumab	Clinical remission was defined as a total Mayo score of ≤ 2 and no subscore > 1. Deep remission included endoscopic and histologic improvement (defined as neutrophil infiltration in < 5% of crypts, no crypt destruction, and no erosions, ulcerations, or granulation tissue) at week 8	Imperial College London	36192482 (2022)20	364
GSE16879	NCT00639821	Mucosal Gene Expression Defects in Patients With Inflammatory Bowel Disease Before and After First Infliximab Therapy	Infliximab	Completed mucosal healing was defined as a decrease to Mayo endoscopic subscore of 0 or 1 with a decrease to grade 0 or 1 on the histological score for UC	KU Leuven	19956723 (2009)21	24
GSE23597	NCT00036439	A Randomized, Placebo-controlled, Double-blind Trial to Evaluate the Safety and Efficacy of Infliximab in Patients With Active Ulcerative Colitis	Infliximab	Clinical response was defined as a decrease in the Mayo score of at least 3 points and at least 30%, with an accompanying decrease in the subscore for rectal bleeding of at least 1 point or an absolute rectal bleeding subscore of 0 or 1	Janssen R&D	21448149 (2011)22	32
GSE73661	NA	NA	Infliximab	Histological mucosal healing was defined as a Geboes score of 0 or 1 Endoscopic mucosal healing was defined as a Mayo endoscopic subscore of 0 or 1	KU Leuven	27802155 (2018)23	23
GSE92415	NCT00487539	A Phase 2/3 Multicenter, Randomized, Placebo-controlled, Double-blind Study to Evaluate the Safety and Efficacy of Golimumab Induction Therapy, Administered Subcutaneously, in Subjects With Moderately to Severely Active Ulcerative Colitis	Golimumab (CNTO 148)	Clinical response was defined as a decrease from baseline in the Mayo score ≥ 30% and ≥ 3 points, accompanied by either a rectal bleeding subscore of 0 or 1 or a decrease from baseline in the rectal bleeding subscore ≥ 1	Janssen R&D	23735746 (2014)24	59

Abbreviations: GEO, Gene Expression Omnibus; NA, not available; PMID, PubMed identifier; UC, ulcerative colitis.

2.3. Data and statistical analysis

Public datasets were imported using GEOquery²⁵ from the GEO^26–28 database and merged using the Bioconductor MultiAssayExperiment²⁹ package. Raw data were analyzed and visualized in R (version 4.3.0, http://www.r-project.org/).³⁰ To correct for multiple testing, the false discovery rate was estimated from P-values adjusted using the Benjamini and Hochberg method.³¹ In all cases, adjusted P < .05 was considered statistically significant. The association between clusters and other variables was tested using the Chi-Square test. Gene Set Variation Analysis (GSVA)³² and Gene Set Enrichment Analysis³³ were used to score samples for enrichment of ~5200 Molecular Signatures Database (MSigDB) v 7.4 signatures^34,35 and were evaluated for correlation to the sample RHI score using Pearson correlation. Data from the anti-TNF therapy meta-analysis cohort were normalized to remove batch effects and variation using sva combat.³⁶ Data were visualized using ggplot 2 package³⁷ in R to create radar charts with mean values of the cluster members shown by the line,³⁸ and unsupervised clustering defined Euclidean distance with complete-linkage using the ComplexHeatmap package.³⁹ Clusters were evaluated using the “gap” statistic from the clusGap function of the cluster package.⁴⁰

3. Results

3.1. Overview of results

In this section, we first report the gene sets that are correlated with histological disease activity using RHI from all samples within the andecaliximab dataset. Eleven Reactome gene sets that were moderately correlated with RHI had low correlation to each other, and represented biologically relevant pathways were used to evaluate each of the study aims. These data are reported as follows: (1) baseline sample differences within the andecaliximab dataset characterize UC disease heterogeneity, (2) correlations of baseline clinical and demographic data from the andecaliximab dataset as well as of published datasets identify the uniqueness of the 11 Reactome gene sets, and (3) cluster memberships predict response to different therapies in 2 independent transcriptomic datasets. A summary of the analytical workflow is outlined in Figure S1.

3.2. C2 gene sets are correlated with RHI score

To investigate whether histological disease activity was associated with RNA signature/gene set enrichment scores, we first evaluated a large compendium of gene sets defined by regulatory pathways, curated sets, computational-derived, ontology, oncology, immunologic, cell type, or hallmark sets to determine which best correlated with RHI score. We calculated gene set scores from all samples (pretreatment baseline and week 8, n = 286 samples) in the UC phase 2 andecaliximab dataset. As shown in Figure S2, many gene sets (C2 [curated], C4 [computational], C5 [ontology], C7 [immunologic], and C8 [cell type signature]) showed right-skewed distributions, indicating an enrichment of higher correlated signatures with RHI score. Although many signatures within each gene set correlated with RHI score (r ≥ 0.40), the C2 gene sets showed the highest correlation to RHI score (r = 0.46). We note that many of the signature sets that have high correlation with RHI were not easily interpreted in the context of known biology (such as computational gene sets, regulatory gene sets, and cell signature gene sets). In contrast, the Reactome pathways (within the C2 group), which also had similar high correlation values, are defined by biologically relevant signaling and metabolic pathways and processes.⁴¹ Therefore, the remainder of the analysis was performed on the Reactome pathways that were highly correlated with RHI.

3.3. A subset of Reactome gene sets define the MARS signatures

While the individual correlations between gene sets and RHI score were modest (r = 0.40 ± 0.02, average correlation calculated by using absolute scores ± SD), we hypothesized that together a small set of these signatures may be useful to describe UC disease heterogeneity. To investigate this hypothesis, we selected Reactome gene sets for the same reasons described in Section 3.2, which showed the highest absolute correlation with RHI score (abs(r) > 0.40) and excluded those with high correlation among themselves (r > 0.85). This approach yielded 11 Reactome gene sets, hereafter referred to as the MARS signatures. This group contains 4 signatures associated with metabolism, 2 signatures associated with the immune system, and 1 each associated with disease, hemostasis, cellular response to stimuli, signal transducers, and protein localization (Table 2). The genes in each signature are listed in Table S1, and the network of each gene/protein is shown in Figure S3.

Table 2.

Reactome gene sets within the MARS signatures.

Correlation with RHI score (r)	C2-CP:Reactome gene set	Parent pathway	Subpathway 1	Subpathway 2	Subpathway 3	GSEA-MSigDB Link
0.46	Regulation of IFNα signaling	Immune system	Cytokine signaling in immune system	Interferon signaling	Interferon-alpha/beta signaling	https://reactome.org/content/detail/R-HSA-912694
0.45	IL-4 and IL-13 signaling	Immune system	Cytokine signaling in immune system	Signaling by interleukins		https://reactome.org/content/detail/R-HSA-6785807
0.42	Platelet adhesion to exposed collagen	Hemostasis				https://reactome.org/content/detail/R-HSA-75892
0.41	Potential therapeutics for SARS	Disease	Infectious disease	SARS-COV infections		https://reactome.org/content/detail/R-HSA-9679191
0.40	Tryptophan catabolism	Metabolism	Metabolism of amino acids and derivatives			https://reactome.org/content/detail/R-HSA-71240
0.40	UPR	Cellular responses to stimuli	Cellular responses to stress			https://reactome.org/content/detail/R-HSA-381119
–0.42	Aspartate and asparagine metabolism	Metabolism	Metabolism of amino acids and derivatives			https://reactome.org/content/detail/R-HSA-8963693
–0.41	Branched-chain amino acid catabolism	Metabolism	Metabolism of amino acids and derivatives			https://reactome.org/content/detail/R-HSA-70895
–0.43	Regulation of FZD by ubiquitination	Signal transduction	Signaling by WNT	TCF dependent signaling in response to WNT		https://reactome.org/content/detail/R-HSA-4641263
–0.44	Sulfur amino acid metabolism	Metabolism	Metabolism of amino acids and derivatives			https://reactome.org/content/detail/R-HSA-1614635
–0.48	Peroxisomal protein import	Protein localization				https://reactome.org/content/detail/R-HSA-9033241

Abbreviations: C2, curated gene sets; CP, canonical pathway; FZD, frizzled receptor; IFNα, interferon-alpha; IL, interleukin; RHI, Robarts Histopathology Index; SARS, severe acute respiratory syndrome; SARS-COV, SARS coronavirus; TCF, T-cell factor; UPR, unfolded protein response; WNT, wingless-type MMTV integration site family.

3.4. The MARS signature reveals patient heterogeneity and associations between gene signatures and histological disease activity

We characterized the baseline samples from patients in the andecaliximab dataset to determine if there were notable differences between patient samples based on the MARS enrichment scores. Baseline clinical and demographic characteristics of these patients were similar: aged 18 to 75 years with moderately to severely active UC defined as Mayo Clinical Score (MCS) ≥ 2, a rectal bleeding score ≥ 1, a stool frequency score ≥ 1, and a Physician Global Assessment of 2 or 3. However, at the level of the MARS enrichment scores, we identified heterogeneity between patient samples exemplified by 3 main clusters (seen as columns of the heatmap in Figure 1A). Samples in cluster 1, comprising 38% of the samples, predominantly displayed low levels of enrichment scores for the signatures negatively correlated with RHI score, hereafter referred to as the “non-disease” gene scores, but high levels of the signatures positively correlated with RHI score, hereafter referred to as the “disease-related” signatures. Although a few samples deviated from this pattern, no subclusters or obvious patterns were discernible. In contrast, samples in cluster 2 showed higher levels of non-disease signature values and lower levels of disease-related signature values. Cluster 3 displayed a mixture with no discernible pattern (Figure 1A). The radar plots of these clusters are generally described, respectively, as (1) right-shifted representing higher disease-related signature scores than non-disease scores, (2) left-shifted, representing higher non-disease scores than disease-related scores, and (3) narrow, representing a mixture of scores that average out to a circle near the center (Figure 1B).

Figure 1.

Andecaliximab dataset baseline clustering revealed differences in gene expression in patient samples. (A) RNA from mucosal biopsy baseline samples from patients with ulcerative colitis in the andecaliximab study were scored with the MARS signatures (right side of heatmap) and sorted using unsupervised hierarchical clustering at both the sample and gene signature level. Values are scaled 1 to –1 on the signature level (row) with red as high and blue as low. (B) Radar plots further depict the mean signature value for each cluster from panel (A). The chart is organized such that each signature is a different axis of the center point, with high values radiating farthest from the center and low scores near the center (scaled 0% to 100% of all sample scores). The radar plots of these clusters are generally described, respectively, as (1) right-shifted representing higher disease-related signature scores than non-disease scores, (2) left-shifted, representing higher non-disease scores than disease-related scores, and (3) narrow, representing a mixture of scores that average out to a circle near the center. AAAM, aspartate and asparagine metabolism; BCAAC, branched-chain amino acid catabolism; FZD, frizzled receptors; I4AI1S, interleukin 4 and interleukin 13 signaling; MARS, metabolism response to stress; PATEC, platelet adhesion to exposed collagen; PPI, peroxisomal protein import; PTFS, potential therapeutics for SARS; ROFBU, regulation of FZD by ubiquitination; ROIS, regulation of interferon-alpha signaling; SAAM, sulfur amino acid metabolism; SARS, severe acute respiratory syndrome; TC, tryptophan catabolism; UPR, unfolded protein response.

3.5. The MARS signature scores do not correlate with clinical or demographic data

We utilized individual patient-level data listed in Table 3 from the andecaliximab study to investigate the sample and/or patient characteristics associated with cluster membership. We examined demographic characteristics (age, sex, and race), clinical characteristics (disease duration, treatment group, prior anti-TNF therapy use or failure, or prior vedolizumab use or failure) as well as disease activity, including MCS and subscores, Geboes subscores, and RHI score. In all cases, sample clusters did not show any statistically significant enrichment for any of these metrics. These results suggest that the signatures identified by correlation to RHI score reveal distinct patient molecular heterogeneity that is not observable by other clinical endpoints or patient characteristics data.

Table 3.

Patient demographics and clinical characteristics from the andecaliximab dataset.

Characteristic	P-value	Adjusted P-value
Response flag	.493	.653
Response type flag	1.000	1.000
Primary endpoint flag	.456	.632
Actual treatment for period 01	.247	.486
Analysis visit	NA	NA
Study identifier	.456	.632
Unique subject identifier	.456	.632
Age	.734	.890
Sex	.342	.580
Race	.246	.486
Randomized population flag	1.000	1.000
Full analysis set population flag	1.000	1.000
Biologic activity analysis pop flag	.083	.350
Completed period 1 induction flag	.021	.248
Completed period 1 maintenance flag	.746	.890
Completed period 2 maintenance flag	.088	.350
Description of planned arm	.247	.486
Description of actual arm	.247	.486
Planned treatment for period 01	.247	.486
Planned treatment for period 02	.199	.486
Actual treatment for period 01	.247	.486
Actual treatment for period 02	.199	.486
Baseline BMI (kg/m2) category	.758	.890
Total doses of exposure—induction	.160	.486
Total doses of exposure—DB treatment	.786	.904
Total doses of exposure—OL treatment	.907	.988
Concomitant use of corticosteroid	.405	.632
Prior TNF-alpha antagonists	.231	.486
Duration of UC (years)	.994	1.000
Smoking status	.045	.340
TNF-alpha antagonist trt failure flag	.285	.542
Immunomodulators prior use flag	.010	.227
Immunomodulators concomitant use flag	.994	1.000
Immunomodulators trt failure flag	.026	.248
Corticosteroids prior use flag	.676	.842
Corticosteroids trt failure flag	.877	.973
Vedolizumab prior use flag	.056	.340
Vedolizumab trt failure flag	.183	.486
Aminosalicylates concomitant use flag	.312	.544
Budesonide concomitant use flag	.448	.632
Unique identifier	.456	.632
EBS clinical remission	.493	.653
Endoscopic subscore	.057	.340
Mayo clinical score	.102	.350
Mayo clinical remission	.310	.544
Mayo clinical response	.839	.948
Partial Mayo clinical score	.554	.705
Physician global assessment for partial MCS	.536	.696
Rectal bleeding subscore	.305	.544
Stool frequency subscore	.063	.340
Grade 0 (structural architectural change; Geboes)	.211	.486
Grade 1 (chronic inflammatory infiltrate; Geboes)	.029	.248
Grade 2A (lamina propria eosinophils; Geboes)	.011	.227
Grade 2B (lamina propria neutrophils; Geboes)	.002	.147
Grade 3 (neutrophils in epithelium; Geboes)	.023	.248
Grade 4 (crypt destruction; Geboes)	.240	.486
Grade 5 (erosion or ulceration; Geboes)	.103	.350
Total Geboes histologic score	.409	.632
Geboes mucosal healing	.072	.340
Geboes mod5	.067	.340
Geboes RHI	.099	.350

Abbreviations: BMI, body mass index; DB, double-blind; EBS, endoscopy/bleeding/stool; MCS, Mayo Clinical score; NA, not available; OL, open-label; RHI, Robarts Histopathology Index; TNF, tumor necrosis factor; trt, treatment; UC, ulcerative colitis.

3.6. The comparison of MARS signatures to other published signatures identifies some similarities

To evaluate how the MARS signatures correlate with published signatures, we visualized in a heatmap the generated enrichment scores for 13 previously described published signatures (Table 4) using samples from baseline and follow-up visits of the andecaliximab dataset and evaluated the correlation between these values and those from the MARS signatures. We then visualized the correlations as a heatmap in Figure 2. Overall, we identified similarities between the MARS signatures and the other published signatures. Of note, the MARS “Interleukin 4 (IL-4) and Interleukin 13 (IL-13) Signalling” signature values were highly correlated with genes that were downregulated in patients with UC in response to infliximab (IFXDn WK4 6 GSE73661; r = 0.94). The MARS signature “Peroxisomal protein import” was highly correlated with genes that were downregulated (UCmeta8Dn; r = 0.86), and “Platelet adhesion to exposed collagen” was correlated with genes that were upregulated (UCmeta8Up; r = 0.87). UCmeta8Dn and UCmeta8Up contain genes that are regulated in patients with UC compared to controls, based on a meta-analysis of 8 publicly available UC datasets.¹⁶ The remaining 9 MARS signatures from this study showed lower correlations with each other or with the additional 13 public signatures.

Table 4.

Summary of published signatures used to evaluate MARS signatures.

GEO series accession number(s)	Signature name	Drug treatment	Indication	Sample type	Organizational name of the dataset provider	PMID (publication year)	UC sample number (n)
GSE107593	UC32_TNF	Anti-TNF therapy	UC	Intestinal biopsy	Pfizer Inc.	33907256 (2021)9	48
GSE92415	MPS_Telesco	Golimumab	UC	Intestinal biopsy	Janssen R&D	29981298 (2018)42	183
NA	MIRI_DEGSEG	Mirikizumab	UC	Intestinal biopsy	Eli Lilly and Company	36881820 (2023)43	444
GSE193677	bMIS_IBD	None	IBD	Intestinal biopsy	MSSM	36109152 (2023)10	421
GSE38713	RGS	None	UC	Intestinal biopsy	IDIBAPS	22605655 (2013)12	15
GSE73661	IFXDn_WK4_6_GSE73661	Infliximab	UC	Intestinal biopsy	KU Leuven	27802155 (2018)23	26
GSE73661	VedoDn_Wk52_GSE73661	Vedolizumab	UC	Intestinal biopsy	KU Leuven	27802155 (2018)23	41
Multiple sourcesa	UCmeta8Up	None	UC	Intestinal biopsy	Multiple sourcesa	34521888 (2021)16	251
GSE73661	IFXUp-GSE73661	Infliximab	UC	Intestinal biopsy	KU Leuven	27802155 (2018)23	26
Multiple sourcesa	UCMeta8Dn	None	UC	Intestinal biopsy	Multiple sourcesa	34521888 (2021)16	251
GSE73661	VedoUp_Wk52_GSE73661	Vedolizumab	UC	Intestinal biopsy	KU Leuven	27802155 (2018)23	41
GSE73661	UC81VEDO	Vedolizumab	UC	Intestinal biopsy	KU Leuven	33907256 (2021)9	53
GSE131032	PauloUC	NA	UC	Mouse intestinal biopsy	Karolinska Institute and University Hospital	31253778 (2019)44	26

^aThere are multiple GEO IDs and providers of these datasets. Refer to the publication reference for greater details.

Abbreviations: bMIS IBD, biopsy molecular inflammation score for patients with inflammatory bowel disease; GEO, Gene Expression Omnibus; MARS, metabolism and response to stress; NA, not available; PMID, PubMed identifier; TNF, tumor necrosis factor; UC, ulcerative colitis.

Figure 2.

Comparison of MARS signatures with 13 public UC signatures. The heatmap depicts the correlation between the MARS signature enrichment scores and enrichment scores for 13 other public UC signatures. Rows and columns are ordered by pairwise Pearson correlation coefficient using hierarchical clustering. Dark red and dark blue indicate strong positive and negative correlations, respectively, while lighter shades of red/blue indicate weaker correlations. Pearson correlation values are indicated as values within the cells. Details of the source of signatures are provided in Table 4. Published UC dataset PMIDs (year): 33907256 (2021)⁹; 29981298 (2018)⁴²; 36881820 (2023)⁴³; 36109152 (2023)¹⁰; 22605655 (2013)¹²; 27802155 (2018)²³; 34521888 (2021)¹⁶; 31253778 (2019).⁴⁴ PMID, PubMed Identifier; UC, ulcerative colitis.

3.7. MARS cluster membership is associated with response to ustekinumab

Due to the lack of efficacy of andecaliximab in the trial, we could not evaluate the ability of the MARS signature scores to predict response to andecaliximab. Alternatively, we used an independent dataset from the UNIFI study, which included 364 baseline mucosal biopsy samples from patients with UC who were treated with the anti-IL-12/23 therapy ustekinumab. We computed the enrichment scores for the 11 MARS gene signatures on all samples and performed hierarchical clustering and heatmap visualization, as described in Section 2.3. These analyses clustered the baseline samples into 4 main groups described from left to right on the heatmap (Figure 3A): cluster 1 had lower scores for non-disease and higher scores for disease-related, cluster 2 was generally low for non-disease and disease-related scores, cluster 3 had high non-disease/low disease-related scores, and cluster 4 was a mixture of score types. The radar plots of these clusters are generally described, respectively, as (1) right-shifted (indicating higher disease-related scores), (2) narrow (indicating lower scores for many signatures), (3) left-shifted (indicating higher non-disease scores), and (4) mixed (Figure 3B). Analysis of the clinical response rates of these clusters showed that cluster 1 (right-shifted in the radar plot) had 4- to 5-times lower mucosal healing rates at week 8 (4.3%) compared to cluster 2 (23%), cluster 3 (22%), and cluster 4 (19%) (adjusted P = .001) (Figure 3C).

Figure 3.

Cluster membership predicts response to ustekinumab. (A) Baseline samples (n = 364) in the UNIFI trial (NCT02407236) were grouped into 4 clusters. Green indicates mucosal healing at week 8, and red indicates lack of mucosal healing at week 8. (B) Radar plots summarize the mean score for each signature per cluster. (C) Proportion of mucosal healing responders at week 8, per cluster (4.3% vs 23%, 22%, and 19%, for clusters 1, 2, 3, and 4, respectively, adjusted P = .001). Clusters defined by MARS signatures identified patients with ~4- to 5-fold lower mucosal healing rates at follow-up. AAAM, aspartate and asparagine metabolism; BCAAC, branched-chain amino acid catabolism; FZD, frizzled receptors; I4AI1S, interleukin 4 and interleukin 13 signaling; MARS, metabolism response to stress; PATEC, platelet adhesion to exposed collagen; PPI, peroxisomal protein import, PTFS, potential therapeutics for SARS; ROFBU, regulation of FZD by ubiquitination; ROIS, regulation of interferon-alpha signaling; SAAM, sulphur amino acid metabolism; SARS, severe acute respiratory syndrome; TC, tryptophan catabolism; UPR, unfolded protein response.

3.8. Cluster membership is associated with response to anti-TNF therapies

To determine whether the ability of MARS signatures to identify non-responders was more generalized and could be applied to patients from other cohorts and treated with other therapies, we evaluated a meta-cohort containing 4 studies including patients with UC who were treated with anti-TNF therapies (Table 1). We performed the same GSVA scoring and analyzed the 11 MARS signatures using unsupervised clustering and heatmap visualization as described in Section 2.3. From baseline samples (n = 138), 4 main sample clusters were described from left to right on the heatmap: cluster 1, containing the fewest samples, exhibited high non-disease and low disease-related scores; clusters 2 and 3 had a mixture, with high/low of both non-disease and disease-related scores; and cluster 4 had low non-disease and high disease-related scores (Figure 4A). The radar plots of these clusters are generally described as left-shifted for cluster 1 and narrow for cluster 2 (lower scores for many signatures). Cluster 3 was mixed, with high scores for some non-disease signatures (peroxisomal protein import and branched-chain amino acid), but also high disease-related signatures, including IL-4 and IL-13 signaling, tryptophan catabolism, potential therapeutics for Severe Acute Respiratory Syndrome, and unfolded protein response (UPR). Cluster 4 was right-shifted (high disease-related signature scores) (Figure 4B). The response rate for these patients at follow-up (week 4, 6, or 8, depending on cohort) was high for clusters 1 and 2 (69% and 70%, respectively), but moderate in clusters 3 and 4 (35% and 40%, respectively, adjusted P = .013) (Figure 4C). Similar to the ustekinumab cohort, the majority of future anti-TNF therapy non-responders were right-shifted, with low non-disease and high disease-related scores.

Figure 4.

Cluster membership predicts response to anti-TNF therapies at baseline. (A) Baseline mucosal biopsy samples (n = 138) in the anti-TNF therapy meta-analysis cohort (GSE16879, GSE23597, GSE73661, and GSE92415) were scored for the MARS signatures. Response at week 4, 6, or 8 is indicated by green lines (non-response is red). (B) Radar plots summarize the mean values of each signature by cluster. (C) Proportion of patients in each cluster who responded at follow-up (week 4, 6, or 8, depending on cohort). The response rate was very high for clusters 1 and 2 (69% and 70%, respectively) but moderate for clusters 3 and 4 (35% and 40%, respectively). AAAM, aspartate and asparagine metabolism; BCAAC, branched-chain amino acid catabolism; FZD, frizzled receptors; I4AI1S, interleukin 4 and interleukin 13 signaling; MARS, metabolism response to stress; PATEC, platelet adhesion to exposed collagen; PPI, peroxisomal protein import; PTFS, potential therapeutics for SARS; ROFBU, regulation of FZD by ubiquitination; ROIS, regulation of interferon-A signaling; SAAM, sulfur amino acid metabolism; SARS, severe acute respiratory syndrome; TC, tryptophan catabolism; UPR, unfolded protein response.

Identification of a cluster enriched for treatment non-responders, which starts with low non-disease and high disease-related scores, prompted us to investigate what happens to the MARS scores after treatment, both in patients who respond and those who do not respond. We calculated the change in MARS scores from baseline to follow-up in all patients in the anti-TNF therapy meta-analysis cohort. Clustering and heatmap visualization revealed 2 main clusters (Figure 5A). Cluster 1 patients had an increase in some of the disease-related scores compared to non-disease scores, as exemplified by the radar plot right-shift (Figure 5B). Cluster 2 included patients who had an increase in non-disease scores and a decrease in disease-related scores, as shown by a left-shifted radar plot (Figure 5B). Patients in cluster 2 had a higher proportion of responders compared to cluster 1 (24% for cluster 1, 64% for cluster 2, adjusted P = .001) (Figure 5C). Therefore, responders were more commonly associated with a left-shifted change, which corresponds to a decrease in disease-related response scores and an increase in non-disease scores.

Figure 5.

Change in MARS scores from baseline to follow-up (n = 89) in the anti-TNF therapy meta-analysis cohort. (A) The change in MARS signature scores between baseline and follow-up (week 4, 6, or 8, depending on cohort) identified 2 main clusters. Response at week 4, 6, or 8 is indicated by green lines (non-response is red). (B) Radar plots show these 2 clusters as right-shifted and left-shifted, respectively. (C) Proportion of responders in clusters 1 and 2. Patients in cluster 2 had a higher proportion of responders compared to cluster 1 (64% and 24%, respectively). AAAM, aspartate and asparagine metabolism; BCAAC, branched-chain amino acid catabolism; FZD, frizzled receptors; I4AI1S, interleukin 4 and interleukin 13 signaling; MARS, metabolism response to stress; PATEC, platelet adhesion to exposed collagen; PPI, peroxisomal protein import; PTFS, potential therapeutics for SARS; ROFBU, regulation of FZD by ubiquitination; ROIS, regulation of interferon-A signaling; SAAM, sulfur amino acid metabolism; SARS, severe acute respiratory syndrome; TC, tryptophan catabolism; TNF, tumor necrosis factor; UPR, unfolded protein response.

Lastly, rather than evaluating the change from baseline in MARS scores, we evaluated the absolute level of the MARS scores at the follow-up visit in this meta-analysis cohort. Unsupervised clustering of follow-up visit MARS scores revealed 2 clusters: cluster 1 with low non-disease/high disease-related response and cluster 2 with high non-disease/low disease-related response (Figure 6A). By radar plots, cluster 1 was right-shifted and cluster 2 was left-shifted (Figure 6B). Cluster 1 contained 16% of responders which was significantly lower (adjusted P = .009) than the proportion of responders in cluster 2 (57%) (Figure 6C). Thus, patients who retained low non-disease and high disease-related scores were more likely to be non-responders. In contrast, patients who had high non-disease and low disease-related scores at follow-up had a high response rate.

Figure 6.

MARS signature scores for follow-up samples in the anti-TNF therapy meta-analysis cohort. (A) Follow-up data for patients treated with anti-TNF therapies clustered with MARS signatures revealed 2 clusters. Response at week 4, 6, or 8 is indicated by green lines (non-response is red). (B) Radar plots depict these clusters as right-shifted (cluster 1) and left-shifted (cluster 2). (C) Proportion of patients who responded, by cluster. Cluster 1 contained a significantly lower proportion of responders compared to cluster 2 (16% vs 57%, respectively; adjusted P = .009). AAAM, aspartate and asparagine metabolism; BCAAC, branched-chain amino acid catabolism; FZD, frizzled receptors; I4AI1S, interleukin 4 and interleukin 13 signaling; MARS, metabolism response to stress; PATEC, platelet adhesion to exposed collagen; PPI, peroxisomal protein import; PTFS, potential therapeutics for SARS; ROFBU, regulation of FZD by ubiquitination; ROIS, regulation of interferon-A signaling; SAAM, sulfur amino acid metabolism; SARS, severe acute respiratory syndrome; TC, tryptophan catabolism; UPR, unfolded protein response.

4. Discussion

Heterogeneity in patient populations has been suggested to be the main reason for variation in patient response observed in clinical trials.⁴⁵ In this study, we use transcriptomic data derived from mucosal biopsies to identify RNA signatures that correlate with histological disease activity. Using this approach, 11 Reactome gene sets (MARS gene signatures) were shown to characterize patient heterogeneity, contain novel pathways related to UC pathogenesis, and identify patients with increased response to therapy. The use of molecular biomarkers from disease sites (eg, mucosal biopsies) may be the next step to obtain higher-resolution information that can identify clinically relevant differences among patients.

Our analysis of multiple clinical datasets consistently identified 2 to 3 main patterns of scores: high non-disease/low disease, low non-disease/high disease, and mixed. Since these groups are treatment agnostic, they are likely, in part driven by the genes within each of the pathways (metabolism, stress, and inflammation) represented by MARS signatures. Although functional validation is required to provide mechanistic insight, literature has shown that several of the pathways represented by the MARS signatures have been associated with UC disease activity and progression.

Diab et al.⁴⁶ performed metabolomic profiling on mucosal biopsies from patients with UC and found several alterations of the amino acid metabolism, tryptophan metabolism, and the alanine, aspartate, and glutamate metabolism to be associated with endoscopic disease activity. In particular, they found that glutamic acid and asparagine were discriminative between treatment-naïve patients with UC (newly diagnosed active disease) vs healthy controls and patients with UC in remission. Interestingly, some metabolic pathways associated with endoscopic disease activity are represented in our non-disease gene signatures identified from patients with moderately to severely active disease. With respect to non-disease genes and treatment response, a recent study by Yang et al.⁴⁷ reported that response to anti-TNF therapy was significantly enriched in metabolism-related pathways identified using a novel predictive gene signature called logOR_Score, which is similar to our findings with the MARS signatures and response to anti-TNF therapy, such that responders had high non-disease gene signatures.

Disease-related genes, particularly those related to the UPR and platelet adhesion from the MARS signatures, are related to UC pathogenesis. Endoplasmic reticulum stress (ERS) in the mucosa induces cellular damage from reactive oxygen species and an inflammatory response, activating UPR genes to mitigate cellular damage.⁴⁸ Regulation of ERS with novel drugs may be a potential therapeutic opportunity.⁴⁹ Patients with UC are also known to have increased platelet activation and aggregation,⁵⁰ which have been shown to decrease with the use of anti-TNF therapies.⁵¹ The molecular mechanisms involved in this process may be interesting to further explore in the context of responders to treatment as has been performed in other disease indications.

Gene sets for “IL-4 and IL-13 Signalling and Peroxisomal Protein Import” were significantly correlated to other public signatures, which is not surprising since both are involved with inflammatory processes of the epithelium. The cytokines IL-4 and IL-13 are commonly involved in the regulation of immune responses and share common receptors and signaling pathways.⁵² Several treatments for UC directed at the IL-4α receptor and IL-13 signaling have been investigated,⁵³ such as anrukinzumab⁵⁴ and tralokinumab.⁵⁵ Anti-TNF therapies have also been shown to reduce IL-13 production in patients with UC through the IL-13α2 receptor.⁵⁶ Peroxisomal protein import and peroxisomal-related pathways are associated with an inflammatory response and have been shown to be upregulated in the blood of patients with IBD who responded to vedolizumab, compared with baseline,⁵⁷ and downregulated in inflamed UC mucosa compared with controls.⁵⁸

Amongst the group identified, some samples were clearly different from the stereotypical patterns. While we do not expect the current clustering approaches to robustly classify these small subgroups of patients confidently, there are potentially multiple levels of differences that may be relevant for disease-related responses. For example, a small group of patients in cluster 4 of the UNIFI study dataset were non-responders, and when compared to responders, showed visual differences in scores for several signatures (most dramatically for aspartate and asparagine metabolism, regulation of frizzled receptor by ubiquitination, and tryptophan catabolism), which could be hypothesized to contribute to their non-response. The difference in the number of ustekinumab responders, when measuring the MARS scores at baseline, suggests that the MARS scores can identify patient heterogeneity that is not explained by clinical data. In particular, patients who have higher disease-related scores and lower non-disease scores at baseline may be less likely to achieve treatment response. It remains to be determined whether such pathways are critical nodes that could be therapeutic targets for effective treatment.

Additionally, using longitudinal samples from several anti-TNF therapy–treated cohorts, we demonstrated that responders had decreased disease-related genes and increased non-disease genes during follow-up, which is consistent with known UC disease pathophysiology and suggests that clusters identified by the MARS signatures may identify a difficult-to-treat subtype of patients with UC. Patients who responded had a combination of high non-disease/low disease-related scores and continued to further increase their non-disease score and reduce their disease-related score. However, by evaluating a score that is dependent on the starting value, the largest changes will occur in patients with very extreme scores at baseline. Consequently, the patients at these extreme levels will have larger changes but are less likely to respond since they initially have more severe disease. We evaluated this possibility by examining the absolute scores of patients at follow-up, and as expected, observed that responders appeared to be in the relatively high non-disease/low disease-related score state.

Several limitations to our study should be acknowledged. Firstly, moderate correlations observed between histological RHI and Geboes scores are likely a result of the biopsies used for histological scoring in the andecaliximab trial being separate from those used for RNA analysis and a result of histological scoring by definition using a qualitative, spatial scoring system that incorporates many structural and cellular components. Secondly, RNA-sequencing analysis is a bulk method, which can miss small components by averaging out the RNA levels. The use of spatial transcriptomics is one way to overcome this limitation and better define the molecular components contributing to histological scores. Thirdly, the robustness of the clustering in small cohorts, variations in study designs between cohorts or analytics, made it difficult to define some clusters. The availability of more recent, larger transcriptomic datasets is necessary to further define the clusters. Lastly, the clustering of patients using the MARS signatures described in this paper is qualitative, and in its current form, could not be used in clinical trials or clinical practice. However, this is an intentional strategy to minimize the loss of information when a predictive marker is distilled to a single cutoff value.

The development of a validated biomarker(s) for use in clinical trials or patient treatment decisions is a difficult, complicated, and lengthy process that has yet to be achieved in IBD, with the only biomarkers currently being used in clinical and trial practice being C-reactive protein and fecal calprotectin (although not completely validated).⁵⁹ We recognize that the work described herein is the first step in the discovery of a set of biomarkers that requires additional steps to allow their use in clinical trials or clinical practice. Future work would require validation in additional datasets and a prospective study. As well, the assay would require simplification, such as the development of a qPCR panel using a reduced number of genes that correlate with the larger MARS gene set. Additionally, the identification of a robust scoring system with predefined cutoffs to prospectively classify patients into each group would need to be developed.

In conclusion, we have developed a novel, multi-dimensional scoring system to molecularly characterize previously undefined heterogeneity in patients with UC and to identify patients less likely to respond to therapy. This exploratory approach broadens our knowledge of these principles and has the potential to define clinical trial populations, enrich for clinical responders, and identify difficult-to-treat populations for therapeutic development in the future.

Author contributions

Bryan Linggi: study conception and design, data analysis and interpretation, literature review, and curation of the manuscript.

Melissa Filice: literature review, drafting of original manuscript, and critical review of the manuscript for important intellectual content.

Bruno Sangiorgi, Michelle I. Smith, Wendy Teft, and Niels Vande Casteele: study conception and design, data interpretation, and critical review of manuscript for important intellectual content.

Vipul Jairath and Christopher Ma: data interpretation, and critical review of manuscript for important intellectual content.

All authors approved the final version of the manuscript for submission.

Funding

None declared.

Conflicts of interest

B.L. is a former employee of Alimentiv Inc. and is currently an employee of AnaptysBio Inc.

M.F., B.S., M.I.S., and W.T. are employees of Alimentiv Inc.

V.J. has received consulting/advisory board fees from AbbVie, Alimentiv Inc., Arena Pharmaceuticals, Asahi Kasei Pharma, Asieris, Astra Zeneca, Bristol Myers Squibb, Celltrion, Eli Lilly, Ferring, Flagship Pioneering, Fresenius Kabi, Galapagos, GlaxoSmithKline, Genentech, Gilead, Janssen, Merck, Mylan, Pandion, Pendopharm, Pfizer, Protagonist, Prometheus, Reistone Biopharma, Roche, Sandoz, Second Genome, Takeda, Teva, TopiVert, Ventyx, and Vividion; and has received speaker’s fees from AbbVie, Ferring, Galapagos, Janssen Pfizer Shire, Takeda, and Fresenius Kabi.

C.M. has received consulting fees from AbbVie, Alimentiv Inc., Amgen, AVIR Pharma, Bristol Myers Squibb, Ferring, Fresenius Kabi, Janssen, McKesson, Mylan, Takeda, Pfizer, and Roche; has received speaker’s fees from AbbVie, Amgen, AVIR Pharma Inc., Alimentiv Inc., Ferring, Janssen, Takeda, and Pfizer; and has received research support from Pfizer.

N.V.C. is a former employee of AcelaBio, Inc.

Data availability

Publicly available microarray gene expression datasets that were used to predict response to ustekinumab (GSE206285) and anti-TNF therapies (GSE16879, GSE23597, GSE73661, and GSE92415) were obtained from the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) data repository: https://www.ncbi.nlm.nih.gov/geo/. The andecaliximab (GS-5745) dataset was provided by Gilead Sciences, Inc. by permission, and will be shared on request to the corresponding author with permission of Gilead Sciences, Inc. Remaining data are available on request and will be shared on reasonable request to the corresponding author.