Crohn’s Patient Serum Proteomics Reveals Response Signature for Infliximab but not Vedolizumab

Abstract

Background

Crohn’s disease is a chronic inflammatory bowel disease that affects the gastrointestinal tract. Common biologic families used to treat Crohn’s are tumor necrosis factor (TNF)-α blockers (infliximab and adalimumab) and immune cell adhesion blockers (vedolizumab). Given their differing mechanisms of action, the ability to monitor response and predict treatment efficacy via easy-to-obtain blood draws remains an unmet need.

Methods

To investigate these gaps in knowledge, we leveraged 2 prospective cohorts (LOVE-CD, TAILORIX) and profiled their serum using high-dimensional isobaric-labeled proteomics before treatment and 6 weeks after treatment initiation with either vedolizumab or infliximab.

Results

The proportion of patients endoscopically responding to treatment was comparable among infliximab and vedolizumab cohorts; however, the impact of vedolizumab on patient sera was negligible. In contrast, infliximab treatment induced a robust response including increased blood-gas regulatory response proteins, and concomitant decreases in inflammation-related proteins. Further analysis comparing infliximab responders and nonresponders revealed a lingering innate immune enrichments in nonresponders and a unique protease regulation signature related to clotting cascades in responders. Lastly, using samples prior to infliximab treatment, we highlight serum protein biomarkers that potentially predict a positive response to infliximab treatment.

Conclusions

These results will positively impact the determination of appropriate patient treatment and inform the selection of clinical trial outcome metrics.

Key Messages.

• What is already known?

• Infliximab and vedolizumab are 2 common therapies for Crohn’s disease, but matching patient, treatment, and dose is not easily predictable. Patient serum is often used for monitoring treatment efficacy, yet little is known about serum markers that robustly predict infliximab or vedolizumab response, or its appropriateness for monitoring.

• What is new here?

• We profiled 2 large longitudinal clinical monitoring cohorts using isobaric-labeled proteomics to identify response predictors and underlying biological differences in vedolizumab and infliximab responders and nonresponders. Infliximab-treated patients displayed differences in the activation of underlying immunological processes when comparing responders and nonresponders.

• How can this study help patient care?

• These results will inform future clinical trial designs.

Introduction

Crohn’s disease (CD) is a major subtype of inflammatory bowel disease (IBD). Recent meta-analyses of CD have reported increasing incidence rates, which continue to drive treatment burdens and hospitalization costs.^1-3 While CD’s etiology remains unclear, extensive patient profiling efforts noted increased active immune response and concomitant intestinal and systemic inflammation. More specifically, a pathologic increase in tumor necrosis factor alpha (TNF-α) and excessive infiltration of activated leukocytes into diseased tissue have been noted. To combat active disease flares and prevent long-term tissue damage, therapeutic biologics were developed and have revolutionized IBD treatment. Infliximab (IFX) is a well-known TNF-α-targeting monoclonal that limits TNF-α’s ability to contribute to systemic inflammation-related networks and is generally well-tolerated by patients.⁴ Vedolizumab (VDZ), in contrast, targets integrin α₄β₇ and limits the ability of activated adaptive immune cells to infiltrate intestinal mucosa.⁵

Extensive clinical data show IFX and VDZ are capable of inducing and maintain remission in CD patients.^6,7 Nevertheless, the currently accepted treatment targets, such as endoscopic remission, are only reached in 30% to 40% of patients.^8,9 Furthermore, identifying optimal patient treatments often takes months to years and requires cycling through several options, with a significant number of relapses between treatment switching. This selection is further hindered by current minimally to moderately predictive biomarkers such as C-reactive protein, albumin, and α4β7 serum levels.^10,11 Interestingly, it appears the initial biologic agent has the highest likelihood of success, with second-line choices being consistently less effective, making initial selections critical to disease management and patient outcomes.¹² Currently, the differences in systems-level effects between responders and nonresponders to IFX and VDZ are not entirely characterized, possibly contributing to suboptimal treatment selection. To fill this gap in knowledge, previous groups have utilized serum proteomics using various technologies to profile the effects of IFX and VDZ treatment, including prediction of relapse and connections to the microbiome.^13–17 However, many prior studies were partially hampered either by small cohorts limiting statistical power, use of prior-generation low-sensitivity mass spectrometers hampered by slow scan speeds and decreased dynamic range, or the use of predetermined target panels that hinder the unbiased discovery of relevant features. As such, characterizing IFX and VDZ’s effect on circulating proteins would benefit significantly from a more complete proteomic profiling. Furthermore, questions regarding the ability of serum to monitor the effects of targeted therapeutics such as VDZ remain largely unanswered. Lastly, identifying predisposing pathways and features that predict response or nonresponse to VDZ or IFX remains an outstanding challenge to the field.

To address these shortcomings, we profiled pre- and post-treatment serum in biologic-naïve early Crohn’s disease patients collected from 2 cohorts prescribed VDZ (LOVE-CD cohort, n = 29, clinical trial ID: NCT02646683) or IFX (Tailorix cohort, n = 36, clinical trial ID: NCT01442025) using isobaric multiplex-labeling technology paired with modern orbitrap-based mass spectrometry. (These trials were previously approved by their respective human research boards.) We show that IFX treatment induces significant group-level effects on the serum proteome, while VDZ does not. We further identified response-specific signatures in IFX cohort serum reflected in orthogonal measurements. Lastly, we use machine learning to classify responders and nonresponders prior to treatment with moderate success and reveal clear enrichments in protein classes predicting response, which may be of clinical relevance. Overall, these results both provide novel insights into systems-level events contributing to response to commonly prescribed therapeutics and simultaneously guide the development of diagnostic panels and CD treatment.

Materials and Methods

Experimental Design

Serum of biologic-naïve adult patients (≥18 years) with early CD (<5 years) and documented endoscopic disease activity (SES-CD for LOVE, CDEIS for TAILORIX, see Table 1) was collected. For this study, 2 independent cohorts of prospectively collected samples and data were used. The first cohort included patients that were treated with IFX in the TAILORIX clinical trial (2012-2015, clinical trial ID: NCT01442025). A detailed description of the trial design and clinical outcomes was published previously.⁸ The second cohort included patients that were treated with VDZ in the LOVE CD clinical trial (2016-2019, clinical trial ID: NCT02646683). A detailed description of the trial was published previously.⁹ Due to a disbalance in patient numbers, serum from an additional 4 VDZ patients that were included in the IBD biobank at the Amsterdam UMC, location AMC, that met the previously mentioned criteria were included. For all patients, a week 0 (pretreatment) and week 6 (on-treatment) serum sample were collected. After baseline, a second endoscopy was performed to assess treatment efficacy after 14 weeks (TAILORIX) or 24 weeks (LOVE). All endoscopies were centrally read. Response to treatment was defined by a ≥50% reduction in the endoscopic disease activity score (Simple Endoscopic Score for CD [SED CD] in the LOVE and the Crohn’s Disease Endoscopic Index of Severity [CDEIS] for TAILORIX). Available clinical data included patient characteristics (age and gender), disease characteristics (medical and surgical history, Montreal phenotype classification), and current and prior treatments.

Table 1.

Demographic and clinical characteristics.

	Vedolizumab (LOVE) (n = 29)		Infliximab (TAILORIX) (n = 36)
	Responder (n = 15)	Nonresponder (n = 14)	Responder (n = 16)	Nonresponder (n = 20)
Age, mean (SD)	32.7 (12.8)	30.0 (10.8)	33.1 (11.6)	31.8 (15.4)
Male, n (%)	6 (40)	8 (57.1)	6 (37.5)	9 (45)
Caucasian, n (%)	14 (93.3)	13 (92.9)	DNC	DNC
Montreal class, n (%)
L1	6 (40)	1 (7.1)	2 (12.5)	5 (25)
L2	4 (26.7)	3 (21.4)	5 (31.3)	3 (15)
L3	3 (20)	10 (71.4)	9 (56.3)	12 (60)
B1	14 (93.3)	11 (78.6)	13 (81.3)	10 (50)
B2	1 (15)	3 (21.4)	1 (6.3)	9 (45)
B3	-	-	2 (12.5)	1 (5)
CD disease duration, median (IQR, months)	15.0 (6.0-30.0)	23.5 (8.3-46.0)	6.0 (1.0-46.3)	6.5 (0.0-52.5)
Smoking, n (%)
Active smoker	4 (26.7)	5 (35.7)	3 (18.8)	6 (30)
Ex smoker	4 (26.7)	3 (21.4)	3 (18.8)	4 (20)
Never smoked	7 (46.7)	6 (42.9)	10 (62.5)	10 (50)
Active perianal fistula, n (%)	-	-	4 (25)	3 (15)
Baseline concomitant medication, n (%)	9 (60.0)	13 (92.9)	DNC	DNC
Previous CS use, n (%)	15 (100)	14 (100)	DNC	DNC
Previous IMM use, n (%)	9 (60.0)	7 (50.0)	6 (37.5)	4 (20)
Previous anti-TNF use, n (%)	3 (20.0)	6 (42.9)	-	-
CRP (mg/L), median (IQR)	7.5 (2.0-13.1)	11.6 (1.4-24.6)	23.5 (8.3-33.5)	26 (9.8-65.5)
SES CD at baseline, median (IQR)	6.0 (4.0-7.0)	14 (8.5-18.0)	DNC	DNC
CDEIS at baseline, median (IQR)	DNC	DNC	10 (7.4-15.5)	11 (6.4-18.8)
Fecal calprotectin at baseline, median (IQR)	DNC	DNC	1800 (1613-1800)	1800 (1388.5-1800)
Albumin, mean (SD)	44.0 (3.6)	41.8 (4.5)	36.7 (5.4)	36.5 (5.2)

Summarized demographic data of LOVE and Tailorix cohorts. Acronyms: CS, corticosteroids; IMM, immunomodulators; CRP, C-reactive protein; CDEIS, Crohn’s Disease Endoscopic Index of Severity; SES CD, Simple Endiscopic Score Crohn’s Disease, DNC, data not collected.

Summary statistics of included cohort subjects.

Sample Processing

Frozen patient sera were shipped overnight directly from Amsterdam UMC, location AMC to University of California San Diego on adequate dry ice and subsequently placed in a −80°C freezer until further processing, never being thawed until processing. Samples were first thawed on ice and then vortexed thoroughly; and 50 uL was aliquoted into a fresh 2 mL tube. To this, 200 uL processing buffer was added (7% sodium dodecyl sulfate, 6M urea, 50 mM TRIS, Roche PhosStop tablet [4906845001], Roche protease inhibitor tablet [11836170001], pH 8.1). Each sample was vortexed and sonicated at 20% power for 5 seconds on, 5 seconds off for 3 cycles on ice. Samples were then reduced (5 uL 500 mM Dithiothreitol) for 30 minutes at 47°C and alkylated (15 uL 500 mM iodoacetamide) for 45 minutes at room temp in the dark. To each sample, 1.5 mL of “binding buffer” was added (90% Methanol, 10% triethylammonium bicarbonate [TEAB], pH 7.1). Each sample was then loaded into an individual well of a Protifi S-trap 96-well plate (C02-96well-1), and wells were washed 5 times with binding buffer. After removing the binding buffer, samples were subjected to trypsin digestion by adding 5 ug of trypsin resuspended in 125 uL 50 uM TEAB for 3 hours at 47°C. Samples were then eluted sequentially by adding 150 uL 50 mM TEAB, then 5% formic acid (FA), and lastly 50% acetonitrile (ACN)/5% FA in a centrifuge spinning at 1000 relative centrifugal force at room temp. Samples were then desalted using Waters Seppak c18 columns (186000308). Peptides were then dried down in a SpeedVac (Thermo SpeedVac). Peptide abundance was ascertained using Thermo-Fisher PepQuant kits; 50 ug of each sample was aliquoted and dried down in a clean 96-well plate. Additionally, a “bridge” sample/channel was made composed of equal proportions (ug) of all samples for normalization purposes. To control for batch effects, all samples were randomly assigned to a multiplex group and channel prior to labeling (15 samples + 1 bridge channel). Specific information on groupings is deposited in a project upload to MassIVE repository (MSV000091105, doi:10.25345/C5NS0M73C, PW: TAILORIX_REV). Each sample was then labeled using TMTpro 16-plex 5 mg kit according to manufacturer recommended directions. Briefly, each 5-mg channel included was resuspended to approximately 250 uL with dry acetonitrile. From this stock, 7 uL of the appropriately assigned channel was added to 50 ug of the peptide sample. After quenching, samples were mixed into their respective 16-plex and dried down. Each multiplexed sample was then desalted, dried, and fractionated via reverse phase high pH HPLC (Ultimate 3000, Thermo Scientific) into 12 fractions (1 mL per fraction, Biobasic c18 column) over the course of 72 minutes. Half of each fraction was taken and dried down for analysis via mass spectrometry.

Mass Spectrometry Data Acquisition

Each dried fraction was resuspended in 20 uL of 5% ACN, 5% FA, and loaded into a Thermo Scientific Easy-nLC 1000 HPLC autosampler with aqueous phase A (reagent = 97% water, 2.9% ACN, and 0.125% FA), while organic phase B was 99.9% ACN, and 0.125% FA. Samples were loaded onto a house pulled and packed column with an inner diameter of 100 um and outer diameter of 360 um. The column was initially packed with 0.5 cm of 5 um c4 (frit), followed by 0.5 cm of 3 um C18, and finally packed with 29 cm 1.8 um c18. Each fraction was subjected to a gradient with the following parameters: total run time = 180 minutes, ramping from 6% ACN (t = 0) to 25% ACN (t = 165), then ramping to 100% ACN for 5 minutes (t = 170).

Mass spectrometry data were acquired on a Thermo Orbitrap Fusion in positive mode using the following settings: MS1 was acquired in data-dependent mode using a scan range of 500 to 1200 m/z and resolution of 60 000 in centroid mode. For MS1, the maximum dwell time was 250 ms, with automatic gain control (AGC) = 2 × 10⁵, and the number of scans = TopN mode. For MS2, ions were first isolated in the quadrupole with an isolation window of 0.7 m/z, with subsequent fragmentation data was acquired using CID in the ion trap using a decision tree method with 2 branches: z = 2 ions within the 600 to 1200 m/z range, while z = 3 or 4 were acquired between 500 and 1200 m/z, a minimum intensity of 5e3, and a maximum injection time of 50 ms. These data were used for subsequent fragment sequence identification. The lower threshold for ion fragmentation was 5 × 10⁴. Selected ions were fragmented using CID in the linear ion trap in centroid mode. Primary scan was set as “rapid scan rate” and AGC = 1 × 10⁴. Quantification based on MS3 scans was performed in the Orbitrap using higher-energy collisional dissociation fragmentation with synchronous precursor selection (SPS) = TRUE, SPS precursors N = 10. These reporter ion detection events occurred with 60 000 resolution and an AGC of 1 × 10⁵ with maximum ion injection time of 250 ms.

Data Processing

The resulting mass spectra raw files were first searched using Proteome Discoverer 2.1 using the built-in SEQUEST search algorithm. A single database was used to search potential spectra Uniprot Swiss-Prot Homo sapiens (taxon ID 9606, downloaded January 2021). Target-decoy searching at both the peptide and protein level was employed with a strict FDR cutoff of 0.01 using the Percolator algorithm built into Proteome Discoverer 2.1. Enzyme specificity was set to full tryptic with static peptide modifications set to carbamidomethylation (+57.0214 Da) and, when appropriate, TMTpro (+304.2071 Da). Dynamic modifications were set to oxidation (+15.995 Da) and N-terminal protein acetylation (+42.011 Da) and phosphorylation (+79.996 Da). Only high-confidence proteins (q < 0.01) were used for analysis.

Peptide spectral matches (PSM) generated using the Proteome Discoverer Pipeline were then further processed using a custom in-house peptide-to-protein abundance script, which mimics the method Proteome Discoverer uses to roll PSM-level information into proteins, using additional quality metrics to assure protein quantitative quality including isolation interference (values >30 filtered) filtering and signal-to-noise filtering (values <10 filtered). These values were then normalized against median values of bridge channels (themselves already a median of all bridge channels). These values were then once more transformed using Log2 transformation prior to any statistical analysis. After quality and sparsity filtering (quantitative values in >40% of samples), remaining missing values were imputed using the missForest R package, which has shown to provide a good balance of accuracy and speed. All RAW and MZML mass spectrometry files are available on massive.ucsd.edu (MSV000091105, doi:10.25345/C5NS0M73C, PW: TAILORIX_REV) along with accompanying results files and metadata necessary for future meta-analyses. Additionally, we have provided the chromatograms of 5 randomly selected files for easy assessment of quality (Supplementary Figure 4).

Statistical Rationale

All statistical analyses were done using 2 different computing environments, where appropriate; R statistical language was used for a majority of analyses using matrixTests package (v0.1.9), with multiple hypothesis testing added using “false discovery rate” (also known as Benjamini-Hochberg correction) using Hmisc package (v4.6). Univariate analysis was done using GraphPad Prism v9.

Results

Patient Demographics and Clinical Characteristics

Serum samples from a total of 65 patients were used from 2 prospective clinical trial cohorts: LOVE CD (VDZ, n = 29) and TAILORIX (IFX, n = 36). For each patient a baseline (pretreatment) and week 6 (during treatment, hereafter referred to as post-treatment for comparison clarity) serum sample was analyzed. Patients received 5 mg/kg at weeks 0, 2. and 6 as induction treatment followed by maintenance treatment with 5 mg/kg every 8 weeks. Patients were randomized into 3 maintenance regimens where the infliximab dose could be increased to a maximum dose of 10 mg/kg according to either clinical and/or biological criteria of no response. Patients received 300 mg of VDZ at week 0, 2, and 6 as induction treatment followed by maintenance treatment with 300 mg of VDZ every 8 weeks. An additional VDZ infusion was administered at week 10 if patients had no clinical response (Table 1).

IFX Treatment Induces a Robust Proteomic Response in Sera While VDZ Does Not

Based on the well-characterized differences in mechanism of action and VDZ’s target selectivity, we hypothesized a proteomic IFX signature would likely be more observable in serum than VDZ. To test this, we subjected pre- and post-treatment serum from patients to multiplexed mass spectrometry–based proteomics.

Overall, we identified 651 high-confidence (FDR < 0.01) proteins from the entire cohort. After additional quality filtering for sparsity (quantitative values present in >40% of samples), 468 proteins were left for subsequent analyses (Supplementary Figure 1A). Due to the inherent properties of multiplexing and the random sample grouping for data acquisition, quantitative values for all proteins were present in both VDZ and IFX-treated cohorts in at least 1 patient per subgroup. These steps resulted in only highly scored proteins present in most samples in each subgroup which were included in further analyses.

We first compared whole-group profiles of pre- and post-treatment serum from patients treated with either IFX or VDZ. We observed significant differences between baseline and 6 weeks after IFX treatment regimens (PERMANOVA, q < 0.001, Figure 1A). In contrast, VDZ post-treatment patients were not significantly distinguishable from baseline (PERMANOVA, q = 0.97). This trend was further exemplified in the number of significantly altered proteins in the IFX group compared with virtually no change seen in VDZ patients post-treatment (Figure 1B). Even without filtering for significance cutoffs, VDZ treatment revealed the change of protein abundance using repeated measures (post-treatment to pretreatment) was largely centered around zero, while IFX displayed a wider distribution of scores (Supplementary Figure 1B). This finding was independent of response rates that were equally distributed for both IFX and VDZ. These results suggest a VDZ signature is not easily discernible using serum, or at least not within the first 6 weeks of treatment. Given these results, we focused on IFX-respondent changes in sera.

Figure 1.

Infliximab, but not Vedolizumab, significantly alters patient serum profiles. A, Principle coordinate analysis (PCoA) of IFX and VDZ pre/post treatment. Statistics calculated are generated using PERMANOVA. B, Volcano plots comparing pre- and post-treatment INF and VDZ differential abundance. Color applied using cutoff of P < .05 and ∆post-pre > 0.1. C, Gene ontology enrichments and associated string-based protein-protein interaction networks generated using significantly up- and downregulated (P < .05) proteins post-treatment IFX patients. D, Univariate analyses of selected proteins identified as differentially abundant in INF vs VDZ. Student’s t test with Welch’s correction used to generated statistics, ^*P < .05, ^**P < .01, ^****P < .0001.

To gain an understanding of pathways most affected by IFX, we subjected a subset of proteins (q < 0.05, Log2 fold-change > 0.5) to gene enrichment analysis. This filtered list consisted of 10 proteins increased after IFX treatment, and 18 decreased post-treatment. Independent of response, IFX induced robust changes in gas exchange- and pH balance-related proteins (Figure 1C). Among these, several hemoglobin subunits (HBA2, HBB, HBD) as well as Carbonic anhydrase 1 and 2 were increased, with their paired expression strongly correlated (Spearman rho = 0,89, FDR = 4e-8, Figure 1D). Expanding this network with known inflammation-related cytokines IL-6, TNF-alpha, and STAT3 revealed a highly connected hub with serum albumin connecting inflammatory cytokines and upregulated gas exchange proteins, with the cytokines segregated from the enriched proteins (Supplementary Figure 1C). Serum albumin levels have previously been reported as inversely proportional to IL-6 and TNF-alpha abundance.¹⁸ We observed post-IFX serum albumin levels were significantly higher compared with post-VDZ, an alteration also reflected by Fetuin-A, another serum carrier class protein (Supplementary Figure 1D). When combined with the network analysis, this result further confirms the connection between albumin and circulating inflammatory cytokines. Further reinforcing this, proteins strongly associated with increased inflammation such as calprotectin, C-reactive protein, and complement-related proteins were all downregulated post-IFX. These enrichments were similar to comparisons between week 6 IFX and VDZ, confirming the minimal changes that occurred in VDZ-treated patients’ sera (Supplementary Figure 1E).

In sum, these results suggest despite similar responder and nonresponder proportions within each cohort, only IFX induced a robust sera-detectable proteomic response within 6 weeks after treatment initiation. This consisted of a significant increase in gas transport and pH altering proteins and a decrease in inflammatory proteins. Despite these robust overall alterations post-IFX, a significant proportion of patients failed to respond endoscopically, and thus we next tested if IFX responders and nonresponders displayed unique proteomic signatures.

IFX Responders Display Minimal But Meaningful Profile Differences Compared With Nonresponders

Only 44% of our IFX-treated cohort had significant endoscopically confirmed positive response (Table 1). We hypothesized the robust response previously described in IFX patients was largely, but not entirely, driven by responders. Further, to our knowledge differences in responder and nonresponder sera profiles of IFX patients are not currently reported in the literature. To address these topics, we compared serum profiles of IFX responders to nonresponders.

In contrast to our hypothesis, comparing pre- and post-treatment samples for responders and nonresponders revealed both responders and nonresponders exhibited significant alterations in protein abundance post-IFX (Figure 2A). Groupwise statistics further revealed post-IFX treated responders and nonresponder profiles did not significantly differ (PERMANOVA q = 0.59, Pseudo-F = 0.91). This was further confirmed by comparing individual protein abundances from post-IFX treatment responders to nonresponders (Figure 2B). Repeated-measures comparisons of responders and nonresponders (pre-post Δresponders - Δnonresponders, Supplementary Figure 2A) revealed both groups displayed a reduction in inflammation-related proteins. However, the magnitude of decrease was greater in responders compared with nonresponders (Figure 2C). In line with this, gene ontology (GO)-term enrichment networks generated using proteins whose abundance was significantly greater in nonresponders vs. responders revealed enrichments strongly associated with inflammation, confirming their nonresponder status (Figure 2D). Conversely, comparing terms that increased to a greater degree in responders revealed minimal GO-term network connectivity but strong cluster of orthologous genes (COG) enrichments for cell-cell junction/adhesion motifs, suggesting increased endothelial or epithelial integrity post-IFX (Supplementary Figure 2B).

Figure 2.

Infliximab responders display differential immune profiles compared with nonresponders. A, Volcano plots comparing pre- and post-treatment differential abundance in responders and nonresponders. Color applied using cutoff of P < .05 and ∆post-pre > 0.1. B, Volcano plots comparing post-treatment only abundances in responders and nonresponders. Color applied using cutoff of P < .05 and ∆post-pre > 0.1. C, String-based protein-protein interaction networks generated using proteins enriched in post-treatment nonresponder patients compared with responders. D, Comparison of selected protein abundance in responders and nonresponders.

Identification of Serum Proteins Signaling IFX Patient Response

Currently, several groups have proposed predictive biomarkers, indicating a positive response to IFX administration in IBD research and systematic reviews.^19–25 To test whether previously proposed biomarkers were useful as serum biomarkers, as well as identify any new biomarkers, we profiled pre-IFX responders using machine-learning pipelines.

We first subjected pre-IFX samples to ensemble feature selection to identify features most predictive of response.²⁶ After algorithm rankings, we filtered features for those with >0.5 summed-selection score, resulting in 10 filtered proteins (Figure 3A). Among these highly ranked features, enrichments were seen in immunoglobulin fragments and lipid transport-related apolipoproteins (SAA, APOC, and APOA). Comparisons of these features revealed minimal but significant differences between these proteins (Figure 3B-C). Interestingly, all immunoglobulin fragments listed were significantly increased in nonresponders. Expanding this analysis to include all immunoglobulin fragments identified in pre-IFX samples, we observed that a majority were present in relatively equal proportions in responders and nonresponders; however, a subset also displayed differential expression, with a majority of these being increased in nonresponders (Supplementary Figure 3A). This result may suggest nonresponders have a unique adaptive immune response, which may not be significantly affected by TNF-α-modulating therapies. With respect to the highly ranked lipid transport-related proteins, SAA and APOC4 were significantly higher in responders, while APOA4 was significantly increased in nonresponders (Supplementary Figure 3B,C). Within-patient APOC4:APOA4 ratio-metric comparisons revealed responders and nonresponders were significantly different (Figure 3D). Beyond their utility in classification tasks, this suggests that certain classes of proteins, and their associated networks, may predispose patients to response or nonresponse. Local network analysis of each protein suggests APOC4 is directly involved in VLDL particle clearance (GO:0034447, FDR = 1e-6) and related to lysosomal enzyme secretion (GO:0090182, FDR = 1e-4). In contrast, APOA4 negatively regulates VLDL clearance (GO:0010903, FDR = 3e-4) and instead has direct network effects on chylomicron remodeling (GO:0034371, FDR = 3e-14), suggesting an opposing functional network effect (Supplementary Figure 3D). This may suggest differences in underlying lipid biology pathways between responders and nonresponders; however, confirming or expanding these results are beyond the scope of the study.

Figure 3.

Predictors of INF response. A, Ensemble feature selection scores generated using 5 different criteria generated using EFS algorithm. B, Univariate analyses of selected proteins identified by feature selection. Student’s t test with Welch’s correction used to generated statistics, ^*P < .05, ^**P < .01. C, Univariate analyses of selected immunoglobulin fragments identified by feature selection. Student’s t-test with Welch’s correction used to generated statistics, ^*P < .05, ^**P < .01. D, Univariate analyses of APOC/APOA ratio comparing responders and nonresponders. Student’s t-test with Welch’s correction used to generated statistics, ^*P < .05. E, Test set feature importance scores, as generated by permutation-based feature importance testing. F, ROC curve generated using the 3-protein trained model and K-nearest neighbors algorithm.

While the biological rationale of these findings remains to be discovered, their selection suggests they can classify responders and nonresponders. To test this, we split each group into training and testing groups (75% training, 25% testing) and trained 5 separate classifiers with these features and selected the algorithm with the highest accuracy (K-nearest neighbors). After testing their ability to classify the test set, we used recursive feature selection to identify features diminishing model accuracy. We observed that 3 features positively contributed to the model (Figure 3E). Using only these 3 features increased model accuracy significantly (Figure 3F). However, the current cohort size is very small for machine-learning standards; and lack of a validation cohort to test these features warrants significant levels of caution for any further applicability.

Beyond their moderate ability to identify responders and nonresponders, the selected features further highlight potential subtle biologically relevant differences, which should be investigated in a more mechanistic fashion.

Discussion

Serum (and blood in general) is one of the most often profiled biomaterials due to its ease of collection and ability to capture biologically relevant post-treatment responses. In line with this, we observed a robust response from IFX-treated patient sera, with significant decreases in inflammation-related proteins and a concomitant increase in gas and pH-related proteins. In contrast, VDZ treatment revealed virtually no proteomic profile changes pre- and post-treatment. Interestingly, both IFX and VDZ induced a positive response as observed by orthogonal clinical metrics. Given the minimal changes to VDZ patient proteomic profiles, serum’s utility as a monitoring strategy must be used with caution. More specifically, both therapeutic mechanism-of-action and expected response times need to be considered when designing disease monitoring or profiling studies using sera. In line with this, IFX’s anti-TNF-α mechanism would likely result in a systemic response, as TNF-α is not sequestered in any specific tissue or compartment. In contrast to IFX, VDZ is a highly targeted monoclonal that binds integrin α₄β₇, impeding lymphocyte infiltration into intestinal tissue and subsequent inflammation.⁵ Given this, it is reasonable to hypothesize VDZ treatment would not substantially affect a patient’s serum proteome profile (at least early into treatment regiments), a hypothesis suggested by these findings. These findings also suggest VDZ does not alter imbalances in systemic inflammatory conditions, an important clinical consideration for triaging patient treatment options. To address this and further uncover biologically relevant biomarkers of response, studies should consider pairing serum or plasma with additional biomaterial such as stool or biopsies when profiling treatment response. However, it is critical to note that this lack of response may also reflect a lack of VDZ response durability, and thus future work should also test if monitoring serum closer to treatment start is feasible. Alternatively, noted differences in treatment effectiveness onset between VDZ and other biologics have been previously noted, suggesting time points later in treatment cycles would be useful to assay.²⁷

Focusing on IFX treatment results, we revealed both responders and nonresponders display a robust decrease in inflammation-related proteins, along with an increase in gas transport. While the reduction in proteins that are directly or indirectly affected by TNF-α signaling would be expected, the biological rationale for the increased abundance of gas transport-related proteins (CA1,2 and hemoglobin subunits) is not as clear. Some evidence suggests decreased levels of CA1 in UC patients; however, to our knowledge evidence in CD is yet to be presented.²⁸ Additional evidence suggests peptides derived from CA1 are protective in vitro and in animal models of IBD.^29,30 Lastly, evidence shows TNF-α and CA1 are concomitantly decreased in aflatoxin models, suggesting shared pathways; however, their coregulation is less defined in the context of IBD.³¹ It is tempting to hypothesize this may be evidence of IFX-respondent wound healing; however, no significant differences between responders and nonresponders were present with respect to these proteins, suggesting they are treatment-specific but not indicative of response, and further research must be done establishing the biological rationale as well as whether the finding is exploitable for diagnostic or therapeutic purposes.

In contrast to the clear IFX response we observed, comparisons between post-treatment responders and nonresponders were dampened in overall significant identifications, the degree of that significance, and magnitude of changes. Despite this, we observed several key differences in nonresponders. For instance, despite decreases in many inflammation-related proteins in both responders and nonresponders, there was a clear trend towards greater decrease in responders compared with nonresponders. This result was also reflective of direct CRP measurements, where on average, nonresponders had greater levels detected, suggesting serum-based proteomics recapitulates direct measurements, with the added benefit of having a systems-level view on subjects. We also observed increased abundance of several proteins in responders. No strong gene ontology networks were generated from these proteins; however, COG analysis revealed a strong network composed of proteins associated with cell-cell junction involvement. Despite being detected in serum, this may reflect an environment in the process of healing. Unfortunately, no tissue-based orthogonal measurements were made (eg, immunohistochemistry, etc.), and thus these results remain to be confirmed. Additionally, while we attempted to obtain additional clinical biomarker measurements for the time points assayed in the current study, they were not available (no measurements taken at timepoints we assayed) to correlate to proteomic data.

Despite minimal differences, several features were predictive of response to IFX. While these algorithms do not purposefully enrich for biologically connected features, we observed clear enrichments present in the subset. First, these features were dominated by immunoglobulins. Interestingly, responders and nonresponders had similar abundances of immunoglobulin fragments prior to IFX treatment. However, the unique cluster of fragments increased in nonresponders suggest nonresponders harbor multiple contributors to a single inflammatory output, and thus, while a decrease in several inflammatory markers was observed, the additional pathologic contributor was never addressed. The second and more arguably discernable emergent pattern was that of proteins associated with lipid biology. Here, SAA and APOC4 were increased in responders, while APOA4 was increased in nonresponders. Exploratory analysis of their protein-protein interactions revealed several unique and potentially opposing networks. Tumor necrosis factor-α affects many metabolic pathways including lipid transport and synthesis.^32,33 Several studies have compared levels of various lipid metabolism and transport proteins in IBD patients to healthy controls and identified several abnormalities that pointed toward dysregulation in IBD.³⁴ A prior study also identified apolipoprotein alterations in IBD using proteomics; however, this study was significantly smaller, used less sensitive mass spectrometers, and identified fewer proteins.¹⁴ However, to our knowledge this is the first study to identify that different apolipoproteins may predict IFX response and be indicative of underlying biologically relevant conditions contributing to a lack of response. As such, future work should further confirm these initial findings in a larger cohort and further confirm the apolipoprotein classes and isoforms, as proteomic identifications can be subject to incorrect isoform calling.

While many of our findings comport with known IBD biology and therapeutic mechanism of action, there are several limitations to acknowledge. Firstly, given the very large sample size needed for most machine-learning tasks, independent validation and testing cohorts were not included, and therefore the generalizability of these results to the broader Crohn’s population is still an open question. Next, the lack of VDZ sera response signature could be interpreted in several ways. For instance, given that this study characterized only 2 time points 6 weeks apart, it may be the case VDZ response signatures were present at later time points after longer treatment periods and thus were simply not captured. Indeed, treatment response to VDZ was determined significantly after the time points assayed, leaving room for response signatures to arise closer to response determination. As such, a longer longitudinal profiling is warranted to observe (1) if VDZ signatures arise later in treatment, and (2) the durability of the captured profiles for IFX. Additionally, since signatures were captured from nondepleted sera, many low-abundance signals are likely masked by the abundance of proteins such as albumin and immunoglobulins. This is despite attempts to control the dynamic range using mass spectrometry-based developments in speed and resolution. Therefore, it is possible VDZ signatures may be more easily discernable when these high-abundant features are depleted and low-abundant features are enriched, a common technique used for mass spectrometry-based serum proteomics. Furthermore, while broad metadata was available, longitudinal endoscopic information on the inflammation state was not collected beyond the initial time point, limiting associative studies between proteomic and metadata features. As such, greater longitudinal tracking of serum and endoscopically derived information paired with serum collection would be of significant utility for discernment and validation of features closely tied to disease scoring. Lastly, as with all high-dimensional data, additional validation is warranted, and data presented herein would be complemented by metabolomic or cell-free profiling of transcripts or miRNA.

Despite these limitations, our results significantly enhance compendium of knowledge guiding research towards precision treatment options for IBD patients by revealing treatment-specific, group-level proteomic shifts. Further, we show a large portion of IFX treatment’s proteome shift is not uniquely indicative of treatment outcome. However, a subset of proteins did distinguish and classify responders and nonresponders, and these proteins point towards proteins whose role in IBD remains poorly characterized and warrants more targeted follow-up. This work can undoubtedly be used to guide more targeted mechanistic studies and inform clinical trials on biospecimen selection and profiling modality.

Supplementary Data

Supplementary data is available at Inflammatory Bowel Diseases online.

Funding Sources

D.J.G. and C.G.G are supported by National Institute of Health/National Institute of Diabetes and Digestive and Kidney Diseases 1R01DK131005 and SDDRC P30. C.G.G. is also funded by Institutional Research and Career Development Award (K12, K12GM068524). P.S.D. is supported by National Institute of Health/National Institute of Diabetes and Digestive and Kidney Diseases U planning grant—DK126626. B.V. is supported by the Clinical Research Fund (KOOR) at the University Hospitals Leuven and the Research Council at the KU Leuven.

Conflicts of Interest

G.D. has served as advisor for Abbvie, Agomab Therapeutics, Alimentiv, Applied Molecular Therapeutics, AstraZeneca, Bristol Meiers Squibb, Boehringer Ingelheim, Cytoki, Celltrion, Eli Lilly, Exeliom Biosciences; Ferring, Galapagos, Glaxo Smith Kline, Gossamerbio, Pfizer, Polpharms, Immunic, Johnson and Johnson, ProciseDx, Prometheus biosciences, Prometheus laboratories, Progenity, Protagonist, Seres, Takeda, Tillotts and Versant. P.S.D. is a current consultant for Abbvie, Takeda, Janssen, Pfizer, Addiso, Abivax, BMS, and Lilly. B.V. receives research support from AbbVie, Biora Therapeutics, Pfizer, Sossei Heptares, and Takeda, receives speaker’s fees from Abbvie, Biogen, Bristol Myers Squibb, Celltrion, Chiesi, Falk, Ferring, Galapagos, Janssen, MSD, Pfizer, R-Biopharm, Takeda, Truvion, and Viatris, and receives consultancy fees from Abbvie, Alimentiv, Applied Strategic, Atheneum, Biora Therapeutics, Bristol Myers Squibb, Galapagos, Guidepoint, Mylan, Inotrem, Ipsos, Janssen, Progenity, Sandoz, Sosei Hepatares, Takeda, Tillots Pharma and Viatris. Funding had no role in the study design, analysis, or results herein.

Data Availability

All RAW and MZML mass spectrometry files are available on massive.ucsd.edu (MSV000091105, doi:10.25345/C5NS0M73C, PW: TAILORIX_REV) along with accompanying results files and metadata necessary for future meta-analyses.