If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
This cross-cohort study aimed to (1) determine a network-based molecular signature that predicts the likelihood of inadequate response to the tumor necrosis factor-ɑ inhibitor (TNFi) therapy, infliximab, in ulcerative colitis (UC) patients, and (2) address biomarker irreproducibility across different cohort studies. Whole-transcriptome microarray data were derived from biopsies of affected colon tissue from 2 cohorts of infliximab-treated UC patients (training N = 24 and validation N = 22). Response was defined as endoscopic and histologic healing at 4-6 weeks and 8 weeks, respectively. From the training cohort, genes with RNA expression that significantly correlated with clinical response outcomes were mapped onto the Human Interactome network map of protein-protein interactions to identify a largest connected component (LCC) of proteins indicative of infliximab response status in UC. Expression levels of transcripts within the LCC were fed into a probabilistic neural network model to generate a classifier that predicts inadequate response to infliximab. A classifier predictive of inadequate response to infliximab was generated and tested in a cross-cohort, blinded fashion; the AUC was 0.83 and inadequate response was predicted with a 100% positive predictive value and 64% sensitivity. Genes separately identified from the 2 cohorts that correlated with response to infliximab appeared distinct but mapped onto the same network region of the Human Interactome, reflecting a common underlying biology of response among UC patients. Cross-cohort validation of a classifier predictive of infliximab response status in UC patients indicates that a molecular signature of non-response to TNFi therapies is present in patients’ baseline gene expression data. The goal is to develop a diagnostic test that predicts which patients will have an inadequate response to targeted therapies and define new targets and pathways for therapeutic development.
Selecting the right drug for each patient results in faster recovery, less pain, and improved quality of life, particularly in progressive diseases such ulcerative colitis (UC). However, response to targeted therapies is only 20–50%.
Translational Significance
We developed a classifier using machine learning and network medicine approaches and validated that it is predictive of infliximab response in a cross-cohort analysis. Network analyses suggest this molecular signature reflects a biology of infliximab response common among UC patients. Further development of such a test could decrease the time to treatment response and allow patients to return to their normal lives sooner.
INTRODUCTION
Ulcerative colitis (UC) is a chronic, relapsing disease characterized by diffuse mucosal inflammation of the colon.
UC is part of a larger spectrum of chronic relapsing diseases of the intestinal tract classified as inflammatory bowel disease (IBD), which also includes Crohn's disease (CD). IBD is a growing health problem, and the estimated prevalence is 568 cases per 100,000 persons in the US and 827 cases per 100,000 persons in Europe.
UC is diagnosed based on clinical presentation and endoscopic evidence of inflammation in the rectum that extends proximally into the colon. Clinical manifestations of active disease include bloody diarrhea, rectal urgency, abdominal pain, weight loss, and malaise.
which often requires use of a targeted therapy. Approved targeted therapies include anti-integrin ɑ4β7 (e.g. vedolizumab), anti-interleukin-12 of 23 (eg, ustekinumab), tumor necrosis factor inhibitor (TNFi; eg, adalimumab, infliximab and golimumab) and Janus kinase inhibitor (JAKi; eg, tofacitinib) therapies.
Selecting the right drug for individual patients from day 1 of treatment results in faster recovery, less pain, and improved quality of life, particularly in chronic progressive diseases such as UC. However, response to targeted therapies in UC are as low 20%–50%. Clinical response in UC clinical trials is defined as a decline in Mayo score of ≥3 points and either a ≥30% relative decrease from baseline with at least a 1 point decrease in rectal bleeding or a rectal bleeding score of 0 or 1.
This definition of response is less stringent than the treatment goals of clinical remission (Mayo score of 0-2) and mucosal healing (endoscopic score of 0-1). This highlights the urgent need to develop precision medicine tools for UC patients so that a therapy suitable to each patient's disease biology can be prescribed.
Analysis of the map of human disease biology called the Human Interactome can be used to interpret a patients’ unique molecular signature in order to identify which therapy will work for which patient based on each individual's unique biology.
Clinical Validation of a blood-based predictive test for stratification of response to tumor necrosis factor inhibitor therapies in rheumatoid arthritis patients.
Analysis of the topology and dynamics of the Human Interactome can reveal the underlying biological processes regulating many of the most common and difficult to treat diseases. This has resulted in the ability to discover novel targets, reprioritize known targets, and develop new biomarkers to predict drug response.
Clinical Validation of a blood-based predictive test for stratification of response to tumor necrosis factor inhibitor therapies in rheumatoid arthritis patients.
Application of this technology is especially useful in human complex diseases such as autoimmune conditions like rheumatoid arthritis and UC. The current goal is to develop diagnostic tests to predict which patients will have an inadequate response to targeted therapies and define new drug targets and pathways for novel therapeutic development.
Twenty-four patients with active UC, refractory to corticosteroids or immunosuppression, underwent colonoscopy with biopsies from diseased colon within a week prior to the first intravenous infusion of 5 mg infliximab per kg body weight. Response to infliximab was defined as endoscopic and histologic healing at 4-6 weeks after first infliximab treatment (8 patients as responders and 16 patients as inadequate responders). This data also included 6 healthy controls. Total RNA was isolated from colonic mucosal biopsies, labelled, and hybridized to Affymetrix Human Genome U133 Plus 2.0 Arrays.
Twenty-two patients underwent colonoscopy with biopsy before infliximab treatment. Response to infliximab was defined as endoscopic and histologic healing at week 8 (12 patients as responders and 11 patients as inadequate responders). For 1 patient, data from 2 samples taken at different timepoints were available. RNA was isolated from pre-infliximab biopsies, labeled and hybridized to Affymetrix Human Genome U133 Plus_2.0 Array.
This study was performed in accordance with the principles outlined in the Declaration of Helsinki.
Microarray analysis
The 2 datasets were downloaded using GEOquery R package. Before-treatment gene expression data were extracted by setting the visit time point to baseline. Probe IDs were converted to gene Entrez ID using the hgu133plus2.db database. The 2 datasets were merged by the common probe IDs. Batch effects were removed using ComBat from the sva R package.
Clinical Validation of a blood-based predictive test for stratification of response to tumor necrosis factor inhibitor therapies in rheumatoid arthritis patients.
contains experimentally determined physical interactions between proteins. These interactions include, regulatory, metabolic, signaling, and binary interactions. The Human Interactome amalgamates data from more than 300 thousand interactions among 18 thousand proteins.
Identification of classifier genes
For all genes, the Pearson correlation between the gene expression values and the response to treatment was determined.
The signal-to-noise ratio of each gene correlation was calculated by randomly shuffling the response outcome 100 times. Selected genes were then mapped onto the consolidated Human Interactome and the largest connected component (LCC) was determined.
Classifier design and validation
Genes identified as discriminatory between responders and inadequate responders to infliximab that were in the LCC were used as features of a probabilistic neural network.
The classifier training included in-cohort validation using a leave-one-sample-out cross-validation where the classifier was blind to the response outcome of that left-out patient. The classifier was trained using the default smoothing parameter (σ = 0.8). The classifier was validated on the validation cohort where the training cohort was used for feature selection and classifier training and the validation cohort was used for independent validation.
The classifier provided a probability for each patient reflecting whether or not that individual responded to infliximab treatment. The log-likelihood ratio of response and inadequate response probabilities were used to define a score for each patient and draw the receiver operating characteristic (ROC) curves by comparing the score to actual response outcomes. The area under the curve (AUC) determined the performance of the classifier. In cross-cohort assessment, the trained classifier was blind to the clinical outcomes of the validation cohort patients.
Response module randomization
Response module was comprised of the largest connected component formed by top genes when derived from both training and validation cohorts. None of the shared genes (STC1, PAPPA, SOD2 and HGF) between the 2 cohorts’ top gene sets was a high degree node in the Human Interactome, that could have caused a high degree of perceived connectedness between the LCC genes from the 2 cohorts. Hence, nodes were randomly assigned to both cohorts uniformly at random.
RESULTS
Identification of gene expression features predictive of inadequate response to infliximab
To identify genes whose expression best distinguished responders from inadequate responders to infliximab, 2 publicly available UC patient gene expression datasets were downloaded for which the clinical outcomes data were available.
(response prediction genes). Unlike conventional differential expression methods that look for large fold-changes in gene expression between 2 groups, this analysis investigated small but significant changes – a high signal-to-noise ratio – between the 2 groups (see Material and methods). Genes were ranked by decreasing value of signal-to-noise ratio and the top genes with the highest signal-to-noise ratio were selected as infliximab response discriminatory genes (Fig 1A).
Fig 1Identification of response discriminary genes. (A) Pearson correlation distribution of gene expression values with response outcomes in observed versus randomized gene expression data. The signal-to-noise ratio of actual and randomized pearson correlations were derived by dividing the randomized valued by the observed value. (B) Genes associated to top 123 probes with highest signal-to-noise ratio were mapped on the network resulting in observation of multiple connected components. Larger nodes correspond to genes with higher signal-to-noise ratio ranks, and node colors indicate expression change in responders with respect to inadequate responders. (C) Comparing the observed average shortest path between the genes associated to top 123 probes with the expected average shortest path generated by 100,000 randomizations. (D) Heatmap representing the baseline gene expression values of LCC genes used for classifier training across patients. Red corresponds to higher relative expression values and green corresponds to lower relative expression values.
Refinement of molecular signature genes using the Human Interactome
The top genes from the training cohort for which expression values across patients were significantly correlated to clinical outcome after infliximab treatment were selected and mapped onto the Human Interactome network map of protein-protein interactions (see Material and Methods) (Fig 1B-C). The cut-off of the top 123 probes was empirically determined as the minimum number of genes needed for the LCC size to plateau (Supplementary Figure S1). Although these genes were identified from gene expression data only, proteins encoded by these genes formed several clusters with the largest 1 containing 12 proteins on the Human Interactome. The associated proteins to the set of 123 probes on the Human Interactome was significantly closer to each other than expected by chance (z-score of -2.07) (Fig 1B-C and Table I). Absolute z-scores > 1.6 have been associated with sub-networks of underlying disease biology.
The probabilistic neural networks were trained using the LCC genes and patient data to teach the predictive classifier the appropriate patient outcome (ie, response or inadequate response to infliximab) for each input (ie, gene expression levels of LCC genes) (see Material and Methods).
Blinded, independent cross-cohort validation assessed the performance of the predictive classifier. In this analysis, the classifier that was trained on the known data and outcomes from the training cohort was used to predict the outcomes on the validation cohort, ultimately testing the ability of the classifier to accurately predict inadequate response to infliximab in an unseen patient population. The classifier predicted probabilities were converted to a continuous classifier prediction score using log-likelihood ratio (see Material and Methods). ROC curves, which plot the rate of false positives versus the rate of true positives, were used to assess cross-cohort performance (Fig 2A). An AUC of 0.83 was observed for the classifier predicting inadequate response to infliximab among validation cohort patients. The distribution of classifier prediction scores within the validation cohort showed a significant difference between the classifier prediction scores for responders and inadequate responders (P-value = 0.004) (Fig 2B). Additionally, the cross-cohort positive predictive value (PPV) and sensitivity were estimated (Fig 2C), which are metrics that describe the accuracy of the inadequate response predictions. At a 100% PPV, the classifier had a sensitivity of 64%.
Fig 2Cross-cohort performance of the response prediction classifier. (A) ROC curve. (B) Classifier predicted scores among validation set patients who are true clinical responders (R) or inadequate responders (IR) to infliximab. (C) Accuracy in predicting inadequate responders to infliximab in the validation cohort.
Baseline gene expression profiles of responders more closely resemble healthy controls
To further evaluate the 12-LCC classifier genes, the expression was compared between responders and inadequate responders. In general, the changes in gene expression between these 2 patient groups were small and no gene reached a threshold that indicated a significant differential expression (P-value ≤ 0.05 and enrichment of -log(0.05) ≥ 2.99) (Fig 3A). However, the expression of the 12-LCC classifier genes tended toward higher enrichment scores. One interpretation was that the genes with the greatest fold-change were not necessarily the most relevant to treatment response as they could potentially be downstream of master regulators or be altered by indirect/secondary consequences of disease-relevant signaling processes.
Fig 3Differential gene expression of response discriminatory genes. (A) Volcano plot indicating the differential expression for all genes between responders and inadequate responders to infliximab. The 12 genes (corresponding to 13 probes) in the classifier LCC are highlighted in orange. (B) Comparison of the expression of the 12 LCC classifier genes of responders (teal) or inadequate responders (red) vs. healthy controls. (C) Unsupervised hierarchical cluster analysis of 12 classifier LCC genes illustrating the relative RNA expression between healthy controls, responders and inadequate responders.
Next, the expression of the 12-LCC classifier genes was compared to that of healthy controls, comparing the fold changes relative to responders or to inadequate responders. The patients who were inadequate responders showed the largest divergence in gene expression pattern from that of healthy controls (Fig 3B). Unsupervised hierarchical clustering analysis showed that the baseline expression profiles of the patients who responded to infliximab more closely resembled the expression pattern of healthy controls than did the inadequate responders (Fig 3C).
The UC infliximab response module is a sub-network on the Human Interactome
The Human Interactome can serve as a blueprint to better understand the interconnectivity and underlying biology of the inadequate response prediction genes. Thus, the top 200 genes with the highest signal-to-noise ratio between responders and inadequate responders among the validation cohort data were also determined. When the 200 top genes from the training and validation cohorts were mapped simultaneously onto Human Interactome, the genes were not randomly scattered on the network, but instead significantly interacted with each other (z-score, absolute value of 7.68) forming a common response module LCC (Fig 4) that was significantly larger than the random expectation (139 genes; z-score of 2.09). To account for genes that were shared between the 2 cohort gene lists, including any high-degree node (UBC) on the Human Interactome, a careful randomization was made to estimate the significance of interconnectivity (see Material and Methods). Three proteins in the response module LCC (GK, FFAR2, CEBPB) are direct interaction partners of TNF-ɑ, the protein target of infliximab. Several proteins in the response module LCC were orphan genes that were not previously part of LCCs of the individual cohorts (eg, STC1 and IL7R) yet were integrated into this response module LCC (Fig 4A). These results shows that even though the response discriminatory genes identified from each cohort were apparently distinct with minimal overlap, their protein products tended to interact significantly on the network, reflecting the existence of an underlying disease biology sub-network, or response module, that defined a molecular signature of non-response to infliximab in UC patients.
Fig 4Apparently distinct gene lists mapped onto the same network region of the Human Interactome indicated a common underlying biology of response. (A) Response module: largest connected component formed by the proteins encoded by the response signature genes from the training and validation cohorts. Proteins encoded by training cohort genes are in blue and those encoded by validation cohort genes are in orange. (B) Distribution of LCC size from random expectation.
This study describes a predictive classifier developed using knowledge from the Human Interactome map of protein-protein interactions and a probabilistic neural network machine learning algorithm. The genes correlating with response to infliximab identified from baseline colon biopsy samples were predictive of inadequate response to infliximab in a cross-cohort validation. The patients in this study were all diagnosed with UC, and as such, differences in the biology between these individuals may not manifest in large fold-changes in gene expression. These subtle differences in transcript levels may be overlooked in conventional differential gene expression analyses in favor of genes with a greater observed fold-change between the responder and inadequate responder groups. However, this study identified small but significant changes in gene expression that may contribute to different treatment outcomes. The network-based method to discover biomarkers described in this study ensured that the differentially expressed genes in the classifier were significantly connected to the subnetwork of ulcerative colitis disease biology. This reduces the large number of differentially expressed genes to those most relevant to the biology of treatment response.
When response discriminatory genes identified separately from the training and validation cohorts were compared, the gene sets showed limited overlap in identity but significant overlap on the Human Interactome. Thus, they are unified in a common response module on the Human Interactome. This observation addresses 1 of the major concerns of biomarker irreproducibility; studies evaluating response prediction biomarkers rarely report the same genes. Many studies have reported prognostic indicators of response to TNFi therapies in UC.
A prospective study determined the predictive value of pre-treatment mucosal T cell-related cytokine gene expression profiles in response to infliximab;
expression of transcripts encoding IL-17A and IFN-γ were associated with remission after 3 infliximab infusions (odds ratio or 5.4, P = 0.013 and 5.5, P = 0.011, respectively). These studies developed predictive models using machine learning approaches, calculating mean gene expression values, evaluating the highest fold changes in gene expression and/or taking a pathway-based approach to describe UC disease biology. None of these studies have been developed into a clinical test for care of UC patients. By mapping the response module, network analyses performed in this study enabled identification of biomarkers associated with a specific disease phenotype (inadequate response to infliximab), reduced the noise inherent to gene expression data and eliminated many false positives that can arise from small sample sizes and characteristics specific to demographics of a particular patient cohort. Future analyses and larger cohort studies will explore the use of genes in the aggregated response module to train and validate a TNFi response classifier.
The proteins encoded by the classifier genes (Table I) are implicated in many biological processes including epithelial cell proliferation, response to reactive oxygen species, regulation of apoptotic signaling and cellular responses to lipid metabolism. Bioactive lipid mediators, including prostaglandins, regulate chronic inflammation through cell differentiation and activation, protect against acute epithelial barrier damage and facilitate tissue regeneration.
TNF stimulation induces COX-2 expression in innate immune cells, initiating proinflammatory responses by converting arachidonic acid into prostaglandins and inducing production of other cytokines and chemokines.
Finally, inclusion of genes encoding proteins involved in epithelial cell proliferation and cell apoptosis indicates a role for barrier function in determining TNFi inadequate response.
A permeable barrier is formed by intestinal epithelial cells that normally restricts access to potential antigens in the intestinal lumen and epithelial regeneration is a critical aspect of mucosal healing and regeneration.
Disrupted barrier function contributes to the pathogenesis of inflammatory bowel diseases and TNF inhibitor treatment can improve intestinal barrier function in IBD.
Consistent with the complexity of TNF-ɑ signaling, transcripts predictive of inadequate response to infliximab are similarly diverse.
There is an interaction between genetic, immune, and environmental factors that is evident in the mucosa gene expression profiles of IBD patients compared to healthy controls and in the genetic risk alleles associated with an increased risk of IBD.
The topological and biological properties of the infliximab response module on the Human Interactome suggests that it is possible to determine a molecular signature for inadequate response to TNFi therapies in patients with UC. TNFi therapies have demonstrated efficacy in the treatment of moderate to severe IBD. However, response rates vary, and initially 40%–60% of patients fail to achieve remission with their initial treatment,
Furthermore, a multicenter, retrospective study reported that 55.6% of UC patients who were primary non-responders to infliximab underwent colectomy within 3.2 years, a surgery that costs an estimated $91,767.
Given the need to rapidly manage disease flares and avoid surgery, there is a critical need for a test that can predict which UC patients will benefit from TNFi therapy and who should consider alternative treatment options.
This network-based approach evaluates protein interactions to select genes that reflect the biology of disease at the individual patient level. The cross-cohort validation of the predictive classifier developed using a response module found in the Human Interactome, suggests the existence of a molecular signature in baseline tissue samples that characterizes UC patients who will have an inadequate response to TNFi therapy. Further development of such a test would decrease the time to treatment response, thus allowing patients to get back to their normal, productive lives sooner while decreasing the burden on supportive family members. Furthermore, this method of biomarker discovery and classifier development can be applied across multiple disease areas with complex phenotypes and datasets containing molecular information. The platform described herein opens new, unprecedented opportunities to create new drug response modules, predict drug response in complex diseases, and achieve the goal of treating patients with the most effective treatment for their unique disease biology.
AUTHOR CONTRIBUTION
S.D.G designed the analysis. S.D.G and I.V and carried out the analysis. J.B.W carried out the biological interpretation of the classifier genes. S.D.G, J.B.W, A.S and V.R.A wrote the paper. M.S carried signal-to-noise ratio analysis. All authors have read the journal's authorship agreement and the manuscript has been reviewed by and approved by all named authors.
DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available from the corresponding author upon reasonable request.
ACKNOWLEDGMENTS
The authors would like to thank Lauren Whelton for review of the manuscript and Keith Johnson for providing feedback. All authors have read the journal's policy on disclosure of potential conflicts of interest. IV, JBW, AS, VRA and SDG are full-time employees and shareholders of Scipher Medicine Corporation. MS is a consultant of Scipher Medicine Corporation.
Clinical Validation of a blood-based predictive test for stratification of response to tumor necrosis factor inhibitor therapies in rheumatoid arthritis patients.
00049-4&id=gr1.jpg){kind=link}
00049-4&id=gr2.jpg){kind=link}
00049-4&id=gr3.jpg){kind=link}
00049-4&id=gr4.jpg){kind=link}