Prediction of anti-TNF therapy failure in ulcerative colitis patients by ensemble machine learning: A prospective study

Abstract

Nowadays, anti-TNF therapy remarkably improves the medical management of ulcerative colitis (UC), but approximately 40 % of patients do not respond to this treatment. In this study, we used 79 anti-TNF-naive patients with moderate-to-severe UC from four cohorts to discover alternative therapeutic targets and develop a personalized medicine approach that can diagnose UC non-responders (UCN) prior to receiving anti-TNF therapy. To this end, two microarray data series were integrated to create a discovery cohort with 35 UC samples. A comprehensive gene expression and functional analysis was performed and identified 313 significantly altered genes, among which IL6 and INHBA were highlighted as overexpressed genes in the baseline mucosal biopsies of UCN, whose cooperation may lead to a decrease in the Tregs population. Besides, screening the abundances of immune cell subpopulations showed neutrophils' accumulation increasing the inflammation. Furthermore, the correlation of KRAS signaling activation with unresponsiveness to anti-TNF mAb was observed using network analysis. Using 50x repeated 10-fold cross-validation LASSO feature selection and a stack ensemble machine learning algorithm, a five-mRNA prognostic panel including IL13RA2, HCAR3, CSF3, INHBA, and MMP1 was introduced that could predict the response of UC patients to anti-TNF antibodies with an average accuracy of 95.3 %. The predictive capacity of the introduced biomarker panel was also validated in two independent cohorts (44 UC patients). Moreover, we presented a distinct immune cell landscape and gene signature for UCN to anti-TNF drugs and further studies should be considered to make this predictive biomarker panel and therapeutic targets applicable in the clinical setting.

Keywords: Anti-TNF therapy; Biomarker; Ensemble machine learning; Inflammation; Inflammatory bowel disease; Ulcerative colitis.

Fig. 1

The workflow. The pretreatment biopsies of 35 ulcerative colitis patients were analyzed in a discovery cohort to identify alternative therapeutic targets and predictive biomarkers of anti-TNF response using feature selection and machine learning algorithms. In addition, an independent in silico cohort with 22 UC samples was used to test the predictive performance of multi-mRNA gene sets. Finally, the introduced anti-TNF response predictive panel was experimentally validated on a real-life cohort. UCR: Ulcerative Colitis Responders and UCN: Ulcerative Colitis Non-responders (to anti-TNF therapy).

Fig. 2

Immune cell subpopulation analysis in UC patients prior to anti-TNF therapy. (A) The immune cell landscape of UC patients prior to commencing anti-TNF therapy was identified using the CIBERSORTx database. (B) Neutrophil accumulation and (C) T cell regulatory reduction were significantly observed in UCN patients. UCR: Ulcerative Colitis Responders and UCN: Ulcerative Colitis Non-responders (to anti-TNF therapy).

Fig. 3

Distinct gene expression profiles between UCN and UCR patients. (A) The most significant differentially expressed genes in UCN and UCR patients (*, **, and *** refer to the adjusted p-value <0.05, <0.01, and <0.001, respectively). (B) Pearson correlation analysis, along with the hierarchical clustering of the 313 differentially expressed genes, revealed a remarkable correlation among intragroup samples. UCR: Ulcerative Colitis Responders and UCN: Ulcerative Colitis Non-responders (to anti-TNF therapy).

Fig. 4

Functional enrichment analysis. (A) Enriched gene ontologies and (B) pathways in biopsies of anti-TNF non-responders using GOR and MSigDB databases, respectively. (C) Microarray and (D) qRT-PCR analysis showed the overexpression of IL6 in UCN patients. The * refers to adjusted p-value <0.01, and p-value <0.05. UCR: Ulcerative Colitis Responders and UCN: Ulcerative Colitis Non-responders (to anti-TNF therapy), MEL: Microarray Expression Level.

Fig. 5

Systematic network construction and analysis. (A) The network of crucial genes was constructed using CytoHubba and CentiScape2.2 plugins, containing 51 genes that were communicated through 258 interactions. (B) and (C) represent the KRAS expression among UCR and UCN patients in both discovery and real-life validation cohorts, respectively. No significant differences were identified. UCR: Ulcerative Colitis Responders and UCN: Ulcerative Colitis Non-responders (to anti-TNF therapy), MEL: Microarray Expression Level.

Fig. 6

Predictive potentiality of three-, four-, and five-mRNA panels in microarray datasets based on accuracy and ROC curve using the ensemble machine learning. (A) The accuracy for every single machine and the ensemble machine for three-, four-, and five-featured panels showed improved performance for the ensemble machine learning approach compared to single machines in the discovery cohort. The ROC curve and AUC of the three (blue line), four- (green line), and five-featured (red line) panels in the (B) discovery and (C)in silico impendent test cohorts. The AUC was shown in the plot. ROC: Receiver Operating Characteristic curve. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Fig. 7

Real-life cohort evaluation of predictive biomarkers and the ensemble machine learning method. (A) Experimental validation of the predictive biomarkers in the real-life cohort. The expression levels of biomarker candidates were measured in UCN and UCR patients of Taleghani hospital by RT-qPCR. *, **, and *** refer to the p < 0.05, <0.01, and <0.001, respectively. (B) The accuracy for the single machines and the ensemble machine for three-, four-, and five-featured panels showed improved performance for the ensemble machine learning approach compared to single machines in the real-life cohort. (C) The ROC curve for the three- (blue line), four- (green line), and five-mRNA (red line) gene sets for the discrimination of UCN from UCR patients in the real-life validation. The AUC was shown in the plot. ROC: Receiver Operating Characteristic curve. UCR: Ulcerative Colitis Responders and UCN: Ulcerative Colitis Non-responders (to anti-TNF therapy). (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)