Application Of Proteomics And Deep Learning In The Diagnosis And Differentiation Of Inflammatory Bowel Disease

Background and aimInflammatory bowel disease(IBD)is a chronic nonspecific bowel disease,including ulcerative colitis(UC)and Crohn’s disease(CD).In the past decade,the incidence of IBD has increased globally.The diagnosis and differentiation of IBD mainly rely on clinical manifestations,imaging examinations,endoscopy and pathological examinations.However,these tests are limited in the identification of atypical CD and UC patients,and in the identification of IBD and other intestinal diseases,especially the distinction between UC and ischemic colitis(IC),and CD and intestinal tuberculosis.tuberculosis(ITB)and primary intestinal lymphoma(PIL).Currently,there are no clinically valid markers for their identification.Therefore,this study aims to provide new ideas for the diagnosis and differentiation of IBD through proteomics and deep learning.MethodsIn proteomics research,we firstly collected the intestinal mucosa of CD,UC,and normal control groups.Proteomics was used to screen differentially expressed proteins(DEPs).To further study the relationship between TNC and IBD,we performed enzyme-linked immunosorbent assays in CD,UC,and NC groups to detect the TNC concentration in serum.The relationship between TNC expression and clinical characteristics such as IBD disease activity,stage,and drug response was studied.The function of the protein was studied at the cellular levelSecondly,we collected serum from patients with UC,IC,and NC.After the serum samples were pretreated.Quantitative protein profiling analysis was performed.DEPs were screened between groups,and functional pathway clustering analysis was performed on DEPs.We collected sera from untreated patients with CD,ITB,and PIL.After the serum samples were pretreated.Quantitative protein profiling analysis was performed DEPs were screened between groups,and functional pathway clustering analysis was performed on DEPs.To further screen for candidate proteins,we calculated the AUC value of each DEP between groups,and selected proteins with AUC values≥0.95 as candidate proteinsIn the study of deep learning,patients diagnosed with IBD in our hospital were included,and pathological images of endoscopic biopsies of the colon were collected.The collected pathological images were learned and trained using multi-example learning methods.The IBD classifier and UC/CD classifier were used to learn and predict.The accuracy,sensitivity,specificity,positive predictive value(PPV),negative predictive value(negative)of the deep learning model to distinguish UC and CD predictive value(NPV)and AUC value were calculatedResultsIn intestinal mucosal proteomics studies,there were 102 and 17 DEPs between IBD and NC,CD and UC,respectively.Among DEPs between CD and UC,angiotensin-converting enzyme 1,angiotensin-converting enzyme 2,fatty acid binding protein 6,and fatty acid binding protein 2 were significantly up-regulated in the CD intestinal mucosa Ten DEPs between IBD patients and NC were involved in NAD metabolism and signaling We further verified the relationship between Tenascin-C(TNC)identified in intestinal mucosal proteomics and 3BD.We found that TNC expression was higher in UC and CD intestinal mucosa than in NC.Serum TNC levels were higher in patients with CD and UC than in NC,and there was no difference between UC and CD.Serum TNC levels were positively correlated with clinical activity of CD and UC,with r being 0.35 and 0.26,respectively.The high expression of intestinal mucosal TNC mRNA in patients with UC could predict no response to infliximab treatment,with an AUC of 0.898,a sensitivity of 87.5%,and a specificity of 87.5%.In THP-1 cells stimulated with recombinant TNC,the expression of IL6 was significantly higher than that of the LPS group and the control groupIn the serum proteomics study of UC and IC,there were 228 DEPs in the UC group compared with the NC group;27 DEPs in the IC group compared with the NC group;and 49 DEPs in the UC group compared with the IC group.Among the DEPs of the three groups,we found that only fibrin 3 was different in the pairwise comparisons among the three groups,with a UC/NC ratio of 0.24,a UC/IC ratio of 0.46 and an IC/NC ratio of 0.53.Cluster analysis of pathways between IC and UC showed that DEPs were mainly involved in the platelet degranulation pathway,the initial trigger pathway of complement,and the complement cascade regulatory pathwayIn the serum proteomics study of CD,ITB,and PIL,there were 108,105,and 55 DEPs in the CD/ITB,CD/PIL,and ITB/PIL groups,respectively.The number of DEPs with AUC>0.95 was 12,19,10 in the CD/ITB,CD/PIL,and ITB/PIL groups,respectively.In CD and ITB,DEPs were mainly involved in the antigen cross-presentation pathway,interferon α/β pathway,adaptive immune response,and ER-phagosome pathway.In CD and PIL,DEPs were mainly involved in the regulatory pathway of AU element-rich mRNA stability,AUF1 binding and mRNA destabilization pathway and NOTCH4 signaling pathway;In CD and PIL,DEPs were mainly involved in the proteasome degradation of GLI1 pathway and neutrophil degranulation pathwayIn deep learning studies,the pathological characteristics of UC and CD were compared.Crypt abscesses,crypt deformation or atrophy,goblet cell reduction,and glandular atrophy or reduction were more common in UC patients,while lymphoid hyperplasia and formation of non-caseinous granuloma was more common in patients with CD.For the distinction of UC and CD by pathological images,the deep learning models using U-net to extract crypts and not using U-net to extract crypts had an accuracy of 85%and 78,respectively;AUC was 0.87 and 0.85,respectively.Among them,sensitivity,specificity,PPV,NPV of using U-net to differentiate CD from UC was 80.8%,92.9%,95.5%,72.2%,respectively.Sensitivity,specificity,PPV,NPV of not using U-net to differentiate CD from UC 72.4%,90.9%,95.5%,55.5%.ConclusionBy applying proteomics,we identified differential proteins between IBD and NC,CD,and UC intestinal mucosa;The expression of proteins involved in NAD metabolism and signaling pathways was significantly changed in IBD compared with NC.TNC identified by proteomics can also be used as biomarkers of IBD diagnosis,disease activity,and drug efficacy,and may participate in the development of IBDSerum proteomics studies have also identified serum proteins for the differentiation of UC and IC,differentiation of CD,ITB,and PIL.These proteins may serve as serum biomarkers for their differential diagnosisBy applying deep learning methods,we found that deep learning can discriminate UC and CD based on pathological images.Deep learning could identify the features of crypts and had a higher accuracy in the differential diagnosis of CD and UC.Therefore,deep learning can be used as a new method for CD and UC differentiation.