Supervised machine learning algorithm identified KRT20, BATF and TP63 as biologically relevant biomarkers for bladder biopsy specimens from interstitial cystitis/bladder pain syndrome patients

Abstract

Objectives This study was carried out to identify biomarkers that distinguish Hunner‐type interstitial cystitis from non‐Hunner‐type interstitial cystitis patients. Methods Total ribonucleic acid was purified from 212 punch biopsy specimens of 89 individuals who were diagnosed as interstitial cystitis/bladder pain syndrome. To examine the expression profile of patients’ bladder specimens, 68 urothelial master transcription factors and nine known markers (E‐cadherin, cytokeratins, uroplakins and sonic hedgehog) were selected. To classify the biopsy samples, principal component analysis was carried out. A decision tree algorithm was adopted to identify critical determinants, in which 102 and 116 bladder specimens were used for learning and validation, respectively. Results Principal component analysis segregated tissues from Hunner‐type and non‐Hunner‐type interstitial cystitis specimens in principal component axes 2 and 4. Principal components 2 and 4 contained urothelial stem/progenitor transcription factors and cytokeratins, respectively. A decision tree identified KRT20, BATF and TP63 to classify non‐Hunner‐type and Hunner‐type interstitial cystitis specimens. KRT20 was lower in tissues from Hunner‐type compared with non‐Hunner‐type interstitial cystitis specimens (P < 0.001). TP63 was lower in Hunner’s lesions compared with adjacent mucosa from Hunner‐type interstitial cystitis patients (P < 0.001). Blinded validation using additional biopsy specimens verified that the decision tree showed fairly precise concordance with cystoscopic diagnosis. Conclusion KRT20, BATF and TP63 were identified as biologically relevant biomarkers to classify tissues from interstitial cystitis/bladder pain syndrome specimens. The biologically explainable determinants could contribute to defining the elusive interstitial cystitis/bladder pain syndrome pathogenesis.

Editorial Comment to Supervised machine learning algorithm identified KRT20, BATF and TP63 as biologically relevant biomarkers for bladder biopsy specimens from interstitial cystitis/bladder pain syndrome patients

“…1 The finding of reduced KRT20 in NHIC vs HIC bladder mucosa in this study was in line with that reported previously by Kuo's group, whereas they demonstrated that the expression levels of P63 proteins encoded by the TP63 gene was reduced in HL-adjacent mucosa compared with NHIC or control healthy mucosa (ref #7 of this paper), which was not detected in the gene expression levels in this study. 1 However, the results of these biomarker studies are totally dependent on how accurately the cystoscopic diagnosis was made to identify HL in the first place. As discussed in the supplementary discussion of this study, the machine learning algorithm cannot replace the cystoscopy diagnosis of HL.…”

“…However, the results of these biomarker studies are totally dependent on how accurately the cystoscopic diagnosis was made to identify HL in the first place. As discussed in the supplementary discussion of this study, the machine learning algorithm cannot replace the cystoscopy diagnosis of HL 1 . So far, useful biomarkers for the IC/BPS subtype classification that exceed the cystoscopic diagnosis have not been identified.…”

“…The major interesting finding of this study is that the use of supervised machine learning algorithm helps identify biologically relevant markers such as KRT20, BATF, and TP63 in bladder biopsy specimens, as well as the best predictable combinations of those in gene expression profiles that can differentiate between non‐Hunner‐type interstitial cystitis (NHIC) and Hunner‐type interstitial cystitis (HIC) and between the “adjacent mucosa” and Hunner lesions (HL) of HIC among interstitial cystitis/bladder pain syndrome (IC/BPS) patients 1 . The finding of reduced KRT20 in NHIC vs HIC bladder mucosa in this study was in line with that reported previously by Kuo’s group, whereas they demonstrated that the expression levels of P63 proteins encoded by the TP63 gene was reduced in HL‐adjacent mucosa compared with NHIC or control healthy mucosa (ref #7 of this paper), which was not detected in the gene expression levels in this study 1 …”