Primary Research On Tumor Prognostic Biomarkers Based On Artificial Intelligence

Background In recent years,with the progress and development of sequencing technology and artificial intelligence algorithms(including machine learning and deep learning),more and more omics data have been published.However,due to the lack of scientific and effective integration methods,there are currently few molecular markers and corresponding systems that can be used in clinical practice.Bladder cancer(BC)and renal cell carcinoma(RCC)are the most common malignant tumors in the urogenital system.With the aging of the population and the improvement of people’s living standards,the incidence of BC and RCC has increased.About 70% of BC patients will experience recurrence after transurethral resection of bladder tumor(TURBT).Once BC recurs,the survival rate of patients will become very slim.Meanwhile,about30%-50% of kidney cancers lack early clinical manifestations,and nearly two-thirds of cc RCCs are discovered incidentally.Nearly half of cc RCCs developed metastases during follow-up even after radical surgery.Therefore,the development of molecular markers that can be used for patient risk stratification and prognosis prediction is of great significance for the clinical prognosis of patients.Herein,the present study will use BC and cc RCC as examples to screen and validate the corresponding patient prognosis biomarkers based on multi-omics data of patients(including gene expression profile data,clinical phenotype data,pathological image data,DNA methylation data,etc.)Methods This study is divided into three parts.In the first part,based on R / shiny,we integrate the gene expression data and corresponding clinical data of tumor patients in the existing public database or personalized by users,and use cutting-edge machine learning methods to build an integrated analysis platform to screen,validate,annotate,and translate cancer survival biomarkers.In the second part,based on the whole slide image(WSI)of the TCGA-BLCA cohort,we used artificial intelligence algorithm stacked predictive sparse decomposition(SPSD)to extract cellular morphological biomarkers(CMBs)from the WSIs of BC patients.Based on CMB,we used consensus clustering to divide BC patients into several groups.Furthermore,we combined the clinical phenotype and gene expression data of BC patients to explore the potential biological mechanism and clinical significance of CMB and CMS.Finally,we constructed a nomogram to predict the overall 3-and 5-year overall survival(OS)of patients with BC.In the third part,we integrate the DNA methylation profile of patients with cc RCC and the clinical data of corresponding patients,and try to use artificial intelligence to develop and validate DNA methylation related biomarkers.Results In the first part,we developed the cancer biomarker online platform CBio Explorer(cancer biomarker explorer),which is a user friendly online platform and standalone application that includes five main modules(dimensionality reduction,benchmark experiment,prediction model,clinical annotation and biological annotation).It integrates six cutting-edge machine learning models based on survival analysis for the screening and validation of cancer survival related biomarkers from the molecular level to clinical practice.In addition,CBio Explorer integrates a novel R package’Curated Cancer Prognosis Data’ to review,curate and integrate the gene expression data and corresponding clinical data of 47,210 clinical samples from 268 gene expression studies of 43 common blood and solid tumors.CBio Explorer can be visited at https://cbioexplorer.znhospital.cn:8888/CBio Explorer/.At the same time,we take BC as an example to evaluate the performance of the five major analysis modules of CBio Explorer.In the second part,we used machine learning algorithm SPSD and Cox PH to screen 27 pathological markers(CMBs)at the level of WSI.Meanwhile,we divided the TCGA-BLCA cohort of BC patients into 2 subtypes based on the consensus clustering of the first 30 CMB with the largest variation.Functional enrichment analysis suggested that CMB based BC subtypes might be involved in BC cells through metabolic pathways,thereby affecting the survival of patients with BC.In the third part,we identified a total of 2,628 differential methylation sites between normal renal tissue and cc RCC.Furthermore,we further screened 11 methylation sites closely related to the overall survival of patients,and constructed and validated the corresponding DNA methylation panel to predict the survival of patients with cc RCC.In clinical settings,we constructed a nomogram containing DNA methylation panel and other clinical characteristics such as patient age,gender and pathological stage,and confirmed the reliability of its clinical use through internal and external validation.Finally,we confirmed that DNA methylation panel was superior to other existing predictive molecular markers.Conclusions In the present study,we take BC and cc RCC as examples to develop and validate tumor prognostic markers at multi-omics levels(gene expression,DNA methylation,and whole slide images).At the level of gene expression,we developed an integrated analysis platform of cancer biomarkers CBio Explorer,and integrated a novel R package ’Curated Cancer Prognosis Data’ to review,curate and integrate the gene expression data and corresponding clinical data of 47,210 clinical samples from 268 gene expression studies of 43 common blood and solid tumors.Taking bladder cancer as an example,we evaluated and validated the performance of the major analysis modules of CBio Explorer.At the level of WSI,we extracted CMBs from the WSIs of BC patients using the SPSD algorithm,and defined the corresponding BC prognosis subtype CMS(cellular morphological subtype).In addition,we further confirmed the clinical prognostic significance and corresponding biological mechanism of CMB and CMS for BC patients.At the DNA methylation level,we developed and validated a panel based on DNA methylation,which was proved to be a prognostic biomarker for patients with cc RCC.