Molecular classification of prostate adenocarcinoma by the integrated somatic mutation profiles and molecular network

www.nature.com/scientificreports OPEN Received: 16 December 2016 Accepted: 20 March 2017 Published: xx xx xxxx Molecular classification of prostate adenocarcinoma by the integrated somatic mutation profiles and molecular network Lei Yang1, Shiyuan Wang1, Meng Zhou1, Xiaowen Chen1, Wei Jiang1, Yongchun Zuo2 & Yingli Lv1 Prostate cancer is one of the most common cancers in men and a leading cause of cancer death worldwide, displaying a broad range of heterogeneity in terms of clinical and molecular behavior. Increasing evidence suggests that classifying prostate cancers into distinct molecular subtypes is critical to exploring the potential molecular variation underlying this heterogeneity and to better treat this cancer. In this study, the somatic mutation profiles of prostate cancer were downloaded from the TCGA database and used as the source nodes of the random walk with restart algorithm (RWRA) for generating smoothed mutation profiles in the STRING network. The smoothed mutation profiles were selected as the input matrix of the Graph-regularized Nonnegative Matrix Factorization (GNMF) for classifying patients into distinct molecular subtypes. The results were associated with most of the clinical and pathological outcomes. In addition, some bioinformatics analyses were performed for the robust subtyping, and good results were obtained. These results indicated that prostate cancers can be usefully classified according to their mutation profiles, and we hope that these subtypes will help improve the treatment stratification of this cancer in the future. Prostate cancer is the most non-cutaneous common cancer in males and one of the leading causes of cancer-related deaths worldwide. The incidence and mortality of prostate cancer exhibit a remarkable variety in different parts of the world, and they are highest in the western world1. It is estimated that 220,800 men were diagnosed with prostate cancer and that 27,540 will die of the disease in 2015 in the United States2. Several demographic, clinical and genetic factors, including age, race, family history, genetic susceptibility, and prostate-specific antigen (PSA) level, have contributed to the high incidence of prostate tumors3. Despite the high incidence of these carcinomas, prostate cancer is often an indolent cancer. Many patients who have indolent prostate cancer will remain asymptomatic for many years after diagnosis, and many others can even live for more than ten years with organ-confined disease4. With the emergence and application of new genomic technologies, such as next-generation sequencing and microarray analyses, more molecular and genetic profiles of prostate adenocarcinomas have been generated in recent years. Based on these profiles, we found that prostate adenocarcinomas exhibit a remarkable biological heterogeneity, including alterations of somatic copy number, point mutations, and structural rearrangements, and these genetic heterogeneities may underlie the high variability of clinical outcomes in prostate adenocarcinomas5–10. Given the tremendous biological heterogeneity of prostate tumors, it is critical to determine the appropriate treatment for patients diagnosed with prostate adenocarcinoma. Therefore, understanding the biological heterogeneity of prostate adenocarcinomas is one of the fundamental goals of cancer informatics, and some studies have shown that classification of prostate cancers into clinically and biologically meaningful subtypes can provide more precise outcome predictions, additional information on the selection of optimal therapies, and a better understanding of the heterogeneity1, 11–15. 1 College of Bioinformatics Science and Technology, Harbin Medical University, Harbin, 150081, China. 2The Key Laboratory of Mammalian Reproductive Biology and Biotechnology of the Ministry of Education, Inner Mongolia University, Hohhot, 010021, China. Lei Yang and Shiyuan Wang contributed equally to this work. Correspondence and requests for materials should be addressed to L.Y. (email: ) or Y.Z. (email: yczuo@imu. edu.cn) or Y.L. (email: ) Scientific Reports | 7: 738 | DOI:10.1038/s41598-017-00872-8 1 www.nature.com/scientificreports/ Because the classification of cancers into clinically meaningful subtypes provides insights into the biological properties responsible for tumor progression and guides treatment and prognosis more precisely, more molecular profiles are being used to subtype all types of cancers. Currently, large-scale genomics projects, including The Cancer Genome Atlas (TCGA) Research Network, are producing molecular profiles for thousands of malignancies, rendering the molecular subtyping of distinct malignancies possible. In the past few years, gene expression data have been used to stratify different molecular subtypes of malignancies by several recent studies11, 12, 16, 17. Based on gene expression profiles of 26000 genes, Lapointe et al. first distinguished prostate cancers from normal samples, and further identified three subtypes of prostate cancers by using unsupervised hierarchical clustering. They also found two genes can be used as surrogate markers for tumor subtypes for predicting tumor recurrence16. Markert et al., analyzed a microarray dataset of 281 prostate cancers, and five distinct molecular subtypes were identified by unsupervised clustering. They found that the first subtype was characterized by poor survival outcome, the second subtype was characterized by intermediate survival outcome, and three subtypes were characterized by benign outcome. They also validated their stratification on an independent dataset of 150 tumor samples17. In the work of Tomlins et al., they analyzed the gene expression profiles of prostate cancer for 1577 patients. Three distinct molecular subtyping, including m-ERG+ subtype, m-ETS+ subtype, and m-SPINK1+ subtype were identified in their study, and these molecular subtypes of prostate cancer were supported by transcriptomic and clinical analysis12. In addition, genomics data from multiple assay platforms, including mRNAseq, miRNA-seq, and DNA methylation data, have been integrated by TCGA to stratify more than ten distinct malignancies, and the stratification results have shown that each cancer type can be divided into three or four molecular subtypes1, 18–27. However, the somatic mutation profiles were seldom used by those studies in the area of tumor subtyping because those data are extremely sparse and rarely shared across patients; thus, they could not be easily used like other molecular profiles28–32. Somatic mutations often disrupt the function of mutated genes, providing insights into the mechanisms of tumorigenesis and tumor progression; therefore, stratification tumors with somatic mutation profiles may provide more effective clinical guidance33. Indeed, some prior attempts integrated somatic mutation profiles and molecular networks to stratify various distinct malignancies into meaningful molecular subtype (...truncated)