Comprehensive analysis of lncRNAs reveals candidate prognostic biomarkers in multiple cancer types

Abstract

12Long non-coding RNAs (lncRNAs) are increasingly recognized as functional units in can-13 cer pathways and powerful molecular biomarkers, however most lncRNAs remain un-14 characterized. Here we performed a systematic discovery of prognostic lncRNAs in 9,326 15 patient tumors of 29 types using a proportional-hazards elastic net machine-learning 16 framework. lncRNAs showed highly tissue-specific transcript abundance patterns. We 17 identified 179 prognostic lncRNAs whose abundance correlated with patient risk and im-18 proved the performance of common clinical variables and molecular tumor subtypes. 19 Pathway analysis revealed a large diversity of the high-risk tumors stratified by lncRNAs 20and suggested their functional associations. In lower-grade gliomas, discrete activation 21 of HOXA10-AS indicated poor patient prognosis, neurodevelopmental pathway activation 22 and a transcriptomic similarity to glioblastomas. HOXA10-AS knockdown in patient-de-23 rived glioblastoma cells caused decreased cell proliferation and deregulation of glioma 24 driver genes and proliferation pathways. Our study underlines the pan-cancer potential 25 of the non-coding transcriptome for developing molecular biomarkers and innovative 26 therapeutic strategies. 27 28 cancer patient prognosis with transcript abundance of protein-coding genes and their genetic 61 and epigenetic alterations [22][23][24][25]. However, associations of lncRNAs with cancer patient sur-62 vival and biological function remain largely unexplored. A recent study characterized recurrent 63 hypomethylation patterns affecting a thousand lncRNAs in the TCGA PanCanAtlas cohort and 64 identified the EPIC1 lncRNA as a marker of poor prognosis in a subset of breast cancers [26]. 65Another TCGA study associated mutations and transcript abundance profiles of lncRNAs with 66 regulatory networks and molecular pathways and nominated candidate oncogenic and tumor 67 suppressive lncRNAs, some of which were functionally validated in cancer cell lines [27]. Analy-68 sis of cell-cycle correlated lncRNAs revealed a subset of S-phase enriched lncRNAs whose 69 transcript abundance profiles correlated with patient survival in multiple TCGA cohorts [28]. 70However, those studies did not analyze robust prognostic performance of lncRNAs using ma-71 chine-learning and cross-validation approaches, indicating further potential to systematically dis-72 cover lncRNAs as candidate prognostic biomarkers of multiple cancer types. 73Here we evaluated the transcript abundance profiles of nearly 6,000 lncRNAs as prognostic bi-74 omarkers in human cancers. Using a comprehensive machine-learning analysis, we compiled a 75 robust catalogue of prognostic lncRNAs across nearly 10,000 tumors of 29 types from the 76 TCGA PanCanAtlas project [22, 29]. The majority of our candidate lncRNAs showed improved 77 prognostic potential compared to standard clinical features and molecular tumor subtypes. We 78 associated prognostic lncRNAs with large-scale deregulation of hallmark cancer pathways, re-79 vealing e...