(P 399) SIMILARITY NETWORK FUSION IDENTIFIES DNA METHYLATION SUBTYPES OF LEIOMYOSARCOMA ASSOCIATED WITH CLINICAL OUTCOME

Objective: Our previous studies, confirmed by other groups, have identified 3 transcriptomic subtypes of leiomyosarcoma (LMS). These subtypes are associated with tumor location and differentiation, but their clinical utility remains limited. Also, our previous computational approach failed to confidently assign roughly one-third of LMS into these transcriptomic subtypes. To address these shortcomings, we sought to incorporate DNA methylation profiles of LMS into subtype discovery. We also employed new computational methods to identify the optimal number of features and clusters using an expanded multi-omic dataset. Our team has previously contributed to development of the Similarity Network Fusion (SNF), a new computational method for data integration and construction of patient similarity networks (PMID: 24464287). This method has previously identified molecular subtypes of multiple types of cancer that were associated with clinical outcome. Here, we employed SNF for LMS subtype discovery.

Methods: Whole transcriptome 3’ end sequencing data was previously generated from 99 archival LMS specimens (PMID: 25896974; “Stanford dataset”). Here, we generated new DNA methylation data from a subset of 61 of these specimens that passed quality control filters using EPIC microarrays (Illumina). Transcriptomic and epigenomic data from 79 LMS specimens from the TCGA cohort was used for validation of the clusters identified in the Stanford dataset and for survival analysis. EPIC array data was processed using custom bioinformatics pipeline in R programming language. Next, we developed a custom meta-clustering computational pipeline to perform SNF, Consensus Spectral Clustering and a comprehensive assessment of clustering quality and the optimal number of clusters. This analysis was performed for transcriptomic and epigenomic data separately, and for a combination of thereof. We analyzed the association between cluster assignment and the histopathological features and clinical outcome.

Results: SNF confirmed the presence of 3 transcriptomic subtypes of LMS based on expression levels of 400 most variable genes. As in the previous studies, these transcriptomic subtypes were associated with tumor location, but not with clinical outcome. However, SNF provided an improved transcriptomic subtype classification as it allowed for a reliable cluster assignment for 96/99 cases (97% of specimens having a positive silhouette value), a marked improvement over prior analyses.
Based on epigenomic data alone and on the combination of transcriptomic and epigenomic data, SNF identified 2 distinct clusters of LMS. The SNF epigenomic clusters were identified based on 200 most variable regions in the Stanford dataset, and also allowed for a highly robust classification of LMS (59/61, 97%, of cases with positive silhouette value). The same set of epigenomic features separated TCGA cases also into 2 SNF clusters. These epigenomic clusters were significantly associated with tumor location (p = 0.0001), grade (p = 0.0001), mitotic index (p = 0.004), as well as recurrence-free survival (p = 0.002) and overall survival (p = 0.003) in the TCGA dataset. On the other hand, the SNF clusters identified based on the combination of transcriptomic and epigenomic data were not associated with any clinicopathological features. This analysis demonstrates that DNA methylation data alone is the most useful for identifying LMS subtypes that are associated with clinical outcome.

Conclusion: We report for the first time the discovery of epigenomic subtypes of LMS and provide a clinical validation of these subtypes in an independent dataset. We demonstrate that SNF algorithm is superior for a robust and reliable identification of LMS subtypes using transcriptomic and epigenomic data. This work warrants further studies on the implementation of molecular subtyping of LMS into clinical practice, as the epigenomic features for subtype classification could be potentially detected in circulating tumor DNA through liquid biopsy testing.