Detecting molecular subtypes from multi-omics datasets using SUMO

“SUMO analysis of TCGA-LGG identifies a cluster of individuals with poor prognosis The 2016 World Health Organization (WHO) classification of diffuse gliomas recommends use of IDHmutation status to separate IDH mutant astrocytoma and oligodendroglioma from IDH wild type astrocytoma, which has a worse prognosis (Louis e…”

“If batch information is known, we use ComBat (Johnson et al., 2007) to adjust for batch effects in this step.”

“We also downloaded two breast cancer datasets from METABRIC.”

“The NMF technique aims to explain the observed data using a small number of basis components by factoring the data into the product of two non-negative matrices; one representing the basis components, and the other containing mixture coefficients (Paatero and Tapper, 1994).”

“We decided to apply SUMO to subtype the LGGs as a case study with the intent to evaluate the robustness and relevance of known and proposed glioma subtypes.”

“To investigate this, we downloaded the xCell scores corresponding to the enrichment of 64 different immune and stromal cell types in these TCGA samples (Aran et al., 2017).”

“Based on the recommendations by Liu et al., 2018, we used progression-free survival for DLBC, TCGT, and THYM.”

“Several studies have associated IDH somatic mutations with a more favorable course of the disease and have identified multiple subtypes with a poor clinical course (Eckel-Passow et al., 2015; Ceccarelli et al., 2016).”

“In the case of methylation data, we use of log2 ratio of the methylated to unmethylated count, also referred to as M-values (Du et al., 2010).”

“…extension of iClusterPlus v.1.26, LRAcluster v.1.0 (Wu et al., 2015), MCCA from PMA package v.1.2.1 (Witten and Tibshirani, 2009), NEMO v.0.1 (Rappoport and Shamir, 2019b), PINSPlus v.2.0.5 (Nguyen et al., 2017), SNF v.2.3 (Wang et al., 2014), and CIMLR v.1.0 (Ramazzotti et al., 2018).”

“Several studies have associated IDH somatic mutations with a more favorable course of the disease and have identified multiple subtypes with a poor clinical course (Eckel-Passow et al., 2015; Ceccarelli et al., 2016).”

“Recent recommendations from the cIMPACT-NOW consortium suggest that EGFR amplification and combined chr7 gain and chr10 loss as well as TERT promoter mutation can be used to diagnose IDH wild type (WT) grade II/III gliomas that are likely to follow a more aggressive clinical disease course (Brat et al., 2018).”

“More sophisticated methods (e.g., iCluster [Shen et al., 2009], iClusterPlus [Mo et al., 2013], and Bayesian consensus clustering [Lock and Dunson, 2013]) model the probabilistic distribution of each data type and infer subtypes by maximization of the likelihood of the observed data.”

“Both initialization and convergence speeds are important factors to consider when formulating the appropriate factorization algorithms (Boutsidis and Gallopoulos, 2008).”

“We used SUMO to integrate the processed level 3 gene expression, DNA methylation, and miRNA expression data for the TCGA-LGG cohort downloaded from the UCSC Xena platform (Goldman et al., 2019) (and preprocessed as described in the Data preprocessing part of the STAR Methods).”

“We use a recent benchmark (Rappoport and Shamir, 2018) and datasets from TCGA and METABRIC (Curtis et al.”

“More sophisticated methods (e.g., iCluster [Shen et al., 2009], iClusterPlus [Mo et al., 2013], and Bayesian consensus clustering [Lock and Dunson, 2013]) model the probabilistic distribution of each data type and infer subtypes by maximization of the likelihood of the observed data.”

“Subtypes 4 and 5 are enriched for G-CIMP-high samples, although subtype 5 is enriched for mutations in ATRX (post hoc test of residuals for c(2) test, p < 10 (5)) and shows a higher enrichment for mast cells (Figure 7), which are known to induce release of selective inflammatory cytokines, such as interleukin-4 (IL-4)…”

“In this work, we used a variant of NMF called Symmetric NMF, in which the decomposition is done on a symmetrical matrix that contains pairwise similarity values between the data points instead of being done directly on the data points (Kuang et al., 2012).”

“MGMT promoter methylation is associated with a better response to alkylating chemotherapy, suggesting that individuals assigned to subtype 1 are more likely to respond to temozolomide (Rivera et al., 2010).”

“Our results support the proposed inclusion of additional diagnosis criteria, in particular chr7 gain and chr10 loss, which could lead to reclassification of several LGGs into GBMs (Stichel et al., 2018).”

“More sophisticated methods (e.g., iCluster [Shen et al., 2009], iClusterPlus [Mo et al., 2013], and Bayesian consensus clustering [Lock and Dunson, 2013]) model the probabilistic distribution of each data type and infer subtypes by maximization of the likelihood of the observed data.”

“The NE subtype has been related previously to the tumor margin, where increased normal NE tissue is likely to be detected (Gill et al., 2014).”

“A more recent method, NEMO (Rappoport and Shamir, 2019b), calculates an average similarity matrix and then detects the clusters using spectral clustering.”

“Given that each assay reveals a snapshot of certain cellular aspects of a disease, integrative analyses are often necessary for a complete understanding of its molecular etiology and for discovering its molecular subtypes and biomarkers (Prasad et al., 2016).”

“Throughout the study, whenever appropriate, we compare SUMO v.0.2.6 with the iClusterBayes extension of iClusterPlus v.1.26, LRAcluster v.1.0 (Wu et al., 2015), MCCA from PMA package v.1.2.1 (Witten and Tibshirani, 2009), NEMO v.0.1 (Rappoport and Shamir, 2019b), PINSPlus v.2.0.5 (Nguyen et al.,…”

“…extension of iClusterPlus v.1.26, LRAcluster v.1.0 (Wu et al., 2015), MCCA from PMA package v.1.2.1 (Witten and Tibshirani, 2009), NEMO v.0.1 (Rappoport and Shamir, 2019b), PINSPlus v.2.0.5 (Nguyen et al., 2017), SNF v.2.3 (Wang et al., 2014), and CIMLR v.1.0 (Ramazzotti et al., 2018).”

“PINSPlus results varied significantly in terms of survival and enrichment of clinical labels.”

“These results are in agreement with other published studies that show that AP-1 binds to demethylated regions in G-CIMP-low tumors, but we find this to be true for all samples assigned to subtype 2 (Souza et al., 2018).”

“The cIMPACT-NOW (the Consortium to Inform Molecular and Practical Approaches to CNS Tumor Taxonomy) initiative was established to evaluate and recommend changes to future CNS tumor classifications based on consensus review of novel diagnostically relevant data (Louis et al., 2017).”

“We evaluated the solutions with 2–19 clusters according to the proportion of ambiguously clustered pairs (PACs) (Şenbabaoğlu et al., 2014) and the cophenetic correlation (Hutchins et al., 2008; see STAR Methods for details).”

“We used ELMER (Silva et al., 2019) in an unsupervised mode to compare subtype 2 tumors with the other LGGs. ELMER identified 382 probes overlapping putative distal enhancers that were hypomethylated in subtype 2 samples (adjusted p < 0.01; methylation difference between means of the groups, >0.3).”

“This group of individuals with a poor prognosis is enriched for those with mesenchymal tumors that are known to lead to worse outcomes (Cancer Genome Atlas Research Network, 2011).”

“Therefore, FOS proteins have been implicated as regulators of cell proliferation, differentiation, and transformation (Mehta and Lo Cascio, 2018).”

“The largest study of diffuse grade II-III-IV gliomas to date used TumorMap (Newton et al., 2017) to integrate gene expression and DNA methylation data from around 1,000 individuals, and they, too, identified IDH status as the primary driver of two macro-clusters (Ceccarelli et al., 2016).”

“NMF has been successfully used as a clustering method in image and pattern recognition (Leuschner et al., 2019), text-mining (Chen et al., 2015), and bioinformatics (Wang et al., 2015).”

“More specifically, we find that the expression of FOSL1, which contributes to regulation of placenta development, is significantly higher in subtype 2 tumors, and higher expression of the gene is associated with a worse prognosis (Kubota et al., 2015).”

“NMF has been successfully used as a clustering method in image and pattern recognition (Leuschner et al., 2019), text-mining (Chen et al., 2015), and bioinformatics (Wang et al., 2015).”

“The NMF technique aims to explain the observed data using a small number of basis components by factoring the data into the product of two non-negative matrices; one representing the basis components, and the other containing mixture coefficients (Paatero and Tapper, 1994).”

“NMF has been successfully used as a clustering method in image and pattern recognition (Leuschner et al., 2019), text-mining (Chen et al., 2015), and bioinformatics (Wang et al., 2015).”

“xCell scores are derived from gene expression, but we observe similar results in analysis of methylation data using MIRA (Lawson et al., 2018).”

“For example, cosine similarity has been shown to be a better metric for the calculation of similarity between single cells in the single-cell sequencing for transposase accessible chromatin (scATAC-seq) (Cai et al., 2018).”

“In this study, we report the p values of the log rank test using exact permutation tests because p values based on the c(2) approximation are highly inaccurate in evaluating clustering solutions on real cancer datasets (Rappoport and Shamir, 2019a).”

“As well, the range of adjusted rand index values (the performance metric used) was typically observed to be narrower for SUMO, suggesting less variability in expected performance.”

“Articles such as Sienkie-wicz et al. [68] (Fig.”

“Consensus clustering has also been successfully applied by the recently published SUMO [62], an integrative clustering algorithm that starts by computing several unimodal PSNs by using a scaled-normalized Euclidean kernel similar to the one exploited by SNF [3].”

“The residuals from this regression were then used for down-stream analysis.”

“1 To understand the heterogeneity among gliomas, tremendous efforts have been made to identify the different adult-type diffuse types, 2-4 which can be broadly classified as lower-grade gliomas (LGGs) according to a lower proliferative activity or glioblastoma (GBM) which exhibits high malignancy…”

“Advantages over SOTA methods: unlike techniques tailored to specific tasks (e.g. clustering with MaDDA (Vitali et al. 2018) or SUMO (Sienkiewicz et al. 2022), classification with DeepIMV (Lee and van der Schaar 2021) or MOGDx (Ryan et al. 2024)), the miss-SNF algorithm produces a general-purpose integrated matrix suit…”

“In the previewed publications, NMF [76-78] or NMTF [41,79] were used to enhance clustering approaches.”

“Similarity Network Fusion (SNF) [ 43], SUMO [ 44], and NEMO [ 45] exemplify this approach ( Fig.”