Abstract
22 min readArticle Figures and data Abstract eLife digest Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Somatic copy number alterations (CNAs) are a hallmark of cancer, but their role in tumorigenesis and clinical relevance remain largely unclear. Here, we developed CNApp, a web-based tool that allows a comprehensive exploration of CNAs by using purity-corrected segmented data from multiple genomic platforms. CNApp generates genome-wide profiles, computes CNA scores for broad, focal and global CNA burdens, and uses machine learning-based predictions to classify samples. We applied CNApp to the TCGA pan-cancer dataset of 10,635 genomes showing that CNAs classify cancer types according to their tissue-of-origin, and that each cancer type shows specific ranges of broad and focal CNA scores. Moreover, CNApp reproduces recurrent CNAs in hepatocellular carcinoma and predicts colon cancer molecular subtypes and microsatellite instability based on broad CNA scores and discrete genomic imbalances. In summary, CNApp facilitates CNA-driven research by providing a unique framework to identify relevant clinical implications. CNApp is hosted at https://tools.idibaps.org/CNApp/. eLife digest In most cases, human cells contain two copies of each of their genes, yet sometimes this can change, an effect called copy number alteration (CNA). Cancer is a genetic disease and thus, studying the DNA from tumor samples is crucial to improving diagnosis and choosing the right treatment. Most tumors contain cells with CNAs; however, the impact of CNAs in cancer progression is poorly understood. CNAs can be studied by examining the genome of tumor cells and finding which regions display an unusual number of copies. It may also be possible to gather information about different cancer types by analyzing the CNAs in a tumor, but this approach requires the analysis of large amounts of data. To aid the analysis of CNAs in cancer cells, Franch-Expósito, Bassaganyas et al. have created an online tool called CNApp, which is able to identify and count CNAs in genomic data and link them to features associated with different cancers. The hope is that a better understanding of the effect of CNAs in cancer could help better diagnose cancers, and improve outcomes for patients. Potentially, this could also predict what type of treatment would work better for a specific tumor. Besides, by using a machine-learning approach, the tool can also make predictions about specific cancer subtypes in order to facilitate clinical decisions. Franch-Expósito, Bassaganyas et al. tested CNApp using previously existing cancer data from 33 different cancer types to show how CNApp can help the interpretation of CNAs in cancer. Moreover, CNApp can also use CNAs to identify different types of bowel (colorectal) cancer in a way that could help doctors to make decisions about treatment. Together these findings show that CNApp provides an adaptable and accessible research tool for the study of cancer genomics, which could provide opportunities to inform medical procedures. Introduction The presence of somatic copy number alterations (CNAs) is a ubiquitous feature in cancer. In fact, the distribution of CNAs is sufficiently tissue-specific to distinguish tumor entities (Ried et al., 2012), and allows the identification of tumors responsive to particular therapies (Cairncross et al., 2013; Davoli et al., 2017). Moreover, high levels of CNAs, which result from chromosome instability, are generally associated with high-grade tumors and poor prognosis (Hieronymus et al., 2018; Sansregret et al., 2018; Smith and Sheltzer, 2018; Stopsack et al., 2019). Two main subtypes of CNAs can be discerned: broad CNAs, which are defined as whole-chromosome and chromosomal arm-level alterations, and focal CNAs, which are alterations of limited size ranging from part of a chromosome-arm to few kilobases (Krijgsman et al., 2014; Zack et al., 2013). Recently, it has been uncovered that while focal events mainly correlate with cell cycle and proliferation markers, broad aberrations are mainly associated with immune evasion markers, suggesting that tumor immune features might be determined by mechanisms related to overall gene dosage imbalance rather than specific actionable genes (Buccitelli et al., 2017; Davoli et al., 2017; Taylor et al., 2018). Furthermore, it has been shown that broad CNAs involving whole-chromosome arms may confer high risk of lethal disease in prostate cancer (Stopsack et al., 2019). Nevertheless, the precise role of CNAs in tumor initiation and progression, as well as their clinical relevance and therapeutic implications in most cancer types remain still poorly understood. Interpretation and visualization of CNAs is time-consuming and very often requires complex analyses with clinical and molecular information. Well-established CNA algorithms, such as the gold-standard circular binary segmentation, define the genomic boundaries of copy number gains and losses based on signal intensities or read depth obtained from array comparative genomic hybridization and SNP-array or next-generation sequencing data, respectively (Olshen et al., 2004). However, the tumor-derived genomic complexity may cause an under- or overestimation of CNAs. This complexity is represented by tumor purity, tumor aneuploidy, and intratumor heterogeneity, which imply high levels of subclonal alterations. Thus, recent segmentation methods improved the accuracy to identify copy number segments in tumor samples either by considering the B allele frequency (BAF), such as ExomeCNV (Sathirapongsasuti et al., 2011), Control-FREEC (Boeva et al., 2012) and SAAS-CNV (Zhang and Hao, 2015), or through adjusting by sample purity and ploidy estimates, such as GAP (Popova et al., 2009), ASCAT (Van Loo et al., 2010) and ABSOLUTE (Carter et al., 2012). However, the state-of-the-art computational approach for CNA analysis in cancer is GISTIC2.0 (Mermel et al., 2011), which is a gene-centered probabilistic method that enables to define the boundaries of recurrent putative driver CNAs in large cohorts (Beroukhim et al., 2010). Nevertheless, despite ongoing progress on identifying CNAs, to our knowledge none of the existing software packages is readily available for integrative analyses to unveil their biological and clinical implications. To address this issue, we developed CNApp, the first open-source application to quantify CNAs and integrate genomic profiles with molecular and clinical variables. CNApp is a web-based tool that provides the user with high-quality interactive plots and statistical correlations between CNAs and annotated variables in a fast and easy-to-explore interface. In particular, CNApp uses purity-corrected genomic segmented data from multiple genomic platforms to redefine CNA profiles, to compute CNA scores based on the number, length and amplitude of broad and focal genomic alterations, to assess differentially altered genomic regions, and to perform machine learning-based predictions to classify tumor samples. To exemplify the applicability and performance of CNApp, we used publicly available segmented data from The Cancer Genome Atlas (TCGA) to (i) measure the burden of global, broad, and focal CNAs as well as generate CNA profiles in a pan-cancer dataset spanning 33 cancer types, (ii) identify cohort-based recurrent CNAs in hepatocellular carcinoma and compare them with previously reported data, and (iii) assess predicting models for colon cancer molecular subtypes and microsatellite instability status based on CNA scores and specific genomic imbalances. CNApp is hosted at https://tools.idibaps.org/CNApp/ and the source code is freely available at GitHub (Franch-Expósito, 2020 ; copy archived at https://github.com/elifesciences-publications/CNApp). Results Implementation CNApp comprises three main sections: 1- Re-Seg and Score: re-segmentation, CNA scores computation, variable association and survival analysis, 2- Region profile: genome-wide CNA profiling, CNA frequencies, correlation profiles and descriptive regions, and 3- Classifier model: machine learning classification model predictions (Figure 1). Each of these sections and their key functions are described below. The input file consists of a data frame with copy number segments provided by any segmentation algorithm. Mandatory fields and column headers are sample name (ID), chromosome (chr), start (loc.start) and end (loc.end) genomic positions, and the log2 ratio of the copy number amplitude (seg.mean) for each segment. Section one incorporates the correction for tumor purity (i.e. fraction of tumor cells in the sample) to measure the actual magnitude of CNAs. Thus, when available, the input file will also include sample purity estimations (purity) and BAF values (BAF), which correct the accuracy of CNA calls and provide copy number neutral loss-of-heterozygosity (CN-LOH) events. Ploidy values, if known, might also be indicated as an independent variable. Annotation of variables can be included in the input file (tagged in every segment from each sample) or by uploading an additional file indicating new variables per sample. Figure 1 with 1 supplement see all Download asset Open asset CNApp workflow. The diagram depicts the overall processes performed by CNApp and indicates the output for each section. Re-Seg and score: re-segmentation, CNA scores computation, variable association and survival analysis First, CNApp applies a re-segmentation approach to adjust for amplitude divergence due to technical variability and correct for estimated tumor purity. Default re-segmentation settings include minimum segment length (100 Kbp), minimum amplitude (seg.mean) deviation from segment to zero (0.16), maximum distance between segments (1 Mb), maximum amplitude (seg.mean) deviation between segments (0.16), and maximum BAF deviation between segments (0.1). These parameters can be customized by the user to better adjust the re-segmentation for each particular dataset. Re-segmented data are then used to calculate the broad (BCS), focal (FCS) and global (GCS) CNA scores, which provide three different quantifications of CNA levels for each sample. To compute these scores, CNApp classifies and weights CNAs based on their length and amplitude. For each sample, BCS is computed by considering broad (chromosome and arm-level) segment weights according to the amplitude value. Likewise, calculation of FCS takes into account weighted focal CNAs corrected by the amplitude and length of the segment. Finally, GCS is computed by considering the sum of normalized BCS and FCS, providing an overall assessment of the CNA burden. To assess the reliability of CNA scores, we compared each score with the corresponding fraction of altered genome using a TCGA pan-cancer set of 10,635 samples. Both BCS (ranging from 0 to 44) and FCS (values ranging from 5 to 2466) highly correlated with the fraction of altered genome by broad and focal copy number changes, respectively (Spearman's rank correlation for BCS = 0.957 and for FCS = 0.938) (Figure 1—figure supplment 1A and B—source data 1). As expected, GCS (values ranged from −1.93 to 12.60) highly correlated with the fraction of altered genome affected by both broad and focal CNAs (Spearman's rank correlation for GCS = 0.963 (Figure 1—figure supplment 1C—source data 1). Parametric and non-parametric statistical tests are used to establish associations between CNA scores and annotated variables from the input file. Additionally, Kaplan-Meier survival curves are computed using either CNA scores or additional variables. Region profile: genome-wide CNA profiling This section transforms segmented data (either re-segmented data from section one or original segments uploaded by the user) into genomic region profiles to allow sample-to-sample comparisons. Different genomic windows can be selected to compute the genomic profiles (i.e. chromosome arms, half-arms, cytobands, sub-cytobands or 40–1 Mb windows). All segments, or either only broad or only focal can be selected for this analysis. Length-relative means are computed for each window by considering amplitude values from those segments included in each specific window. Default cutoffs for low-level copy number gains and losses (i.e., |0.2|) are used to infer CNA frequencies. Genomic profiles are presented in genome-wide heatmaps to visualize general copy number patterns. Up to six annotation tracks can be added and plotted simultaneously allowing visual comparison and correlation between CNA profiles and different variables, including the CNA scores obtained in section 1. CNA frequency summaries by genomic region and by sample are represented as stacked bar plots. Correlation values and hierarchical clusters are optional. Importantly, assessing differentially altered regions between sample groups might contribute to discover genomic regions associated with annotated variables and thus unveil the biological significance of specific CNAs. To do so, CNApp interrogates descriptive regions associated with any sample-specific annotation variable provided in the input file. Default statistical significance is set to p-value lower than 0.1. However, p-value thresholds can be defined by the user and adjusted p-value is optional. A heatmap plot allows the visualization and interpretation of which genomic regions are differentially altered between sample groups. By selecting a region of interest, box plots and stacked bar plots are generated comparing seg.mean values and alteration counts in Student's t-test and Fisher's test tabs, respectively. Additionally, genes comprised in the selected genomic region are indicated. Classifier model: Machine learning classification model predictions This section allows the user to generate machine learning-based classifier models by choosing a variable to define sample groups and one or multiple classifier variables. To do so, CNApp incorporates the randomForest R package (Liaw and Wiener, 2002). The model construction is performed 50-times, in which training sets of 75% of samples are used in each permutation and the remaining 25% are classified by applying the trained model. By default, annotation variables from the input file are loaded and can be used either by defining sample groups or as a classifier. If Re-Seg and Score and/or Region profile sections have been previously completed, the user can upload data from these sections (i.e. CNA scores and genomic regions). Predictions for the model performance are generated and the global accuracy is computed along with sensitivity and specificity by group. Classifier models can be useful to point out candidate clinical or molecular variables to classify sample subgroups. Nevertheless, the output from the Classifier model would need to be further validated using alternative prediction algorithms and/or independent datasets. A summary of the data distribution and plots for real and model-predicted groups are visualized. A table with prediction rates throughout the 50-times iteration model and real tags by sample is displayed and can be downloaded. Characterization of cancer types based on CNA scores First, we evaluated the ability of CNApp to analyze and classify cancer types according to CNA scores, and assessed whether CNApp was able to reproduce specific CNA patterns across different cancer types. To do so, by using CNApp default parameters we obtained re-segmented data, CNA scores and cancer-specific CNA profiles for 10,635 tumor samples spanning 33 cancer types from the TCGA pan-cancer dataset. The distribution of BCS, FCS and GCS confirmed the existence of distinct CNA burdens across cancer types (Figure 2A—source data 1). While cancer types such as acute myeloid leukemia (LAML), thyroid carcinoma (THCA) or thymoma (THYM) showed low levels of broad and focal events (GCS median values of −1.67 for LAML, −1.68 for THCA, and −1.52 for THYM), uterine carcinosarcoma (UCS), ovarian cancer (OV) and lung squamous cell carcinoma (LUSC) displayed high levels of both types of genomic imbalances (GCS median values of 2.55, 2.44, and 0.97 for UCS, OV, and LUSC, respectively). Some cancer types displayed a preference for either broad or focal CNAs. For example, kidney chromophobe (KICH) tumors showed the highest levels of broad events (median BCS value of 27), while focal CNAs in this cancer type were very low (median FCS value of 49). In contrast, breast cancer (BRCA) samples displayed high FCS values (median FCS value of 150), while BCS values were only intermediate (median BCS value of 7). Overall correlations between CNA scores were assessed by computing Spearman's rank test, obtaining values of 0.59 between BCS and FCS, 0.90 between BCS and GCS, and 0.85 between FCS and GCS. In addition, we further assessed the correlation between BCS and FCS for each individual BCS value. While tumors with low BCS displayed a positive correlation between broad and focal alterations, tumors did not maintain such correlation in higher BCS values (Figure 2—figure supplement 1A and B). This correlation between BCS and FCS is maintained across the 33 cancer types (Figure 2—figure supplement 1C). Figure 2 with 2 supplements see all Download asset Open asset Analysis of the TCGA pan-cancer dataset and clustering by tumor type. CNApp outputs to characterize pan-cancer 10,635 samples including 33 TCGA cancer types. (A) Broad, Focal and Global CNA scores (BCS, FCS and GCS, respectively) distribution across the 33 cancer types (Figure 2—source data 1). (B) Genome-wide chromosome arm CNA profile heatmap for 10,635 samples considering broad and focal events. Annotation tracks for FCS, BCS and GCS are presented. (C) Arm regions frequencies as percentages relative to the TCGA pan-cancer dataset (red for gains and blue for losses). (D) Heatmap plot showing 20 out of the 33 TCGA cancer type profile correlations, by Pearson's method, hierarchically clustered by tissue of origin. Gastrointestinal, gynecological and squamous cancers are clustering consistently in their respective groups. Figure 2—source data 1 Broad, Focal and Global CNA scores across the pan-cancer cohort. https://cdn.elifesciences.org/articles/50267/elife-50267-fig2-data1-v2.xlsx Download elife-50267-fig2-data1-v2.xlsx Subsequent analysis aimed at generating genome-wide patterns for each cancer type based on chromosome-arm genomic windows and the overall corresponding frequencies. In agreement with previous studies (Beroukhim et al., 2010), cancer type-specific patterns of genomic gains and losses determined the tissue-of-origin (Figure 2B). Additionally, we found that chromosome arms altered in more than 25% across all samples were 1q, 7 p, 7q, 8q and 20q for copy number gains, and 8 p and 17 p for copy number losses. Conversely, chromosome arms affected by CNAs in less than 10% of all cancer types included chromosome arms 2q and 19 p (Figure 2C). By using a subset of 20 out of the 33 cancer types for which tumor type information was available, we asked CNApp to compute the average arm-region for each cancer type to assess if they clustered according to their CNA profiles (Figure 2—figure supplement 2A). Our analysis showed that correlation values resulting from Pearson's test hierarchically clustered according to the tissue-of-origin from the tumor. Gastrointestinal (colon, rectum, stomach and pancreatic), gynecological (ovarian and uterine) and squamous (cervical, head and neck, and lung) cancers clustered together based on specific CNA profiles for each group (Figure 2D). Intriguingly, correlation profiles using 5 Mb windows and only considering focal alterations showed a very similar degree of clustering based on the tissue of origin (Figure 2—figure supplement 2B&C). Identification of recurrent CNAs in hepatocellular carcinoma: Benchmark with other available tools Next, we attempted to test the ability of CNApp to identify recurrent broad and focal CNAs in a large cohort, and to assess the impact of the customizable parameters to describe CNA profiles. For that reason, we chose to perform CNA analysis of 370 samples from TCGA corresponding to the Liver Hepatocellular Carcinoma (LIHC) cohort. The pattern of recurrent broad and focal CNAs identified by GISTIC2.0 in the TCGA study (Ally et al., 2017) was similar to earlier reports, confirming the suitability of this cohort and the consistent identification of a CNA profile for hepatocellular carcinoma (HCC) (Chiang et al., 2008; Guichard et al., 2012; Schulze et al., 2015; Totoki et al., 2014; Wang et al., 2013). By applying the default parameters of CNApp to the LIHC dataset and selecting chromosome arms as genomic regions to assess broad events, we consistently found copy number gains at 1q (56%) and 8q (46%), and copy number losses at 8 p (62%) and 17 p (47%) as the most frequent alterations (Figure 3A). These CNAs are the same as those identified by GISTIC2.0; however, frequencies were slightly lower (Supplementary file 1). Similarly, GISTIC2.0 detected significant gains with rates between 25–40% on eight additional chromosome-arms, including 5 p, 5q, 6 p, 20 p, 20q, 7 p, 7q, and 17q, which were also identified by CNApp, but in 20–30% of the samples. Likewise, GISTIC2.0 detected significant broad deletions at a frequency between 20% and 40% on 18 additional chromosome-arms, of which 4q, 6q, 9 p, 13q, 16 p, and 16q losses were observed at frequencies ≥ 20% by CNApp, and the rest of them displayed rates between 10% and 20%. Therefore, the identification of CNAs in CNApp is very consistent with those described by GISTIC2.0. Differences in frequencies might be expected due to the lower copy number amplitude thresholds used by GISTIC2.0 in comparison with CNApp default cutoffs (|0.1| vs. |0.2|, corresponding to ~2.14/1.8 copies vs. 2.3/1.7 copies, respectively). Indeed, previous reports analyzing CNAs in other HCC cohorts and using greater copy number thresholds, showed frequencies of alterations much more similar to those estimated by CNApp (Chiang et al., 2008; Guichard et al., 2012; Schulze et al., 2015; Wang et al., 2013). To assess the impact of modifying CNApp amplitude thresholds, we analyze the same dataset dropping the cutoff to |0.1|. By doing so, the overall number of broad CNAs increased, reaching frequency values similar or even higher than those reported by GISTIC2.0 (Figure 3B and Supplementary file 1). Of note, such drop from |0.2| to |0.1| might result in the identification of subclonal genomic imbalances, which are very frequent in tumor samples (McGranahan and Swanton, 2017), and it would also be of utility to compensate for low tumor purities. To evaluate the impact of other customizable parameters of CNApp to the results, we also tested whether the identification of broad events was affected by: (i) the relative length to classify a segment as arm-level alteration, and (ii) the re-segmentation provided by CNApp. As expected, increasing the percentage of chromosome arm required to classify a CNA segment as arm-level (from ≥50% to≥70%) or skipping the re-segmentation step led to an underestimation of some broad events, whereas decreasing the percentage of chromosome arm (from ≥50% to≥40%) resulted in the opposite (Figure 3—figure supplement 1A–C and Supplementary file 1). Figure 3 with 1 supplement see all Download asset Open asset Identification of recurrent broad and focal CNAs. Calculation of broad and focal CNA frequencies using several parameters in CNApp in order to describe the genomic landscape of LIHC. (A) CNApp frequencies for chromosome arm regions using default cutoffs, corresponding to 2.3/1.7 copies for gains and losses, respectively. (B) CNApp frequencies for chromosome arm regions relaxing cutoffs to make them equivalent to those of GISTIC2.0. (C) CNApp frequencies of focal events using default thresholds and sub-cytobands genomic regions. (D) Frequencies of focal events from moderate- to high-amplitude levels using sub-cytobands genomic regions. As far as focal CNAs are concerned, CNApp and GISTIC2.0 use different strategies to quantify their recurrence. Therefore, the comparison between the two methods was evaluated in a more indirect manner. GISTIC2.0 generates minimal common regions (also known as 'peaks') that are likely to be altered at high frequencies in the cohort, which are scored using a Q-value and may present a wide variety of genomic lengths (Mermel et al., 2011). Instead, CNApp allows dividing the genome in windows of different sizes, calculating average copy number amplitudes for all segments included within each window. We reasoned that considering the length of GISTIC2.0 reported 'peaks', CNApp might also be capable of identifying recurrent focal altered regions by dividing the genome in smaller windows. To test our hypothesis, we asked CNApp to calculate the frequency of focal gains and losses by dividing the genome by sub-cytobands. As a result, CNApp consistently localized the most frequently altered sub-cytobands, including gains at 1q21.3 (25%), 8q24.21 (17%, MYC), 5p15.33 (13%, TERT), 11q13.3 (12%, CCND1/FGF19) and 6p21.1 (11%, VEGFA), and losses at 13q14.2 (20%, RB1), 1p36.11 (18%, ARID1A), 4q35.1 (17%, IRF2) and 9p21.3 (14%, CDKN2A), which are in agreement with previous studies in HCC (Figure 3C and Supplementary file 2) (Chiang et al., 2008; Guichard et al., 2012; Schulze et al., 2015; Wang et al., 2013). Compared to GISTIC2.0, CNApp reported 14 of the 27 significant amplifications and 14 of the 34 significant deletions at rates > 10%, and the remaining alterations displaying rates between 4–10% (Supplementary file 3) (Wang et al., 2013). Most importantly, regions with the highest frequency detected by CNApp showed a good match with lowest GISTIC2.0 Q-residual values, indicating that the most significant 'peaks' identified by GISTIC2.0 were actually included in the most recurrently altered sub-cytobands reported by CNApp. Recurrent focal alterations occur at lower frequencies than broad events (Beroukhim et al., 2010). Previous studies describing the genomic landscape of HCC mostly focused on focal high-amplitude CNAs (>3 copies for gains and <1.3 copies for losses), thus reporting lower frequencies than those estimated by CNApp using default thresholds (Chiang et al., 2008; Guichard et al., 2012; Schulze et al., 2015). In our analysis, excluding the low-level alterations and evaluating only the moderate and high-amplitude events (≥3 and≤1 copies), amplifications reached maximum rates of 11%, whereas high-level losses only reached ~2% (Figure 3D and Supplementary file 2). Top recurrent focal gains involved sub-cytobands 1q21.3 (11%), 8q24.21 (11%, MYC), 11q13.3 (7%, CCND1/FGF19), and 5p15.33 (5%, TERT). Recurrent losses estimated at ~2% of the samples included 13q14.2 (RB1), 9p21.3 (CDKN2A), 4q35.1 (IRF2), and 8p23.1. Although slight discrepancies between frequencies might be explained by minimal variability in the copy number threshold, CNApp results are highly consistent with previous reports (Chiang et al., 2008; Guichard et al., 2012; Schulze et al., 2015). In order to assess genomic differences between independent sample sets, CNApp determines significance based on regression analysis and statistical tests such as Student's t-test or Fisher's exact test, which allow association of genomic regions based on the seg.mean value or the presence of alterations with specific samples, respectively. To test the suitability of our method, we analyzed a subset of 100 randomly selected samples from the LIHC (n = 50) and COAD (n = 50) TCGA cohorts and we compared our results with those obtained by the recently published tool CoNVaQ (Larsen et al., 2018). Fourteen CNAs (5 gains and 9 losses) showing significant differences between COAD and LIHC were identified by CoNVaQ with a Q < 0.05. The lengths of these CNAs ranged from 38 to 133 Mb, with an average of 67 Mb, suggesting that they were mainly broad events. Consequently, in CNApp, we selected genomic windows corresponding to chromosome arms, and used the default CNA cut-offs (i.e. |0.2|) and maximum number of base pairs allowed for merging adjacent segments (i.e. 1,000,000 bp) for comparison. By doing so, 9 out of 14 events identified by CoNVaQ were also detected by CNApp with an adjusted p<0.05 (Fisher's exact test). For the remaining five events, four of them were found with an adjusted p=0.06–0.1, thus suggesting a good correlation between CNApp and CoNVaQ when considering broad CNAs. Classification of colon cancer according to CNA scores and genomic regions A proposed taxonomy of colorectal cancer (CRC) includes four consensus molecular subtypes (CMS), mainly based on differences in gene expression signatures (Guinney et al., 2015). Briefly, CMS1 includes the majority of hypermutated tumors showing microsatellite instability (MSI), high CpG island methylator phenotype (CIMP), and low levels of CNAs; CMS2 and CMS4 typically comprise microsatellite stable (MSS) tumors with high levels of CNAs; and finally, mixed MSI status and low levels of CNAs and CIMP are associated with CMS3 tumors. A representative cohort of 309 colon cancers from the TCGA Colon Adenocarcinoma (COAD) cohort (Cancer et al., 2012) with known CMS classification (CMS1, N = 64; CMS2 N = 112; CMS3 N = 51; CMS4 N = 82) and MSI status (MSI, N = 72; MSS, N = 225) was analyzed by using CNApp. In agreement with Guinney and colleagues, survival curves generated by CNApp indicated that CMS1 patients after relapse showed the worst survival rates as compared to CMS2 patients (Figure 4—figure supplement 1A) (Guinney et al., 2015). Next, we asked CNApp to perform the re-segmentation step using the default copy numb
Discussion(0)
No comments yet. Be the first to comment.