Briefings in Bioinformatics, 2022, 23(5), 1–9 https://doi.org/10.1093/bib/bbac161 Case Study Refinement of computational identification of somatic copy number alterations using DNA methylation microarrays illustrated in cancers of unknown primary Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 Pedro Blecua† , Veronica Davalos† , Izar de Villasante, Angelika Merkel, Eva Musulen, Laia Coll-SanMartin and Manel Esteller Corresponding author: Manel Esteller, Josep Carreras Leukaemia Research Institute (IJC), Badalona, Barcelona, Catalonia 08916, Spain. Tel.: +34-93-557-28-00; Fax: +34-93-282-84-51; E-mail: mesteller@carrerasresearch.org † Pedro Blecua and Veronica Davalos contributed equally to this work. Abstract High-throughput genomic technologies are increasingly used in personalized cancer medicine. However, computational tools to maximize the use of scarce tissues combining distinct molecular layers are needed. Here we present a refined strategy, based on the R-package ‘conumee’, to better predict somatic copy number alterations (SCNA) from deoxyribonucleic acid (DNA) methylation arrays. Our approach, termed hereafter as ‘conumee-KCN ’, improves SCNA prediction by incorporating tumor purity and dynamic thresholding. We trained our algorithm using paired DNA methylation and SNP Array 6.0 data from The Cancer Genome Atlas samples and confirmed its performance in cancer cell lines. Most importantly, the application of our approach in cancers of unknown primary identified amplified potentially actionable targets that were experimentally validated by Fluorescence in situ hybridization and immunostaining, reaching 100% specificity and 93.3% sensitivity. Keywords: somatic copy number alterations, DNA methylation, gene amplification, actionable target identification, cancers of unknown primary Introduction [4], that allow to detect SCNAs quantitatively in cancer. Deoxyribonucleic acid (DNA) methylation alterations, Our approach, hereafter referred as conumee-KCN , particularly the epigenetic inactivation of tumor sup- refines conumee’s calling of SCNAs by estimating a pressor genes, are a hallmark of human tumors and are dynamic sample-dependent threshold for different copy increasingly used as biomarkers and drug targets [1–3]. number states while accounting for tumor purity-, intra- For this reason, epigenomic tools that are cost-effective, sample- and copy number state-associated variation. such as DNA methylation microarrays, are gaining Commonly, SCNAs are detected by a fixed threshold momentum for translational purposes [1–3]. In recent (e.g. >0.3) or by n SDs from the sample mean/median years, as a potential alternative to the use of SNP-based and are not further distinguished [4–6]. Previous studies SNP Array 6.0 (SNP6) arrays, several approaches to detect showing associations between high amplifications and genome-wide Somatic Copy Number Alterations (SCNAs) high drug response rates have encouraged us to refine from Infinium Human Methylation 450K/EPIC arrays the SCNA calling from DNA methylation arrays to have been developed [4–6] and are applied in several further be able to distinguish between gains, moderate fields, including cancer research [i.e. 7–9]. Herein, we amplifications and high amplifications. For instance, the present a refined strategy to predict SCNAs using 450K highest responses to crizotinib are observed among non- DNA methylation microarrays [10], based on conumee small cell lung cancer patients harboring high levels Pedro Blecua, PhD, is an associate researcher at the Josep Carreras Leukaemia Research Institute in Spain. Veronica Davalos, PhD, is an associate researcher at the Josep Carreras Leukaemia Research Institute in Spain. Izar de Villasante, MSc, is a bioinformatician at the Josep Carreras Leukaemia Research Institute in Spain. Angelika Merkel, PhD, is a head of Bioinformatics Unit at the Josep Carreras Leukaemia Research Institute in Spain. Eva Musulen, MD, PhD, is an associate researcher at the Josep Carreras Leukaemia Research Institute in Spain. She also works as pathologist at the Hospital Universitari General de Catalunya-Grupo Quirónsalud in Spain. Laia Coll-SanMartin is a PhD student at the Josep Carreras Leukaemia Research Institute in Spain. Manel Esteller, MD, PhD, is an investigator of the Cancer Epigenetics group at the Josep Carreras Leukaemia Research Institute in Spain, and the Centro de Investigación Biomédica en Red de Cancer (CIBERONC) in Spain. He is also a professor of the Institucio Catalana de Recerca i Estudis Avançats (ICREA) in Spain; and the Physiological Sciences Department, School of Medicine and Health Sciences, University of Barcelona (UB) in Spain. Received: October 24, 2021. Revised: March 30, 2022. Accepted: April 10, 2022 © The Author(s) 2022. Published by Oxford University Press. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License (https://creativecommons.org/licenses/ by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact journals.permissions@oup.com 2 | Blecua et al. of MET amplification [11]. Similarly, high responses to samples of CUPs were available in house from our FGFR inhibitors are detected in high-level clonal FGFR2- previous study [13]. For all 450K datasets, we used amplified gastric cancer patients [12]. the R-package minfi [21] for processing and quality The dual detection of DNA methylation and copy num- control. We applied quantile normalization and a ber alterations using only the epigenomic platform could minimum detection threshold with P-value < 0.01 per be particularly relevant in the clinical setting, consider- probe and a maximum of 10% of failed probes per ing the common limitations in the amount of available sample. Sex chromosomes and cross-reactive probes tissue as well as the short timing usually available for were excluded from the analysis, and probes that were oncology patients to receive the best tailored individual located +/− 10 bases away from known SNPs were also treatment. We highlight the application of our algorithm filtered out. in a cohort of 211 cases of Cancer of Unknown Pri- mary (CUPs), previously profiled by our group using DNA Improved quantitative SCNA calling from 450K methylation arrays [13]. The intrinsic features of CUPs, data based on conumee Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 including their early dissemination, aggressive clinical We based our SCNA calling strategy on conumee [4], course and lack of evident pharmacological targets [14– a popular tool for SCNA calling from DNA methyla- 16], provide an ideal scenario to demonstrate the advan- tion arrays. Conumee calculates copy number alterations tages of identifying potentially druggable somatic ampli- based on the signal intensity (I) ratio between tumor fied targets to broaden the therapeutic alternatives to and normal samples (R = ITumor /INormal ) using the follow- treat this orphan tumor type and to eventually improve ing steps: (1) normalized intensity values of both the its dismal clinical outcome. ‘methylated’ and ‘unmethylated’ signals are added; (2) a ‘best-linear-fit’ is performed between the query sample (tumor, in our case) and all normal healthy samples Materials and methods ([20], in our case); (3) the log2 -ratio of probe intensities Affymetrix SNP6 microarray processing between the two is calculated; (4) from those median Affymetrix SNP6 array data (.CEL files) from The log2 -ratio of all probes intensities for predefined genomic Cancer Genome Atlas (TCGA) were downloaded from bins are derived (log2 [R]) and (5) copy number alterations Genomic Data Commons (GDC) Data Portal (https:// are estimated as difference of these ratios from the mean portal.gdc.cancer.gov/) and were processed in our pre- of all bins (mean(log2 [R]): vious study [17]. In brief, SNP6 array data were processed together, quantile-normalized and median-polished with cn = log2 [R]–mean(log2 [R]). Affymetrix power tools. Genotyping was performed with the Birdseed algorithm. PennCNV [18] was employed to Note that the intercept, mean(log2 [R]), represents the generate log2 R ratio and B-allele frequencies. Total- copy number neutral state (baseline), since ITumor = INormal . and allele-specific somatic copy number calls, as well (6) Final segments are derived from the predefined bins as purity estimation, were generated with ASCAT using circular binary segmentation (DNAcopy package [19]. [22]). Finally, SCNAs can be detected when the copy number estimate (cn) exceeds a certain threshold, either 450K DNA methylation microarrays processing a fixed value or measured as units of SD. We refined this approach by (1) estimating dynamic Illumina Infinium HumanMethylation450 (450K) Bead- thresholds for different types of SCNAs (‘copy number Chip array data from TCGA was downloaded from states’) and (2) incorporating tumor purity, intra- GDC Data Portal (https://portal.gdc.cancer.gov/) using sample variability and copy-number-state-dependent the R package TCGAbiolinks. From the 7009 TCGA noise. samples with matched genotyping (SNP6 array) and To define a threshold for copy number estimates (cn) DNA methylation (450K) data, 442 samples across above which we denote a certain copy number state 18 cancer types were used as our training cohort (CN) (‘Amp10’, ‘Amp’, ‘Gain’, ‘HetLoss’ or ‘HomDel’, see (Supplementary Table S1 available online at https:// Results), we assume that this threshold is array (a) depen- academic.oup.com/bib), and 151 samples across 11 dent and proportional to the tumor purity (ρ), the intra- cancer types were used as our validation cohort sample variation of the array (sd(log2 [Ra ])) and a constant (Supplementary Table S2 available online at https:// (KCN ) that reflects the copy number-state-associated academic.oup.com/bib). The 450K data from 75 normal biological/technical variation: samples across 18 different tissues were also retrieved from TCGA. The 450K data from 96 whole-blood (WB) TCN = mean(log2 [Ra ]) + KCN ∗ ρ ∗ sd(log2 [Ra ]). samples from healthy individuals were available from a previous study [20]. In addition, 450K data from cancer Tumor purity (ρ) is calculated using the ‘ABSOLUTE’ cell lines (CCLs) were downloaded from the Cancer method within the ‘RF_Purify’ R package [23]. Cell Line Encyclopedia (CCLE, Broad Institute) using Similarly, we denote that the difference between inten- the R-package GEOquery Gene Expression Omnibus sity ratios of segments with copy number alterations (GEO) database (GSE68379). The 450K data from 211 (log2 [RCN ]) and the overall array mean (mean(log2 [Ra ]) Computational identification of copy number by DNA methylation | 3 varies with the tumor purity (ρ) and the intra-sample were performed according to standard procedures variation of the array (sd(log2 [Ra ]): [24–27]. ρ ∗ sd(log2 [Ra ]) ∼ log2 [RCN ]–mean(log2 [Ra ]). Hybridizations were analyzed using a standard flu- orescence microscope (Leica DM5500 B Fluorescence For each copy number state (CN), we perform a linear microscopy, Leica Biosystems Newcastle Ltd, Newcastle regression and retrieve KCN : upon Tyne, UK Nikon Eclipse 50i, Tokyo, Japan, or Zeiss,  Oberkochen, Germany) equipped with appropriate filter ρ ∗ sd(log2 [Ra ]) ∼ 0 + log2 [RCN ]–mean(log2 [Ra ])), sets. Acquisition and processing of digital images were y = 0 + Bx, performed using CytoVision Imaging System (Leica B = y/x, Biosystems Newcastle Ltd) or the ISIS FISH Imaging KCN = 1/B → x/y, System (MetaSystems, Altlussheim, Germany). ∗y = ρ ∗ sd(log2 [Ra ]), x = log2 [RCN ]–mean(log2 [Ra ]), KCN = (log2 [RCN ]–mean(log2 [Ra ]))/ρ ∗ sd(log2 [Ra ]) Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 Immunohistochemistry (IHC) An example of linear regression is shown in FFPE tissue sections were analyzed using standard Supplementary Figure S1A available online at https:// IHC techniques. The primary antibodies used were: academic.oup.com/bib. We have termed our method anti-CCND1 (Cyclin D1 P211F11 clone, PA0046), anti-c- ‘conumee-KCN ’. MYC antibody (c-MYC Y69 clone, 3PR00355) and anti- BCL6 (BLC-6 LN22 clone, PA0204) from Leica Biosystems Software (Leica Biosystems Newcastle Ltd); anti-CCNE1 (HE12, We run conumee [4] (v. 1.26.0) with default parameters sc-247, Santa Cruz Biotechnology, Dallas, TX, USA) using the option ‘exclude’ [excluded over 10 000 genomic and anti-HER2/ERBB2 [PATHWAY HER-2/neu (4B5), regions corresponding to common copy number poly- TA9145, Ventana, Roche Diagnostics, Basel, Switzerland]. morphisms in the otherwise healthy population (arising Immunostainings were performed automatically using from both tissue and blood samples), as described by automatized protocol ‘F’ on Leica BOND-MAX platform the authors in the conumee online manual]. For the (Leica Biosystems) except for HER2, where the Bench- comparative benchmark, we run ChAMP [5] (v.2.22.) and Mark ULTRA IHC/ISH system (Roche Diagnostics) was cnAnalysis450k [6] (v.0.99.26) with default parameters, used. Positive staining for anti-CCND1, anti-CCNE1, and according to the authors instructions, using a mini- anti-c-MYC and anti-BCL6 antibody was localized in mum frequency of >0.3 and an effect size of >0.15 to call the nucleus of the neoplastic cells. The evaluation amplifications, respectively. was semiquantitative (weak, moderate or intense), indicating the percentage of positive nuclei. Positive Statistical analyses staining for anti-HER2/ERBB2 antibody was localized in All statistical analyses were performed under the R the membrane of the neoplastic epithelial cells. The package, R version 4.1.2 (2021-11-01)—‘Bird Hippie’. evaluation was performed according to the criteria The P-values, displayed in Supplementary Table S8 described in the CAP/ASCO guidelines [25]. For each and Supplementary Figure S5B, available online at staining, an external positive control was included. https://academic.oup.com/bib, were calculated with a two-sided Fisher exact test. Results Experimental allidation of potential We developed a refined strategy, based on conumee [4], actionable targets that allows us to dynamically define a threshold to quan- Fluorescence in situ hybridization (FISH) titatively call copy number variations from 450K DNA Formalin-fixed, paraffin-embedded (FFPE) tissue sec- methylation data in cancer (Figure 1A). The algorithm tions were analyzed using standard FISH techniques was trained using 442 tumor samples publicly available using the following commercial probes: MYC IQFISH from TCGA, representing 18 cancer types with sample- Break-Apart Probe (Agilent Technologies, G111623-2, matched genotyping (SNP6 array) and DNA methylation Santa Clara, CA, USA), CCND1 IQFISH Break-Apart Probe (450K array) data (Figure 1A). (Agilent Technologies, G111622-2), PIK3CA Spectrum For our method, we define five copy number states Green FISH Probe Kit/CEP3 SpectrumOrange (Vysis, (CN): (1) Homozygous deletions ‘HomDel’ (complete loss 06N10-001/06J36-003, Abbott Laboratories, Chicago, of all the alleles) = 0 copies; (2) Hemizygous deletions IL, USA), LSI MET SpectrumRed FISH probe Kit/CEP7 ‘HetLoss’ (loss of one allele) = 1 copy; (3) Gains ‘Gain’ = 3–4 SpectrumGreen (Vysis, 06N05-020/06J37-007), LSI BCL6 copies; (4) Moderate amplification ‘Amp’ = 5–9 copies and (ABR) Dual Color, Break Apart Rearrangement Probe (5) High amplification ‘Amp10’ = ≥10 copies. We used as (Vysis, 01N23-020), PathVysion LSI HER-2/neu Spectru- reference a list of 94 genes that include 83 frequently mOrange/CEP17 SpectrumGreen (Vysis, 02J01/06J37-017 amplified genes [28] and 11 frequently deleted genes and CCNE1 BAC-Spectrum Red labeled probe (RP11- [29, 30] across cancers (Supplementary Table S3 avail- 104J24). Preparation of slides, hybridization and analysis able online at https://academic.oup.com/bib). 4 | Blecua et al. Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 Figure 1. Computational prediction of SCNA from DNA methylation arrays using conumee-KCN . (A) Workf low for the developed strategy to refine SCNA detection, ‘conumee-KCN ’, trained over 442 primary tumor samples from TCGA, across 18 cancer types, with matched genotyping (SNP6 array) and DNA methylation array (450K) data. Our thresholding strategy refines conumee outputs and quantitatively calls SCNA from 450K arrays by considering tumor purity ρ (RF_Purity) and rigorous estimation of copy-number-state (CN)-dependent constants KCN . ∗ A list of 94 genes frequently amplified or deleted in cancer was used as reference to define the copy number states. Thus, by using calibrated KCN ’s and considering tumor purity, intra-sample variability and copy-number-state-dependent noise, thresholds for each CN can be estimated for each 450K profiled sample to accurately identify SCNA. (B) Benchmarking of our strategy (conumee-KCN ) against conumee (fixed threshold of 0.3), cnAnalysis450k and ChAMP in an independent, validation set consisting of 151 TCGA samples, with matched genotyping (SNP6 array) and DNA methylation array (450K) data. True positive (TP) and false-positive (FP) rates of 450K-derived calls versus SNP6-derived calls (ASCAT) for amplifications are depicted, showing the improved performance of our approach. (C) TP and FP rates of conumee-KCN versus ASCAT in the TCGA validation set for the three amplification copy number states (Amp10, Amp and Gain). (D) Representative examples of gene amplifications in two samples from the TCGA validation cohort. Thresholds estimated by conumee-KCN for Amp and Amp10 are depicted (dotted grey lines). TP = #True positives/#True positives + #False Negatives; FP = #False positives/#False positives + #True Negatives. As a first step, for each sample of our TCGA training to calculate for each of the samples and each of the copy cohort (Supplementary Table S1 available online at number state: (1) the intra-sample variation of the signal https://academic.oup.com/bib), we called SCNA from intensity ratio (sd(log2 [Ra ]), (2) the mean signal intensity genotyping data using ASCAT (see Materials and Meth- ratio (mean(log2 (RCN )) for all genes in the corresponding ods) and created gene subsets for each copy number gene list and (3) the global array mean signal intensity state (CN) based on the respective calls. Next, we used (mean((log2 [Ra ]). To broaden the applicability of this tool conumee and the corresponding DNA methylation data to any tissue type, we used WB samples from healthy Computational identification of copy number by DNA methylation | 5 individuals [20] as normal control, as is often done in lung adenocarcinoma sample, predicted by our approach large scale studies such as TCGA (GISTIC and ASCAT). We from the TCGA validation cohort, are shown in Figure 1D. verified the suitability of this type of samples as normal The performance of conumee-KCN was further assessed counterpart by testing for significant DNA methylation in CCLs available from the CCLE (Broad Institute). The differences between our sample set (96 samples from higher concordance with SNP6-derived calls (TP = 0.247), 48 men and 48 women, all young adults, to avoid biases in comparison with conumee (TP = 0.109), corroborates regarding age and gender) and a set of 75 normal samples the superiority of our approach (Supplementary Figure S3 spanning 18 tissues. In 97% (1269) of the 1307 CpGs, used available online at https://academic.oup.com/bib). to infer the copy number state of the 94 genes frequently Once we confirmed the robustness of our strategy, we amplified or deleted in cancer, we observed no significant applied it to a set of CUP samples aimed to illustrate differences (Supplementary Table S4 available online at its usefulness in the clinical setting (Figure 2A). First, we https://academic.oup.com/bib). applied conumee to the DNA methylation arrays from In parallel, we estimated the tumor purity (ρ) of 211 CUP cases previously profiled by our group [13]. Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 each sample using the RF_Purify software [23], results Next, using the KCN ’s calibrated from 442 TCGA primary that were strongly correlated with the purity estimates tumors (Figure 1A) along with tumor purity estimates from genotyping data (SNP6 array) derived with ASCAT from RF_Purity (Supplementary Table S5 available online (R2 = 0.573; R = 0.76; Supplementary Figure S2 avail- at https://academic.oup.com/bib), we annotated the dif- able online at https://academic.oup.com/bib). Finally, ferent copy number states (CN): Amp10, Amp, Gain, Het- implementing the above estimates, we performed a Loss or HomDel. Aiming to identify gene amplification linear regression and extracted the copy number-state- events with a relevant role in CUP pathogenesis, we dependent constants KCN required for our thresholding dropped gains as it is not uncommon for cancer genomes strategy (see Materials and Methods). This workflow is to undergo whole genome doubling [31] and selected summarized in Figure 1A. We additionally estimated those genes that were amplified (CN = Amp or Amp10) mean KCN ’s from 10 000 randomizations with 442 sam- in at least 5 CUPs (2.4% of 211 CUPs; Figure 2A). Consid- ples each, drawing from the total set of 7009 TCGA sam- ering the clinical potential of our findings, we focused ples from 18 cancer types (Supplementary Figure S1B on gene amplifications given the progress of inhibitory available online at https://academic.oup.com/bib). The drugs for cancer treatment, and restricted the analysis obtained KCN values highly ref lect the values derived to protein coding genes (as annotated by Gencode_v34), from the 442 TCGA samples used as training set. also taking into account that 450K arrays are enriched Next, using a validation cohort consisting of an in CpGs located in promoters and gene bodies [10]. After independent set of 151 TCGA tumor samples across this filtering process, a total of 1159 candidate genes 11 cancer types (Supplementary Table S2 available amplified in CUPs were selected for further analysis online at https://academic.oup.com/bib), we bench- (Figure 2A; Supplementary Table S6 available online at mark the performance of our thresholding strategy https://academic.oup.com/bib). against conumee using a fixed threshold (>0.3) and Going one step further with the CUP cohort, in two other commonly used tools, namely, ChAMP [5] order to identify actionable gene amplifications with and cnAnalysis450k [6] (Figure 1B). We compared the a potential relevance in the clinical setting and thus concordance between SCNA calls from SNP6 genotyping broaden the therapeutic opportunities for this dismal arrays (ASCAT) and SCNA calls from 450k methylation tumor type, we used publicly available data and online arrays using the respective approaches, focusing on tools to further explore the 1159 amplified genes, amplifications (Gains, Amp and Amp10), the center of including the Cancer Genome Interpreter (https://www. our additional analyses. Figure 1B shows how, while the cancergenomeinterpreter.org/) [32], Precision Oncology call concordance [true-positive (TP) rate] is similar for Knowledge Base (https://www.oncokb.org) [33], Clin- the three standard approaches (TP = 0.051 for conumee- ical Interpretations of Variants in Cancer (https:// fixed threshold 0.3, TP = 0.043 for ChAMP and TP = 0.060 civicdb.org/) and Clinical Trials (www.clinicaltrials.gov). for cnAnalysis450k), our strategy conumee-KCN sensibly The inclusion criteria to identify potential action- improves SCNA prediction with a TP = 0.218, demonstrat- able targets were as follows (Figure 2A): (1) Is that ing the superiority of our method in comparison with gene a known or predicted driver in cancer?; (2) Is previously developed tools. there a drug (preclinical or clinical) to target that Moreover, a significant improvement of our strategy gene? and (3) Is the amplification of that gene a in comparison with previous methods [4–6] is that it potential biomarker for the drug response?. Preclin- allows to detect SCNAs quantitatively to distinguish ical evidences, early and late clinical trials and FDA Gains, Amps or Amp10s. As shown in Figure 1C, the con- guidelines were also considered (Figure 2A). According cordance of Amps (TP = 0.224) and Amp10s (TP = 0.236) to these criteria, we identified 15 potential action- further support the added value of our approach. able targets (Figure 2A; Supplementary Table S7 avail- Representative examples of ERBB2 amplification in a able online at https://academic.oup.com/bib), includ- breast cancer (BRCA) sample and MET amplification in a ing well-recognized oncogenes such as MYC, CCND1, 6 | Blecua et al. Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 Figure 2. Epigenomic-based computational prediction of SCNA in CUP using conumee-KCN . (A) Workf low for the detection of SCNA in CUPs using conumee-KCN and further identification of gene amplifications with clinical relevance. (B) Experimental validation of conumee-KCN predicted SCNA in CUP patient samples. Representative images of copy number state validation by FISH and protein expression by IHC are shown for three well-recognized oncogenes: MYC (c-MYC), CCND1 (Cyclin D1) and ERBB2 (HER2). ∗ Orange/green break-apart FISH probes were used. ERBB2 (Supplementary Figure S4 available online at (Supplementary Figure S5, Supplementary Table S8 avail- https://academic.oup.com/bib), BCL6, PIK3CA, MET and able online at https://academic.oup.com/bib). CCNE1. Furthermore, we found some of these genes Finally, to demonstrate the robustness of our strat- to be significantly co-amplified, not only those within egy in a bona fide manner, we sought to experimen- the same amplicon (e.g. role of amplicon 8q24 in tally validate our predictions using the gold standard cancer [34], Supplementary Figure S4 available online at approach to assess copy number: FISH. The results of https://academic.oup.com/bib) but also genes located the experimental validation of 21 cases are summa- on different chromosomes [i.e. FGF3/FGF4 (11q13.3) rized in Supplementary Figure S6 available online at co-amplified with PIK3CA/PIK3CB (3q26.32/3q22.3)] https://academic.oup.com/bib. FISH assays confirmed Computational identification of copy number by DNA methylation | 7 the predicted SCNA calls in 95.2% of cases (14/14 purity into account, among current DNA methylation- amplification events and 6/7 non-amplifications). In based SCNA calling tools, this is unique for conumee- terms of the confusion matrix, we obtained a sensitivity KCN . Incorporation of tumor purity in the thresholding of 93.3% and a specificity of 100%. Representative function could improve SCNA prediction, considering examples of amplified and non-amplified CUP cases that cancer-cell intrinsic features could be hidden in for three well-recognized oncogenes, ERBB2 (HER2), biological samples with a low proportion of tumor cells, MYC (c-MYC) and CCND1 (Cyclin D1), are shown in which is not uncommon in clinical setting. Figure 2B. Importantly, in comparison to non-amplified The usefulness of conumee-KCN was illustrated in CUP, cases, these gene amplification events were clearly a heterogeneous group of metastatic tumors that lack associated with protein upregulation detected by IHC an identifiable primary tumor despite a standardized (Figure 2B). Additional validations of BCL6 and MET are diagnostic work-up. Most CUP patients (80–85%) have shown in Supplementary Figure S7 available online at an unfavorable prognosis with a dismal survival of 3– https://academic.oup.com/bib. Thus, our experimental 6 months despite empirical chemotherapy treatments Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 results confirmed the SCNA predictions provided by our [16, 36, 37]. Thus, CUP management is an unmet medical DNA methylation microarray-based strategy conumee- need. Using our epigenomic-based computational pre- KCN herein presented. dictions of SCNA in CUPs, we identified many potential drug-actionable targets, including well-recognized onco- genes, such as MYC, CCND1, ERBB2, BCL6, PIK3CA, MET Conclusion and Discussion and CCNE1. Noticeably, some of these genes had been The occurrence of shared molecular alterations in previously identified in CUPs with very similar amplifi- different tumor types has supported the emergence cation frequencies [38–43], confirming the power of our of tissue-agnostic approaches, which have led to a approach. More relevant, successful examples of CUP revolutionary paradigm shift in drug development and cases harboring MET [43, 44] or ERBB2 [45] amplifications cancer treatment. Novel precision oncology trial designs, with favorable response to the corresponding target ther- such as basket trials, which eligibility is based on apies (crizotinib or trastuzumab, respectively) reinforce the presence of specific genetic alterations, irrespec- the importance of intensifying the research in this field. tive of histology, are opening bright opportunities for Currently, an ongoing clinical trial is comparing the effi- patients with tumors harboring actionable alterations. cacy of molecularly guided therapies versus platinum- The tissue-agnostic arsenal of FDA-approved drugs is based standard chemotherapy in CUP patients [38]. increasing, and considering this scenario, optimization The dual use of DNA methylation arrays to identify of computational strategies to maximize the identifi- both epigenomic and genomic SCNA alterations is a cation of actionable alterations that guide the targeted significant advantage particularly evident in the clinical treatment choice is paramount. The more actionable practice, considering the need to maximize the use alterations identified, the more tools for precision of scarce tissues and the limited time to obtain the medicine. most comprehensive molecular information to effi- Here, we described conumee-KCN , an improved strat- ciently guide therapeutic decision-making. In addition, a egy based on conumee [4], to detect SCNA in cancer genome-wide approach, as the herein described, boosts using DNA methylation microarrays as a surrogate the possibility of identifying co-amplification events, for genotyping arrays. Our approach shows significant which might open new venues for combination of improvement over similar approaches, namely, con- targeted therapies in CUPs in a tissue agnostic manner. umee (standard) [4], ChAMP [5] and cnAnalysis450k Intriguingly, preclinical studies have shown that the [6], that previously have been reported to have low combined use of FGFR and PI3K inhibitors resulted reliability [35]. Moreover, the refinement of our approach in an enhanced efficacy in comparison with single estimates a dynamic threshold for each copy number treatments [46, 47]. Furthermore, although we focused state, allowing differentiation between gains, moderate the CUP analysis on the most promising actionable amplifications and high amplifications, a functionality targets considering the current evidences, further that does not exist in the currently available tools. studies of the remaining candidates identified by our Thus, with conumee-KCN , the user can estimate sample- conumee-KCN approach could reveal genes or pathways dependent thresholds for Gains, Amps or Amp10s by with key roles in CUP pathogenesis that might open using calibrated KCN ’s and considering tumor purity novel therapeutics opportunities for this dismal cancer and intra-sample variability to accurately identify SCNA type. from any 450K profiled sample. Considering previous The robustness of conumee-KCN is reflected by the studies showing highest drug responses among patients high sensitivity (93.3%) and specificity (100%) obtained harboring high levels of amplifications of the target gene through the experimental validation by FISH. In this [i.e. 11, 12], this feature of conumee-KCN is an added value regard, selective tissue-agnostic treatment of CUPs that could better guide treatment decision-making in focusing on the gene targets identified in this study healthcare of cancer patients. Furthermore, whereas might improve the very scarce landscape of effective genotyping-based SCNA callers generally take tumor therapies against this tumor type. Importantly, our 8 | Blecua et al. approach is in principle f lexible and generalizable to Acknowledgements other DNA methylation microarrays (e.g. EPIC Methyla- We thank Jonathan Garcia Barrasa for the technical tion Arrays) as well as potentially useful for any cancer support. type. Thus, the application of conumee-KCN in further biocomputational studies is warranted. Funding Sarah Jennifer Knott Foundation Research Award (to Data availability M.E.) Health Department PERIS project no. SLT/002/16/ Data and code to reproduce key components of Figure 1 00374 and AGAUR project no. 2017SGR1080 of the are publicly available at https://github.com/izarvillasa Catalan Government (Generalitat de Catalunya); Min- nte/SCNA.git. SNP6 array data to determine total Copy isterio de Ciencia e Innovación; Agencia Estatal de Number, as well as 450K DNA methylation microarrays, Investigación; European Regional Development Fund were obtained from TCGA at https://portal.gdc.cancer. project no. RTI2018-094049-B-I00; Cellex Foundation; ‘la Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 gov/. CUP and WB 450K array data were obtained from Caixa’ Banking Foundation (LCF/PR/GN18/51140001). [13, 20], respectively. Competing interests References M.E. is a consultant to Ferrer International and Quima- 1. Berdasco M, Esteller M. Clinical epigenetics: seizing opportuni- ties for translation. Nat Rev Genet 2019;20(2):109–27. tryx. The remaining authors declare that they have no 2. Mancarella D, Plass C. Epigenetic signatures in cancer: proper competing interests. controls, current challenges and the potential for clinical trans- lation. Genome Med 2021;13(1):23. Key Points 3. Locke WJ, Guanzon D, Ma C, et al. DNA methylation cancer • In the era of precision medicine, maximizing the use of biomarkers: translation to the clinic. Front Genet 2019;10:1150. commonly scarce tumor tissue to offer the most compre- 4. Hovestadt V, Zapatka M. Conumee: enhanced copy-number var- hensive molecular landscape from one biological sample iation analysis using Illumina DNA methylation arrays. R pack- is a critical need. age version 190. http://bioconductor.org/packages/conumee/. • We improved SCNA prediction from DNA methylation 5. Feber A, Guilhamon P, Lechner M, et al. Using high-density DNA arrays with our conumee-KCN strategy to further distin- methylation arrays to profile copy number alterations. Genome guish gains, moderate amplifications and high amplifi- Biol 2014;15(2):R30. cations, a tool potentially useful for any cancer type. 6. Knoll M, Debus J, Abdollahi A. cnAnalysis450k: an R package for • The added value of our strategy is significant, particu- comparative analysis of 450k/EPIC Illumina methylation array larly since none of the existing approaches using DNA derived copy number data. Bioinformatics 2017;33(15):2266–72. methylation arrays stratify gene amplifications, a fea- 7. Gao Y, Widschwendter M, Teschendorff AE. DNA methylation ture especially relevant in clinical setting. patterns in normal tissue correlate more strongly with breast • The usefulness of conumee-KCN strategy was illustrated cancer status than copy-number variants. EBioMedicine 2018;31: in samples from CUP patients, an aggressive tumor type 243–52. with dismal prognosis. 8. Haider Z, Landfors M, Golovleva I, et al. DNA methylation • The identification of 15 candidate actionable targets and copy number variation profiling of T-cell lymphoblastic amplified in CUPs highlights the potential of our leukemia and lymphoma. Blood Cancer J 2020;10(4):45. approach in the clinical setting. 9. Wang X, Grasso CS, Jordahl KM, et al. Copy number alterations • The conumee-KCN SCNA predictions in CUPs were vali- are associated with metastatic-lethal progression in prostate dated by experimental procedures (FISH, IHC), demon- cancer. Prostate Cancer Prostatic Dis 2020;23(3):494–506. strating a very high sensitivity (93.3%) and specificity 10. Sandoval J, Heyn H, Moran S, et al. Validation of a DNA methy- (100%). lation microarray for 450,000 CpG sites in the human genome. Epigenetics 2011;6(6):692–702. 11. Camidge DR, Otterson GA, Clark JW, et al. Crizotinib in patients with MET-amplified NSCLC. J Thorac Oncol 2021;16(6):1017–29. 12. Pearson A, Smyth E, Babina IS, et al. High-level clonal FGFR Supplementary Data amplification and response to FGFR inhibition in a translational Supplementary data are available online at https:// clinical trial. Cancer Discov 2016;6(8):838–51. academic.oup.com/bib. 13. Moran S, Martinez-Cardus A, Sayols S, et al. Epigenetic profiling to classify cancer of unknown primary: a multicentre, retrospec- Authors’ contributions tive analysis. Lancet Oncol 2016;17(10):1386–95. 14. Kato S, Alsafar A, Walavalkar V, et al. Cancer of unknown pri- P.B., I.d.V. and A.M. took care of algorithm design, mary in the molecular era. Trends Cancer 2021;7(5):465–77. implementation and computational analysis; V.D. was in 15. Laprovitera N, Riefolo M, Ambrosini E, et al. Cancer of unknown charge of target identification and curation and sample primary: challenges and progress in clinical management. Can- preparation; E.M. and L.C.-S. were in charge of sample cers (Basel) 2021;13(3):451. preparation and handling and experimental results 16. Rassy E, Pavlidis N. Progress in refining the clinical management curation. P.B., V.D. and M.E. wrote the manuscript. All of cancer of unknown primary in the molecular era. Nat Rev Clin authors read, edited and approved the final manuscript. Oncol 2020;17(9):541–54. Computational identification of copy number by DNA methylation | 9 17. Riaz N, Blecua P, Lim RS, et al. Pan-cancer analysis of bi-allelic 32. Tamborero D, Rubio-Perez C, Deu-Pons J, et al. Cancer genome alterations in homologous recombination DNA repair genes. Nat interpreter annotates the biological and clinical relevance of Commun 2017;8(1):857. tumor alterations. Genome Med 2018;10(1):25. 18. Wang K, Li M, Hadley D, et al. PennCNV: an integrated hidden 33. Chakravarty D, Gao J, Phillips SM, et al. OncoKB: precision oncol- Markov model designed for high-resolution copy number vari- ogy knowledge base. JCO Precis Oncol 2017;1. ation detection in whole-genome SNP genotyping data. Genome 34. Raeder MB, Birkeland E, Trovik J, et al. Integrated genomic anal- Res 2007;17(11):1665–74. ysis of the 8q24 amplification in endometrial cancers identi- 19. Van Loo P, Nordgard SH, Lingjaerde OC, et al. Allele-specific copy fies ATAD2 as essential to MYC-dependent cancers. PLoS One number analysis of tumors. Proc Natl Acad Sci U S A 2010;107(39): 2013;8(2):e54873. 16910–5. 35. Kilaru V, Knight AK, Katrinli S, et al. Critical evaluation of copy 20. Voisin S, Almen MS, Zheleznyakova GY, et al. Many obesity- number variant calling methods using DNA methylation. Genet associated SNPs strongly associate with DNA methylation Epidemiol 2020;44(2):148–58. changes at proximal promoters and enhancers. Genome Med 36. Fizazi K, Greco FA, Pavlidis N, et al. Cancers of unknown 2015;7:103. primary site: ESMO clinical practice guidelines for diagno- Downloaded from https://academic.oup.com/bib/article/23/5/bbac161/6582004 by guest on 31 January 2023 21. Aryee MJ, Jaffe AE, Corrada-Bravo H, et al. Minfi: a flexi- sis, treatment and follow-up. Ann Oncol 2015;26(Suppl 5): ble and comprehensive bioconductor package for the anal- v133–8. ysis of Infinium DNA methylation microarrays. Bioinformatics 37. Conway AM, Mitchell C, Kilgour E, et al. Molecular character- 2014;30(10):1363–9. isation and liquid biomarkers in carcinoma of unknown pri- 22. Olshen AB, Venkatraman ES, Lucito R, et al. Circular binary mary (CUP): taking the ‘U’ out of ‘CUP’. Br J Cancer 2019;120(2): segmentation for the analysis of array-based DNA copy number 141–53. data. Biostatistics 2004;5(4):557–72. 38. Ross JS, Sokol ES, Moch H, et al. Comprehensive genomic pro- 23. Johann PD, Jager N, Pfister SM, et al. RF_Purify: a novel tool filing of carcinoma of unknown primary origin: retrospective for comprehensive analysis of tumor-purity in methylation molecular classification considering the CUPISCO study design. array data based on random forest regression. BMC Bioinform Oncologist 2021;26(3):e394–402. 2019;20(1):428. 39. Gatalica Z, Xiu J, Swensen J, et al. Comprehensive analysis of 24. Ventura RA, Martin-Subero JI, Jones M, et al. FISH analysis for cancers of unknown primary for the biomarkers of response the detection of lymphoma-associated chromosomal abnormal- to immune checkpoint blockade therapy. Eur J Cancer 2018;94: ities in routine paraffin-embedded tissue. J Mol Diagn 2006;8(2): 179–86. 141–51. 40. Clynick B, Dessauvagie B, Sterrett G, et al. Genetic characterisa- 25. Wolff AC, Hammond MEH, Allison KH, et al. Human epidermal tion of molecular targets in carcinoma of unknown primary. J growth factor receptor 2 testing in breast cancer: American Soci- Transl Med 2018;16(1):185. ety of Clinical Oncology/College of American Pathologists clin- 41. Varghese AM, Arora A, Capanu M, et al. Clinical and molecular ical practice guideline focused update. J Clin Oncol 2018;36(20): characterization of patients with cancer of unknown primary in 2105–22. the modern era. Ann Oncol 2017;28(12):3015–21. 26. Noonan SA, Berry L, Lu X, et al. Identifying the appropriate 42. Loffler H, Pfarr N, Kriegsmann M, et al. Molecular driver alter- FISH criteria for defining MET copy number-driven lung ade- ations and their clinical relevance in cancer of unknown pri- nocarcinoma through oncogene overlap analysis. J Thorac Oncol mary site. Oncotarget 2016;7(28):44322–9. 2016;11(8):1293–304. 43. Ross JS, Wang K, Gay L, et al. Comprehensive genomic profiling 27. Simons A, Shaffer LG, Hastings RJ. Cytogenetic nomenclature: of carcinoma of unknown primary site: new routes to targeted changes in the ISCN 2013 compared to the 2009 edition. Cyto- therapies. JAMA Oncol 2015;1(1):40–9. genet Genome Res 2013;141(1):1–6. 44. Palma NA, Ali SM, O’Connor J, et al. Durable response to Crizo- 28. Santarius T, Shipley J, Brewer D, et al. A census of amplified and tinib in a MET-amplified, KRAS-mutated carcinoma of unknown overexpressed human cancer genes. Nat Rev Cancer 2010;10(1): primary. Case Rep Oncol 2014;7(2):503–8. 59–64. 45. Lee KK, Kim M, Kim KM, et al. Next-generation sequencing for 29. Pertesi M, Ekdahl L, Palm A, et al. Essential genes shape cancer better treatment strategy of cancer of unknown primary (CUP). genomes through linear limitation of homozygous deletions. Ann Oncol 2019;30:v766. Commun Biol 2019;2:262. 46. Holzhauser S, Lukoseviciute M, Andonova T, et al. Targeting 30. Cheng J, Demeulemeester J, Wedge DC, et al. Pan-cancer anal- fibroblast growth factor receptor (FGFR) and phosphoinositide 3- ysis of homozygous deletions in primary tumours uncov- kinase (PI3K) Signaling pathways in medulloblastoma cell lines. ers rare tumour suppressors. Nat Commun 2017;8(1):1221. Anticancer Res 2020;40(1):53–66. https://doi.org/10.1038/s41467-017-01355-0. 47. Holzhauser S, Lukoseviciute M, Papachristofi C, et al. Effects 31. Quinton RJ, DiDomizio A, Vittoria MA, et al. Whole-genome of PI3K and FGFR inhibitors alone and in combination, and doubling confers unique genetic vulnerabilities on tumour cells. with/without cytostatics in childhood neuroblastoma cell lines. Nature 2021;590(7846):492–7. Int J Oncol 2021;58(2):211–25.