SCIENCE ADVANCES | RESEARCH ARTICLE CANCER Copyright © 2020 The Authors, some Matching cell lines with cancer type and rights reserved; exclusive licensee subtype of origin via mutational, epigenomic, American Association for the Advancement and transcriptomic patterns of Science. No claim to original U.S. Government Works. Distributed Marina Salvadores1, Francisco Fuster-Tormo1,2, Fran Supek1,3* under a Creative Commons Attribution Cell lines are commonly used as cancer models. The tissue of origin provides context for understanding biological NonCommercial mechanisms and predicting therapy response. We therefore systematically examined whether cancer cell lines License 4.0 (CC BY-NC). exhibit features matching the presumed cancer type of origin. Gene expression and DNA methylation classifiers trained on ~9000 tumors identified 35 (of 614 examined) cell lines that better matched a different tissue or cell type than the one originally assigned. Mutational patterns further supported most reassignments. For instance, cell lines identified as originating from the skin often exhibited a UV mutational signature. We cataloged 366 “golden set” cell lines in which transcriptomic and epigenomic profiles strongly resemble the cancer type of origin, further proposing their assignments to subtypes. Accounting for the uncertain tissue of origin in cell line panels can change the interpretation of drug screening and genetic screening data, revealing previously unknown genomic determinants Downloaded from http://advances.sciencemag.org/ on February 11, 2021 of sensitivity or resistance. INTRODUCTION However, even if the genetic identity of the cell line is correct, its Cell lines are an important research tool, often used in place of properties may not match the cancer type it is meant to model. In primary cells and intact organisms to study biological processes. particular, the tissue of origin might be incorrect. One way in which Cell lines are used for various applications such as testing drug this error could arise is that tumors thought to originate in a cer- metabolism and cytotoxicity, study of gene function, generation of tain tissue might be metastatic lesions originating from a distal site artificial tissues, and synthesis of biological compounds (1). In (8). Cell lines derived from these tumors would have a different tissue/ cancer research, cell lines derived from tumors are commonly used cell type identity than that assigned at isolation, constituting a case as models, because they are presumed to carry the genomic and of mislabeling. It is conceivable that, also in the case of primary tu- epigenomic alterations that arise in tumors (2). To understand the mors, ambiguous histological or anatomical features may cause the response of tumors to therapy, many studies have linked genetic tumor type or subtype to be misdiagnosed and therefore also for a and/or epigenetic alterations with drug response across cell line cell line derived from that tumor. Furthermore, the process of es- panels, generating datasets such as the Genomics of Drug Sensitivity tablishing the culture might select for a rare cell type that is not in Cancer (GDSC) (3) and the Cancer Cell Line Encyclopedia (CCLE) representative of the tumor isolate on the whole, meaning that the (4). These efforts have advanced our understanding of tumor biology derived cell line would again effectively be mislabeled with a differ- by generating a massive resource of genomic, transcriptomic, ent cell type (9). In addition to the initial changes upon adaptation epigenomic, and drug response data for hundreds of cell lines (2). to culture, cell lines evolve over time due to selection and due to ge- As a model for cancer, cell lines are cost-effective, convenient, netic drift, potentially diverging from the characteristics of the origi- and amenable to high-throughput screening (1, 2). However, a major nating tissue (1). question associated with the use of cell lines is whether they are repre- Tissue and/or cell type is a key determinant of response of cul- sentative of the cancer they are meant to model, which may be tured cells to a variety of experimental conditions, including drug complicated by issues of misidentification (1, 2, 5). exposure and genetic perturbation (10, 11). Therefore, having accu- Misidentified cell lines may lead to inconsistent conclusions across rate information on the tissue and cell type identity of a tumoral cell studies using the affected cell lines. For instance, the cell lines re- line is important for interpreting the experimental results obtained ferred to as HEp-2 and INT 407 in the literature are commonly using these cell lines. cross-contaminated with HeLa (cervical cancer) cells, rather than Recent work has examined cell line panels of certain cancer being laryngeal cancer and normal intestinal epithelium cells, re- types, showing discrepancies between the features of cell lines and spectively (6, 7). Because of the potential for contamination, demon- corresponding tumor types or subtypes. For example, a gene ex- strating cell line identity via genetic markers is now a routine pression analysis of lung tumors and lung cell lines (9) suggested quality-control step. Current resources based on large-scale cancer that some lung adenocarcinoma cell lines did not resemble adeno- cell panels are therefore largely unaffected by this issue (4). carcinoma tumors but instead clustered with other lung tumor subtypes (small cell and squamous cell tumors). A study of high-grade serous ovarian cancer cell lines that used gene expression, driver gene 1 Institute for Research in Biomedicine (IRB Barcelona), The Barcelona Institute of mutations, and copy number alteration (CNA) data reported that Science and Technology, Barcelona, Spain. 2MDS Research Group, Institut de Recerca two frequently used cell lines showed poor genetic similarity to Contra la Leucèmia Josep Carreras, Institut Català d'Oncologia-Hospital Germans molecular profiles of this ovarian cancer subtype (12). A study of a Trias i Pujol, Universitat Autònoma de Barcelona, Badalona, Spain. 3Institució Catalana de Recerca i Estudis Avançats (ICREA), Barcelona, Spain. panel of renal cancer cell lines compared their CNA to kidney *Corresponding author. Email: fran.supek@irbbarcelona.org tumors, finding that some cell lines used as models of the clear cell Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 1 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE carcinoma more closely resemble papillary renal cell cancer (13). tumor and cell line data, we quantile-normalized the data and applied These examples highlight the need to systematically identify the cell ComBat, a batch effect correction method (19), which is highly lines whose genotype and/or molecular phenotypes do not resem- performant compared to other related methods (20). In brief, ComBat ble the characteristics of the matched human tumor type. A major estimates parameters for location and scale adjustment of each batch challenge in the use of human tumor molecular data to classify cell (TCGA and CL in our case) for each gene. Then, it removes the lines are the widespread global changes in gene regulation between variability that is particular to the CL but not present in TCGA, cell lines and tumors that arise in cell culture conditions. while retaining the intra-dataset variability of the tumors, which We performed a systematic analysis that aligned mRNA expression should presumably be evident in both the tumor and in the cell line and DNA methylation data between ~600 cancer cell lines and ~9000 datasets. tumors from 22 different cancer types, adjusting for global differ- A principal components analysis (PCA) in the data (before and ences in transcriptomes and epigenomes. Classifiers trained on after adjustment) suggests that there were strong global differences human tumor mRNA and DNA methylation profiles were used to between TCGA and CL, which are largely removed by our approach identify those cell lines whose genomic and epigenomic profiles are (Fig. 1A and fig. S1A). To quantify this, we calculated the effect size highly consistent with human tumors of their declared cancer type (Cohen’s d statistic) for each gene/probe between TCGA and CL of origin. Conversely, we used the same classifiers to identify those before and after adjustment. It can be observed that these differences cell lines that might be mislabeled with an incorrect cancer type or are reduced to the minimum after adjustment (fig. S1C). In addition, that might have diverged from their original tissue and/or cell type we trained a classification model that predicts the CL versus TCGA identity. Our data suggest that tens of cell lines might be epigenetically origin of the data points based on GE and MET (Fig. 1B and fig. Downloaded from http://advances.sciencemag.org/ on February 11, 2021 and/or genetically not consistent with their stated tissue or cell type S1B). The model is able to distinguish CL versus TCGA perfectly of origin, which is an important consideration for experiments that when using the preadjustment datasets [area under the receiver use these cell lines. We demonstrate this by reanalyzing associations operating characteristic (ROC) curve (AUC) = 1/area under the between drug sensitivity and genetic variation in a large panel of cell precision-recall (PR) curve (AUPRC) = 1], while the post-adjustment lines. After explicitly accounting for putative cases of cell lines with datasets do not perform better than random (~0.5 for AUC and ~0 mislabeled tissue identity, many previously unknown associations for AUPRC), suggesting that the cell type–specific signal has been of genes with drug sensitivity or resistance were revealed. largely removed. Last, we tested the optimal number of features (genes/probes) using tumor classifiers and calculating the accuracy in the cell line data (table S1A); we selected 5000 features with the RESULTS highest SD for subsequent analyses. Identification of tissue/cell type of origin for cell lines by Once the data were aligned, we set out to determine which cell a joint analysis with tumors lines have tissue identity not matching the declared tissue of origin The tissue that originated a tumor is well known to be a major (henceforth, “suspect set”) and, conversely, which cell lines have determinant of drug responses, including drugs targeted to specific largely retained their tissue identity (henceforth, “golden set”), by genetic mutations, both in vitro (10, 11) and in vivo (14, 15). Tissue comparing against a large set of tumors from 17 tissues in the TCGA of origin is an important factor in shaping the networks of genetic (Fig. 1C). Using TCGA data, we derived one-versus-rest classifica- interactions in cancer (16), and it also determines the phenotypes tion models (using ridge regression), separately for the GE and the resulting from genetic perturbation (11). Therefore, ascertaining MET data. These two data types were used because they yielded the tissue/cell type identity of cell lines is relevant for interpreting higher accuracies for the one-versus-rest setting than the four different results of various experiments. For this reason, here, we have mutation-based classifier types we tested (fig. S1G); the mutation-­ systematically examined the global features of the transcriptome based classifiers were nonetheless useful for subsequent validation and epigenome to identify the tissue of origin of tumoral cell lines. analyses (see below). Some pairs of cancer types were considered During the process of adaptation to cell culture, the cells under- jointly based on their overall similarity, for example, stomach go changes in gene regulation that affect many genes (17, 18). The adenocarcinoma (TCGA code: STAD) and esophageal adenocarci- global alterations in gene expression and DNA methylation mean noma (subset of samples from TCGA code: ESAD); see Materials that it is not straightforward to directly compare cell line transcrip- and Methods for a full list. Our study examines solid cancer types tomes and epigenomes with data obtained from tumors. To adjust and blood cancers in separate analyses. for these cell culture–associated shifts, we introduce a computational Differential tumor purity across the TCGA does not have a methodology—HyperTracker—which can unify transcriptome, notable bearing on our tissue classification: Accuracies of models epigenome, and mutational data across tumors and cell lines and from higher-purity tumor samples were similar to models from provide robust predictions of tissue, cell type, and subtype identity. other tumor samples (fig. S1D). Moreover, introducing GE data of In particular, we collected gene expression [RNA sequencing healthy tissues [from Genotype-Tissue Expression (GTEx)] into a (RNA-Seq)] and DNA methylation data (microarrays) for 9681 and joint analysis with the TCGA tumor data did not further improve 9039 human tumors, respectively (TCGA), and additionally for the GE classifier (fig. S1F). GE data from healthy tissues, considered 614 cell lines (CL) of various solid cancer types and 69 CL of blood by themselves, were less informative for assigning cancer cell lines tumors. For gene expression data (GE), we examined transcript-per-­ to tissues of origin than GE data from tumors (fig. S1E), consistent million (TPM) normalized counts for the 12,419 genes, where with a tumoral origin of the cell lines. RNA-Seq data could be linked between cell lines and tumors. For DNA Next, we obtained predictions of cancer type identity for each methylation data (MET), we examined beta values for 10,141 probes cell line. For every cancer type, we split TCGA data randomly into from methylation arrays after selecting a single probe per gene pro- training and testing sets, and we used the calculated PR curve of the moter with the highest variance across the dataset. To align human TCGA testing dataset to obtain the precision score for every cell line Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 2 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE A B C Downloaded from http://advances.sciencemag.org/ on February 11, 2021 Fig. 1. Methodology for data alignment and cancer type classification. (A) Principal component (PC) 1 and PC2 of a PCA, in the gene expression (GE) data, before adjustment for batch effects (raw data) and after adjustment [quantile normalization (QN) + ComBat] [see fig. S1 for PCA of DNA methylation data (MET)]. Colors represent the dataset sources (GDSC and CCLE are two sources for the cell line data, and TCGA is the source for the tumor data). (B) ROC curves for classifying tumors versus cell lines in the data before adjustment (orange) and after adjustment (blue) for GE and MET. (C) Schematic overview of the HyperTracker methodology. First, we systematically identified possible mislabeled cell lines using GE and MET data, independently. Second, we used various types of mutation-based data to corroborate the predictions. Third, we further validated the cell lines (CL) suspected to originate from skin using independent data, such as drug sensitivity. CT, cancer type. (details in Materials and Methods; all TCGA-derived precision lines, 60%). For these cell lines, two independent types of evidence— score values are listed in table S1, B and C). The higher the precision, transcriptomes and epigenomes—support that they match their the more likely the cell line is to belong to that particular cancer expected cancer type well, suggesting that these cell lines would be type. As expected, most of the cancer type labels of the cell lines preferred experimental models. Further, we designated as the “silver match the declared tissue of origin of that cell line—they tend to set” those cell lines that have precision ≥0.7 for one classifier (either cluster at high precision values for the cognate cancer type (red dots GE or MET but not both) (n = 131 of 614, 21%). From the remaining in Fig. 2A and fig. S2). However, among these many correctly clas- 117 cell lines, we selected as suspect set those CL that exhibit a pre- sified cell lines (red dots), there are some with similarly high preci- cision of ≥0.7 for both GE and for MET, but for a different cancer sion scores, but which were originally annotated as belonging to type than declared for that cell line (n = 43 of 614, 7% of analyzed another cancer type (Fig. 2A, blue dots with labels shown). A cluster- cell lines) (Fig. 2C and fig. S4C). This set of cell lines may consist ing analysis of the GE and MET values for the genes with the highest either of mislabeled cell lines, where the cancer type of origin is dif- weights in the classification models (Fig. 2B and fig. S3) showed ferent than initially thought, or of heavily diverged cell lines, where that the samples generally cluster by cancer type, but not by CL versus the genomic and/or epigenomic alterations accumulating during TCGA label. Moreover, we observed that the suspect cell lines (i.e., culture have overridden the original cancer type identity. Notably, cell lines with highly confident precision scores to a different cancer cell line cross-contamination issues (21) cannot commonly underlie type) tend to cluster with the newly assigned cancer type, rather than the trends we observe, because the repositories that provided GE with the original one (Fig. 2B). and MET data have used genetic markers to ascertain the identity of In further analyses, we designated as the golden set those cell the cell lines (4). The fact that two classifiers based on independent lines that have precision ≥0.7 (see fig. S4 and Materials and Methods data types—one transcriptomic and one epigenomic—reached the for threshold selection) for both GE and, independently, MET in same predictions adds confidence that these are bona fide cases of their originally declared cancer type (n = 366 of 614 examined cell mistaken tissue/cell type identity. In case of blood cancer classifiers, Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 3 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE A SKCM Rest KIRC Rest CRAD Rest 0.00 0.00 TCGA_precision MET TCGA_precision MET TCGA_precision MET 0.25 0.25 0.25 0.50 0.50 0.50 RF48 0.75 LN229 0.75 0.75 MDST8 VAESBJ COLO741 SW626 ES2 SW1710 1.00 GCT SW684 1.00 1.00 1.00 0.75 0.50 0.25 0.00 1.00 0.75 0.50 0.25 0.00 1.00 0.75 0.50 0.25 0.00 TCGA_precision GE TCGA_precision GE TCGA_precision GE B GE MET C Downloaded from http://advances.sciencemag.org/ on February 11, 2021 COLO741 TCGA_precision > 0.7 GCT ES2 CT MDST8 SW684 Golden set Silver set TCGA_precision > 0.7 CT RF48 Suspected set GE or MET 67 (11%) Undetermined CT DT Norm value Cancer type (CT) Data type (DT) SKCM GYNE ESTAD GDSC TCGA −4 −2 0 2 4 CRAD SARC CCLE Fig. 2. Detection of cell lines suspected to be mislabeled with a different cancer type. (A) TCGA-based precision scores for 614 cell lines were calculated in the MET and GE cancer type classifiers (one-versus-rest) and for 69 blood cancer cell lines. The higher the precision, the higher the confidence that the sample belongs to that particular cancer type (here, showing cases of SKCM, KIRC, and CRAD from left to right; see fig. S2 for the other cancer types). The cell lines that were originally annotated as the cancer type that is being tested are shown in red, and the rest in blue. (B) Heat map showing the 25 genes (GE) and CpG probes (MET) with the highest absolute values of ridge regression coefficients for each of the cancer types in the plot in one-versus-rest classifiers. The suspected skin cell lines are labeled. The cancer types shown are the suspected cancer type [melanoma (SKCM) in this case] and, additionally, the originally declared cancer types of the suspected cell lines [here, esophagus and stomach cancer (ESTAD), sarcoma (SARC), colorectal cancer (CRAD), and ovarian and uterus cancer (GYNE)]. See fig. S3 for the heat maps for the rest of the suspected cell lines. (C) Overview of the results from the systematic mislabeling testing of all cell lines. Cell lines with a TCGA_precision ≥ 0.7 to its original cancer type in (i) both in GE and in MET are assigned to the golden set group and (ii) either in GE or in MET are assigned to the silver set. If, however, the TCGA-based precision ≥ 0.7 to a different cancer type in GE and in MET, the cell line is assigned to the suspect set. which were highly accurate in distinguishing myeloid and lymphoid validation setting, we applied these genomic classifiers to a problem lineages, there were no cell lines suspected to be mislabeled between of one-versus-one classification, where we contrasted the originally the lineages (fig. S2B). assigned cancer type versus the newly proposed cancer type for each reassignment. We found that these one-versus-one classifiers based Validation of individual examples of suspected mislabeled on genomic data had satisfactory accuracy with whole-exome cell lines using genomic classifiers sequencing (WES) datasets that we used [fig. S5; notably, our recent We detected 43 cell lines that bear both transcriptomic and epigenomic work (22) suggests that whole-genome sequences, when available, features of a different cancer type than the one they were originally would provide more powerful classifiers that draw on regional muta- annotated with. We next turned to support individual examples of tion density (RMD) patterns]. reassigned tissue identity by analyzing independent data. In partic- For the 43 examples of suspect cell line tissue identity, we first ular, we used genomic sequence–based classifiers, which are able to derived one-versus-one classification models separately for GE and predict the tissue of origin based on somatic mutation patterns MET and prioritized reassignments that were consistently observed (22, 23), in particular, the trinucleotide mutation spectra and the across multiple runs of the classification algorithm. We randomized presence of oncogenic mutations and CNA profiles (22, 23). In this the labels to obtain a background model of expected values of the Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 4 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE consistency score for the reclassification (Fig. 3B and fig. S6A). higher confidence. Reassuringly, this list contains two cell lines that From the 43 suspect cell lines, 35 are consistently reassigned to the have been previously shown to be misclassified: SW626, which was other tissue (score > 10) (Fig. 3A and fig. S6B). Next, we calculated initially annotated as ovarian cancer but later discovered to be the same score for the genomic classifiers (based on mutations and derived from colon cancer (24), and COLO741, which was originally CNA, as described above) on these 35 suspect cell lines (Fig. 3A). thought to be a colon adenocarcinoma cell line but later shown to Of these, approximately two-thirds (n = 22 cell lines) received originate from a melanoma (25). The fact that these two known high support for the new tissue label by one or more genomic examples were detected and reassigned to the correct cancer type classifiers [Fig. 3A; consistency score ≥ 15, corresponding to provides evidence that our HyperTracker method is overall reliable. randomization-based false discovery rates (FDRs) of 0, 0, and 18% The two plausible reasons why a cell line thought to originate for the CNA, OGM, and MS96 classifiers, respectively; Fig. 3B]. from one cell type would be reassigned to a different cell type are (i) These data suggest that 22 cell lines are candidates for assignment to that, at the time of isolation, the cell line was not of the type that it another cancer type, based on converging evidence from the levels was thought to be (mislabeling) and (ii) that, during prolonged cell of the genome, epigenome, and transcriptome, which provides culture, the cell line diverged greatly and now resembles another A Genomic B Observed Random A204 0 20 20 1 10 BRCA SARC G402 19 19 NA NA NA BRCA SARC Suspected CT Suspected CT Downloaded from http://advances.sciencemag.org/ on February 11, 2021 0.10 0.09 Original CT Density Density Original CT H1650 20 20 0 20 0 BRCA LUAD 0.06 H1666 3 20 20 18 0 BRCA LUAD 0.05 0.03 OACM51 20 20 0 6 0 BRCA ESTAD 0.00 0.00 RL952 16 18 15 0 0 BRCA GYNE 0 5 10 15 20 0 5 10 15 20 SKUT1 0 19 NA NA NA BRCA SARC CNA MS96 SW626 20 20 20 19 8 CRAD GYNE C 20 SW1710 H1395JHH1DU145 H SW626_2 10 20 6 15 2 ESTAD GYNE H1650 CORL105 R H226 CGTHW1 SW626 W C33A 20 20 20 19 20 GYNE CESC C33A CAL51 20 20 20 6 20 GYNE BRCA H1666 Likely mislabeled SW626_2 Mislabeled JHH1 20 20 20 20 20 GYNE LIHC 15 SKUT1_2 0 20 NA NA NA GYNE SARC DANG SW1710 20 20 0 20 0 KIRC BLCA ECC10 SNU182 G401 MS96 VAESBJ 19 20 NA NA NA KIRC SARC 10 CGTHW1 19 1 12 19 20 LGGBM THCA FU97 ECC10 20 1 1 13 0 LGGBM ESTAD ECC12 20 0 0 2 2 LGGBM ESTAD OACM51 CAL51 G401 19 0 11 10 11 LGGBM KIRC 5 Likely transdif. Likely mislabeled MDST8 18 20 17 0 0 LGGBM CRAD RPMI2650 RPMI2650 20 14 0 2 0 LGGBM SCC ECC12 A204 SNU182 20 1 0 11 2 LGGBM LIHC MDST8 0 RL952 HS578T CORL105 2 20 20 20 0 LIHC LUAD 0 5 10 15 20 FU97 20 20 2 8 0 LIHC ESTAD CNA H1395 20 0 15 20 0 LIHC LUAD D 5.0 E DU145 20 20 20 20 20 LUAD PRAD Signature 7 (UV) − Jarvis et al. 2.5 0.75 H226 20 20 20 19 0 MESO SCC RF48 SW684 SW684 GCT ln IC50 (trametinib) HS578T 20 20 17 0 0 SARC BRCA GCT 0.0 0.50 DANG 20 20 20 13 19 SCC PAAD MDST8 −2.5 COLO741 20 20 NA 13 19 SKCM CRAD ES2 0.25 ES2 19 20 NA 0 11 SKCM GYNE −5.0 ES2 GCT 5 20 NA NA NA SKCM SARC RF48 MDST8_2 20 20 NA 20 20 SKCM CRAD −7.5 MDST8 0.00 RF48 19 0 NA 0 19 SKCM ESTAD −5.0 −2.5 0.0 2.5 5.0 0.0 0.1 0.2 0.3 ln IC50 (dabrafenib) Signature 7 (UV) − Petljak et al. SW684 0 20 NA 20 20 SKCM SARC SKCM Rest Suspected Not suspected be l la New E ET A 6 M la ina S9 G N G l l be M C r ig O M O Fig. 3. Additional evidence supporting tissue identity of the suspected mislabeled cell lines. (A) Prediction consistency score (0 to 20) for each suspected cell line for 20 runs of one-versus-one classifiers that predicted suspected versus original cancer type in GE, MET, CNA, trinucleotide mutation spectrum (MS96), and oncogenic mutations (OGM). A value of 20 means that the cell line is predicted as suspected consistently in the 20 runs of the classification algorithm, and a value of 0 means that it is predicted as original cancer type. (B) Histograms of the consistency scores for CNA and MS96 classifiers for the models based on actual data and a baseline expectation on randomized data. (C) Prediction scores for MS96 and CNA for the suspected cell lines. Colors represent the suspected cancer type [see column “new” in (A)]. Gray dots represent the random values. (D) Drug sensitivity (IC50) for mutant BRAF-targeting drugs dabrafenib and trametinib for 614 cell lines. Cell lines originally labeled as skin cancer are shown in turquoise, and skin-suspected cell lines are marked with a square and their name. (E) Burden of UV-associated mutation signature 7 (estimated from two different sources) in 614 cell lines. Cell lines originally labeled as skin cancer are shown in turquoise, and skin-suspected cell lines are marked with a square and the name label. Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 5 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE cell type (transdifferentiation). Our data allow us to examine how MDST8), providing further supporting evidence that these are of prevalent each case is: Mislabeling is expected to be reflected equally skin, likely melanoma, origin. in both the epigenome and the genome, while transdifferentiation is In conclusion, out of six cell lines suspected to originate from expected to be reflected more strongly in the (presumably more skin, four were confirmed by the UV mutational signatures and malleable) epigenome, and less so in the genome, which retains the two were additionally confirmed by the drug sensitivity to BRAF/MEK mutations from the original tumor. We suggest that mislabeling at (mitogen-activated protein kinase kinase) inhibitors. This notable isolation is a much more common scenario (Fig. 3C, many reassigned example demonstrates how the transcriptome- and epigenome-based cell lines are in the upper right corner). However, it is possible that tissue/cell type classifiers are able to link cultured human cell lines there exist individual examples of cell lines that have effectively with their correct cancer type of origin. transdifferentiated during culture, because their genomic features In addition to these examples of skin cell lines, we were able to are consistent with the original tissue identity, while the epigenomic support several other cancer type reassignments using drug sensi- features are consistent with another tissue (Fig. 3C, lower left corner; tivity (results are summarized in table S1D) (32). The DANG cell e.g., the RPMI2650 and OACM51 cell lines are possible candidates). line is consistent with squamous cell carcinoma of the lung or of the head and neck (SCC), rather than with its original assignment of Validation of cell lines suspected to originate from the skin pancreatic adenocarcinoma (notably, this reassignment is also ob- From the previous analysis, we identified six cell lines that are reas- served with multiple genomic classifiers; Fig. 3A). Similarly, SW1710 signed from various cancer types to skin cancer. We note that, of may be a kidney, rather than a bladder, cell line, based on the original skin cancers, the TCGA study contains only melanoma but not the reassignment via transcriptome and epigenome, supported by the Downloaded from http://advances.sciencemag.org/ on February 11, 2021 nonmelanoma skin cancers, so we were not able to distinguish mutation patterns (Fig. 3A) and additionally supported in the global between cell type identities of different skin cancers. analysis of drug responses (table S1D). We note that such analyses To further support that these cell lines are skin cells, we performed of drug screening data can be applied to distinguish only certain an independent analysis based on mutational signatures. Large-scale pairs of tissues and not all reassignments can be reliably validated in analyses of trinucleotide mutation spectra across human tumors this test (see AUC scores in table S1D). have revealed at least 30 different types of mutational signatures (26). Of these, signature 7 (C>T changes in CC and TC contexts) Identification of subtypes for cell lines using was associated with exposure to ultraviolet (UV) light and is highly multi-omics analyses abundant in sun-exposed melanoma tumors (27). The same signa- Tumors are heterogeneous, and major differences exist between tures were recently estimated in cancer cell lines (28, 29), which tumor samples of the same cancer type. To manage this variability, enabled us to use the existence of UV-linked signature 7 to examine cancer types are subdivided on the basis of their molecular charac- whether these cell lines originated from the skin. On the basis of teristics, including global patterns in gene expression and DNA mutational burden of signature 7, the known melanoma cell lines methylation (33–35). However, with the exception of a few tumor are clearly separated from the rest (Fig. 3E), meaning the approach types, prominently breast cancer, molecular subtypes are still being can distinguish skin-derived cells. Among the melanoma cell lines established or refined, often with the goal of better predicting with high mutational burden of signature 7, we found four of five of disease progression in response to particular treatments. Because the suspected cell lines (Fig. 3E), in particular, GCT, SW684, ES2, drug screens and genetic screens performed on cell line panels have and MDST8 are very likely skin cell lines, and not sarcoma, sarcoma, the goal of serving as models for actual tumors, it is useful to be able ovarian cancer, or colorectal cancer, respectively, as originally thought. to transfer the subtype assignments from tumors to cell lines, thereby For the sixth suspected cell line COLO741, the mutational signature establishing which cell lines are the most appropriate model for data are not available; however, COLO741 has previously been each cancer subtype. reported to express skin-specific genes (25). Previously, molecular subtypes from tumors have been inferred The RF48 cell line (originally considered stomach, here putatively in cell lines using different strategies, often based on gene expres- reassigned to skin) does not exhibit the UV signature or the DNA sion, for example, in breast (36), colorectal (37), and renal cancer methylation patterns of skin; therefore, a highly confident call cannot (13). In a recent pan-cancer study, subtypes have been assigned to a be made. Nonetheless, a pattern of cancer driver mutations in RF48 set of 600 cell lines (38). suggests that it is not a stomach cell line (Fig. 3A). Previous work Our approach to assign subtypes to cell lines is to apply the same based on gene expression proposed a lymphoid origin for RF48 (30). strategies that yielded accurate cancer type classifiers: first, the inte- Next, we sought to substantiate these findings using drug sensi- gration of transcriptomic and epigenomic data to boost confidence tivity data. In particular, two drugs (dabrafenib and trametinib) that in the predictions, and second, careful adjustment of the datasets to target mutant BRAF are approved for treating melanoma in the make them comparable between TCGA tumors and cell lines (fig. S1). clinic. These drugs are known to have poor efficacy in other cancer An important consideration here is the lack of known labels needed types bearing BRAF mutations, such as in colon cancers (31). to assess accuracy; thus, assignments should be treated as tentative. Therefore, sensitivity to these drugs adds confidence that we are However, for breast cancer cell lines, the subtype labels are available looking at a melanoma cell line (note that the converse does not and can be used as a benchmark (36). necessarily hold here: resistance does not imply that the cell line is We examined subtypes proposed for 15 cancer types in the not a melanoma). Therefore, we compared the median inhibitory TCGA and generated subtype classifiers (see Materials and Methods) concentration (IC50) of these two drugs for all cell lines (Fig. 3D). As for each cancer type. The combination of both data types (GE and expected, many melanoma cell lines cluster at low values of IC50 for MET) achieved a higher cross-validation accuracy in the TCGA the two drugs, meaning that these cells are sensitive to the drug. (median AUPRC across cancer types: 0.81) than GE (0.76) or MET This includes two of five of our suspected cell lines (ES2 and (0.72) individually. Therefore, we used the combined datasets to Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 6 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE generate subtype classifiers and propose assignments of the cell silver set cell lines, 10 significant, previously unknown associations lines to cancer subtypes. Most were uniquely assigned to a single were found (Fig. 4B). For instance, we detected that SMAD4-mutant subtype (fig. S7A); we used only those in further analysis. As a cell lines are more resistant to piperlongumine, a natural product that benchmark, we calculated the accuracy for the breast cancer cell exerts antitumor properties via multiple pathways (41). Mutations lines with subtypes available (fig. S7B): The median AUPRC across in the tumor suppressor gene EP300 were associated with higher sen- subtypes for CL is 0.83. This suggests acceptable performance in sitivity to three drugs in pancreatic cancer cell lines (Fig. 4D). The obtaining tentative subtype assignments for cell lines in all 15 cancer observation that more associations were found despite using a types, which we provide in table S1E. This resource is complemen- somewhat lower number of cell lines (thus less statistical power) tary to a recent set of subtype predictions for nine cancer types emphasizes the importance of focusing on the cell lines that more based on transcriptomes (38). closely model the tissue and/or cell type of origin of the cognate tumor. Next, we examined whether the relative prevalence of subtypes is Colorectal cancer provides an illustrative example of the impor- similar between tumors and cell line panels of the same cancer type. tance of removing nonrepresentative cell lines from drug screening Cell line panels of some cancer types have good representation of efforts. In the original study (10), 50 colorectal adenocarcinoma subtypes, for instance, lung squamous cell cancer, head and neck (CRAD) cell lines were tested. Of those, we strongly suspected that squamous cell cancer, lung adenocarcinoma, and gastric/esophageal MDST8 derives from skin. To test the influence of this individual cancers (fig. S7C). However, the converse is the case for liver, skin, mislabeled cell line, we have performed association testing with all and thyroid cancer cell lines, where a single subtype predominates colorectal cell lines and again after excluding MDST8. For the asso- in cell line panels, unlike in tumors (statistics are listed in table S1F). ciation between dabrafenib response and BRAF mutation status, we Downloaded from http://advances.sciencemag.org/ on February 11, 2021 In addition, we observe suboptimal representation (half of the observed that CRAD cell lines in general (i.e., irrespective of BRAF tumor subtypes are not represented) in the kidney, bladder, and mutation) are not sensitive to dabrafenib, except for MDST8, which brain cancer cell line panels, when considering the 463 cell lines we is strongly sensitive to dabrafenib (Fig. 4A). The FDR of the analysis analyzed. This suggests that, in some cancer types more than others, of variance (ANOVA) analysis when using all (allegedly) colorectal the commonly used cell line panels do not represent the diversity of cell lines is significant at 6%, while when removing MDST8 the FDR molecular subtypes in tumors well. One possible reason for this is for BRAF-dabrafenib association in colon becomes nonsignificant the relative ease of culturing certain subtypes compared to others (2). (45%). Therefore, in this case, the presence of a single mislabeled cell line is sufficient to cause the appearance of an association Accounting for mislabeled cell lines reveals associations between a drug and a feature, which is likely not relevant for this in drug screening data cancer type. This is evidenced in the clinical responses of patients: We detected 35 cell lines that may have a tissue or cell type identity BRAF-mutant melanoma patients respond well to dabrafenib, while different than the one originally assigned to them. Because the cell colorectal tumors with the same BRAF V600 mutation are not type is an important determinant of drug response in cancer cell sensitive to BRAF or MEK inhibitor monotherapy (31). lines and in tumors (10), we hypothesized that the inclusion of this new tissue information when searching for genetic determinants of Accounting for mislabeled cell lines in genetic screening drug sensitivity may change the results. In a comprehensive study, data analyses Iorio et al. (10) searched for associations between drug response Motivated by the associations revealed by reanalyzing the drug and cancer functional events (CFEs): recurrent mutations, CNA, screening data, we asked whether the same extends to genetic and hypermethylation events present in human tumors. Here, we screening data in cancer cell lines, because results in genetic screens used GDSCTools (39) to reproduce the results of that study, however, also depend on cell lineage (11). To further investigate, we analyzed after filtering the cell lines to those that better represent the cancer CRISPR screening data from Project Score and Project Achilles (see type in question. In particular, we repeated the same analysis using Materials and Methods), from which 347 cell lines overlap our tested for each tissue (i) all the cell lines, (ii) only the cell lines in the golden cell lines. Then, we applied the same association testing method, set (G), and (iii) as a less stringent filtering criterion, only the cell which was, however, underpowered, because the number of avail- lines in the golden set and silver set (G&S). In addition, as controls, able overlapping cell lines was smaller. Nonetheless, in colorectal we included a random subset of cell lines that matches (iv) the number and ovarian cancer (which have the largest number of cell lines in of cell lines in golden set (r_G) and (v) the number in “golden and this dataset), we observed that by focusing only on the golden set silver set” combined (r_G&S). and/or silver set, the number of recovered associations increased For most of the cancer types, we observed that one of the filtered (as a control, there were no increases in randomly chosen cell line subsets recovered a higher number of significant [at FDR ≤ 25%, as subsets of the same size; fig. S8B and table S1H). applied in the original study (10)] associations of CFE with drug To illustrate the importance of removing suspect cell lines in sensitivity or resistance than were recovered using all cell lines (fig. gene dependency screenings, we provide two examples of associa- S8A and table S1G). For instance, for glioblastoma, using the golden tions that were originally not significant because of the presence of set cell lines, we found 23 associations, which were not recovered a mislabeled cell line. For ovarian cancer, the presence of SW626 from the entire cell line panel or from the random subset controls [mislabeled cell line confirmed by the literature (24)] prevents find- (Fig. 4B). For example, this recovers the positive association of ing the association between MED8 dependency and a copy number CDKN2A loss with camptothecin sensitivity (Fig. 4C), which was gain in the region containing ASXL1 [“cnaOV72” in (10)] as signifi- previously reported in an independent analysis of the NCI-60 cell cant (fig. S9A). Similarly, for colorectal cancer, the presence of line panel screening data (40). Similarly, for pancreatic adenocarci- MDST8 (mislabeled cell line confirmed by the UV mutational noma, benefits were observed by focusing on cell lines that resemble signature) prevents detection of the association between TUBB4B the corresponding cancer type better: Using only the golden set plus dependency and a copy number gain in the region containing STK4 Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 7 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE CRAD all (n = 44) CRAD w/o MDST8 (n = 43) A 7.5 7.5 B GBM Resistant ANOVA ANOVA 0.05 FDR = 6% FDR = n.s. (45%) 30 Sensitive 20 5.0 5.0 10 1 ln IC50 dabrafenib ln IC50 dabrafenib 0 l G G &S &S Al 2.5 2.5 r_ G G r_ PAAD 0.0 0.0 0.01 10 5 0.38 −2.5 −2.5 MDST8 0 BRAF mut Wild type BRAF mut Wild type l G G &S &S Al r_ G G r_ All − GBM (n = 35) Golden set − GBM (n = 27) C CI−1040 Tipifarnib Tipifarnib 5 loss_cnaGBM2_(CDKN2A) loss_cnaGBM45 4 loss_cnaGBM45 JNK−9L Afatinib Afatinib Downloaded from http://advances.sciencemag.org/ on February 11, 2021 loss_cnaGBM101 loss_cnaGBM68_(CDKN2C,FAF1) 4 loss_cnaGBM68_(CDKN2C,FAF1) BMS−345541 −Log10 (P value) −Log10 (P value) 3 gain_cnaGBM123_(EGFR) 3 2 2 1 1 0 0 −2 −1 0 1 2 3 −2 0 2 Signed effect size Signed effect size All − PAAD (n = 29) Golden & silver sets − PAAD (n = 21) D CHIR−99021 4 SN−38 EP300_mut OSU−03012 TP53_mut SMAD4_mut 3 GSK1070916 3 CDKN2A_mut −Log10 (P value) −Log10 (P value) BMS−754807 TP53_mut 2 2 1 1 0 0 −2 −1 0 1 2 −3 −2 −1 0 1 2 Signed effect size Signed effect size n.s. Resistant Sensitive Fig. 4. Drug sensitivity association testing using high-confidence sets of cell lines. (A) Drug sensitivity (IC50) to dabrafenib in all colorectal (CRAD) cell lines (left) and all CRAD cell lines except MDST8 (right), which is suspected of being skin cancer. Cell lines with a BRAF mutation and without (wild type) are compared. ANOVA FDR for this association (dabrafenib sensitivity with BRAF mutation status) is shown in blue for both datasets. Horizontal line is shown at 0, because score < 0 implies sensitivity to the drug. Dots and error bars represent the mean and SEM. (B) Number of significant associations between “CFEs” (includes mutations and CNAs in cancer genes) and drug associations detected (at FDR 25%) in the ANOVA test for all cell lines (“all”), cell lines in the golden set (“G”), cell lines in the golden plus silver sets (“G&S”), random subset of cell lines that match the number in the golden set (“r_G”), and random subset of cell lines that match the number in the golden plus silver sets (“r_G&S”). For the random subsets, the number of significant associations is calculated from 10 random selections and median shown. P values for a sign test (one-tailed) between the number of associations in the G/G&S and in r_G/r_G&S are shown. See fig. S7 for the remaining cancer types. (C) Differential sensitivity of drugs was analyzed by ANOVA for all brain cancer cell lines (left) and the brain cancer cell lines in the golden set only (right). Each point is an association between the sensitivity of a drug and a genetic feature (CFE). (D) Differential sensitivity of drugs was analyzed by ANOVA for all pancreatic (PAAD) cell lines (left) and PAAD cell lines in the golden and silver set only (right). Each point is an association between the sensitivity of a drug and a genetic feature (CFE). n.s., not significant. (“cnaCOREAD32”) (fig. S9B). Next, a significant association between growth factor receptor (EGFR) protein and is clinically indicated WRN dependency and MLL2 (also known as KMT2D) gene muta- for EGFR-mutated lung cancer; however, in EGFR-altered glioblas- tion is recovered only with the filtered cell lines in ovarian cancer toma, afatinib is generally not considered to elicit a response (43). (fig. S9C). This WRN-MLL2 association has been recently reported Consistently, afatinib sensitivity was associated with EGFR alterations using a different set of cell lines (from Project Score) (42) that par- in lung cancer previously (10), as well as in our reanalysis (FDR lung tially overlap our set. G&S sets = 0.6%), but not in the brain cell line panel (all cell lines, Last, our reanalyses of drug screening and genetic screening data FDR ≥ 25%). However, using the focused (golden set) of brain cancer cell revealed an interesting association independently supported in lines revealed a significant association (ANOVA FDR = 15%; fig. S9D) both drug and genetic data. The drug afatinib inhibits the epidermal between afatinib sensitivity and a different genetic alteration: copy Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 8 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE number loss in a region at 1p32.3 containing the CDKN2C and ferences between cell lines and tumors, which might have resulted, FAF1 genes [“cnaGBM68” in (10)]. The same loss at 1p32.3 is asso- e.g., from impurities in tumors or from altered expression of cell ciated with sensitivity to genetic knockout of EGFR in brain cell line cycle–related genes in cell lines (38). Parallel analyses of different panels in two independent large-scale genetic screens (Project Scores omics data provide increased confidence in our inferences, which and Project Achilles; fig. S9, F and G) and to pharmacological inhibi- suggested, remarkably, that 5.7% (35 of total 614 considered cell tion in another drug screen (PRISM; fig. S9E). The meta-analysis of the lines) exhibit significant transcriptomic and epigenomic features of two drug screens and two genetic screens suggests high strength a different tissue/cell type than the declared cell type of origin. For of combined evidence (P = 0.00094, Fisher’s method of combining 3.6% (22 cell lines), these reassignments to a different cancer type P values) linking the CNA loss at 1p32.3 (chr1: 51169045-51472178) were additionally supported in at least one type of genomic evidence. with sensitivity to pharmacological or genetic EGFR inhibition in This increased confidence that these were examples of cell lines brain cells, suggesting a candidate marker for follow-up work. with mislabeled (or, less likely, diverged) tissue/cell type identity. In summary, the presence of cell lines with dubious or incorrect Notable examples are cell lines GCT, SW684, ES2, and MDST8 that labels of tissue identity may strongly affect association studies of we predict to originate from the skin, based on the presence of drug or CRISPR screening data in two different ways. First, the the UV mutational signature, in addition to strong evidence in presence of mislabeled cell lines can cause the appearance of spurious transcriptome/epigenome data. These cases are reminiscent of the associations that do not reflect the biology of the cancer type of recent reports of UV mutational signatures found in some cases of pre- interest. Second, the presence of mislabeled or divergent cell lines sumed lung cancers, suggesting that they may instead be metastases can prevent the recovery of true associations. originating from sun-exposed skin (44). Downloaded from http://advances.sciencemag.org/ on February 11, 2021 In interpreting our data, an important consideration is that the cancer sample types in TCGA may not necessarily reflect the full diver- DISCUSSION sity of rarer subtypes within a cancer type, which may cause some Cell lines are commonly used as models for tumors; however, it is ambiguous predictions. For instance, the ECC10 and ECC12 cell lines an open question how to best apply the available cell line panels to are assigned to STAD cancer type (stomach adenocarcinoma) when learn about cancer biology. The availability of genomic data from matched with TCGA tumors. These cell lines originate from gastric large tumor cohorts and from cell line panels has spurred efforts to small cell neuroendocrine carcinomas. This might explain why, in our find which cell lines are closer to tumors by their transcriptomic analysis, gene expression patterns point toward brain tissues, while (9, 37, 38) and/or genomic features (12, 13), presumably making better mutational features suggest stomach cancer. In such cases of disagree- models, and which are more distant from examples of actual tumors, ment between different types of features, a future use of a more ex- thus making less good models. haustive set of reference tumor data, which includes rarer cancer types, Our work addresses a different question: We attempt to detect may resolve the ambiguity and improve confidence in predictions. the cancer type (i.e., tissue and/or cell type) that originated the cell The genomic classifiers we used here were based on whole-exome line to ascertain whether this matches the declared origin of the cell sequences and were overall less powerful than the transcriptome/ line. A mismatch may conceivably stem from the sampling step, for DNA methylation classifiers in our data (figs. S1G and S5). Recent instance, a metastasis might have a different tissue of origin than work by us and others (22, 23) suggest that analyzing whole-genome thought at the time of surgical collection. The work-up after collecting sequences of these cell lines would permit use of additional, highly the tumor sample may have inaccurately assigned the cancer type, predictive features based on RMD of chromosomal domains. This based on unclear histological or anatomical features. Another possi- may provide further genomic evidence for the identity of the cell of bility is that the mismatch might stem from the step of adaptation to origin for the 35 suspected cell lines. Furthermore, targeted experi- cell culture, where a minority cell type that is not representative of mental follow-up work on these cell lines will provide further the tumor prevails over other tumoral cells. We consider these to be evidence to support or refute our predictions, which are based on cases of cell line mislabeling during isolation. In addition, we would global, multi-omics analyses. also detect cases where the cell line might have acquired some prop- Knowing the correct tissue-of-origin label for a cell line is im- erties of a different tissue/cell type during extended periods in culture; portant, because this has a strong bearing on the response of the cell however, our analyses (Fig. 3C) suggest that this is a less common line to drug treatment and to genetic perturbation. We demonstrate occurrence, although individual examples cannot be ruled out. the implications of this general principle to analyses of drug and This phenomenon of tissue/cell type mislabeling is distinct from genetic screening data: By accounting for suspect cell lines, the power well-known and widespread cell line misidentification issues (21), to discover determinants of sensitivity to pharmacological and to where one cell line (often HeLa) was mistakenly used in place of genetic perturbation may increase substantially for some cancer another cell line originating from a different individual, commonly types, such as brain, lung, and pancreatic cancers. Therefore, when due to cross-contamination. The cell line panels that provided data designing future screening efforts, it is important not only to increase used in our analyses (GDSC and CCLE) have authenticated their the number of cell lines to gain more power but also to focus on the cell lines (4, 42); thus, misidentification/cross-contamination cannot cell lines that best reflect the tissue and/or cell type of interest. underlie our observations of mislabeling of the cancer type of origin. [We note that there were rare cases of misidentified cells reported in these panels (42); however, these do not overlap our results.] MATERIALS AND METHODS Methodologically improving over previous work, we introduce Omics data collection and preparation the HyperTracker framework that performs global analyses, which DNA methylation data independently examine transcriptomic, epigenomic, and mutational We downloaded DNA methylation data as beta values (platform: features. In addition, we carefully adjust for the known bulk dif- Illumina HumanMethylation450) from the GDC Data Portal (45) Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 9 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE for TCGA samples and from GDSC (3) for CL samples. We filtered commonly grow in suspension, making them less likely to be out all probes outside promoter regions and probes with not avail- confounded with solid tumors. For further analysis, we merged the able (NA) values in more than 100 samples. For the probes in pro- cancer types that were overall similar: HNSC with LUSC and ESCC moter regions, we selected only one probe per gene that had the (jointly known as SCC), GBM with LGG (LGGBM), STAD with highest SD across samples. We transformed the beta values to m-values ESAD (ESTAD), and OV with UCEC (GYNE). (log2 ratio of the intensities of methylated probe versus unmethylated The identification of the cell line samples was performed by the probe). In total, this yielded 10,141 features for 942 CL samples laboratories generating the cell line databases, using short tandem and 8453 TCGA samples. repeat analysis (4, 42). They reported a few commonly misidentified Gene expression data cell lines: Ca9-22, RIKEN, MKN28, KP-1N, OVMIU, and SK-MG-1 We downloaded gene expression data as TPM from GDC Data (42). These cell lines do not overlap with our suspected samples, and Portal (45) for TCGA samples, from GDSC (3) and the CCLE (4) additionally, the misidentification does not affect tissue or cancer for CL samples, and from GTEx Portal for healthy data. We filtered type of origin. out genes with NA values in more than 100 samples and selected the overlapping genes between the four sources. We removed low Data alignment between tumors and cell lines expressed genes (TPM < 1 in 90% of the samples) and applied a For the alignment of TCGA and CL data, we first applied quantile square-root transformation to the TPM data. In total, we have normalization (R package preprocessCore 1.46.0) and next applied 12,419 features for 942 CL samples and 9149 TCGA samples. ComBat (R package sva 3.32.1), a batch effect correction method. For both DNA methylation (MET) and gene expression (GE), We used ComBat as if our dataset was the TCGA and CL data com- Downloaded from http://advances.sciencemag.org/ on February 11, 2021 we created datasets of different sizes: 1000, 2000, 3000, 5000, and bined, and the batch effects labels were whether a sample belongs to 8000 features by selecting the genes/probes with the highest SD TCGA or CL (for MET) or a sample belongs to TCGA, GDSC, or across TCGA samples. In addition, we downloaded tumor purity data CLLE (for GE). We applied this method for GE, MET, CNA, MS96, from Aran et al. (46) and used consensus measurement of purity and RMD. For validation, we calculated a PCA, subsampling TCGA estimations (CPE) for creating the groups of high- versus low-purity data to match the number of CL samples (stratified by cancer types). samples. High-purity samples were defined as CPE > 0.75, and for In addition, we calculated Elastic Net classifiers to predict (in the the lower-purity group, we took a subset of samples (same number processed dataset) TCGA versus CL and calculated the AUC and of samples that are in high-purity subset) with the lowest purity. AUPRC to check whether the process of alignment is being successful CNA data or not. In addition to the chosen adjustment method, we tested other We downloaded CNA data (computed by gene) from GDC Data approaches based on canonical correlation analysis, partial least Portal (45) for TCGA samples and from DepMap (47) for CL sam- squares, and PCA, which did not exceed accuracy of ComBat and ples. In total, we have 20,491 features for 942 CL samples and 9188 therefore were not examined further. TCGA samples. To reduce the dataset, we selected CNA events in 299 known cancer driver genes (48). Cancer type classifiers Mutation data For the TCGA dataset, we generated ridge regression model for pre- For human tumors, we downloaded mutation data as WES MC3 dicting the cancer type in a one-versus-rest manner (using cv.glmnet dataset (49) from the GDC Data Portal for TCGA samples. For cell lines, function with alpha = 0 and family = binomial, R package glmnet aligned short reads (bam files) were obtained from the European 2.0.18). To calculate the accuracy, we trained classifiers in the TCGA Genome-phenome Archive (ID number: EGAD00001001039). Variant dataset and tested in the CL dataset. In particular, we calculated the calling was performed using Strelka (version 2.8.4) with default pa- AUC and the AUPRC for each cancer type versus the rest (all the rameters. Variant annotation was performed using ANNOVAR rest of cancer types combined). (version 2017-07-16). In samples where Strelka was unable to run, a TCGA_precision score realignment was performed using Picard tools (version 2.18.7) to For each cell line, we calculated a TCGA_precision score of belong- convert the bams to FASTQ, and following that, the alignment was ing to a particular cancer type. For this, we divided the TCGA data performed by executing bwa sampe (version 0.7.16a) with default into two datasets (training and testing) of the same size, keeping the parameters. The resulting bam files were sorted and indexed using cancer type proportions. For each cancer type, we trained classifiers Picard tools. To account for germline variants, we removed all in the TCGA training dataset, and we introduced the cell lines one mutations that were present in the gnomAD v2.1.1 database (50) at by one with the testing data and calculated the PR curve (TCGA an allele frequency of ≥0.001 in any of the populations. Last, using testing + 1CL). We set the cell line precision score for that specific the filtered somatic mutations, we calculated three sets of mutational cancer type as the precision at the threshold where the cell line is features: RMD, mutation spectra (MS96), and oncogenic mutations situated in the PR curve. Overall, for every cell line, we obtained (OGM) as described by Salvadores et al. (22). RMD features did not 17 precision scores, 1 for each possible cancer type. We repeated exhibit high accuracy when applied to exome-sequencing data and this procedure five times and calculated the median precision for so were not considered further in this analysis. every case to get more robust values. In addition, when training for For the cell line samples, we matched their cancer types to the one cancer type (label = 1) versus the rest of cancer types combined TCGA cancer types using the cell line metadata from GDSC (3) and (label = 0), we made some exceptions and removed those cancer manually annotated those that did not have a TCGA cancer type types that are similar, and therefore, the classifier is not good at separat- label using Cellosaurus (51). Next, we selected the cell lines from ing them (e.g., when we calculated precision for ESTAD, we removed solid tumors that had a matching cancer type in TCGA, yielding a from the rest CRAD and PAAD, all hidden cases in table S1I). This total of 614 cell lines from 22 cancer types. Blood cancers (LAML is conservative with respect to reassigning cell lines to another cancer and DLBC) are tested separately, because the cell lines there type; however, some resolution is traded off, because the more Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 10 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE closely related cancer types are, by design, not distinguished. We molecular subtype classifications that have been described by have further attempted to reclassify cell lines within these hidden TCGA-related publications [mRNA, DNA methylation, protein, tissues and the combined ones. However, when using one-versus-one miRNA, CNA, integrative (iCluster), and others]. From those, we classifiers, the accuracy is not good enough for distinguishing the selected the “subtype_selected,” which is the classification that was two cancer types in the cell lines. We selected TCGA-based preci- chosen as the representative one in that cancer type (usually mRNA sion score ≥ 0.7 as a threshold to separate the different sets (golden, or integrative; see documentation of PanCancerAtlas_subtypes func- silver, and suspected sets), because this cutoff value approximately tion for the complete list). We combined the GE and MET datasets maximizes the F1 score for cell line classification (fig. S4A). At the to predict subtypes. For these data, we generated ridge regression TCGA-based precision threshold ≥0.7, the FDRs estimated on the model for predicting the subtypes in a one-versus-rest manner original cell line labels for gene expression and DNA methylation (using cv.glmnet function with alpha = 0 and family = binomial, classifiers would be 28 and 22%, respectively (fig. S4A); because the R package glmnet 2.0.18) within each cancer type. We trained models original labels are not always correct, these FDR estimates are con- in TCGA, and we predicted subtypes for the cell lines. In addition, servative. Also, by visual inspection of the distribution densities of we used cell line’s subtypes for breast cancer from a previous paper putatively correct predictions (originally labeled as the matching (36) to calculate the confusion matrix and the AUPRC. cancer type) and putatively incorrect predictions (originally labeled We performed a chi-square test (R package stats 3.6.0) and as another cancer type) for cell line tissue labels, one can appreciate calculated the Cramer’s V statistic (R package lsr 0.5) for checking how a TCGA_precision ≥70% threshold separates well the two whether the proportion of subtypes between TCGA and CL is main- groups (fig. S4B). tained for each cancer type. Downloaded from http://advances.sciencemag.org/ on February 11, 2021 Once we have a list of suspected cell lines, we have an “original” cancer type and a “suspected” cancer type. Therefore, we generated Drug and CRISPR screening data one-versus-one classifiers (original versus suspected) using TCGA We downloaded drug sensitivity and CFE data from Iorio et al. (10). dataset (balancing the classes), and for each suspected cell line, we CFEs are a collection of recurrent mutations, CNA, and hypermethyla- checked whether it is predicted as original or suspected. We repeated tion events present in human tumors (10). We used GDSCTools this prediction 20 times and counted the number of times a cell line (39) to search for associations between the drugs and the CFEs in is predicted as suspected. Therefore, we defined a consistency score every cancer as they did. In particular, we performed this analysis (range between 0 and 20) for every cell line, where 0 means never using for each tissue (i) all the cell lines, (ii) only the cell lines in the predicted as suspected and 20 always predicted as suspected. As a golden set (G), and (iii) only the cell lines in the golden and silver set control, we repeated the same procedure with randomized cancer (G&S). In addition, as controls, we included a random subset of all type labels, providing estimates of false discovery for different con- cell lines matching (iv) the number of cell lines in the golden set sistency score thresholds (Fig. 3B and fig. S6A). We calculated this (r_G) and (v) the number in golden and silver set combined (r_G&S). prediction score for GE, MET, CNA, OGM, and MS96 datasets. For We counted the number of significant hits (at FDR ≤ 25%) for each calculating the FDR at a score ≥ 15, we applied the following for- of the cancer types. For the controls, we repeated the subsampling mula: FDR = FP/(FP + TP), where FP is the number of cell lines 10 times and took the median of significant hits. We compared the with score ≥ 15 in the randomized data and TP is the number of cell number of hits for all the cell lines (same as in Iorio et al. study) with lines with score ≥ 15 in the actual data. the number of hits for the different subsets of cell lines according to our grouping. In addition, we performed a sign test (R package Independent validation BSDA 1.2.0) comparing the significant hits in the G/G&S subsets versus We downloaded drug sensitivity for the CL from the GDSC the significant hits over 10 runs in the random G/random G&S and database (3). From the provided drugs, we selected trametinib and calculated the P value for all cancer types (alternative = “less”). dabrafenib, U.S. Food and Drug Administration–approved drugs Similarly, we downloaded gene dependency data from Project for melanoma treatment. We compared IC50 values for these two Score (42) and Project Achilles (52) processed with the Project drugs for all cancer types. Score pipeline and combined them. From a total of 696 unique cell We downloaded mutational signatures from cell lines available lines, 357 overlap with the 614 cell lines tested with our method. For from Jarvis et al. (29) and Petljak et al. (28), and we compared the those 357 tested cell lines, we repeated the same procedure as de- exposures of all cell lines for signature 7 (UV light). In Petljak et al. scribed above for the drug sensitivity data. dataset, signature 7 is divided into signature 7a, b, c, and d. Therefore, we used the sum of exposures across all four subtypes of signature 7. SUPPLEMENTARY MATERIALS We downloaded another set of drug screening data (PRISM 19Q3) Supplementary material for this article is available at http://advances.sciencemag.org/cgi/ (32) for the CL dataset. For the suspected cell lines, we generated content/full/6/27/eaba1862/DC1 one-versus-one classifiers (using cv.glmnet function with alpha = 0 View/request a protocol for this paper from Bio-protocol. and family = binomial, R package glmnet 2.0.18) for predicting original versus suspected cancer type, based on the drug sensitivity REFERENCES AND NOTES data. We performed 20 runs of each case and counted how many 1. G. Kaur, J. M. Dufour, Cell lines: Valuable tools or useless artifacts. Spermatogenesis 2, 1–5 (2012). times it is predicted as suspected (consistency score, which can vary 2. A. Goodspeed, L. M. Heiser, J. W. Gray, J. C. Costello, Tumor-derived cell lines as molecular from 0 to 20). In addition, we calculated the AUC for each classifier. models of cancer pharmacogenomics. Mol. Cancer Res. 14, 3–13 (2016). 3. W. Yang, J. Soares, P. Greninger, E. J. Edelman, H. Lightfoot, S. Forbes, N. Bindal, D. Beare, J. A. Smith, I. R. Thompson, S. Ramaswamy, P. A. Futreal, D. A. Haber, M. R. Stratton, Subtype classifiers C. Benes, U. McDermott, M. J. Garnett, Genomics of Drug Sensitivity in Cancer (GDSC): We downloaded cancer subtypes for the TCGA samples from the R A resource for therapeutic biomarker discovery in cancer cells. Nucleic Acids Res. 41, package TCGAbiolinks 2.12.6, which comprises many available D955–D961 (2013). Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 11 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE 4. J. Barretina, G. Caponigro, N. Stransky, K. Venkatesan, A. A. Margolin, S. Kim, C. J. Wilson, 22. M. Salvadores, D. Mas-Ponte, F. Supek, Passenger mutations accurately classify human J. Lehár, G. V. Kryukov, D. Sonkin, A. Reddy, M. Liu, L. Murray, M. F. Berger, J. E. Monahan, tumors. PLoS Comput. Biol. 15, e1006953 (2019). P. Morais, J. Meltzer, A. Korejwa, J. Jané-Valbuena, F. A. Mapa, J. Thibault, E. Bric-Furlong, 23. W. Jiao, G. Atwal, P. Polak, R. Karlic, E. Cuppen, A. Danyi, J. de Ridder, C. van Herpen, P. Raman, A. Shipway, I. H. Engels, J. Cheng, G. K. Yu, J. Yu, P. Aspesi, M. de Silva, K. Jagtap, M. P. Lolkema, N. Steeghs, G. Getz, Q. D. Morris, L. D. Stein; PCAWG Tumor Subtypes and M. D. Jones, L. Wang, C. Hatton, E. Palescandolo, S. Gupta, S. Mahan, C. Sougnez, Clinical Translation Working Group, A deep learning system can accurately classify R. C. Onofrio, T. Liefeld, L. MacConaill, W. Winckler, M. Reich, N. Li, J. P. Mesirov, primary and metastatic cancers based on patterns of passenger mutations. Nat. Commun. S. B. Gabriel, G. Getz, K. Ardlie, V. Chan, V. E. Myer, B. L. Weber, J. Porter, M. Warmuth, 11, 214494 (2019). P. Finan, J. L. Harris, M. Meyerson, T. R. Golub, M. P. Morrissey, W. R. Sellers, R. Schlegel, 24. M. T. Furlong, C. D. Hough, C. A. Sherman-Baust, E. S. Pizer, P. J. Morin, Evidence L. A. Garraway, The Cancer Cell Line Encyclopedia enables predictive modelling for the colonic origin of ovarian cancer cell line SW626. J. Natl. Cancer Inst. 91, 1327–1328 of anticancer drug sensitivity. Nature 483, 603–607 (2012). (1999). 5. R. J. Geraghty, A. Capes-Davis, J. M. Davis, J. Downward, R. I. Freshney, I. Knezevic, 25. E. Medico, M. Russo, G. Picco, C. Cancelliere, E. Valtorta, G. Corti, M. Buscarino, C. Isella, R. Lovell-Badge, J. R. W. Masters, J. Meredith, G. N. Stacey, P. Thraves, M. Vias; Cancer S. Lamba, B. Martinoglio, S. Veronese, S. Siena, A. Sartore-Bianchi, M. Beccuti, M. Mottolese, Research UK, Guidelines for the use of cell lines in biomedical research. Br. J. Cancer 111, M. Linnebacher, F. Cordero, F. Di Nicolantonio, A. Bardelli, The molecular landscape 1021–1046 (2014). of colorectal cancer cell lines unveils clinically actionable kinase targets. Nat. Commun. 6, 6. L. Vaughan, W. Glänzel, C. Korch, A. Capes-Davis, Widespread use of misidentified cell line 7002 (2015). KB (HeLa): Incorrect attribution and its impact revealed through mining the scientific 26. L. B. Alexandrov, P. H. Jones, D. C. Wedge, J. E. Sale, P. J. Campbell, S. Nik-Zainal, literature. Cancer Res. 77, 2784–2788 (2017). M. R. Stratton, Clock-like mutational processes in human somatic cells. Nat. Genet. 47, 7. J. Neimark, Line of attack. Science 347, 938–940 (2015). 1402–1407 (2015). 8. B. B. Campbell, N. Light, D. Fabrizio, M. Zatzman, F. Fuligni, R. de Borja, S. Davidson, 27. N. K. Hayward, J. S. Wilmott, N. Waddell, P. A. Johansson, M. A. Field, K. Nones, M. Edwards, J. A. Elvin, K. P. Hodel, W. J. Zahurancik, Z. Suo, T. Lipman, K. Wimmer, A.-M. Patch, H. Kakavand, L. B. Alexandrov, H. Burke, V. Jakrot, S. Kazakoff, O. Holmes, C. P. Kratz, D. C. Bowers, T. W. Laetsch, G. P. Dunn, T. M. Johanns, M. R. Grimmer, C. Leonard, R. Sabarinathan, L. Mularoni, S. Wood, Q. Xu, N. Waddell, V. Tembe, I. V. Smirnov, V. Larouche, D. Samuel, A. Bronsema, M. Osborn, D. Stearns, P. Raman, G. M. Pupo, R. De Paoli-Iseppi, R. E. Vilain, P. Shang, L. M. S. Lau, R. A. Dagg, S.-J. Schramm, Downloaded from http://advances.sciencemag.org/ on February 11, 2021 K. A. Cole, P. B. Storm, M. Yalon, E. Opocher, G. Mason, G. A. Thomas, M. Sabel, B. George, A. Pritchard, K. Dutton-Regester, F. Newell, A. Fitzgerald, C. A. Shang, S. M. Grimmond, D. S. Ziegler, S. Lindhorst, V. M. Issai, S. Constantini, H. Toledano, R. Elhasid, R. Farah, H. A. Pickett, J. Y. Yang, J. R. Stretch, A. Behren, R. F. Kefford, P. Hersey, G. V. Long, J. Cebon, R. Dvir, P. Dirks, A. Huang, M. A. Galati, J. Chung, V. Ramaswamy, M. S. Irwin, M. Aronson, M. Shackleton, A. J. Spillane, R. P. M. Saw, N. López-Bigas, J. V. Pearson, J. F. Thompson, C. Durno, M. D. Taylor, G. Rechavi, J. M. Maris, E. Bouffet, C. Hawkins, J. F. Costello, R. A. Scolyer, G. J. Mann, Whole-genome landscapes of major melanoma subtypes. Nature M. S. Meyn, Z. F. Pursell, D. Malkin, U. Tabori, A. Shlien, Comprehensive analysis 545, 175–180 (2017). of hypermutation in human cancer. Cell 171, 1042–1056.e10 (2017). 28. M. Petljak, L. B. Alexandrov, J. S. Brammeld, S. Price, D. C. Wedge, S. Grossmann, 9. C. Virtanen, Y. Ishikawa, D. Honjoh, M. Kimura, M. Shimane, T. Miyoshi, H. Nomura, K. J. Dawson, Y. S. Ju, F. Iorio, J. M. C. Tubio, C. C. Koh, I. Georgakopoulos-Soares, M. H. Jones, Integrated classification of lung tumors and cell lines by expression profiling. B. Rodríguez-Martín, B. Otlu, S. O’Meara, A. P. Butler, A. Menzies, S. G. Bhosle, K. Raine, Proc. Natl. Acad. Sci. U.S.A. 99, 12357–12362 (2002). D. R. Jones, J. W. Teague, K. Beal, C. Latimer, L. O’Neill, J. Zamora, E. Anderson, N. Patel, 10. F. Iorio, T. A. Knijnenburg, D. J. Vis, G. R. Bignell, M. P. Menden, M. Schubert, N. Aben, M. Maddison, B. L. Ng, J. Graham, M. J. Garnett, U. McDermott, S. Nik-Zainal, E. Gonçalves, S. Barthorpe, H. Lightfoot, T. Cokelaer, P. Greninger, E. van Dyk, H. Chang, P. J. Campbell, M. R. Stratton, Characterizing mutational signatures in human cancer cell H. de Silva, H. Heyn, X. Deng, R. K. Egan, Q. Liu, T. Mironenko, X. Mitropoulos, lines reveals episodic APOBEC mutagenesis. Cell 176, 1282–1294.e20 (2019). L. Richardson, J. Wang, T. Zhang, S. Moran, S. Sayols, M. Soleimani, D. Tamborero, 29. M. C. Jarvis, D. Ebrahimi, N. A. Temiz, R. S. Harris, Mutation signatures including APOBEC N. Lopez-Bigas, P. Ross-Macdonald, M. Esteller, N. S. Gray, D. A. Haber, M. R. Stratton, in cancer cell lines. JNCI Cancer Spectr. 2, pky002 (2018). C. H. Benes, L. F. A. Wessels, J. Saez-Rodriguez, U. McDermott, M. J. Garnett, A landscape 30. J. Ji, X. Chen, S. Y. Leung, J.-T. A. Chi, K. M. Chu, S. T. Yuen, R. Li, A. S. Y. Chan, J. Li, of pharmacogenomic interactions in cancer. Cell 166, 740–754 (2016). N. Dunphy, S. So, Comprehensive analysis of the gene expression profiles in human 11. E. Kim, M. Dede, W. F. Lenoir, G. Wang, S. Srinivasan, M. Colic, T. Hart, A network of human gastric cancer cell lines. Oncogene 21, 6549–6556 (2002). functional gene interactions from knockout fitness screens in cancer cells. Life Sci. Alliance 2, 31. R. B. Corcoran, C. E. Atreya, G. S. Falchook, E. L. Kwak, D. P. Ryan, J. C. Bendell, O. Hamid, e201800278 (2019). W. A. Messersmith, A. Daud, R. Kurzrock, M. Pierobon, P. Sun, E. Cunningham, S. Little, 12. S. Domcke, R. Sinha, D. A. Levine, C. Sander, N. Schultz, Evaluating cell lines as tumour K. Orford, M. Motwani, Y. Bai, K. Patel, A. P. Venook, S. Kopetz, Combined BRAF and MEK models by comparison of genomic profiles. Nat. Commun. 4, 2126 (2013). inhibition with dabrafenib and trametinib in BRAF V600–mutant colorectal cancer. J. Clin. Oncol. 13. R. Sinha, A. G. Winer, M. Chevinsky, C. Jakubowski, Y.-B. Chen, Y. Dong, S. K. Tickoo, 33, 4023–4031 (2015). V. E. Reuter, P. Russo, J. A. Coleman, C. Sander, J. J. Hsieh, A. A. Hakimi, Analysis of renal 32. S. M. Corsello, R. T. Nagari, R. D. Spangler, J. Rossen, M. Kocak, J. G. Bryan, R. Humeidi, cancer cell lines from two major resources enables genomics-guided cell line selection. D. Peck, X. Wu, A. A. Tang, V. M. Wang, S. A. Bender, E. Lemire, R. Narayan, P. Montgomery, Nat. Commun. 8, 15165 (2017). U. Ben-David, Y. Chen, M. G. Rees, N. J. Lyons, J. M. McFarland, B. T. Wong, L. Wang, 14. P. Jonsson, C. Bandlamudi, M. L. Cheng, P. Srinivasan, S. S. Chavan, N. D. Friedman, N. Dumont, P. J. O’Hearn, E. Stefan, J. G. Doench, H. Greulich, M. Meyerson, F. Vazquez, E. Y. Rosen, A. L. Richards, N. Bouvier, S. D. Selcuklu, C. M. Bielski, W. Abida, D. Mandelker, A. Subramanian, J. A. Roth, J. A. Bittker, J. S. Boehm, C. C. Mader, A. Tsherniak, T. R. Golub, O. Birsoy, L. Zhang, A. Zehir, M. T. A. Donoghue, J. Baselga, K. Offit, H. I. Scher, Non-oncology drugs are a source of previously unappreciated anti-cancer activity. E. M. O’Reilly, Z. K. Stadler, N. Schultz, N. D. Socci, A. Viale, M. Ladanyi, M. E. Robson, bioRxiv 730119 [Preprint]. 9 August 2019. https://doi.org/10.1101/730119. D. M. Hyman, M. F. Berger, D. B. Solit, B. S. Taylor, Tumour lineage shapes BRCA-mediated 33. Cancer Genome Atlas Network, Comprehensive molecular portraits of human breast phenotypes. Nature 571, 576–579 (2019). tumours. Nature 490, 61–70 (2012). 15. S. Kopetz, J. R. Desai, E. T. Chan, J. R. Hecht, P. J. O’dwyer, D. M. Maru, V. K. Morris, F. Janku, 34. M. Ceccarelli, F. P. Barthel, T. M. Malta, T. S. Sabedot, S. R. Salama, B. A. Murray, A. Dasari, W. Chung, J. P. J. Issa, P. Gibbs, B. P. James, G. Powis, K. B. Nolop, O. Morozova, Y. Newton, A. Radenbaugh, S. M. Pagnotta, S. Anjum, J. Wang, G. Manyam, S. Bhattacharya, L. L. Saltz, Phase II pilot study of vemurafenib in patients with metastatic P. Zoppoli, S. Ling, A. A. Rao, M. Grifford, A. D. Cherniack, H. Zhang, L. Poisson, BRAF-mutated colorectal cancer. J. Clin. Oncol. 2015, 4032–4038 (2015). C. G. Carlotti, D. P. da C. Tirapelli, A. Rao, T. Mikkelsen, C. C. Lau, W. K. A. Yung, R. Rabadan, 16. S. Park, B. Lehner, Epigenetic epistatic interactions constrain the evolution of gene J. Huse, D. J. Brat, N. L. Lehman, J. S. Barnholtz-Sloan, S. Zheng, K. Hess, G. Rao, expression. Mol. Syst. Biol. 9, 645 (2013). M. Meyerson, R. Beroukhim, L. Cooper, R. Akbani, M. Wrensch, D. Haussler, K. D. Aldape, 17. M. Lukk, M. Kapushesky, J. Nikkilä, H. Parkinson, A. Goncalves, W. Huber, E. Ukkonen, P. W. Laird, D. H. Gutmann, S. Anjum, H. Arachchi, J. T. Auman, M. Balasundaram, S. Balu, A. Brazma, A global map of human gene expression. Nat. Biotechnol. 28, 322–324 (2010). G. Barnett, S. Baylin, S. Bell, C. Benz, N. Bir, K. L. Black, T. Bodenheimer, L. Boice, 18. K. E. Varley, J. Gertz, K. M. Bowling, S. L. Parker, T. E. Reddy, F. Pauli-Behn, M. K. Cross, M. S. Bootwalla, J. Bowen, C. A. Bristow, Y. S. N. Butterfield, Q.-R. Chen, L. Chin, J. Cho, B. A. Williams, J. A. Stamatoyannopoulos, G. E. Crawford, D. M. Absher, B. J. Wold, E. Chuah, S. Chudamani, S. G. Coetzee, M. L. Cohen, H. Colman, M. Couce, F. D’Angelo, R. M. Myers, Dynamic DNA methylation across diverse human cell lines and tissues. T. Davidsen, A. Davis, J. A. Demchok, K. Devine, L. Ding, R. Duell, J. B. Elder, Genome Res. 23, 555–567 (2013). J. M. Eschbacher, A. Fehrenbach, M. Ferguson, S. Frazer, G. Fuller, J. Fulop, S. B. Gabriel, 19. W. E. Johnson, C. Li, A. Rabinovic, Adjusting batch effects in microarray expression data L. Garofano, J. M. Gastier-Foster, N. Gehlenborg, M. Gerken, G. Getz, C. Giannini, using empirical Bayes methods. Biostatistics 8, 118–127 (2007). W. J. Gibson, A. Hadjipanayis, D. N. Hayes, D. I. Heiman, B. Hermes, J. Hilty, K. A. Hoadley, 20. C. Chen, K. Grennan, J. Badner, D. Zhang, E. Gershon, L. Jin, C. Liu, Removing batch effects A. P. Hoyle, M. Huang, S. R. Jefferys, C. D. Jones, S. J. M. Jones, Z. Ju, A. Kastl, A. Kendler, in analysis of expression microarray data: An evaluation of six batch adjustment methods. J. Kim, R. Kucherlapati, P. H. Lai, M. S. Lawrence, S. Lee, K. M. Leraas, T. M. Lichtenberg, PLOS ONE 6, e17238 (2011). P. Lin, Y. Liu, J. Liu, J. Y. Ljubimova, Y. Lu, Y. Ma, D. T. Maglinte, H. S. Mahadeshwar, 21. S. P. J. M. Horbach, W. Halffman, The ghosts of HeLa: How cell line misidentification M. A. Marra, M. McGraw, C. McPherson, S. Meng, P. A. Mieczkowski, C. R. Miller, G. B. Mills, contaminates the scientific literature. PLOS ONE 12, e0186281 (2017). R. A. Moore, L. E. Mose, A. J. Mungall, R. Naresh, T. Naska, L. Neder, M. S. Noble, A. Noss, Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 12 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE B. P. O’Neill, Q. T. Ostrom, C. Palmer, A. Pantazi, M. Parfenov, P. J. Park, J. S. Parker, V. Chandan, J. Cheville, J. A. Copland, M. Farnell, T. Flotte, N. Giama, T. Ho, M. Kendrick, C. M. Perou, C. R. Pierson, T. Pihl, A. Protopopov, A. Radenbaugh, N. C. Ramirez, J.-P. Kocher, K. Kopp, C. Moser, D. Nagorney, D. O’Brien, B. P. O’Neill, T. Patel, G. Petersen, W. K. Rathmell, X. Ren, J. Roach, A. G. Robertson, G. Saksena, J. E. Schein, S. E. Schumacher, F. Que, M. Rivera, L. Roberts, R. Smallridge, T. Smyrk, M. Stanton, R. H. Thompson, J. Seidman, K. Senecal, S. Seth, H. Shen, Y. Shi, J. Shih, K. Shimmel, H. Sicotte, S. Sifri, M. Torbenson, J. D. Yang, L. Zhang, F. Brimo, J. A. Ajani, A. M. A. Gonzalez, C. Behrens, T. Silva, J. V. Simons, R. Singh, T. Skelly, A. E. Sloan, H. J. Sofia, M. G. Soloway, X. Song, J. Bondaruk, R. Broaddus, B. Czerniak, B. Esmaeli, J. Fujimoto, J. Gershenwald, C. Guo, C. Sougnez, C. Souza, S. M. Staugaitis, H. Sun, C. Sun, D. Tan, J. Tang, Y. Tang, L. Thorne, A. J. Lazar, C. Logothetis, F. Meric-Bernstam, C. Moran, L. Ramondetta, D. Rice, A. Sood, F. A. Trevisan, T. Triche, D. J. Van Den Berg, U. Veluvolu, D. Voet, Y. Wan, Z. Wang, P. Tamboli, T. Thompson, P. Troncoso, A. Tsao, I. Wistuba, C. Carter, L. Haydu, P. Hersey, R. Warnick, J. N. Weinstein, D. J. Weisenberger, M. D. Wilkerson, F. Williams, L. Wise, V. Jakrot, H. Kakavand, R. Kefford, K. Lee, G. Long, G. Mann, M. Quinn, R. Saw, R. Scolyer, Y. Wolinsky, J. Wu, A. W. Xu, L. Yang, L. Yang, T. I. Zack, J. C. Zenklusen, J. Zhang, K. Shannon, A. Spillane, J. Stretch, M. Synott, J. Thompson, J. Wilmott, H. Al-Ahmadie, W. Zhang, J. Zhang, E. Zmuda, H. Noushmehr, A. Iavarone, R. G. W. Verhaak, Molecular T. A. Chan, R. Ghossein, A. Gopalan, D. A. Levine, V. Reuter, S. Singer, B. Singh, N. V. Tien, profiling reveals biologically discrete subsets and pathways of progression in diffuse T. Broudy, C. Mirsaidi, P. Nair, P. Drwiega, J. Miller, J. Smith, H. Zaren, J.-W. Park, N. P. Hung, glioma. Cell 164, 550–563 (2016). E. Kebebew, W. M. Linehan, A. R. Metwalli, K. Pacak, P. A. Pinto, M. Schiffman, 35. Y. Liu, N. S. Sethi, T. Hinoue, B. G. Schneider, A. D. Cherniack, F. Sanchez-Vega, L. S. Schmidt, C. D. Vocke, N. Wentzensen, R. Worrell, H. Yang, M. Moncrieff, C. Goparaju, J. A. Seoane, F. Farshidfar, R. Bowlby, M. Islam, J. Kim, W. Chatila, R. Akbani, R. S. Kanchi, J. Melamed, H. Pass, N. Botnariuc, I. Caraman, M. Cernat, I. Chemencedji, A. Clipca, C. S. Rabkin, J. E. Willis, K. K. Wang, S. J. McCall, L. Mishra, A. I. Ojesina, S. Bullman, S. Doruc, G. Gorincioi, S. Mura, M. Pirtac, I. Stancul, D. Tcaciuc, M. Albert, I. Alexopoulou, C. S. Pedamallu, A. J. Lazar, R. Sakai, S. J. Caesar-Johnson, J. A. Demchok, I. Felau, A. Arnaout, J. Bartlett, J. Engel, S. Gilbert, J. Parfitt, H. Sekhon, G. Thomas, D. M. Rassl, M. Kasapi, M. L. Ferguson, C. M. Hutter, H. J. Sofia, R. Tarnuzzer, Z. Wang, L. Yang, R. C. Rintoul, C. Bifulco, R. Tamakawa, W. Urba, N. Hayward, H. Timmers, A. Antenucci, J. C. Zenklusen, J. Zhang, S. Chudamani, J. Liu, L. Lolla, R. Naresh, T. Pihl, Q. Sun, Y. Wan, F. Facciolo, G. Grazi, M. Marino, R. Merola, R. de Krijger, A.-P. Gimenez-Roqueplo, A. Piché, Y. Wu, J. Cho, T. DeFreitas, S. Frazer, N. Gehlenborg, G. Getz, D. I. Heiman, J. Kim, S. Chevalier, G. McKercher, K. Birsoy, G. Barnett, C. Brewer, C. Farver, T. Naska, M. S. Lawrence, P. Lin, S. Meier, M. S. Noble, G. Saksena, D. Voet, H. Zhang, B. Bernard, N. A. Pennell, D. Raymond, C. Schilero, K. Smolenski, F. Williams, C. Morrison, J. A. Borgia, N. Chambwe, V. Dhankani, T. Knijnenburg, R. Kramer, K. Leinonen, Y. Liu, M. Miller, M. J. Liptay, M. Pool, C. W. Seder, K. Junker, L. Omberg, M. Dinkin, G. Manikhas, D. Alvaro, S. Reynolds, I. Shmulevich, V. Thorsson, W. Zhang, R. Akbani, B. M. Broom, A. M. Hegde, M. C. Bragazzi, V. Cardinale, G. Carpino, E. Gaudio, D. Chesla, S. Cottingham, M. Dubina, Downloaded from http://advances.sciencemag.org/ on February 11, 2021 Z. Ju, R. S. Kanchi, A. Korkut, J. Li, H. Liang, S. Ling, W. Liu, Y. Lu, G. B. Mills, K.-S. Ng, A. Rao, F. Moiseenko, R. Dhanasekaran, K.-F. Becker, K.-P. Janssen, J. Slotta-Huspenina, M. Ryan, J. Wang, J. N. Weinstein, J. Zhang, A. Abeshouse, J. Armenia, D. Chakravarty, M. H. Abdel-Rahman, D. Aziz, S. Bell, C. M. Cebulla, A. Davis, R. Duell, J. B. Elder, J. Hilty, W. K. Chatila, I. Bruijn, J. Gao, B. E. Gross, Z. J. Heins, R. Kundra, K. La, M. Ladanyi, A. Luna, B. Kumar, J. Lang, N. L. Lehman, R. Mandt, P. Nguyen, R. Pilarski, K. Rai, L. Schoenfield, M. G. Nissan, A. Ochoa, S. M. Phillips, E. Reznik, F. Sanchez-Vega, C. Sander, N. Schultz, K. Senecal, P. Wakely, P. Hansen, R. Lechan, J. Powers, A. Tischler, W. E. Grizzle, R. Sheridan, S. O. Sumer, Y. Sun, B. S. Taylor, J. Wang, H. Zhang, P. Anur, M. Peto, K. C. Sexton, A. Kastl, J. Henderson, S. Porten, J. Waldmann, M. Fassnacht, S. L. Asa, P. Spellman, C. Benz, J. M. Stuart, C. K. Wong, C. Yau, D. N. Hayes, J. S. Parker, D. Schadendorf, M. Couce, M. Graefen, H. Huland, G. Sauter, T. Schlomm, R. Simon, M. D. Wilkerson, A. Ally, M. Balasundaram, R. Bowlby, D. Brooks, R. Carlsen, E. Chuah, P. Tennstedt, O. Olabode, M. Nelson, O. Bathe, P. R. Carroll, J. M. Chan, P. Disaia, P. Glenn, N. Dhalla, R. Holt, S. J. M. Jones, K. Kasaian, D. Lee, Y. Ma, M. A. Marra, M. Mayo, R. K. Kelley, C. N. Landen, J. Phillips, M. Prados, J. Simko, K. Smith-McCune, S. VandenBerg, R. A. Moore, A. J. Mungall, K. Mungall, A. G. Robertson, S. Sadeghi, J. E. Schein, K. Roggin, A. Fehrenbach, A. Kendler, S. Sifri, R. Steele, A. Jimeno, F. Carey, I. Forgie, P. Sipahimalani, A. Tam, N. Thiessen, K. Tse, T. Wong, A. C. Berger, R. Beroukhim, M. Mannelli, M. Carney, B. Hernandez, B. Campos, C. Herold-Mende, C. Jungk, A. D. Cherniack, C. Cibulskis, S. B. Gabriel, G. F. Gao, G. Ha, M. Meyerson, S. E. Schumacher, A. Unterberg, A. von Deimling, A. Bossler, J. Galbraith, L. Jacobus, M. Knudson, T. Knutson, J. Shih, M. H. Kucherlapati, R. S. Kucherlapati, S. Baylin, L. Cope, L. Danilova, D. Ma, M. Milhem, R. Sigmund, A. K. Godwin, R. Madan, H. G. Rosenthal, C. Adebamowo, M. S. Bootwalla, P. H. Lai, D. T. Maglinte, D. J. V. D. Berg, D. J. Weisenberger, J. T. Auman, S. N. Adebamowo, A. Boussioutas, D. Beer, T. Giordano, A.-M. Mes-Masson, F. Saad, S. Balu, T. Bodenheimer, C. Fan, K. A. Hoadley, A. P. Hoyle, S. R. Jefferys, C. D. Jones, T. Bocklage, L. Landrum, R. Mannel, K. Moore, K. Moxley, R. Postier, J. Walker, R. Zuna, S. Meng, P. A. Mieczkowski, L. E. Mose, A. H. Perou, C. M. Perou, J. Roach, Y. Shi, M. Feldman, F. Valdivieso, R. Dhir, J. Luketich, E. M. M. Pinero, M. Quintero-Aguilo, J. V. Simons, T. Skelly, M. G. Soloway, D. Tan, U. Veluvolu, H. Fan, T. Hinoue, P. W. Laird, C. G. Carlotti, J. S. D. Santos, R. Kemp, A. Sankarankuty, D. Tirapelli, J. Catto, K. Agnew, H. Shen, W. Zhou, M. Bellair, K. Chang, K. Covington, C. J. Creighton, H. Dinh, E. Swisher, J. Creaney, B. Robinson, C. S. Shelley, E. M. Godwin, S. Kendall, C. Shipman, H. Doddapaneni, L. A. Donehower, J. Drummond, R. A. Gibbs, R. Glenn, W. Hale, Y. Han, C. Bradford, T. Carey, A. Haddad, J. Moyer, L. Peterson, M. Prince, L. Rozek, G. Wolf, J. Hu, V. Korchina, S. Lee, L. Lewis, W. Li, X. Liu, M. Morgan, D. Morton, D. Muzny, R. Bowman, K. M. Fong, I. Yang, R. Korst, W. K. Rathmell, J. L. Fantacone-Campbell, J. Santibanez, M. Sheth, E. Shinbrot, L. Wang, M. Wang, D. A. Wheeler, L. Xi, F. Zhao, J. A. Hooke, A. J. Kovatich, C. D. Shriver, J. DiPersio, B. Drake, R. Govindan, S. Heath, T. Ley, J. Hess, E. L. Appelbaum, M. Bailey, M. G. Cordes, L. Ding, C. C. Fronick, L. A. Fulton, B. V. Tine, P. Westervelt, M. A. Rubin, J. I. Lee, N. D. Aredes, A. Mariamidze, V. Thorsson, R. S. Fulton, C. Kandoth, E. R. Mardis, M. D. McLellan, C. A. Miller, H. K. Schmidt, A. J. Bass, P. W. Laird, Comparative molecular analysis of gastrointestinal R. K. Wilson, D. Crain, E. Curley, J. Gardner, K. Lau, D. Mallery, S. Morris, J. Paulauskis, adenocarcinomas. Cancer Cell 33, 721–735.e8 (2018). R. Penny, C. Shelton, T. Shelton, M. Sherman, E. Thompson, P. Yena, J. Bowen, 36. R. M. Neve, K. Chin, J. Fridlyand, J. Yeh, F. L. Baehner, T. Fevr, L. Clark, N. Bayani, J. M. Gastier-Foster, M. Gerken, K. M. Leraas, T. M. Lichtenberg, N. C. Ramirez, L. Wise, J.-P. Coppe, F. Tong, T. Speed, P. T. Spellman, S. DeVries, A. Lapuk, N. J. Wang, W.-L. Kuo, E. Zmuda, N. Corcoran, T. Costello, C. Hovens, A. L. Carvalho, A. C. de Carvalho, J. L. Stilwell, D. Pinkel, D. G. Albertson, F. M. Waldman, F. McCormick, R. B. Dickson, J. H. Fregnani, A. Longatto-Filho, R. M. Reis, C. Scapulatempo-Neto, H. C. S. Silveira, M. D. Johnson, M. Lippman, S. Ethier, A. Gazdar, J. W. Gray, A collection of breast cancer D. O. Vidal, A. Burnette, J. Eschbacher, B. Hermes, A. Noss, R. Singh, M. L. Anderson, cell lines for the study of functionally distinct cancer subtypes. Cancer Cell 10, 515–527 P. D. Castro, M. Ittmann, D. Huntsman, B. Kohl, X. Le, R. Thorp, C. Andry, E. R. Duffy, (2006). V. Lyadov, O. Paklina, G. Setdikova, A. Shabunin, M. Tavobilov, C. McPherson, R. Warnick, 37. J. Ronen, S. Hayat, A. Akalin, Evaluation of colorectal cancer subtypes and cell lines using R. Berkowitz, D. Cramer, C. Feltmate, N. Horowitz, A. Kibel, M. Muto, C. P. Raut, A. Malykh, deep learning. Life Sci. Alliance 2, e201900517 (2019). J. S. Barnholtz-Sloan, W. Barrett, K. Devine, J. Fulop, Q. T. Ostrom, K. Shimmel, Y. Wolinsky, 38. K. Yu, B. Chen, D. Aran, J. Charalel, C. Yau, D. M. Wolf, L. J. van ‘t Veer, A. J. Butte, A. E. Sloan, A. D. Rose, F. Giuliante, M. Goodman, B. Y. Karlan, C. H. Hagedorn, J. Eckman, T. Goldstein, M. Sirota, Comprehensive transcriptomic analysis of cell lines as models J. Harr, J. Myers, K. Tucker, L. A. Zach, B. Deyarmin, H. Hu, L. Kvecher, C. Larson, R. J. Mural, of primary tumors across 22 tumor types. Nat. Commun. 10, 3574 (2019). S. Somiari, A. Vicha, T. Zelinka, J. Bennett, M. Iacocca, B. Rabeno, P. Swanson, M. Latour, 39. T. Cokelaer, E. Chen, F. Iorio, M. P. Menden, H. Lightfoot, J. Saez-Rodriguez, M. J. Garnett, L. Lacombe, B. Têtu, A. Bergeron, M. McGraw, S. M. Staugaitis, J. Chabot, H. Hibshoosh, GDSCTools for mining pharmacogenomic interactions in cancer. Bioinformatics 34, A. Sepulveda, T. Su, T. Wang, O. Potapova, O. Voronina, L. Desjardins, O. Mariani, 1226–1228 (2018). S. Roman-Roman, X. Sastre, M.-H. Stern, F. Cheng, S. Signoretti, A. Berchuck, D. Bigner, 40. O. N. Ikediobi, M. Reimers, S. Durinck, P. E. Blower, A. P. Futreal, M. R. Stratton, E. Lipp, J. Marks, S. McCall, R. McLendon, A. Secord, A. Sharp, M. Behera, D. J. Brat, A. Chen, J. N. Weinstein, In vitro differential sensitivity of melanomas to phenothiazines is based K. Delman, S. Force, F. Khuri, K. Magliocca, S. Maithel, J. J. Olson, T. Owonikoko, A. Pickens, on the presence of codon 600 BRAF mutation. Mol. Cancer Ther. 7, 1337–1346 (2008). S. Ramalingam, D. M. Shin, G. Sica, E. G. V. Meir, H. Zhang, W. Eijckenboom, A. Gillis, 41. Y. Yamaguchi, T. Kasukabe, S. Kumakura, Piperlongumine rapidly induces the death E. Korpershoek, L. Looijenga, W. Oosterhuis, H. Stoop, K. E. van Kessel, E. C. Zwarthoff, of human pancreatic cancer cells mainly through the induction of ferroptosis. Int. J. Oncol. C. Calatozzolo, L. Cuppini, S. Cuzzubbo, F. DiMeco, G. Finocchiaro, L. Mattei, A. Perin, 52, 1011–1022 (2018). B. Pollo, C. Chen, J. Houck, P. Lohavanichbutr, A. Hartmann, C. Stoehr, R. Stoehr, 42. F. M. Behan, F. Iorio, G. Picco, E. Gonçalves, C. M. Beaver, G. Migliardi, R. Santos, Y. Rao, H. Taubert, S. Wach, B. Wullich, W. Kycler, D. Murawa, M. Wiznerowicz, K. Chung, F. Sassi, M. Pinnelli, R. Ansari, S. Harper, D. A. Jackson, R. McRae, R. Pooley, P. Wilkinson, W. J. Edenfield, J. Martin, E. Baudin, G. Bubley, R. Bueno, A. D. Rienzo, W. G. Richards, D. van der Meer, D. Dow, C. Buser-Doepner, A. Bertotti, L. Trusolino, E. A. Stronach, S. Kalkanis, T. Mikkelsen, H. Noushmehr, L. Scarpace, N. Girard, M. Aymerich, E. Campo, J. Saez-Rodriguez, K. Yusa, M. J. Garnett, Prioritization of cancer therapeutic targets using E. Giné, A. L. Guillermo, N. V. Bang, P. T. Hanh, B. D. Phu, Y. Tang, H. Colman, K. Evason, CRISPR–Cas9 screens. Nature 568, 511–516 (2019). P. R. Dottino, J. A. Martignetti, H. Gabra, H. Juhl, T. Akeredolu, S. Stepa, D. Hoon, K. Ahn, 43. M. Westphal, C. L. Maire, K. Lamszus, EGFR as a target for glioblastoma treatment: K. J. Kang, F. Beuschlein, A. Breggia, M. Birrer, D. Bell, M. Borad, A. H. Bryce, E. Castle, An unfulfilled promise. CNS Drugs 31, 723–735 (2017). Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 13 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE 44. J. D. Campbell, A. Alexandrov, J. Kim, J. Wala, A. H. Berger, C. S. Pedamallu, S. A. Shukla, E. Korpershoek, L. Looijenga, W. Oosterhuis, H. Stoop, K. E. van Kessel, E. C. Zwarthoff, G. Guo, A. N. Brooks, B. A. Murray, M. Imielinski, X. Hu, S. Ling, R. Akbani, M. Rosenberg, C. Calatozzolo, L. Cuppini, S. Cuzzubbo, F. DiMeco, G. Finocchiaro, L. Mattei, A. Perin, C. Cibulskis, A. Ramachandran, E. A. Collisson, D. J. Kwiatkowski, M. S. Lawrence, B. Pollo, C. Chen, J. Houck, P. Lohavanichbutr, A. Hartmann, C. Stoehr, R. Stoehr, J. N. Weinstein, R. G. W. Verhaak, C. J. Wu, P. S. Hammerman, A. D. Cherniack, G. Getz, H. Taubert, S. Wach, B. Wullich, W. Kycler, D. Murawa, M. Wiznerowicz, K. Chung, C. G. A. R. Network, M. N. Artyomov, R. Schreiber, R. Govindan, M. Meyerson, Distinct W. J. Edenfield, J. Martin, E. Baudin, G. Bubley, R. Bueno, A. De Rienzo, W. G. Richards, patterns of somatic genome alterations in lung adenocarcinomas and squamous cell S. Kalkanis, T. Mikkelsen, H. Noushmehr, L. Scarpace, N. Girard, M. Aymerich, E. Campo, carcinomas. Nat. Genet. 48, 607–616 (2016). E. Giné, A. L. Guillermo, N. Van Bang, P. T. Hanh, B. D. Phu, Y. Tang, H. Colman, K. Evason, 45. R. L. Grossman, A. P. Heath, V. Ferretti, H. E. Varmus, D. R. Lowy, W. A. Kibbe, L. M. Staudt, P. R. Dottino, J. A. Martignetti, H. Gabra, H. Juhl, T. Akeredolu, S. Stepa, D. Hoon, K. Ahn, Toward a shared vision for cancer genomic data. N. Engl. J. Med. 375, 1109–1112 (2016). K. J. Kang, F. Beuschlein, A. Breggia, M. Birrer, D. Bell, M. Borad, A. H. Bryce, E. Castle, 46. D. Aran, M. Sirota, A. J. Butte, Systematic pan-cancer analysis of tumour purity. V. Chandan, J. Cheville, J. A. Copland, M. Farnell, T. Flotte, N. Giama, T. Ho, M. Kendrick, Nat. Commun. 6, (2015). J.-P. Kocher, K. Kopp, C. Moser, D. Nagorney, D. O’Brien, B. P. O’Neill, T. Patel, G. Petersen, 47. Cancer Cell Line Encyclopedia Consortium; Genomics of Drug Sensitivity in Cancer F. Que, M. Rivera, L. Roberts, R. Smallridge, T. Smyrk, M. Stanton, R. H. Thompson, Concortium, Pharmacogenomic agreement between two cancer cell line data sets. M. Torbenson, J. D. Yang, L. Zhang, F. Brimo, J. A. Ajani, A. M. A. Gonzalez, C. Behrens, Nature 528, 84–87 (2015). J. Bondaruk, R. Broaddus, B. Czerniak, B. Esmaeli, J. Fujimoto, J. Gershenwald, C. Guo, 48. M. H. Bailey, C. Tokheim, E. Porta-Pardo, S. Sengupta, D. Bertrand, A. Weerasinghe, A. J. Lazar, C. Logothetis, F. Meric-Bernstam, C. Moran, L. Ramondetta, D. Rice, A. Sood, A. Colaprico, M. C. Wendl, J. Kim, B. Reardon, P. K.-S. Ng, K. J. Jeong, S. Cao, Z. Wang, P. Tamboli, T. Thompson, P. Troncoso, A. Tsao, I. Wistuba, C. Carter, L. Haydu, P. Hersey, J. Gao, Q. Gao, F. Wang, E. M. Liu, L. Mularoni, C. Rubio-Perez, N. Nagarajan, I. Cortés-Ciriano, V. Jakrot, H. Kakavand, R. Kefford, K. Lee, G. Long, G. Mann, M. Quinn, R. Saw, R. Scolyer, D. C. Zhou, W.-W. Liang, J. M. Hess, V. D. Yellapantula, D. Tamborero, A. Gonzalez-Perez, K. Shannon, A. Spillane, J. Stretch, M. Synott, J. Thompson, J. Wilmott, H. Al-Ahmadie, C. Suphavilai, J. Y. Ko, E. Khurana, P. J. Park, E. M. Van Allen, H. Liang; MC3 Working Group; T. A. Chan, R. Ghossein, A. Gopalan, D. A. Levine, V. Reuter, S. Singer, B. Singh, N. V. Tien, Cancer Genome Atlas Research Network, M. S. Lawrence, A. Godzik, N. Lopez-Bigas, T. Broudy, C. Mirsaidi, P. Nair, P. Drwiega, J. Miller, J. Smith, H. Zaren, J.-W. Park, N. P. Hung, J. Stuart, D. Wheeler, G. Getz, K. Chen, A. J. Lazar, G. B. Mills, R. Karchin, L. Ding, E. Kebebew, W. M. Linehan, A. R. Metwalli, K. Pacak, P. A. Pinto, M. Schiffman, Comprehensive characterization of cancer driver genes and mutations. Cell 173, L. S. Schmidt, C. D. Vocke, N. Wentzensen, R. Worrell, H. Yang, M. Moncrieff, C. Goparaju, Downloaded from http://advances.sciencemag.org/ on February 11, 2021 371–385.e18 (2018). J. Melamed, H. Pass, N. Botnariuc, I. Caraman, M. Cernat, I. Chemencedji, A. Clipca, 49. K. Ellrott, M. H. Bailey, G. Saksena, K. R. Covington, C. Kandoth, C. Stewart, J. Hess, S. Ma, S. Doruc, G. Gorincioi, S. Mura, M. Pirtac, I. Stancul, D. Tcaciuc, M. Albert, I. Alexopoulou, K. E. Chiotti, M. McLellan, H. J. Sofia, C. Hutter, G. Getz, D. Wheeler, L. Ding, S. J. Caesar-Johnson, A. Arnaout, J. Bartlett, J. Engel, S. Gilbert, J. Parfitt, H. Sekhon, G. Thomas, D. M. Rassl, J. A. Demchok, I. Felau, M. Kasapi, M. L. Ferguson, C. M. Hutter, H. J. Sofia, R. Tarnuzzer, R. C. Rintoul, C. Bifulco, R. Tamakawa, W. Urba, N. Hayward, H. Timmers, A. Antenucci, Z. Wang, L. Yang, J. C. Zenklusen, J. Zhang, S. Chudamani, J. Liu, L. Lolla, R. Naresh, T. Pihl, F. Facciolo, G. Grazi, M. Marino, R. Merola, R. de Krijger, A.-P. Gimenez-Roqueplo, A. Piché, Q. Sun, Y. Wan, Y. Wu, J. Cho, T. DeFreitas, S. Frazer, N. Gehlenborg, G. Getz, D. I. Heiman, S. Chevalier, G. McKercher, K. Birsoy, G. Barnett, C. Brewer, C. Farver, T. Naska, J. Kim, M. S. Lawrence, P. Lin, S. Meier, M. S. Noble, G. Saksena, D. Voet, H. Zhang, N. A. Pennell, D. Raymond, C. Schilero, K. Smolenski, F. Williams, C. Morrison, J. A. Borgia, B. Bernard, N. Chambwe, V. Dhankani, T. Knijnenburg, R. Kramer, K. Leinonen, Y. Liu, M. J. Liptay, M. Pool, C. W. Seder, K. Junker, L. Omberg, M. Dinkin, G. Manikhas, D. Alvaro, M. Miller, S. Reynolds, I. Shmulevich, V. Thorsson, W. Zhang, R. Akbani, B. M. Broom, M. C. Bragazzi, V. Cardinale, G. Carpino, E. Gaudio, D. Chesla, S. Cottingham, M. Dubina, A. M. Hegde, Z. Ju, R. S. Kanchi, A. Korkut, J. Li, H. Liang, S. Ling, W. Liu, Y. Lu, G. B. Mills, F. Moiseenko, R. Dhanasekaran, K.-F. Becker, K.-P. Janssen, J. Slotta-Huspenina, K.-S. Ng, A. Rao, M. Ryan, J. Wang, J. N. Weinstein, J. Zhang, A. Abeshouse, J. Armenia, M. H. Abdel-Rahman, D. Aziz, S. Bell, C. M. Cebulla, A. Davis, R. Duell, J. B. Elder, J. Hilty, D. Chakravarty, W. K. Chatila, I. de Bruijn, J. Gao, B. E. Gross, Z. J. Heins, R. Kundra, K. La, B. Kumar, J. Lang, N. L. Lehman, R. Mandt, P. Nguyen, R. Pilarski, K. Rai, L. Schoenfield, M. Ladanyi, A. Luna, M. G. Nissan, A. Ochoa, S. M. Phillips, E. Reznik, F. Sanchez-Vega, K. Senecal, P. Wakely, P. Hansen, R. Lechan, J. Powers, A. Tischler, W. E. Grizzle, C. Sander, N. Schultz, R. Sheridan, S. O. Sumer, Y. Sun, B. S. Taylor, J. Wang, H. Zhang, K. C. Sexton, A. Kastl, J. Henderson, S. Porten, J. Waldmann, M. Fassnacht, S. L. Asa, P. Anur, M. Peto, P. Spellman, C. Benz, J. M. Stuart, C. K. Wong, C. Yau, D. N. Hayes, D. Schadendorf, M. Couce, M. Graefen, H. Huland, G. Sauter, T. Schlomm, R. Simon, M. D. W. Parker, A. Ally, M. Balasundaram, R. Bowlby, D. Brooks, R. Carlsen, E. Chuah, P. Tennstedt, O. Olabode, M. Nelson, O. Bathe, P. R. Carroll, J. M. Chan, P. Disaia, P. Glenn, N. Dhalla, R. Holt, S. J. M. Jones, K. Kasaian, D. Lee, Y. Ma, M. A. Marra, M. Mayo, R. K. Kelley, C. N. Landen, J. Phillips, M. Prados, J. Simko, K. Smith-McCune, S. VandenBerg, R. A. Moore, A. J. Mungall, K. Mungall, A. G. Robertson, S. Sadeghi, J. E. Schein, K. Roggin, A. Fehrenbach, A. Kendler, S. Sifri, R. Steele, A. Jimeno, F. Carey, I. Forgie, P. Sipahimalani, A. Tam, N. Thiessen, K. Tse, T. Wong, A. C. Berger, R. Beroukhim, M. Mannelli, M. Carney, B. Hernandez, B. Campos, C. Herold-Mende, C. Jungk, A. D. Cherniack, C. Cibulskis, S. B. Gabriel, G. F. Gao, G. Ha, M. Meyerson, S. E. Schumacher, A. Unterberg, A. von Deimling, A. Bossler, J. Galbraith, L. Jacobus, M. Knudson, T. Knutson, J. Shih, M. H. Kucherlapati, R. S. Kucherlapati, S. Baylin, L. Cope, L. Danilova, D. Ma, M. Milhem, R. Sigmund, A. K. Godwin, R. Madan, H. G. Rosenthal, C. Adebamowo, M. S. Bootwalla, P. H. Lai, D. T. Maglinte, D. J. Van Den Berg, D. J. Weisenberger, S. N. Adebamowo, A. Boussioutas, D. Beer, T. Giordano, A.-M. Mes-Masson, F. Saad, J. T. Auman, S. Balu, T. Bodenheimer, C. Fan, K. A. Hoadley, A. P. Hoyle, S. R. Jefferys, T. Bocklage, L. Landrum, R. Mannel, K. Moore, K. Moxley, R. Postier, J. Walker, R. Zuna, C. D. Jones, S. Meng, P. A. Mieczkowski, L. E. Mose, A. H. Perou, C. M. Perou, J. Roach, M. Feldman, F. Valdivieso, R. Dhir, J. Luketich, E. M. M. Pinero, M. Quintero-Aguilo, Y. Shi, J. V. Simons, T. Skelly, M. G. Soloway, D. Tan, U. Veluvolu, H. Fan, T. Hinoue, C. G. Carlotti, J. S. D. Santos, R. Kemp, A. Sankarankuty, D. Tirapelli, J. Catto, K. Agnew, P. W. Laird, H. Shen, W. Zhou, M. Bellair, K. Chang, K. Covington, C. J. Creighton, H. Dinh, E. Swisher, J. Creaney, B. Robinson, C. S. Shelley, E. M. Godwin, S. Kendall, C. Shipman, H. Doddapaneni, L. A. Donehower, J. Drummond, R. A. Gibbs, R. Glenn, W. Hale, Y. Han, C. Bradford, T. Carey, A. Haddad, J. Moyer, L. Peterson, M. Prince, L. Rozek, G. Wolf, J. Hu, V. Korchina, S. Lee, L. Lewis, W. Li, X. Liu, M. Morgan, D. Morton, D. Muzny, R. Bowman, K. M. Fong, I. Yang, R. Korst, W. K. Rathmell, J. L. Fantacone-Campbell, J. Santibanez, M. Sheth, E. Shinbrot, L. Wang, M. Wang, D. A. Wheeler, L. Xi, F. Zhao, J. A. Hooke, A. J. Kovatich, C. D. Shriver, J. DiPersio, B. Drake, R. Govindan, S. Heath, T. Ley, J. Hess, E. L. Appelbaum, M. Bailey, M. G. Cordes, L. Ding, C. C. Fronick, L. A. Fulton, B. Van Tine, P. Westervelt, M. A. Rubin, J. I. Lee, N. D. Aredes, A. Mariamidze, Scalable open R. S. Fulton, C. Kandoth, E. R. Mardis, M. D. McLellan, C. A. Miller, H. K. Schmidt, science approach for mutation calling of tumor exomes using multiple genomic R. K. Wilson, D. Crain, E. Curley, J. Gardner, K. Lau, D. Mallery, S. Morris, J. Paulauskis, pipelines. Cell Systems 6, 271–281.e7 (2018). R. Penny, C. Shelton, T. Shelton, M. Sherman, E. Thompson, P. Yena, J. Bowen, 50. K. J. Karczewski, L. C. Francioli, G. Tiao, B. B. Cummings, J. Alföldi, Q. Wang, R. L. Collins, J. M. Gastier-Foster, M. Gerken, K. M. Leraas, T. M. Lichtenberg, N. C. Ramirez, L. Wise, K. M. Laricchia, A. Ganna, D. P. Birnbaum, L. D. Gauthier, H. Brand, M. Solomonson, N. A. Watts, E. Zmuda, N. Corcoran, T. Costello, C. Hovens, A. L. Carvalho, A. C. de Carvalho, D. Rhodes, M. Singer-Berk, E. M. England, E. G. Seaby, J. A. Kosmicki, R. K. Walters, J. H. Fregnani, A. Longatto-Filho, R. M. Reis, C. Scapulatempo-Neto, H. C. S. Silveira, K. Tashman, Y. Farjoun, E. Banks, T. Poterba, A. Wang, C. Seed, N. Whiffin, J. X. Chong, D. O. Vidal, A. Burnette, J. Eschbacher, B. Hermes, A. Noss, R. Singh, M. L. Anderson, K. E. Samocha, E. Pierce-Hoffman, Z. Zappala, A. H. O’Donnell-Luria, E. V. Minikel, P. D. Castro, M. Ittmann, D. Huntsman, B. Kohl, X. Le, R. Thorp, C. Andry, E. R. Duffy, B. Weisburd, M. Lek, J. S. Ware, C. Vittal, I. M. Armean, L. Bergelson, K. Cibulskis, K. M. Connolly, V. Lyadov, O. Paklina, G. Setdikova, A. Shabunin, M. Tavobilov, C. McPherson, R. Warnick, M. Covarrubias, S. Donnelly, S. Ferriera, S. Gabriel, J. Gentry, N. Gupta, T. Jeandet, R. Berkowitz, D. Cramer, C. Feltmate, N. Horowitz, A. Kibel, M. Muto, C. P. Raut, A. Malykh, D. Kaplan, C. Llanwarne, R. Munshi, S. Novod, N. Petrillo, D. Roazen, V. Ruano-Rubio, J. S. Barnholtz-Sloan, W. Barrett, K. Devine, J. Fulop, Q. T. Ostrom, K. Shimmel, Y. Wolinsky, A. Saltzman, M. Schleicher, J. Soto, K. Tibbetts, C. Tolonen, G. Wade, M. E. Talkowski, The A. E. Sloan, A. De Rose, F. Giuliante, M. Goodman, B. Y. Karlan, C. H. Hagedorn, J. Eckman, Genome Aggregation Database Consortium, B. M. Neale, M. J. Daly, D. G. MacArthur, J. Harr, J. Myers, K. Tucker, L. A. Zach, B. Deyarmin, H. Hu, L. Kvecher, C. Larson, R. J. Mural, Variation across 141,456 human exomes and genomes reveals the spectrum of S. Somiari, A. Vicha, T. Zelinka, J. Bennett, M. Iacocca, B. Rabeno, P. Swanson, M. Latour, loss-of-function intolerance across human protein-coding genes. bioRxiv 531210 L. Lacombe, B. Têtu, A. Bergeron, M. McGraw, S. M. Staugaitis, J. Chabot, H. Hibshoosh, [Preprint]. 13 August 2019. https://doi.org/10.1101/531210. A. Sepulveda, T. Su, T. Wang, O. Potapova, O. Voronina, L. Desjardins, O. Mariani, 51. A. Bairoch, The Cellosaurus, a cell-line knowledge resource. J. Biomol. Tech. 29, 25–38 S. Roman-Roman, X. Sastre, M.-H. Stern, F. Cheng, S. Signoretti, A. Berchuck, D. Bigner, (2018). E. Lipp, J. Marks, S. McCall, R. McLendon, A. Secord, A. Sharp, M. Behera, D. J. Brat, A. Chen, 52. R. M. Meyers, J. G. Bryan, J. M. McFarland, B. A. Weir, A. E. Sizemore, H. Xu, N. V. Dharia, K. Delman, S. Force, F. Khuri, K. Magliocca, S. Maithel, J. J. Olson, T. Owonikoko, A. Pickens, P. G. Montgomery, G. S. Cowley, S. Pantel, A. Goodale, Y. Lee, L. D. Ali, G. Jiang, R. Lubonja, S. Ramalingam, D. M. Shin, G. Sica, E. G. Van Meir, H. Zhang, W. Eijckenboom, A. Gillis, W. F. Harrington, M. Strickland, T. Wu, D. C. Hawes, V. A. Zhivich, M. R. Wyatt, Z. Kalani, Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 14 of 15 SCIENCE ADVANCES | RESEARCH ARTICLE J. J. Chang, M. Okamoto, K. Stegmaier, T. R. Golub, J. S. Boehm, F. Vazquez, D. E. Root, methodology; F.S.: conceptualization, funding acquisition, methodology, project W. C. Hahn, A. Tsherniak, Computational correction of copy number effect improves administration, supervision, writing, revision, and editing. Competing interests: The authors specificity of CRISPR-Cas9 essentiality screens in cancer cells. Nat. Genet. 49, 1779–1784 declare that they have no competing interests. Data and materials availability: All data (2017). needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. Raw data can be downloaded from the publicly available databases Acknowledgments: The results published here are, in whole or part, based on data generated as described in Materials and Methods. Additional data related to this paper may be requested by the TCGA Research Network, the Genomics of Drug Sensitivity in Cancer (GDSC) Project, the from the authors. Cancer Cell Line Encyclopedia (CCLE) Project, the Cancer Dependency Map (DepMap) Project (in particular, Project Score), and the Genotype-Tissue Expression (GTEx) Project. In addition, Submitted 12 November 2019 we thank all the members from the Genome Data Science group for insightful discussions. Accepted 1 May 2020 Funding: F.S. is funded by the ERC StG 757700 HYPER-INSIGHT and by MINECO grant Published 1 July 2020 BFU2017-89833-P. M.S. is funded by the FPU2017 fellowship of the Spanish government. We 10.1126/sciadv.aba1862 acknowledge funding from the Severo Ochoa Center of Excellence award to the IRB Barcelona. Author contributions: M.S.: conceptualization, data curation, formal analysis, investigation, Citation: M. Salvadores, F. Fuster-Tormo, F. Supek, Matching cell lines with cancer type and subtype methodology, validation, visualization, writing, revision, and editing; F.F.-T.: data curation and of origin via mutational, epigenomic, and transcriptomic patterns. Sci. Adv. 6, eaba1862 (2020). Downloaded from http://advances.sciencemag.org/ on February 11, 2021 Salvadores et al., Sci. Adv. 2020; 6 : eaba1862 1 July 2020 15 of 15 Matching cell lines with cancer type and subtype of origin via mutational, epigenomic, and transcriptomic patterns Marina Salvadores, Francisco Fuster-Tormo and Fran Supek Sci Adv 6 (27), eaba1862. DOI: 10.1126/sciadv.aba1862 Downloaded from http://advances.sciencemag.org/ on February 11, 2021 ARTICLE TOOLS http://advances.sciencemag.org/content/6/27/eaba1862 SUPPLEMENTARY http://advances.sciencemag.org/content/suppl/2020/06/29/6.27.eaba1862.DC1 MATERIALS REFERENCES This article cites 49 articles, 10 of which you can access for free http://advances.sciencemag.org/content/6/27/eaba1862#BIBL PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions Use of this article is subject to the Terms of Service Science Advances (ISSN 2375-2548) is published by the American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. The title Science Advances is a registered trademark of AAAS. Copyright © 2020 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of Science. No claim to original U.S. Government Works. Distributed under a Creative Commons Attribution NonCommercial License 4.0 (CC BY-NC).