MYC activity inference captures diverse mechanisms of aberrant MYC pathway activation in human cancers

Abstract

MYC is deregulated in more than 50% of all cancers. While MYC amplification is the most common MYC-deregulating event, many other alterations can increase MYC activity. We thus systematically investigated MYC pathway activity across different tumor types. Using a logistic regression framework, we established tumor type-specific, transcriptomic-based MYC activity scores that can accurately capture MYC activity. We show that MYC activity scores reflect a variety of MYC-regulating mechanisms, including MYCL and/or MYCN amplification, MYC promoter methylation, MYC mRNA expression, lncRNA PVT1 expression, MYC mutations and viral integrations near the MYC locus. Our MYC activity score incorporates all of these mechanisms, resulting in better prognostic predictions compared to MYC amplification status, MYC promoter methylation and MYC mRNA expression in several cancer types. In addition, we show that tumor proliferation and immune evasion are likely contributors to this reduction in survival. Lastly, we developed a MYC activity signature for liquid tumors in which MYC translocation is commonly observed, suggesting that our approach can be applied to different types of genomic alterations. In conclusion, we developed a MYC activity score that captures MYC pathway activity and is clinically relevant.

Introduction

Transcription factor (TF) c-MYC (MYC) is among the most critical oncogenes in cancer. MYC activity is tightly regulated in normal cells at transcriptional and post-transcriptional levels (1,2). It is also induced in response to various mitogenic stimuli, including Wnt and EGF (3–6). Aberrant activation of the MYC pathway through MYC amplification is frequently observed in cancer (7). MYC deregulation impacts almost every aspect of tumorigenesis, leading to the promotion of unrestricted cell proliferation, inhibition of cell differentiation, reduction of cell adhesion, and increased genomic instability (8,9).

While MYC amplification is the predominant genetic alterations that hyperactivates MYC and occurs in ~20% of all solid tumors (7), several other oncogenic mechanisms can also lead to aberrant MYC activity, including hyperactivity of growth factor pathways (10), post-translational modifications (11), increased MYC protein stability (12–15), MYC mutations (16,17), MYC translocation (18), enhancer-dependent mechanisms (19,20), changes in MYC binding partners and amplifications of MYC family members MYCN or MYCL (7). These additional MYC-related alterations have important clinical implications. For example, although MYC amplifications are rarely detected in liposarcomas, high MYC protein expression demarcates a subset of patients with poor prognosis (21). In neuroblastoma, MYC enhancer hijacking or MYC translocations occur in a subset of patients in the absence of MYC amplification events (22).

MYC amplification is easily detected by FISH, and therefore has been commonly used as a proxy for enhanced MYC activity if present. However, due to the existence of alternative MYC-activating mechanisms, MYC amplification status does not identify all tumors with aberrant MYC activity, leading to “false negatives”. Additionally, not all samples with MYC amplifications are associated with MYC pathway hyperactivation. Passenger-like amplifications of the MYC gene have been reported in some tumors, as manifested by the low correlation between MYC amplification and overexpression (23).

In line with these issues, several studies have developed gene signatures to infer MYC activity in tumor samples and have found that inferred MYC activity provided a better prognostic biomarker than MYC amplification status (24–29). In breast cancer, Chandriani et al. proposed a MYC gene expression signature predictive of survival (28) and Tanioka et al. identified a MYC signature that predicted worse survival in Jun-deleted luminal breast cancer samples (27). In liver cancer, Dang et al. found that a NELFE-dependent MYC signature was predictive of NELFE/MYC-driven tumors which could not be identified by MYC gene amplification or translocation alone (26). In neuroblastoma, a MYCN signature was proposed that identified a particular patient subset with poor prognosis who were otherwise predicted to be at low- or intermediate-risk for adverse outcome by conventional genomic assays (24). Similarly, a 157-gene signature for MYCN activity inference was shown to be predictive of patient prognosis in neuroblastoma with normal MYCN copy number (25). Lastly, Jung et al recently showed that an 18-gene MYC signature was prognostic in breast cancer, diffuse large B-cell lymphoma (DLBCL), high-grade serous epithelial ovarian cancer, medulloblastoma, neuroblastoma (29). This signature was also capable of identifying a MYCN-related molecular subtype associated with poor prognosis (29). Although these studies have provided important insights into MYC activity, only a limited number of cancer types have been investigated in this manner.

The Cancer Genome Atlas (TCGA) Network has recently published a pan-cancer study that investigated genomic and transcriptomic features associated with MYC across 33 cancer types (7). This study suggests that up to 50% of patients in 24/33 cancer types exhibit MYC-related alterations, including MYCN/MYCL amplifications and mutations in MYC binding proteins. However, each alteration was investigated individually, preventing conclusions about general MYC activity. Thus, given the previous observation that MYC activity more accurately reflects MYC pathway deregulation than MYC amplification status, we argue that a systematic investigation of MYC pathway activity across different cancer types will provide more comprehensive and precise insights in the contribution of MYC deregulation to tumor development and progression.

In this study, we applied a gene signature-based method to calculate MYC activity in tumor samples based on transcriptomic profiles, and systematically investigated the association of MYC activity with critical genomic and clinical factors in 18 cancer types. We show that our MYC activity inference captures a range of MYC aberrations, including amplification of MYCN or MYCL, changes in MYC promoter methylation, MYC mRNA expression and altered expression of lncRNA PVT1. Moreover, we found that MYC activity scores provide a better metric for predicting prognosis compared to MYC amplification and can even identify high-risk samples without MYC amplifications in several cancer types. We show that the poor clinical outcome associated with high-MYC-activity samples might be explained by increased proliferation and immune evasion. Lastly, we define a prognostic DLBCL-specific MYC activity score as a proof-of-principle, suggesting that our framework can be useful in studying a range of genetic alterations.

Materials and Methods

Utilized datasets

All datasets used in this study are publicly available. RNAseq data generated by The Cancer Genome Atlas (TCGA) for urothelial bladder carcinoma (BLCA), Brain Lower Grade Glioma (LGG), Breast invasive carcinoma (BRCA), Cervical squamous cell carcinoma and endocervical adenocarcinoma (CESC), Colon adenocarcinoma (COAD), Esophageal carcinoma (ESCA), Head and Neck squamous cell carcinoma (HNSC), Liver hepatocellular carcinoma (LIHC), Lung adenocarcinoma (LUAD), Lung squamous cell carcinoma (LUSC), Ovarian serous cystadenocarcinoma (OV), Pancreatic adenocarcinoma (PAAD), Prostate adenocarcinoma (PRAD), Rectum adenocarcinoma (READ), Sarcoma (SARC), Skin Cutaneous Melanoma (SKCM), Stomach adenocarcinoma (STAD), Uterine Corpus Endometrial Carcinoma (UCEC), were obtained from TCGA on FireBrowse (gdac.broadinstitute.org/). These datasets included 7033 Level 3 RNAseq samples with matched clinical information and provided RSEM normalized gene expression for 20,502 genes.

Additional gene expression datasets were obtained from PREdiction of Clinical Outcomes from Genomic profiles (PRECOG, https://precog.stanford.edu) (30) as normalized datafiles. Datasets were filtered by tissue type to match TCGA tumor types and only datasets with more than 40 samples and over 20% mortality rate were included in our analysis. This resulted in the inclusion of 58 datasets (see Supp. Table 8). Gene expression datasets were also obtained from Gene Expression Omnibus (GEO) (see Supp. Table 8) for accession numbers). These datasets were provided as normalized expression at the probeset level, in which some genes might be represented by multiple probesets. We converted probeset expression into gene expression values. Specifically, for one-channel arrays, we selected the probeset with the highest hybridization intensity across all samples to represent gene expression. For two-channel arrays, the average expression values of all probesets were calculated to represent gene expression. Datasets from one-channel arrays were further median normalized for each gene to transform intensities into relative expression values.

TCGA MAF files were obtained from https://gdc.cancer.gov/about-data/publications/pancanatlas. TCGA DNA methylation profiles were downloaded from FireBrowse (http://firebrowse.org). Based on the annotation file for Infinium HumanMethylation450K BeadChip platform, CpGs located in the MYC promoter were identified. Methylation levels of CpGs were then averaged to obtain level of MYC promoter methylation. Aneuploidy scores were downloaded as a supplemental file from prior work (31). MYC fusion information was obtained from prior works (32,33). Viral integrations near the MYC locus were obtained from a previous publication (34). Integrations were considered to be near the MYC locus if they took place within the 128,000,000 to 129,600,000 base pair-window on chromosome 8. Macrophage regulation scores were downloaded as a supplemental file from prior work (35).

MYC activity signature generation

MYC/MYN/MYCL copy numbers were obtained from previous work (7), which contained focal copy number data for 8884 TCGA tumor samples. These focal copy number data were estimated based on In Silico Admixture Removal (ISAR) correction (36) using purity and ploidy values as inputs. GISTIC 2.0 (37) was applied to the ISAR-corrected data to identify significant changes in copy number. The resulting values were log2-transformed $\left(\text{log}2\left(\frac{\text{copy}\text{number}}{2}\right)\right)$ and represent relative copy number of focal events, which are alterations that cover less than 50% of the chromosomal arm. Similar to previous work (7), a low minimum amplitude threshold of 0.1, which eliminates low-level artifactual segments, was used to indicate samples with MYC/MYCN/MYCL amplifications.

For each tumor type, we used logistic regression to assess the association between each gene and MYC amplification status as follows: $f\left(X_i,Z_i,G_i, S_i\right)=\frac{1}{1+e^{- \left({\beta }_0+{\beta }_1{X}_i+{\beta }_2{Z}_i+{\beta }_3{G}_i+{\beta }_4{S}_i\right)}}$ where X_i indicates gene expression of gene i, Z_i indicates the patient’s age, G_i indicates the patient’s gender, and S_i indicates the tumor stage of the according sample. The choice of only including age, stage and gender as confounding variables was based on the relatively complete information that was available for these variables. Additional clinical features including lymph node status and tumor grade were considered but the limited availability of this information in some cancer types would lead to the exclusion of these samples, which would limit our sample size. We excluded gender as a covariate in the logistic regression model for cancers dominated by one gender, which included BRCA, CESC, OV, PRAD, and UCEC. The response variable Y was set to 0 for non-MYC amplified samples and 1 for MYC amplified samples. The R glm function was utilized to fit each regression model using family = “binomial”. P-values for β₁ were collected for each logistic regression model, log10 transformed and separated genes in an “up” and “down” list based on the sign of β₁; positive β₁ coefficients classified genes as “up” and negative coefficients as “down”. The “up” gene list represents genes more likely to be upregulated in the presence of MYC amplifications, while “down” genes are more likely to be downregulated. Gene weights were capped at 10 and rescaled between 0 and 1 to generate final weight profiles. An identical procedure was followed for DLBCL using GSE4475 as training dataset and MYC translocation status as dependent variable. All MYC weight profiles are provided in Supp. Table 2.

MYC activity scoring

MYC activity scores were calculated using the rank-based Binding Associated with Sorted Expression (BASE) algorithm (38). This algorithm assesses how similar a patient gene expression profile is to a predefined weight matrix and consequently assigns representative patient-specific scores. Due to its scale-free nature, samples can be compared within a dataset, but not between datasets. For each dataset, the patient gene expression matrix along with the appropriate MYC signature were inputted into BASE. If the gene expression dataset was a one channel microarray and RNAseq dataset, we median normalized the data. The “up” and “down” MYC weight profiles were assessed separately. First, BASE ranks patient gene expression values (g_j) based on expression level. The “up” or “down” MYC weights (w_j) were also ranked to match the order of the ranked patient profile. For example, w₁will represent the weight for the highest ranked gene g₁ in the patient profile. Two cumulative distributions were then generated by a foreground (f) and background (b) function. These functions are given by:

$$\text{f}(j)=\frac{{\sum }_{j=1}^i|g_j{w}_j|}{{\sum }_{j=1}^n|g_j{w}_j|},1\le i\le \text{n}$$

$$\text{b}(j)=\frac{{\sum }_{j=1}^i|g_j\left(1-w_j\right)|}{{\sum }_{j=1}^n|g_j\left(1-w_j\right)|},1\le i\le n$$

These functions assess the distribution of weights within the patient gene expression profile and compare this to the background distribution. If highly ranked genes in a patient profile have high weights in the signature, the foreground function will increase much more rapidly than the background function, representing the enrichment of highly weighted genes. BASE records the maximal deviation between these two distributions to quantify this enrichment. BASE then adjusts this enrichment score by preforming 1000 iterations of randomized MYC signatures to generate a normalized enrichment score for the “up” and “down” profile. Lastly, the enrichment score for the “down” profile is subtracted from the enrichment score of the “up” profile to generate a final MYC activity score. Higher MYC activity scores represent greater MYC activity and lower scores lower activity. We have provided an R script on Code Ocean to facilitate the use of our MYC signatures on other gene expression datasets (https://codeocean.com/capsule/6889906/).

Signature cross-validation

Ten-fold cross-validation was performed during signature generation. Samples were randomly separated into two equally-sized groups. No consideration was given to dividing MYC amplified samples equally. MYC activity signatures were generated on one group using methods described above and signatures were applied to the unseen test group. AUC values evaluating the ability of the MYC signatures in predicting MYC amplification status were calculated and plotted using the pROC R package.

Survival analysis

Survival analyses were performed using the R survival package (version 3.1–8). Log-rank tests were performed to evaluate overall or relapse-free survival probabilities between two groups, using the survdiff function. Kaplan-Meier plots were generated using the survfit function. For all reported KM plots and accompanying log-rank tests, the number of MYC high samples was set to be identical to the number of MYC amplified samples in that cancer type in the TCGA dataset. This was done to allow for direct comparisons between different predictors while avoiding potential differences in power due to differently sized groups. For any comparison within non-MYC amplified samples, we considered the percentage of MYC amplified samples as a cutoff. Cox proportional hazards models were performed on continuous MYC activity scores in a univariate regression model, using the coxph function. Accompanying p-values were obtained from a two-sided Wald test. Forest plots were based on the results of multivariate Cox proportional hazards models in which all variables specified in the figure panels were included.

For the survival analysis in the PRECOG dataset, MYC activity scores were first inferred in a cancer type-specific manner. After computing the inferred MYC activities, we fitted a univariate Cox regression model to measure the association between MYC activity and all-cause or disease-specific mortality (if available). z-scores were extracted from the fitted models and a meta z-score was calculated for MYC scores across microarray datasets of the same cancer type. A meta z-score was calculated using the weighted Stouffer’s z-score method which uses the dataset sample size as weights. Meta p-values were calculated from the meta z-scores by referring to the standard normal distribution. Meta p-values were adjusted for each cancer type using the Benjamini-Hochberg multiple testing correction method.

Comparison of published breast cancer MYC activity signatures

The prognostic relevance of published MYC activity signatures (27–29) for breast cancer were compared using the breast cancer datasets available in PRECOG. The work of Tanioka et al. contained three MYC activity signatures based on previous publications (39–41) which were evaluated independently. MYC activity scores were calculated similarly as described in associated publications. In brief, the average gene expression of ~400 Core MYC signature genes was used as MYC activity score for the signature of Chandriani et al. (2009). The summed ranks of 18 MYC signature genes were normalized by the summed ranks for all genes to represent MYC activity scores from Jung et al. (2017). MYC activity scores for the three MYC signatures reported by Tanioka et al. (2018) were calculated using the fGSEA R package (version 1.12.0) (42). The z-score of gene expression values were calculated across patients for each gene and patient profiles were ranked based on these z-scores to generate a pre-ranked profile. Using these MYC activity scores, meta-p values were calculated for breast cancer dataset in PRECOG identically as described in the “Survival analysis” section.

Gene set enrichment analysis

Pre-ranked Gene Set Enrichment Analysis (GSEA) was performed on MYC activity signatures to annotate signatures with biological functions. Genes in each signature were ranked based on gene weights from high to low. The GSEA software (version 4.1.0) (43) was utilized to calculate enrichments of the C2 collection of pathways (version 7.1). For Suppl Figure 1B, pathways traditionally associated with MYC overactivity including those related to ribosome synthesis, RNA processing, and translation were selected for a side-by-side comparison among cancer types.

Calculation of tumor purity and infiltration scores

Tumor purity scores of TCGA datasets were obtained from previous work (35). ESTIMATE inferences were obtained from http://bioinformatics.mdanderson.org/estimate/ (December, 2017). Immune infiltration scores of specific immune cell types were calculated using our established framework described in (44). In short, immune cell-specific weight profiles and a patient gene expression dataset were inputted into BASE to infer the infiltration of selected immune cells (naïve B cells, memory B cells, CD4 T cells, CD8 T cells, NK cells, and monocytes) identically as described in the “MYC activity inference” section.

Correlations

The Spearman correlation coefficient (SCC) was reported for all correlation analyses as the assumptions underlying the Pearson correlation (i.e., normal distribution, homoscedasticity or linearity) were not met. SCC was calculated using the R function cor and significance was assessed using cor.test. All analyses were conducted in R (version 3.6.1).

Results

MYC amplification status has limitations for functional assessments of MYC in cancer

In addition to MYC amplification, other mechanisms can lead to aberrant activation of the MYC pathway in cancer (Figure 1A). A number of cancer types investigated in our study are not classically associated with MYC amplifications. To exemplify that other mechanisms can contribute to MYC overactivity, we thus decided to focus on such a cancer type and highlighted Skin Cutaneous Melanoma (SKCM) here. Using the TCGA SKCM dataset, we compared the expression level of MYC between MYC-amplified and non-amplified samples. As expected, MYC-amplified samples had significantly higher MYC expression compared to non-MYC amplified samples (p=2E-5, two-sided Wilcoxon Rank Sum, Figure 1B). However, it is notable that 120 out of 385 (31%) of samples without MYC amplifications had higher MYC expression levels than the median MYC expression of MYC amplified samples (Supp. Table 1). This suggests that high MYC expression is driven by other mechanisms than MYC amplification in a significant proportion of tumor samples. Furthermore, 22 out of 77 samples (29%) with MYC amplifications had lower MYC mRNA expression than the median of non-MYC amplified samples (Supp. Table 1), suggesting that not all MYC amplification events are associated with MYC upregulation. For example, in cancer types with high levels of aneuploidy (31), e.g. LUSC and ESCA, a larger proportion of MYC-amplified samples were observed, compared to those cancer types with lower CNVs (Figure 1C). Taken together, these findings indicate that a systematic investigation of MYC activity is necessary to better understand the implications of MYC pathway activation in cancer development and progression.

Figure 1. MYC amplification status may not correctly reflect MYC pathway activity in different cancer types.

A. Summary of potential mechanisms that affect MYC activity. B. Comparison of MYC mRNA expression in samples without and with MYC amplifications in SKCM. Exprs refers to RSEM-normalized gene expression. C. Correlation between the percentage of MYC amplified samples and median aneuploidy scores(31) across cancer types. D. Enrichment of MYC amplified samples within patients with high MYC scores in independent datasets for lung and breast cancer. E. Classification of four MYC groups using SKCM as example. Samples were ranked by MYC activity scores and colored by MYC amplification status. The percentage of MYC amplifications (17% in SKCM) was used as a cutoff to determine high and low MYC activity scores. F. MYC amplification status in samples with high MYC activity scores. H-Amp and H-NonAmp refer to samples with and without MYC amplifications, respectively, that had high MYC activity scores. Tumor types ordered based on the percentage of MYC amplified samples as indicated behind tumor names.

We have previously proposed a computational method to predict sample-specific P53 pathway activity based on the expression of genes that were associated with TP53 gene mutation status (45). Here, we adapted this method to calculate the activity of the MYC pathway in tumor samples. For each cancer type, we compared the expression of genes between MYC-amplified and non-amplified samples. The rationale being that MYC-amplified tumors are enriched for tumors with a hyper-activated MYC pathway, and therefore differential expression of genes from this comparison would reflect the transcriptomic output of MYC pathway activation. We thus defined cancer-type-specific signatures based on differentially expressed genes to gauge MYC activities in tumors (Supp. Figure 1A–B, Supp. Table 2–3, see Methods for details). In the TCGA datasets, a total of 18 cancer types contained at least 20 samples with focal MYC amplifications and had available transcriptomic profiles (Supp. Table 1). Our analysis hereafter focused on these cancer types.

As an initial evaluation of our MYC activity signatures, we applied the signatures to the TCGA datasets and used these signatures to calculate MYC activity scores for each patient. We then examined the relationship between MYC activity scores and MYC amplification status. In all cancer types, MYC scores were significantly higher in patients with MYC amplifications compared to patients without amplifications (Supp. Figure 2A). We performed ten-fold cross-validation in which we generated MYC signatures based on 50% of the data and applied the signature to the unseen test data ten times. This confirmed that MYC amplified samples consistently received higher MYC activity scores (Supp. Figure 2B). To investigate the relation between MYC amplification and MYC activity scores more precisely, we ranked patients based on MYC activity scores and assessed the rank of MYC amplified samples. As expected, most MYC amplified samples received high MYC activity scores and were significantly enriched among the most highly ranked samples (Supp. Figure 2C). Next, we evaluated the prediction accuracy of MYC activity scores in predicting MYC amplification status in independent datasets. MYC activity was significantly higher in patients with MYC amplifications compared to MYC wild-type samples in breast and lung cancer (Figure 1D). In addition, we compared our MYC activity inference with that of published MYC activity signatures to assess consistency with previously developed cancer type-specific MYC activity signatures. We collected five independent MYC activity signatures for breast cancer, inferred MYC activity for each signature (see methods), and correlated the resulting MYC activity scores with each other. We observed high consistency between our MYC activity inference and that of published MYC signatures as indicated by correlation coefficient ranging between 0.40 and 0.92 (Suppl. Figure 2D). Lastly, we validated our MYC activity inference using gene expression data from experimentally altered MYC models. In breast cancer, the MMTV-C-Myc murine model had significantly higher levels of MYC activity compared to normal murine mammary tissue (Supp. Figure 2E, p=0.02, two-sided Wilcoxon Rank Sum). In medulloblastoma, silencing of MYC using an siRNA approach reduced MYC activity compared to untreated and control siRNA treated tumors (Supp. Figure 2E, p=0.049, two-sided student T-test). Thus, our MYC activity signatures can accurately infer MYC activity in tumor samples.

Based on our analysis, we concluded that the majority of samples with MYC amplifications received high MYC activity scores in the TCGA datasets. However, we also noted a fraction of samples in each cancer type with MYC amplifications that were assigned low MYC scores and samples without MYC amplifications with high MYC activity scores. We thus generated four MYC activity groups based on MYC amplification status and MYC activity (Figure 1E, Supp. Table 4). We ranked patients based on MYC activity and used the number of MYC amplified samples as a cutoff for high and low MYC activity. On average, 52% of samples with high MYC activity scores (range 35–83%) were associated with MYC amplifications, whereas the other 48% (range 17–65%) had normal MYC copy number (Figure 1F). In addition, an average of 20% of samples (range 8–34) with low MYC activity scores had MYC amplifications (Supp. Figure 2F). In conclusion, we developed tumor specific MYC activity scores that represent MYC-related processes and identify samples that might not be driven by MYC amplification or samples with amplification-independent MYC alterations.

Alternative mechanisms associated with aberrant MYC activation in MYC non-amplified tumors

Since a large variety of mechanisms can affect MYC activity (7,10–20), we hypothesized that we could identify additional reasons for low or high MYC activity. We first investigated if MYC promoter methylation and/or mRNA expression could contribute to MYC activity, expecting that low methylation levels and high MYC mRNA would be associated with high MYC activity. Indeed, using the four identified MYC activity groups, hypo-methylation of the MYC promoter was observed in samples with high MYC activity in the TCGA SKCM dataset, irrespective of MYC amplification status (Figure 2A, p<0.001, two-sided Wilcoxon rank test) and MYC mRNA was significantly increased in samples receiving high MYC activity scores (Figure 2B, p<0.0001, two-sided Wilcoxon rank test), again irrespective of MYC amplification status. Notably, patients without MYC amplifications but with high MYC activity scores had significantly lower MYC promoter methylation and higher MYC mRNA levels compared to patients with amplifications but low MYC activity scores (Figure 2A, p<0.001; Figure 2B, p<0.0001, respectively, two-sided Wilcoxon rank test). To investigate the combination of promoter methylation and mRNA regulation, we stratified samples into low and high methylation and mRNA groups and assessed MYC activity. MYC mRNA expression seemed to be most strongly associated with MYC activity, as the majority of samples with low MYC mRNA received low MYC activity scores (Figure 2C). This pattern was very similar in patients with MYC amplifications and patients without MYC amplifications (Supp. Figure 3A–B). Expanding this analysis to all 18 cancer types in the TCGA cohort, we observed very similar patterns in about half of the cancer types; MYC activity was negatively correlated with promoter methylation and positively correlated with MYC mRNA expression, regardless of MYC amplification status (Figure 2D). This suggests that even during a MYC amplification event, MYC activity might still be regulated at the MYC locus through epigenetic and transcriptional regulation in a subset of tumor types.

Figure 2. Characteristics of non-MYC-amplified samples.

A. Boxplot of MYC promoter methylation in SKCM stratified by MYC amplification status and MYC activity. Outliers not indicated. B. Boxplot of MYC mRNA expression in SKCM stratified by MYC amplification status and MYC activity. Outliers not indicated. C. MYC activity in SKCM stratified by MYC promoter methylation and MYC mRNA expression. D. Spearman correlation between MYC promoter methylation, MYC mRNA expression with MYC activity in MYC amplified and non-MYC-amplified samples. E. Comparison of MYC activity between patients without (N) and with (Y) MYCL or MYCN amplifications, with low (<) or high (>) MYC promoter methylation, mRNA expression, or PVT1 expression. Low and high MYC promoter methylation, mRNA expression, or PVT1 expression groups were classified identically as shown in Figure 1E but using promoter methylation or expression as sorting feature instead of MYC activity. Numbers indicate number of samples with available data, color indicates median MYC activity. Two-sided Wilcoxon rank test, *: p<0.05, **: p<0.01, ***: p<0.001, ****: p<0.0001.

Next, we were particularly interested in tumors lacking MYC amplifications but exhibiting high MYC activity scores, because of the potential of “MYC addiction” in these tumors. Similar to the MYC addiction phenotype observed in MYC amplified samples, we could potentially identify additional samples that display MYC addiction in the absence of MYC amplification. MYC-addicted tumors are more sensitive to MYC inhibition (46), offering a potential therapeutic avenue for patients exhibiting high MYC activity. As exemplified in Figure 1A, putative mechanisms increasing MYC activity in these tumors include MYCL and MYCN amplifications (7), expression of lncRNA PVT1 (47), MYC translocations (18), viral integrations near the MYC locus (2) and MYC point mutations (16,17). LncRNA PVT1 has become of interest due to its close genomic location to the MYC locus. Studies have indicated a regulatory role of PVT1 in determining MYC expression (47). We first investigated MYCL and MYCN amplifications and expression of PVT1. For each of these alterations, patients were stratified into two groups based on the presence/absence of MYCN or MYCL amplifications or on high or low expression of PVT1. To classify samples into PVT1 low and high groups, we used an identical approach as Figure 1E, but using PVT1 expression as sorting feature instead of MYC activity. We also included MYC promoter methylation and mRNA expression as a reference. We then assessed MYC activity in each of these groups and clustered tumor types based on MYC activity (Figure 2E). Four observations emerged from this analysis. First, MYC mRNA again stood out as major factor related to MYC activity, being significantly associated with high MYC activity in 14 out of 18 tumor types. Second, patients with MYCL and MYCN amplifications had significantly higher MYC activity in 10 out of 18 tumor types (LUSC, COAD, ESCA, UCEC, PRAD, BRCA, LUAD, STES, HNSC and BLCA). BLCA showed an especially interesting pattern, with only MYCL and MYCN amplifications being significantly associated with MYC activity, but not MYC promoter methylation, MYC mRNA or PVT1 mRNA expression. Third, MYC activity was almost uniformly increased in samples with high PVT1 expression (12/18 cancer types), with only UCEC showing lower MYC activity in PVT1-high tumors. Lastly, several cancer types exhibited a somewhat diverse pattern. For example, decreased MYC promoter methylation in PAAD was associated with high MYC activity, but increased MYC mRNA expression also showed high MYC activity, potentially reflecting a counter mechanism in MYC activity regulation. In addition, we also considered that MYCN and MYCL methylation or mRNA expression could affect MYC activity. We thus performed the same analysis as in Figure 2E using MYCN and MYCL promoter methylation and mRNA expression (Supp. Figure 3C). However, no clear pattern seem to arise even though a number of comparisons were statistically significant. Due to the close genomic location of MYC and PVT1, we reasoned that the association between MYC activity and PVT1 expression could be affected by co-amplification of the two genes. However, correlations between PVT1 mRNA and MYC activity were similar in patients without and with MYC amplifications, suggesting that this is an unlikely explanation (Supp. Figure 3D). Collectively, MYCL and MYCN amplifications contribute to MYC activity in a number of cancer types and MYC and PVT1 mRNA expression are associated with MYC activity in most cancer types.

We next investigated MYC translocations, viral integrations near the MYC locus and MYC point mutations. Unlike amplifications, MYC mutations are uncommon in solid tumors. Commonly observed MYC mutations include S161, T58 and S62, of which the last two are relatively often observed in hematological cancers (16,17). Three recurrent alterations (>2 patients) were identified in this cohort; P74x (4 patients), S161L (6 patients), H374R (3 patients). All but P74S mutations and 1 patient with H374R were classified as low MYC activity samples (Supp. Table 5). Accordingly, mutations in S161L have been indicated to interfere with MYC binding to co-activator complexes (7), suggesting that our MYC activity score can accurately predict the effect of mutations on MYC activity. However, sample size is low preventing concrete conclusions. Similarly, only three MYC translocations were identified in this cohort of which only one was classified as non-MYC amplified, preventing further investigation. Lastly, viral integrations of HPV16 and HPV18 near the MYC locus have been described in several cancer types, but seem especially prominent in CESC (48). A total of 17 CESC samples displayed viral integrations (10 HPV16, 5 HPV18, and 2 HPV45) close to the MYC locus, which had significantly higher MYC activity compared to samples without nearby viral integrations (Supp. Figure 3E–F, p=5E-4, two-sided Wilcoxon Rank Sum). However, the majority of these samples (10/17) also had MYC amplifications and the MYC activity of the remaining 7 samples was no longer significantly higher compared to samples without integrations (P>0.05, two-sided Wilcoxon Rank Sum), suggesting that viral integration alone is insufficient to increase MYC activity. None of the other cancer types displayed viral integrations near the MYC locus. In conclusion, we investigated a number of mechanisms that affect MYC activity and found that MYCN or MYCL amplification, MYC mRNA expression and PVT1 mRNA expression are the most predominant mechanisms across cancer types that affect MYC activity that were assessed in this study.

MYC activity score provides a better prognostic metric than MYC amplification status

MYC amplification status has been shown to be a significant prognostic factor in multiple cancer types (24–29). As demonstrated above, the inferred MYC activity can more effectively recapitulate the aberrant activation of the MYC pathway through different mechanisms. We therefore hypothesized that the inferred MYC activity provides a better prognostic metric. Specifically, we examined the association of MYC activity with patient survival time and compared this to the prognostic ability of other MYC alterations, including MYC amplification, MYC mRNA and MYC promoter methylation. We initially selected three cancer types with the highest percentage of death events (percentage of death events >=50%), including ovarian cancer (OV), pancreatic cancer (PAAD), and melanoma (SKCM) for an in-depth prognostic analysis. As shown in Figure 3A, high MYC activity was significantly associated with poor survival in PAAD (HR=2.07, P=0.001, Cox P-H regression). In comparison, MYC amplification, although associated with prognosis, exhibited lower significance (HR=1.93, P=0.03, Cox P-H regression). MYC promoter methylation and mRNA were not significantly associated with survival (p>0.05, Cox P-H regression).

Figure 3. MYC activity score better predicts prognostic compared to MYC amplification or mRNA on the pan-cancer level.

A. KM plots of prognostic value of MYC amplification status, MYC promoter methylation, MYC mRNA and MYC activity in PAAD. B. KM plots of prognostic value of MYC amplification status, MYC promoter methylation, MYC mRNA and MYC activity in SKCM. C. Forest plot indicating the association between survival and clinical variables, MYC amplification, MYC promoter methylation, MYC mRNA expression, and MYC activity assessed by multivariate Cox regression. D. Meta-analysis of the association between MYC activity and survival in the PRECOG datasets using univariate Cox regression. Bars indicate −log10 (BH-adj. p-values). The bar for BLCA was cut to increase resolution of cancer types with lower significance. Dark grey bars indicate significant associations (BH-adj. meta-p < 0.05). E. Examples of datasets from PRECOG indicating that low MYC activity is associated with longer survival in PRAD, LUAD, BRCD, and BLCA. LR-p = log-rank p-value, HR = hazard ratio of univariate Cox regression model, Cox-p = univariate Cox regression p-value.

We observed a similar finding in SKCM, MYC activity being the most significant predictor of survival (p=3E-4, HR=1.68, Cox P-H regression, Figure 3B). Notably, MYC promoter methylation was also significantly associated with prognosis in SKCM (p=0.006, HR=0.53, Cox P-H regression, Figure 3B), exemplifying that information is missed when just considering MYC amplification status. In addition, continuous MYC activity scores were still prognostic in samples without MYC amplifications (Supp. Figure 4A–B). Since MYC amplification, MYC promoter methylation, and MYC mRNA expression are correlated with MYC activity (Figure 1D, 2B), the significant association between MYC activity and prognosis might be explained by these variables. We thus assessed the relationship between MYC activity and prognosis, while adjusting for MYC amplification status, MYC promoter methylation and MYC mRNA expression, as well as other clinical factors. Even after adjusting for these variables, MYC activity was still significantly associated with survival in SKCM (Figure 3C). No significant differences in survival were observed in OV.

In addition, other tumor characteristics, including the expression of E2F TFs and immune cell infiltration, might have affected the relationship between MYC activity and prognosis. MYC is a pivotal regulator of E2F TFs (49,50) and is also a direct target gene of E2F TFs (51), resulting in feedback loops. Increased expression of E2F TFs is negatively associated with prognosis in several cancer types (52), which we potentially captured with our MYC activity scores. The tumor microenvironment is another important factor that affects patient prognosis (44,53), which should be taken into consideration during prognostic analyses. We thus used multivariate Cox P-H regression to assess the relationship between prognosis and MYC activity scores while adjusting for the expression of E2F TFs and immune cell infiltration in SKCM. Using two separate methods to infer immune cell infiltration (53,54), we observed that MYC activity was still significantly associated with survival in both multivariate models (p=0.02 and p=0.046, Cox P-H regression, Suppl. Table 6). In addition, we observed that expression of E2F5 was also consistently associated with survival (p=0.06. and p=0.003, Cox P-H regression, Suppl. Table 6). Thus, these results indicate that additional features within tumor samples do not explain the observed prognostic relationship between MYC activity and patient survival in SKCM.

To extend our prognostic analysis to other cancer types, we selected all TCGA cancer types with survival data. Besides PAAD and SKCM, we observed that UCEC, BRCA, and ESCA also showed a significant relationship between MYC activity and patient survival (Supp. Figure 4C). To confirm and extend these results, we used the large collection of gene expression datasets with associated survival data in PRECOG to assess the prognostic relevance of MYC activity. We inferred MYC activity in each dataset and calculated meta-p values based on Cox P-H model p-values for each cancer type to assess overall significance. We confirmed the relation between MYC activity and prognosis in SKCM and PAAD (meta-p=5E-5, meta-p=3E-2, respectively), both showing poorer survival for patients with high MYC activity (Figure 3D). We also observed a strong relationship between MYC activity and prognosis in prostate cancer (PRAD) (meta-p=5E-35), where all three available datasets showed poor prognosis for high MYC activity patients (Figure 3D, Supp. Table 7). In addition, we also observed prognostic roles for MYC activity in LUAD, BRCA, BLCA, SARC, COAD and HNSC (meta-p=3E-12, 2e-12, 1e-6, 0.004, 0.03, and 0.03, respectively) (Figure 3D–E). Since the prognostic relevance of continuous MYC activity scores are difficult to interpret clinically, we sought to identify if stratifying MYC activity into high and low MYC activity groups could identify high-risk patients. We maintained the percentage of MYC amplified samples from the TCGA and stratified patients into high and low based on the percentage of MYC amplifications observed in the TCGA. MYC high groups had significantly lower survival in the four tested datasets for PRAD, LUAD, BRCA and BLCA (Figure 3E), indicating that MYC activity scores can identify patients with poor prognosis. In conclusion, MYC activity is prognostic in several cancer types and is more prognostic than MYC amplification status in PAAD and SKCM.

High MYC activity is associated with high proliferative rate in both MYC-amplified and non-amplified tumors

We hypothesized that the inverse correlation between MYC activity and patient survival time can be at least partially explained by the function of MYC in regulating cell proliferation. To test this hypothesis, we investigated the correlation between Ki67 expression levels and MYC activity in different cancer types in the TCGA dataset. Expression of Ki67 is often used as indicator of cell proliferation, high Ki67 expression indicating a high proliferative rate. In 9/18 tumor types, patients with MYC amplifications had significantly higher Ki67 expression (Figure 4A, p<0.05, two-sided Wilcoxon rank test), suggesting that the presence of a MYC amplification alters the proliferative potential of a tumor. Next, we stratified patients based on MYC activity and observed a significant difference in Ki67 expression, with samples with higher MYC activities showing higher Ki67 expression (Figure 4B, p<0.05, two-sided Wilcoxon rank test). This indicates that not just MYC amplifications, but also other MYC-related alterations can affect proliferation. Notably, when we further stratified patients by MYC amplification status, we observed that samples with MYC amplifications but low MYC activity scores had significantly lower proliferation rates than samples with MYC amplifications that received high MYC activity scores (Figure 4B, p<0.05, two-sided Wilcoxon rank test). In the same way that MYC gene expression does not accurately recapitulate MYC activity, MKI67 expression might similarly not be equivalent to the level of proliferation. We thus compared our findings to a proliferation signature (55) and observed an identical association between MYC activity and proliferation (Supp. Figure 5A–B). This supports our hypothesis that not all samples with MYC amplifications are driven by MYC. In conclusion, high MYC activity is associated with increased proliferation in several tumor types.

Figure 4. High MYC activity is associated with high proliferative rate of tumor cells.

A. Boxplots comparing Ki67 expression between non-MYC amplified and MYC amplified samples. Only cancer types with a significant difference in MKI67 expression show. B. Boxplots comparing Ki67 expression between non-MYC amp/MYC activity low, MYC amp/MYC activity low, non-MYC amp/MYC activity high, and MYC amp/MYC activity high samples. Two-sided Wilcoxon rank test, *: p<0.05, **: p<0.01, ***: p<0.001, ****: p<0.0001. Outliers not indicated. Exprs refers to RSEM-normalized gene expression.

High MYC activity is associated with an immune evasive microenvironment.

Another factor that determines patient prognosis is the tumor microenvironment (44,53). Previous studies have shown that MYC amplification status is negatively associated with immune infiltration (56–59). We thus investigated the association between MYC activity and immune-related characteristics in the TCGA dataset. First, we assessed the correlation between MYC activity and tumor purity, which indicates the abundance of cells other than tumor cells. Both a computational purity inference method (ESTIMATE) and immunohistochemistry (IHC) assessment of tumor purity revealed predominantly positive correlations between tumor purity and MYC activity (Figure 5A), indicating that as MYC activity increases, the tumor fraction increases as well. Reports have suggested that MYC prompts tumor immune evasion by promoting the expression of immune checkpoints, including CD47 and PD-L1, and the recruitment of suppressive cells (56–59). We indeed confirmed the positive correlation between MYC activity and immune checkpoints CD47 and PD-L1 in a number of cancer types (Supp. Figure 5C).

Figure 5. High MYC activity associated with high immune evasive microenvironment.

A. Spearman correlation between MYC activity and tumor purity determined by ESTIMATE and IHC. Darker and lighter colors indicated significant (p<0.01) and non-significant (p>0.01) correlations. Purple indicates correlations with IHC, orange represents correlations with ESTIMATE. B. Spearman correlation between MYC activity immune infiltration of six immune cell types. C. Volcano plot of Macrophage regulation score correlations with MYC activity. D. Spearman correlation between MYC activity and monocyte (Mono) and macrophage (MФ) chemokines.

To investigate this relationship with the immune microenvironment in more detail, we inferred immune infiltration of six immune cells. In accordance with purity findings, most cancer types had negative correlations with the six immune cell types (Figure 5B). Cytotoxic cells, including NK and CD8 T cells, were especially sparse in tumors with high MYC activity in the majority of tumor types. Contrary to other cancer types, OV showed positive correlations between MYC activity and cytotoxic cells and naive B cells. We hypothesize that this is related to the very high overall CNV burden in OV (Figure 1B), which might have different relations with the immune system compared to other cancer types.

We noticed that several cancer types showed positive correlations between macrophages and MYC activity. Macrophages tend to display an immunosuppressive phenotype in the tumor microenvironment (60), which would be in line with the proposed immunomodulating role of MYC (56–59). We thus hypothesized that the presence of myeloid cells might be tightly linked to high MYC activity. A macrophage regulation score has been proposed that captures the recruitment and differentiation of macrophages within a tumor (61). This score was significantly enriched in OV, SARC, BRCA, and COAD, potentially explaining increased infiltration of monocytes in these tumor types (Figure 5C). To validate this, we assessed the expression of myeloid-attracting chemokines. We consistently observed positive correlations with MYC activity and monocyte chemokines CCL7/CCL22/CCL2 and macrophage chemokines CCL3/CCL4/CCL5 expression in OV, BRCA, SARC, LGG, and READ (Figure 5D). Taken together, these findings suggest that MYC activity confers an immunosuppressive phenotype in most cancer types, except for OV. Immunosuppression might be mediated directly through the expression of immune checkpoints and indirectly through the recruitment of immunosuppressive myeloid cells.

MYC activity captures MYC rearrangements and better predicts prognosis in DLBCL

We have shown in solid tumors that MYC pathway activity can be correctly inferred from gene signatures defined by comparing MYC-amplified with non-amplified tumors. In certain liquid tumors, aberrant MYC activation is also frequently observed but caused by other MYC-related genomic events such as MYC gene rearrangement. For example, MYC rearrangements are the primary genetic event in Burkitt lymphoma (BL) (62). A closely related lymphoma, diffuse large B-cell lymphoma (DLBCL), displays MYC rearrangements in 5–15% of patients (63,64). As a proof of concept, we wanted to evaluate if our strategy on solid tumors could also be applied to other types of MYC alterations. We thus generated a MYC activity signature based on MYC rearrangement status in DLBCL and applied it to independent datasets.

We first evaluated MYC activity in patients with known MYC rearrangements. While MYC-Ig fusions are the most common type of MYC rearrangement in DLBCL, MYC fusion to non-Ig genes has been observed as well (64). We observed that both types of rearrangements received similar MYC activity scores, suggesting that either type of translocation increases MYC activity (Figure 6A). This was also recapitulated in high Area Under the Curve (AUC) scores (AUC = 0.96 and 0.87 respectively), which means that most samples with MYC rearrangements received high MYC activity scores (Figure 6B). We next assessed clinical relevance, as MYC rearrangement in DLBCL is associated with poor prognosis (63–65). We indeed confirmed the prognostic association between MYC rearrangement status and overall survival (Figure 6C–D). Consequently, MYC activity was also significantly associated with survival in the training dataset and was much more significantly associated with survival in the test dataset compared to MYC rearrangement or MYC mRNA expression (Figure 6D). Importantly, we identified a high-risk cohort displaying high MYC activity within non-MYC-rearranged samples (Figure 6C–D), indicating that additional MYC alterations might be involved in DLBCL. In conclusion, we have shown that MYC-rearrangement can be used to define MYC activity scores that are significantly associated with survival, even within patients without detected MYC rearrangements.

Figure 6. MYC activity is prognostic in DLBCL.

A. Boxplots showing MYC activity scores for patients with wild-type (WT) MYC or MYC rearrangements in the Hummel and Sha datasets. Patient with MYC rearrangements in the Hummel dataset were stratified further into non-Ig MYC rearrangements (MYC non-Ig) and Ig MYC rearrangements (MYC Ig). Two-sided Wilcoxon rank test, *: p<0.05, **: p<0.01, ***: p<0.001, ****: p<0.0001. B. AUC curves for Hummel and Sha datasets evaluating the ability of MYC activity scores to predict the presence or absence of MYC rearrangements. C. KM plots for MYC rearrangements, MYC activity scores and MYC activity scores in non-MYC rearranged patients in the Hummel dataset. D. KM plots for MYC rearrangements, MYC activity scores and MYC activity scores in non-MYC rearranged patients in the Sha dataset.

Discussion

MYC is the most frequently amplified oncogene in solid cancers (66) and is known to transcriptionally regulate many essential cellular pathways, including those involved in cell growth and proliferation (8,9). While MYC amplification is the most common deregulatory mechanism that increases MYC activity in tumor cells, several other mechanisms affect MYC activity (7,10–20). Consequently, MYC activity inference based on genomic assessment alone might inadequately reflect true MYC activity. Patients without MYC amplifications but with alternative MYC-increasing mechanisms might be missed, whereas patients with passenger-like MYC amplifications might be inappropriately included. To address this issue, we have developed tumor type specific MYC activity scores that infer MYC activity based on gene expression.

We have shown that the majority of patients with MYC amplifications receive high MYC activity scores, consistent with the notion that MYC is an important driver gene in cancer (7,66). Notably, a number of patients without MYC amplifications received high MYC activity scores, suggesting potential other mechanisms of MYC activation. We investigated MYC promoter methylation, MYC mRNA expression, MYCN/MYCL amplifications, PVT1 mRNA expression, viral integrations near the MYC promoter, MYC translocations and MYC mutations as potential contributors to MYC activity due to their direct mechanisms of affecting MYC. High MYC promoter methylation can reduce MYC gene accessibility and thereby decrease MYC activity. Since the half-life of the MYC protein is short (13), constant MYC mRNA expression is required and increases in mRNA abundance have been used as a proxy of increased MYC activity, which is associated with poor survival in triple negative breast cancer (TNBC) (67) and mantle cell lymphoma (68). We showed that MYC promoter methylation and/or MYC mRNA abundance might represent important levels of regulation in about half or the tumor types that were investigated, since promoter methylation and mRNA expression were consistently negatively and positively correlated with MYC activity, respectively

Both MYCN and MYCL have overlapping functions with MYC (69), and some of the published MYC signatures also capture MYCN amplifications (29), suggesting that amplifications in MYCL or MYCN can resemble MYC amplifications. We found that amplifications in these genes are associated with increased MYC activity in the absence of MYC amplifications in several cancer types. Most cancer types also showed significantly higher MYC activity in patients expressing high levels of PVT1. MYC and PVT1 are commonly co-amplified (70), suggesting a functional benefit of PVT1 gain. Indeed, high expression of PVT1 can increase MYC activity by stabilizing the MYC protein (70). In addition, PVT1 modulates MYC pathway activity by regulating key components downstream of MYC pathway signaling (47). Although numerous studies have examined the interaction of PVT1 and MYC, the detailed mechanisms of the interaction between them remain unclear.

MYC activity was much more prognostic compared to MYC amplification status in SKCM, PAAD, and DLBC, showing the clinical relevance of our study. High MYC activity scores conferred poor prognosis in a number of cancer types, consistent with previous publications (24–29). This might be due to tumor characteristics associated with high MYC activity, including increased cell proliferation and immunosuppression. MYC activity was significantly associated with expression of Ki67, a well-known marker of proliferation. In addition, we showed that MYC activity is positively correlated with monocyte infiltration, as well as myeloid-attracting chemokines in several cancer types. MYC-directed modulation of myeloid cells has been reported (59,71), validating our findings. In addition, MYC is directly involved in the transcription of immune checkpoints PD-L1 and CD47 (59) and interferes with IFNγ signaling (57), which is critical for effective immune cell activation (72). Coactivation of MYC and RAS in lung cancer promotes the secretion of CCL9 and IL-23, which mediate the recruitment of macrophages and exclusion of cytotoxic cells, respectively (71). Deactivation of MYC results in immediate reversal of stromal changes and reduced tumor burden, suggestion a prominent role for MYC in programming an immunosuppressive stroma (71).

Gene expression signatures have been proposed in previous studies to estimate MYC activity in tumor samples (24–29). In breast cancer, five gene signatures have been reported and have shown significant prognostic values. A comparison indicated comparable performance of our BRCA-specific signature with those signatures in terms of prognostic prediction (Supp. Figure 4D). Since we applied an unbiased generic framework for defining MYC signatures, the resulting cancer-specific signatures are more appropriate for systematic analyses of MYC activity and their clinical effect in all cancer types.

As a proof of concept, we also evaluated if other MYC alterations could be used to generate MYC activity scores. We used MYC rearrangements observed in DLBCL as an example and showed that MYC activity scores are a significant predictor of survival, even in patients without MYC rearrangements. This is consistent with prior findings in DLBCL showing that patients without MYC fusions can still present with high MYC mRNA or protein levels (73,74). This finding exemplifies the utility of our approach in capturing the downstream effect of genomic alterations in gene expression profiles. Taken together, our analyses show that MYC activity is more clinically useful than MYC amplification status and that our approach can be expanded to other types of genomic alterations.

Implications.

By using cancer type specific MYC activity profiles, we were able to assess MYC activity across many more tumor types than previously investigated. The range of different MYC-related alterations captured by our MYC activity score can be used to facilitate the application of future MYC inhibitors and aid physicians to pre-select patients for targeted therapy.