Integrative Clinical Genomics of Metastatic Cancer

Introduction

Tumor metastasis is the process in which cancer cells disperse from a primary site and progressively colonize distant organs. In over 90% of cases, metastatic spread of tumor cells is the greatest contributor to deaths from cancer^1,2. With the preponderance of cancer patients enrolled in early stage (Phase I-II) clinical trials harboring metastatic disease², and with the advent of genomic testing of tumors, there remains the promise of matching patients to the right therapy based on comprehensive molecular profiles³ of pathogenic somatic^4,5 and germline⁶ variants, and components of the functional genome, tumor phenotype, and tumor microenvironment afforded by RNA sequencing^7,8. While metastatic tumors share key driver mutations with the primary tumor from which they arose, they often develop new mutations as they evolve during metastasis and treatment⁹. Thus, it is preferable to match patients to potential therapies and clinical trials based on a real-time analysis of their metastatic tumor, rather than archival material of their primary tumor^2,10.

In 2010, we introduced the Michigan Oncology Sequencing (Mi-Oncoseq) Program, an IRB-approved protocol to carry out prospective, integrative exome and transcriptome sequencing of advanced cancer patients⁷, mirroring the efforts of the TCGA project which focused on generating exome and transcriptome sequence¹¹ in primary cancers. It was among the first comprehensive, clinical DNA and RNA sequencing programs offered for cancer patients^8,12–16. The purpose of the Mi-Oncoseq program was to determine the utility of genomic sequencing of tumors and germline coupled with a multi-disciplinary Precision Medicine Tumor Board (PMTB) in the management of advanced cancer patients^7,15. The program transitioned into sequencing in a clinical setting (under CLIA) as part of the Clinical Sequencing Exploratory Research (CSER) consortium in 2013^14–16.

In this study, we carried out clinical grade whole exome (tumor/normal) and transcriptome sequencing (i.e., integrative sequencing) of 500 cancer patients harboring metastatic cancers from over 30 primary sites and biopsied from over 22 organs (abbreviated as the “MET500” cohort). Sequencing matched tumor and normal samples from patients delineated potentially pathogenic germline alterations as well as provided high resolution copy number landscapes. RNA sequencing analysis provided insights into functional gene fusions, transcriptional pathway activation, and a landscape of immune infiltration.

The landscape of molecular aberrations in metastatic cancers

We successfully obtained 537 biopsies from 556 enrolled patients and obtained complete sequencing results on 500 patients with metastatic cancers, representing a 93% success rate. Reasons for failure included lack of tumor content on biopsy (37 cases, 6.6%), biopsy material not available (19 cases, 3.4%; patient declined biopsy, poor physical performance, unable to image site, unsafe for biopsy, insufficient tissue, or enrolled in other clinical trials). The majority of patients, 468 (93.6%) were patients seen at the University of Michigan Comprehensive Cancer Center, however patients from 21 other institutions were also enrolled. The patient demographics were 258 (51.6%) males, 242 females (48.4%), 460 (92%) Caucasian, and 40 (8%) non-Caucasian. The median age of the cohort was 59, with a range of 18 to 86 (Supplementary Table 1 and Extended Data Fig. 1a, b).

Fig. 1a portrays the cancer types (n=20) represented in the MET500 cohort. The top three cancer types in our cohort include 93 (18.6%) metastatic prostate cancers, 91 (18.2%) metastatic breast cancers, and 42 (8.4%) soft tissue sarcomas. There were also 25 (5%) carcinomas of unknown primary (CUP). Fig. 1b highlights the diverse metastatic sites analyzed (n>30) in the MET500 cohort. The most prevalent sites of metastases included 134 liver, 114 lymph node, 46 lung, 42 bone, and 32 abdominal mass/ascites/pleural fluid.

For each patient we performed paired exome sequencing on tumor and germline DNA in order to identify likely pathogenic variants and resolve their somatic or germline origin. Mean target coverages for tumor and normal exomes were 180X and 120X. The average tumor content was 62%. Sequencing metrics are summarized in Supplementary Table 2. Within the targeted regions, we identified an average of 119 somatic mutations per patient. For the majority of cancer types the number of mutations significantly increased in metastases relative to primary tumors in TCGA (Fig. 1c). The difference was more pronounced for tumor types with low mutation burden in the primary stage, e.g. prostate or adrenal cancer. To identify the most recurrent, and hence likely pathogenic, targets of genetic alterations, we performed an integrative analysis of single nucleotide variants (SNV), copy-number variants (CNV), and gene fusions. For each patient and gene, we classified the most recurrently mutated genes as putative tumor-suppressors or oncogenes based on the increase in frequency of inactivating mutations and expert knowledge (Fig. 1d and e, Supplementary Table 3). We found a long-tailed mutational spectrum for both tumor-suppressors and oncogenes. TP53 (266, 53.2%), CDKN2A (80, 16%), PTEN (79, 15.8%) and RB1 (68, 13.6%), were the most frequently altered tumor suppressors, while the most frequently mutated oncogenes included PIK3CA (67, 13.4%), AR (63, 12.6%) and KRAS (51, 10.2%). Overall, tumor suppressors were altered across many cancer types (e.g., TP53 and RB1), while oncogenes were more strongly associated with individual cancer types (e.g., AR or GNAS) (Extended Data Fig. 1c). We further compared the alteration frequencies to those from primary tumors and found that the increase in mutation burden (Fig. 1c) was mirrored by an increase in the frequency of genetic aberrations for the most widely mutated genes (Extended Data Fig. 1d).

Germline variants in metastatic cancer

Through sequencing matched germline DNA, we identified 63 presumed pathogenic germline mutations (PPGM), as defined by ClinVar expert curation, involving 18 genes in 61 individuals (12.2%) (Fig. 2a). These included 30 deleterious missense mutations, 8 nonsense mutations, 20 frameshift mutations, and 5 deleterious splice site mutations (Fig. 2b and Supplementary Table 4). Seventy-five percent of the PPGMs were in genes related to DNA repair with MUTYH (n=10, 16%), BRCA2 (n=9, 14%), CHEK2 (n=9, 14%), and BRCA1 (n=5, 8%) the most common. Outside of DNA repair pathways, we observed PPGMs in APC (n=6, 9.5%), MITF (n=5, 8%), and HOXB13 (n=3, 5%), among other genes. Of the 63 instances of pathogenic alleles identified, 5 alleles were previously unreported, while the remaining alleles have existing ClinVar entries with assigned pathogenic or likely pathogenic significance. Of the 61 individuals with PPGMs identified in this study, 30 (49%) had a somatic second allele aberration within the tumor genome, including LOH and exhibited molecular phenotypes consistent with pathogenicity (Supplementary Table 4). The remaining cases were of carrier status for the identified allele.

Next, we compared the frequencies in genes with PPGMs discovered in the MET500 cohort to the population frequencies in 52,790 individuals compiled by the Exome Aggregation Consortium (ExAC) (http://exac.broadinstitute.org/) excluding TCGA cancer samples¹⁷. The odds of any PPGM in metastatic cancer significantly exceeded the odds found in the populations comprising ExAC (OR = 3.00, 2.28–3.9, P=1 × 10⁻¹³). The genes analyzed and found to be enriched in the metastatic series included BRCA1, BRCA2, APC, CHEK2, MITF, MLH1, NBN and RB1 (Supplementary Table 5).

The gene fusion landscape of metastatic cancer

To identify both activating and inactivating gene fusions we analyzed 868 transcriptome libraries from 496 metastatic tumor RNAs. These fusion junctions involved 12,027 unique gene pairs, an average of 34 gene fusions per tumor, derived from a range of structural aberrations (Fig. 3a). Large differences in fusion-burden were observed across tumor type (Extended Data Fig. 2a). 199 cases (39.8%) harbored at least one putative pathogenic fusion with 138 activating fusions and 103 deleterious fusions (Supplementary Table 6). The activating fusions could be classified as DNA-binding (n=88), protein kinases (n=29), and signal transducers (n=21) (Fig. 3b). The loss-of-function fusions segregated into canonical tumor suppressor genes (n=59), chromatin modifying genes (n=35), and genes involved in cell adhesion (n=9). The most commonly fused tumor suppressor genes were NF1 (n=18), TP53 (n=11), PTEN (n=11), and RB1 (n=6) (Extended Data Fig. 2b). Interestingly, we identified a series of 8 novel fusion pairs in metastatic cancers that we believe are pathogenic (Fig. 3c). These include activated FGFR, BRAF, and ALK fusions with novel partners, extending the range of both fusion partners and cancer types for these clinically targetable fusions.^18–22 Novel gene fusions with functional domains include GREB1-NR4A3 in uterine leiomyosarcoma, POC5-PRKD1 in polymorphous low-grade adenocarcinoma of the tongue, and CIC-CITED1 in undifferentiated high-grade sarcoma. Notch fusions fall into 2 classes, those predicted to be sensitive to gamma-secretase inhibition (e.g., NOTCH2-SPAG17) and fusions which are independent of gamma-secretase processing (e.g., PARS2-NOTCH2).

Transcriptional signatures of metastatic disease

To investigate the potential clinical utility of metastatic expression profiles, we analyzed transcriptomes of the 496 biopsy samples (868 libraries). We first evaluated to what extent tissue and cancer specific gene expression is maintained across metastatic lesions. We used the t-SNE projection²³ to qualitatively visualize the expression of primary cancer marker across the MET500. Compared to primary tumors, metastatic samples were less well separated, more heterogeneous, and did not segregate based on biopsy site, with the exception of liver biopsies (Fig. 4a, Extended Data Fig. 2c–d),. We compared the expression of tissue-specific marker genes derived from 36 normal tissues²⁴ between normal, primary, and metastatic samples, and observed significant dedifferentiation with disease progression (Extended Data Fig. 2e).

Next, we looked at transcriptional signatures associated with perturbed cancer-related genes^25,26. Compared to normal tissues, transcriptional output was increased for most oncogenic signatures (Extended Data Fig. 3a–b), indicating a global shift towards a cancer-related transcriptional program. Unsupervised clustering of signature scores across patients revealed relevant associations between gene sets, and phenotypic similarities among patients (Extended Data Fig. 4). Inference of patient-specific activities^27,28 revealed coordinated changes across curated pathways that coalesce into a small number of principal cancer hallmarks (Extended Data Fig. 5a–b): interferon response, inflammatory response, epithelial to mesenchymal transition (EMT), proliferation, and metabolism. Importantly, these associations were robust to algorithm choice. Compared to normal tissues, metastatic tumors show a global increase in proliferation, stress-response, and metabolism. Conversely, hallmarks of EMT and cancer-immune responses can be either up- or down-regulated (Extended Data Fig. 5a–b). Next, we computationally delineated 25 non-redundant experimental “meta-signatures” (Extended Data Fig. 6). Unsupervised clustering and correlation analysis of meta-signatures revealed four of the canonical cancer hallmarks: immune response, EMT, proliferation, and metabolism (Fig. 4b). Metastatic tumors fall into two main subtypes: an EMT-like subtype associated with inflammation signatures²⁹, and a proliferative subtype associated with increased metabolism and systemic stress. In agreement, we observed mutual exclusivity between curated proliferative and EMT gene sets (Fig. 4c). Interestingly, this trend was less prominent across primary tumors (Extended Data Fig. 7a). Importantly, meta-signature activities were found to be weakly associated with biopsy site (Extended Data Fig. 7b) and primary tissue (Extended Data Fig. 7c), and held independently for common cancer types and biopsy sites (Extended Data Fig. 7d).

The immune microenvironment of metastatic disease

To characterize the phenotype of host-immune responses, we leveraged exome, RNA-seq, and a dedicated assay for T-cell repertoire profiling. Based on immune-cell markers proposed by Yoshihara et al.³⁰, we developed an RNA-seq based score, MImmScore, to assess the magnitude of leukocyte infiltration. We found that MImmScores is negatively correlated with tumor content (Extended Data Fig. 8a), and positively correlated with stromal infiltration (Extended Data Fig. 8b). MImmScores were compared to canonical T-cell expression markers (RNA-seq based) and DNA-based T-cell receptor β CDR3 sequencing, and all three measures are in good agreement (Extended Data Fig. 8c–d). We also discovered that metastatic immune infiltration is strongly determined by tumor type (Fig. 5a) and to a lesser degree by biopsy site (Extended Data Fig. 9a). Cancer types known to be infiltrated in the localized stage (Extended Data Fig. 9b), including kidney cancer^31,32, lung cancer³³, and melanoma³⁴, remained infiltrated at metastatic sites. Less immunogenic types, such as breast and prostate cancer, were generally associated with lower MImmScores in the primary and metastatic stage (Fig. 5a, Extended Data Fig. 9b). Immune infiltration was found to be heterogeneous not only across cancer types, but also within individual cohorts (Extended Data Fig. 9c–d). Strikingly, individual patients with high levels of immune infiltration could be identified even within tumor types that did not thus far respond to immunotherapies.

We hypothesized that metastatic tumors differ not only in the magnitude but also composition, in terms of leukocyte cell types, of tumor infiltrating lymphocytes (TILs) and macrophages (TAMs). Unsupervised clustering revealed groups of samples with significant differences in TIL composition, based on bulk tumor transcriptome data³⁵ (Fig. 5b). Cancers were most strongly typified by the different ratios of M2 to M0 (unpolarized) macrophages (clusters TIL-2,4,6,7) and different CD8+ to CD4+ T-cell ratios (high CD8+ TIL-1, high CD4+ TIL-4). While immunosuppressive M2 macrophages³⁶ were highly prevalent, pro-inflammatory, anti-tumor M1 macrophages were largely absent. A small cluster of samples (TIL-5) was characterized by a dominant ratio of cytotoxic CD8+ T-cells. To assess clonal T-cell expansion, we selected index cases with a high MImmScore and CD8+ T-cell ratio or with low immune infiltration. We ascertained the identity and frequency of T-cell clones by T-cell receptor β (CDR3) deep-sequencing, and found that the estimated numbers of T-cells were dramatically increased in the index cases (Supplementary Table 7). Most importantly, this increase was correlated with a significant expansion of T-cell clones (increased clonality) (Fig. 5c, Extended Data Fig. 9e) and a concomitant decrease in the ratio of Tregs to cytotoxic T-cells (Extended Data Fig. 9f). Highly mutated samples were found to be associated with a larger number of infiltrating T-cells (Fig. 5d) and increased MImmScores (Extended Data Fig. 10a).

Next, we focused on the expression of ligand /receptor pairs on the surface of T-cells and APCs. These molecules are either co-stimulatory and required for T-cell activation, or co-inhibitory as in the case of immune checkpoints (Fig. 5e). The majority of patients were either immunologically silent (clusters Tcell-0, APC-0), or immunologically active (Tcell-1, APC-1), with a highly significant overlap between the independent cluster analyses (Extended Data Fig. 10b–c). Importantly, almost all samples in APC-1 express CD80/CD86 and almost all samples in Tcell-1 express CD28. CD80/CD86 are the ligands for the CD28 receptor and a critical signal for T-cell activation.

Finally, we examined the relationships between the emerging predictive biomarkers for immune therapy and the transcriptomic immune phenotypes. We stratified patients into three categories: immunologically silent, partially active, and fully active. A sample was categorized as completely or partially active if it was a member of all or at least one of the active clusters: TIL-5, APC-1, Tcell-1, respectively. Patients in the active categories exhibited increased levels of expression biomarkers: PD-L1³⁷ (Fig. 5f), HLA³⁸, and granzyme³⁹ and had higher mutational burden (Fig. 5g), which is both a prognostic and predictive marker⁴⁰. Finally, leveraging a predictive signature to immunotherapy in metastatic melanoma⁴¹, we developed a clinical-response score. As expected, immunologically active patients had significantly higher clinical-response scores (Fig. 5h).

Conclusion

Decreases in the cost of sequencing have led to the widespread adoption of integrative sequencing for the study of cancer and precision oncology. Accordingly, our real-time clinical sequencing program was established to explore the practical challenges of clinical translation and at the same time characterize the genomic landscape of advanced cancer. The resulting MET500 cohort represents the first assessment of the genetic and transcriptomic heterogeneity across a wide range of metastatic cancers.

The distribution of mutation frequencies across diverse lineages of metastatic cancers is extremely long-tailed, with relatively few genes mutated at a high rate. We found that 12.2% of our cases harbored potentially pathogenic germline variants, a majority of which (75%) were related to DNA repair pathways. Mutations in DNA repair pathways have therapeutic implications; hypermutated tumors may respond to immune checkpoint inhibitors⁴², while HR deficiency could suggest sensitivity to PARP inhibitors^42,43. The high prevalence of likely pathogenic germline variants suggest that metastatic patients should be considered for genetic counseling and associated germline testing.

By integrating whole exome sequencing with RNA sequencing we were able demonstrate that transcriptome profiling provides clinically important and complementary molecular information. We demonstrate how RNA sequencing can be employed in a clinical context to characterize gene fusions, outlier gene expression, transcriptional pathways and the immune microenvironment. Across the MET500 cohort, 37% cases harbored a putative driver fusion, or an inactivating fusion in a tumor suppressor gene. RNA-seq data played an important role in characterizing the transcriptional networks active in tumor cells as well as the metastatic tumor microenvironment and suggest that metastatic tumors are significantly dedifferentiated, but retain some tissue- and cancer-specific gene expression patterns. We were able to delineate two distinct types of metastases: proliferative and EMT-like. Interestingly, proliferative tumors were associated with increased metabolism and stress response, while EMT-like tumors were associated with inflammation-related signatures.

Particularly valuable in the context of immunotherapy are mechanism-driven biomarkers that delineate discrete immune checkpoints or mechanisms of immune evasion. However, immune biomarkers need to characterize a complex disease state comprising the tumor genotype (e.g., mutational burden), phenotype (e.g., PD-L1 expression), and host response (e.g., presence of CD8+ T-cells). Towards comprehensive immunogenomic profiling, in this study we leveraged DNA and RNA sequencing data, which enabled us to characterize not only the tumor genotype, but also the phenotype of the host-immune response. Our results demonstrate the feasibility of using RNA-seq data to delineate immunologically and potentially clinically distinct subtypes of metastatic tumors, highlighting the potential of clinical RNA sequencing for monitoring the tumor microenvironment and guiding immunotherapeutic approaches.

While this study compares the molecular attributes of a metastatic cancer cohort to those of primary cancer cohorts, it does not utilize matched samples of primary and metastatic biopsies from individual cases. The sequencing of matched samples could illuminate further the processes behind tumor evolution, resistance to therapy, and immune interactions.

In summary, the metastatic solid tumor cohort represented in this study is a powerful complement to studies that have been carried out on primary cancers. Metastatic cancer is a highly heterogeneous disease at the genetic, transcriptomic, and microenvironment level. Significant progress in the treatment of advanced cancer will therefore depend on our ability to learn the therapeutic implications of metastatic heterogeneity and to develop screening methods and clinical trial designs that match patients to the most promising therapies.

METHODS

No statistical methods were used to predetermine sample size. The experiments were not randomized.

Extended Data

Supplementary Material