Cancer of unknown primary
BMJ 2020; 371 doi: https://doi.org/10.1136/bmj.m4050 (Published 07 December 2020) Cite this as: BMJ 2020;371:m4050
- 1Department of Gastrointestinal Medical Oncology, University of Texas MD Anderson Cancer Center, Houston, TX, USA
- 2Division of Hematology/Oncology, Department of Medicine, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA
- Correspondence to: M S Lee mslee2{at}mdanderson.org
ABSTRACT
Cancers of unknown primary (CUPs) are histologically confirmed, metastatic malignancies with a primary tumor site that is unidentifiable on the basis of standard evaluation and imaging studies. CUP comprises 2-5% of all diagnosed cancers worldwide and is characterized by early and aggressive metastasis. Current standard evaluation of CUP requires histopathologic evaluation and identification of favorable risk subtypes that can be more definitively treated or have superior outcomes. Current standard treatment of the unfavorable risk subtype requires assessment of prognosis and consideration of empiric chemotherapy. The use of molecular tissue of origin tests to identify the likely primary tumor site has been extensively studied, and here we review the rationale and the evidence for and against the use of such tests in the assessment of CUPs. The expanding use of next generation sequencing in advanced cancers offers the potential to identify a subgroup of patients who have actionable genomic aberrations and may allow for further personalization of therapy.
Introduction
Cancers of unknown primary (CUPs) are defined as histologically confirmed, metastatic malignancies with a primary tumor site that cannot be identified on standard baseline evaluation.1 These cancers are typically characterized by aggressive and early metastasis and unpredictable patterns of spread.2 CUP is not a single disease; rather, it is a heterogeneous grouping comprising diverse primary tumor types that elude identification with standard evaluation. A systematic review of 884 patients with CUP reported in 12 autopsy cohort studies conducted between 1944 and 2000 reported that a primary tumor site could be identified by autopsy in 644 (73%) patients, with the most common sites being lung (27%), pancreas (24%), hepatobiliary (8%), kidneys (8%), bowel (7%), genitourinary (7%), and stomach (6%).3 The inability to identify such primary sites before death presents a diagnostic and therapeutic dilemma for patients and their physicians. Tissue of origin tests that identify primary tumor type on the basis of transcriptomic analysis with good sensitivity and specificity have been studied and commercialized. However, despite the divergence of chemotherapy and molecularly targeted therapy in most solid tumors, whether identification of the primary tumor site and receipt of site directed treatment improves patients’ outcomes in a meaningful way remains, surprisingly, unclear. Here, we will review the management of CUP and the data on the utility and accuracy of tissue of origin tests and additional next generation sequencing assays in the evaluation of CUP.
Incidence and prevalence
CUPs are estimated to comprise 2-5% of all diagnosed cancers worldwide,456 with an incidence range of 6-16 per 100 000 person years.789 Studies of national cancer registries show a peak in incidence of CUP about 1980 in the US and in the 1990s in Europe and Australia, with subsequent declines in age adjusted incidence rates,4101112 which may reflect both better diagnostic imaging and evolution of standard pathologic assessments such as immunohistochemistry, as well as true declines in incidence rates. However, registry based studies of incidence and prevalence are hampered by inadequate specificity in recording diagnoses, and more detailed auditing may result in reclassification of up to 32% of cases of CUP to an alternative primary tumor.8 In the US, a higher proportion of CUP is found in patients who are older, female, or black or reside in less affluent or less educated counties.10 These data, however, cannot fully distinguish whether this reflects a distinct biology of CUP in these communities or a disparity in the extent of the diagnostic investigation.
Sources and selection criteria
We searched PubMed and Medline for articles written in English in 2005-19, using search terms “occult primary”, “cancer of unknown primary”, and “cancer of unknown origin”. We selected articles for inclusion on the basis of study quality and design, prioritizing randomized controlled trials and prospective and retrospective studies of sufficient size. Given the ongoing pace of clinical research, including recent reports of important randomized clinical trials, we have also included additional sources relevant to this review such as meeting abstracts describing final results of relevant randomized clinical trials.
Presentation and initial evaluation of CUP
Because of the heterogeneity of CUP, patients can present with any number of signs and symptoms that are related to the site of malignant involvement. In the absence of a radiographically detected, or clinically suspected, primary site, a biopsy should be taken from the site that can most safely undergo at least core needle biopsy to ensure that adequate tissue is available for histopathologic, and possibly molecular and genetic, analyses. Immunohistochemical stains are routinely used to attempt to identify the most likely primary site and are done in a tiered fashion to seek to identify the broad type of malignancy (carcinoma, melanoma, lymphoma, or sarcoma) and then identify the subtype and primary site.131415 In most cases of CUP, histopathology is sufficient to determine epithelial source but cannot further define a primary tumor site or histology. About half of cases of CUP have pathologic findings of metastatic adenocarcinoma, 30% have undifferentiated or poorly differentiated carcinoma, and 15% have squamous cell carcinoma. Approximately 5% of CUPs cannot be classified beyond undifferentiated neoplasm, a grouping that comprises a mixture of neuroendocrine carcinomas, lymphomas, germ cell tumors, melanomas, sarcomas, and embryonal malignancies.16
If the site of origin remains uncertain after routine histopathologic analyses, further evaluation should follow the consensus guidelines developed by leading cancer agencies. Figure 1 summarizes subsequent management. The National Comprehensive Cancer Network (NCCN), the European Society of Medical Oncology (ESMO), and the National Institute for Health and Care Excellence (NICE) recommend a basic evaluation in all patients, as described in box 1, with additional investigation considered given details of clinical presentation.171819 For example, upper endoscopy may be useful in selected patients with suspicious signs, symptoms, or laboratory abnormalities suggestive of an upper gastrointestinal primary.
Summary of recommended evaluation for cancer of unknown primary (CUP). ECOG PS=Eastern Cooperative Oncology Group performance status; LDH=lactate dehydrogenase
Recommended management of cancer of unknown primary171819
Initial evaluation for essentially all patients
Thorough medical history, particularly regarding past biopsies or malignancies, removed lesions, and spontaneously regressing lesions
Complete physical examination, including head and neck, skin, lymph node, rectal, genitourinary, pelvic, and breast examination
Laboratory studies including complete blood count, electrolytes, liver function tests, creatinine, calcium, lactate dehydrogenase, and urinalysis
Computed tomography scans of thorax, abdomen, and pelvis, with intravenous contrast unless contraindicated
Mammography for female patients*
Serum prostate specific antigen for men over aged ≥40, especially with adenocarcinoma bone metastases*
Fecal occult blood test
Biopsy (optimally core needle biopsy) of most accessible site
Additional recommended investigations in specific clinical scenarios
Isolated cervical lymph nodes
Head and neck computed tomography. Consider PET-CT (to identify primary head and neck site)
Supraclavicular/axillary lymph nodes in women, or other histopathologic evidence for breast cancer
Breast magnetic resonance imaging and/or ultrasonography, if mammogram unremarkable
Mediastinum
β-hCG and α-fetoprotein (to evaluate germ cell tumor)
Retroperitoneal mass
β-hCG, α-fetoprotein, testicular ultrasonography if male <65 years
hCG=human chorionic gonadotropin; PET-CT=positron emission tomography-computed tomography
*National Institute for Health and Care Excellence guidelines specifically do not recommend these tests unless compatible clinical and/or pathologic features are present
Positron emission tomography (PET) scans are not routinely recommended in evaluation of CUPs in the NCCN, ESMO, or NICE guidelines. Select scenarios exist, however, in which PET should be considered. Specifically, the ESMO and NICE guidelines highlight the potential utility of PET scans in patients with isolated squamous cell carcinoma in cervical lymph nodes. A meta-analysis of 246 patients with biopsy confirmed carcinoma of cervical nodes with occult primary included in seven studies of 18F-fluorodeoxyglucose PET-computed tomography (PET-CT) for detection of primary tumor site showed a primary tumor detection rate of 44% (95% confidence interval 31% to 58%), with excellent sensitivity of 97% (63% to 99%) but specificity of only 68% (49% to 83%).20 In other CUPs, PET-CTs may also be informative of primary tumor site.
However, the overall rate of primary detection with PET scans has been shown be quite low for patients with CUP in other anatomic sites. A meta-analysis of PET-CT in 433 CUP patients from 11 studies found a primary tumor detection rate of 37%, with 84% (78% to 88%) sensitivity and 84% (78% to 89%) specificity.21 A more recent meta-analysis including 1942 patients in 20 studies found a primary tumor detection rate of 40.9% (39.0% to 42.9%).22 However, prospective clinical trials comparing PET-CT with conventional imaging modalities such as conventional computed tomography scans are lacking. In a prospective study of 136 patients newly diagnosed as having CUP with extra-cervical metastases, patients underwent PET-CT along with administration of intravenous and oral contrast to facilitate interpretation of diagnostic computed tomography images. Imaging results were correlated with a standard of reference established by a multidisciplinary team to determine the most likely primary tumor site. PET-CT determined the primary tumor site in 38/135 (28%), whereas conventional computed tomography determined the primary tumor site in 43/135 (32%); no significant differences in sensitivity, specificity, or accuracy were seen.23 Given the lack of prospective trial data showing the clinical benefit of PET-CTs compared with conventional computed tomography imaging, guidelines do not recommend PET-CT scans as standard.
Notably, identifying patients with germ cell tumors and lymphomas or other hematologic malignancies, which have potential for curative therapy, is of paramount importance. NCCN guidelines also prioritize identification of thyroid carcinomas, neuroendocrine tumors, and sarcomas, which have markedly different treatment options and prognoses.
Identification and management of favorable and unfavorable risk groups of CUP
Favorable risk groups
Distinct subsets exist within the heterogeneous grouping of CUP, and their recognition is important as their treatment differs considerably from that of undifferentiated CUP. These more favorable risk subsets encompass approximately 10-20% of cases of CUP and are distinguished on the basis of clinical and pathologic features that provide ample evidence to suggest the site of origin even if it is not actually detected,1824 as described in table 1.
Favorable risk subtypes of cancer of unknown primary (CUP)
Isolated localized lymph node metastases to the neck or to the axilla in a woman are important subsets of CUP, and patients treated using locoregional paradigms of head and neck cancer or breast cancer, respectively, have the opportunity for potentially curative therapy. A systematic review of 24 retrospective studies found that 321/446 (72%) women with isolated axillary nodal metastases, even if no breast primary is found on initial evaluation, are ultimately found to have an occult breast primary when treated definitively with surgical resection.25 Given this high rate of primary tumor detection, these patients should receive therapy with curative intent as indicated for stage II breast cancer.
Similarly, identifying diseases that would have a better prognosis if disease specific therapy is administered is important. For example, in a retrospective study of 42 patients with CUP who were deemed to have a colorectal primary, of whom 32 received first line or second line therapy with a colorectal cancer specific regimen, median survival among patients receiving a colorectal site specific regimen was 27 months. The retrospective nature of the study makes it subject to reporting bias, but these data may suggest favorable prognosis in this subgroup with identifiable colorectal cancer.26 Identifying a site of origin would allow for administration of optimal evidence based, disease specific treatments to patients, so it is reasonable to assume that many patients would have improved outcomes if a site specific regimen can be administered. However, this has not been definitively demonstrated for CUP more broadly.
Unfavorable risk groups
Following evaluation, most patients (80-90%) do not fit any of the favorable risk subgroups as outlined in table 1 and thus comprise an unfavorable subset of disease, with poor prognosis. A retrospective analysis of 49 patients with unfavorable subset CUP and liver metastases showed a median survival of 10 (95% confidence interval 7 to 13) months, and a systematic review of four additional published series in similar cohorts showed a range of median survival from 1.7 to 7.2 months.27 Patient prognosis must be incorporated into any decisions made on use of chemotherapy and choice of chemotherapy. A prognostic model, developed from 150 unselected patients with CUP and then validated with an external dataset of 116 patients, established poor performance status (Eastern Cooperative Oncology Group performance status (ECOG PS) 2-3) and elevated serum lactate dehydrogenase concentration as key independent prognostic variables. A poor risk group was defined as having poor performance or elevated lactate dehydrogenase, with 11% one year survival and 3.9 months median survival (compared with 45% one year survival and 11.7 months median survival in the good prognostic group).28 ESMO guidelines recommend consideration of two drug chemotherapy combinations for patients with ECOG PS 0-1 and normal lactate dehydrogenase, and weighing chemotherapy versus best supportive care for patients with ECOG PS 2 or greater, elevated lactate dehydrogenase, or both.18
For patients with unfavorable subset CUP with adequate performance status, the mainstay of treatment is empiric chemotherapy. Multiple chemotherapy regimens have been tested in phase II studies, but no single regimen has shown superiority. A meta-analysis of 543 patients with unfavorable subset CUP treated with a total of 16 regimens in 10 randomized controlled trials found that no single regimen had a significant benefit compared with any others, with wide confidence intervals. The chemotherapy regimens in these studies included platinum agents, taxanes, vinca alkaloids, fluoropyrimidines, and irinotecan.29 Thus, various suggested or preferred chemotherapy regimens for patients with unfavorable subset CUP are listed in NCCN and ESMO guidelines and are summarized in table 2.
Recommended treatment regimens for cancer of unknown primary (CUP)
Tissue of origin tests
The molecular underpinnings of diverse types of human cancers are increasingly understood, as genomic, transcriptomic, and epigenetic analyses of a range of common and rare tumor types are ongoing. These studies show that distinct tumor types have recognizable differences in gene expression and other molecular features, suggesting that the site of origin of CUP might be elucidated from a tumor sample by analyzing how a given tumor’s genetic and molecular profile compares with the common patterns seen in cancers with known sites of origin. For example, the Cancer Genome Atlas has characterized genomic and gene expression variations within multiple tumor types and is studying differences and similarities across cancer types in the Pan-Cancer Atlas.30 In an updated analysis, 11 286 tumor samples from 33 cancer types were assayed for aneuploidy, DNA methylation, mRNA, and microRNA profiles. Among 10 165 tumors analyzed by mRNA expression profiles, 25 groupings were identified, with tumor type being a primary driving factor for these groupings. Additionally, tumors with similar histology, such as squamous morphology inclusive of cervical, head and neck, and lung primaries, clustered together. Clusters were also identified on the basis of similar tissue or organ system of origin, including neuroendocrine and glioma, melanomas of skin and eye, clear cell and papillary renal carcinomas, hepatocellular and cholangiocarcinomas, a gastrointestinal group of colorectal and stomach adenocarcinomas, and a digestive system group of pancreatic and stomach adenocarcinomas.31 Thus, molecular features such as gene expression or microRNA profiles differ among tumor tissue types and may be used to determine the primary tumor tissue type. Several platforms have been developed and commercialized to apply a classifier to identify the tissue of origin in CUP samples. Although several studies to validate the sensitivity and specificity of these assays have been performed, as will be described in the following sections, no clear gold standard exists to determine the true tissue of origin, so these limitations should be considered in analyzing these results.
92 gene real time polymerase chain reaction based assay
Accuracy
A 92 gene set (CancerTYPE ID, Biotheranostics, San Diego, CA, USA) was initially developed using a training set of 578 tumors representing 39 tumor classes, including a range of epithelial and non-epithelial tumor types. This gene set included 87 genes differentially expressed in the distinct tumor types and five reference genes expressed at a relatively invariable level across disease types. The genes selected were enriched in DNA binding transcription factors and in cell surface receptor proteins. This panel resulted in 87% classification accuracy in a validation cohort.32 A subsequent expansion of the training set to 2206 specimens representing 30 tumor types and 54 histological subtypes was used to enhance the classifier. Comparing a sample of unknown origin against the gene expression profiles for the reference tumors of each subtype within the database provides a probability that the unknown sample is classified as a known tumor type. In an internal validation study, this classifier resulted in 85% sensitivity for tumor type and 87% sensitivity for histological subtype.33 In a separate test set of 187 formalin fixed, paraffin embedded (FFPE) tumor samples, the 92 gene assay had an 83% sensitivity for tumor type.33 A validation study performed in three independent laboratories used the assay on 790 FFPE tumor samples encompassing more than 50 subtypes and found overall sensitivity of 87% (84% to 89%) and specificity ranging from 98% to over 99% for tumor types. For tumor subtyping analysis, sensitivity was 82% (79% to 85%), with specific ranging from 98% to over 99%. No significant decrease in performance was seen on the basis of metastatic tumors, histological grade, or cases with limited tissue.34
Multiple retrospective studies of the 92 gene set classifier have been performed to determine the sensitivity of the assay and correlate results to clinical actionability. A retrospective, multicenter analysis was performed on 501 patients deemed to have CUP, of whom 38 were subsequently found to have a latent primary site that became manifest during routine care. Of these 38 patients with a “gold standard” for the primary site, 28 had sufficient tissue to perform the 92 gene real time polymerase chain reaction (RT-PCR) assay, and 20/28 had sufficient RNA quality and quantity for the 92 gene classifier to predict a site of origin. The assay result was correct in 15/20 (75%) of cases, was incorrect in 3/20 (15%), and was unclassifiable in 2/20 (10%). By comparison, clinical features and immunohistochemistry suggested a primary site in six cases, five of which were correct, and suggested two or more primary sites in 13 cases, eight of which included the correct primary site.35 Although the number of cases was very limited, this study provided early evidence that a more definitive primary tumor site could be identified using the 92 gene classifier compared with standard clinical assessment.
A subsequent study built on this experience by pooling these 20 patients with an additional 151 patients with CUP prospectively studied using the 92 gene classifier. Of the 171 patients, 144 successfully had a single tumor type determined using the 92 gene classifier. The most common tumor sites determined included colorectal in 26/171 (15%), non-small cell lung in 18/171 (11%), breast in 15/171 (9%), hepatocellular in 10/171 (6%), ovary in 9/171 (5%), and pancreas in 9/171 (5%). As an update to the previous test, 18/24 (75%) patients who eventually had a latent primary tumor identified had the tumor type correctly identified by the 92 gene classifier. Additionally, of 52 patients who had a likely primary tumor site identified by immunohistochemistry, the 92 gene panel was concordant in 40 (77%).36 Thus, these studies show that compared with a gold standard method of identifying a primary tumor, the 92 gene classifier is accurate in 75-77% of cases.
The accuracy of the 92 gene classifier was compared against standard immunohistochemical pathology assessment in a prospectively defined, blinded study of 131 high grade, predominantly metastatic tumors with known reference diagnoses, 122 of which were evaluable. FFPE slides were provided to blinded pathologists to perform immunohistochemistry and also used to determine cancer type by the 92 gene classifier. The 92 gene assay had an accuracy of 79% (96/122; 95% confidence interval 71% to 85%) for tumor type, compared with an accuracy of 69% (84/122; 60% to 76%) for immunohistochemistry. The P value for difference in sensitivity was 0.019.37
Clinical utility
The clinical utility of the 92 gene classifier has been evaluated in a few prospective studies. A multicenter, prospective phase II trial enrolled patients with CUP and sufficient tissue for testing with the 92 gene assay. Patients waited two to three weeks before starting therapy, pending results of the 92 gene assay, at which time they were assigned to receive protocol prescribed treatment based on the predicted tissue of origin. The regimens included were standard for the era of enrollment from October 2008 through December 2011, although therapy for many of these diseases has since evolved.38
The primary objective of the study was to assess the efficacy of classifier guided therapy for patients with CUP, as measured by the overall survival compared with historical controls. In total, 289 patients were enrolled, of whom 252 (87%) had the assay successfully performed. Of the patients who had the assay successfully performed, 247/252 (98%) had a tissue of origin predicted; the most common sites predicted were biliary tract in 52/252 (21%), urothelium in 31 (12%), colorectum in 28 (11%), non-small cell lung in 27 (11%), pancreas in 12 (5%), and breast in 12 (5%), with other sites having less than 5% each. Of the 252 patients, 223 (87%) received treatment in the study, of whom 194 (87%) or 67% of the enrolled population received assay directed therapy. The median overall survival for these 194 patients was 12.5 (95% confidence interval 9.1 to 15.4) months. The median survival was 13.4 months in patients with tumor types deemed to have higher rates of response to treatment (colorectal, breast, ovary, kidney, prostate, bladder, non-small cell lung, germ cell, poorly differentiated neuroendocrine, lymphoma, and small cell lung), compared with 7.6 months for those with tumor types with lower rates of response (biliary tract, pancreas, gastroesophageal, liver, sarcoma, cervix, carcinoid, endometrium, mesothelioma, melanoma, skin, thyroid, head and neck, and adrenal) (P=0.04).38 This was thought to reflect improved outcomes compared with median survival of less than 10 months in historical controls receiving empiric therapy for CUP. Of course, systemic therapy options for distinct disease types have evolved markedly since that era, in particular with the evolution of predictive biomarkers for targeted therapies and now immune checkpoint inhibitors, so outcomes in the current landscape of disease need to be studied.
A more recent prospective multicenter observational trial was performed to determine clinical impact, as measured by the composite primary outcome of changes in the patient’s treatment, narrowing of treatment options, or elimination of a treatment option based on the results of the 92 gene assay. The primary outcome was assessed through a survey completed by ordering medical oncologists (n=73) and attending pathologists (n=34). In the study, 444 patients were enrolled from February 2013 through October 2014, and 397 (89%) had sufficient RNA for analysis. Of these, 379 (95%) had tumor type and histological subtype determined by the 92 gene assay. Among 271 assay results from medical oncologists, the 92 gene assay either confirmed or narrowed the diagnosis in a significant proportion of patients but often provided a result that was not initially suspected. In 79/271 (29%) cases, a single site was clinically suspected, and this site was confirmed in 60% of cases, but a previously unsuspected site was indicated by the assay in 39% of cases. In 80/271 (30%) cases, two or more sites were suspected in the differential diagnosis, and the assay narrowed the diagnosis in 66% of cases and provided a previously unsuspected site in 27% of cases. In the remainder of the 112 cases that the oncologist reported were CUP without an a priori suspected primary site, the assay provided a tumor type prediction in 97% of cases. Only 203 of the patients included in this study went on to receive therapy, and among these patients the oncologists reported that the 92 gene assay changed the planned treatment regimen in 47% of cases,39 suggesting a considerable effect of testing on patient care.
Microarray based tissue of origin test
The Tissue of Origin Test (Cancer Genetics) uses gene expression profiling via microarrays to determine the similarity of a tumor sample’s gene expression profile to 15 known tumor types. This test uses a 1550 gene expression profile obtained from microarrays and reports a similarity score, ranging from 0 to 100, for each of the 15 potential tissue types. The assay was trained using 2039 human tumor samples from 15 tissues of origin.40 The analytic performance of this assay was studied at four independent laboratories that each performed the assay on 60 frozen tissue specimens from metastatic and primary tumors and found high reproducibility, with overall concordance of 89.4% (range 87.0-92.5%).41 A subsequent blinded validation study processed 547 frozen tumor samples at four laboratories, with 258 (47%) samples derived from metastatic tissue and the remainder from poorly differentiated or undifferentiated primary tumor tissue. The overall agreement with the reference diagnosis was 87.8% (480/547; 95% confidence interval 84.7% to 90.4%), with a sensitivity of 87.8% (84.7% to 90.4%) and a specificity of 99.4% (98.3% to 99.9%).40 A subsequent retrospective study of 21 fresh frozen CUP samples found that 16 (76%) samples resulted in a positive result in a single tissue, with an indeterminate result in five (24%). This study was limited by lack of a gold standard given that the CUP samples by definition did not have a known primary tumor type.42
Subsequently, validation studies for this assay were done using FFPE specimens. Microarray data for 462 metastatic, poorly differentiated, or undifferentiated FFPE tissues that all had a reference diagnosis were obtained, and an 88.5% (85.3% to 91.3%) concordance was found. The reproducibility was established in three independent laboratories, with 89.3% (133/149) concordance.43 A blinded, multicenter, prospective study sought to compare the accuracy of immunohistochemistry and pathologists’ interpretation of tumor tissues against the microarray assay. FFPE metastatic tumor samples from 160 cases with known reference were prepared and sent to blinded pathologists for standard immunohistochemical analysis and for microarray analysis. The microarray result was concordant with the reference diagnosis in 89.2% of samples, compared with 83.3% per immunohistochemical assessment (odds ratio 2.9, 1.2 to 6.7). In 51 poorly differentiated and undifferentiated tumors, the accuracy of the microarray assay was 94.1%, compared with immunohistochemical accuracy of 79.1% (P=0.016), although results were similar in well and moderately differentiated tumors (microarray accuracy 85.3% versus immunohistochemical accuracy 86.8%; P=0.52).44
Summary of investigational assays
Various additional assays have been studied to identify tissue of origin for CUPs,4546 and these are summarized in table 3. These additional assays are not currently commercially available, but they show the potential of tissue of origin testing. These include a microRNA based assay, initially developed using quantitative RT-PCR,474849 and subsequently refined using a custom designed microarray identifying 42 tumor types with overall assay sensitivity of 85% and specificity of more than 99%.50 The second generation microRNA assay also had 70% (59/84) concordance with initial clinical diagnosis at presentation and 92% (77/84) agreement with final clinical diagnosis in patients who clinically were deemed to have CUP.51 The microarray gene expression database that was used to develop the 92 gene classifier was also used by an independent group to develop a distinct classifier for adenocarcinoma of unknown primary, which identified 70/84 (83%) tumors of known origin.52 Finally, a 10 gene quantitative RT-PCR assay was developed that can identify six tumor types (lung, breast, colon, ovary, pancreas, and prostate)53 and successfully assigned a tissue of origin in 63/104 (61%) patients with CUP, although most samples were identified as lung, pancreas, or colon.54 However, categorizing only six tumor types would not be sufficient to successfully identify primary tumor type in the current era, especially as fairly commonly misidentified tumor types such as biliary and gastroesophageal were not included in the panel.
Summary of tissue of origin tests
Prospective randomized clinical trials using tissue of origin testing
The first prospective randomized study reported was performed at multiple sites in Japan and randomized patients with CUP 1:1 to receive empiric carboplatin-paclitaxel or site specific therapy on the basis of the results of tissue of origin testing.55 However, the molecular classifier used was not one of the commercialized, previously validated classifiers that is available for current use. The molecular assay required a fresh frozen tumor specimen, from which RNA was extracted, and an Affymetric microarray was run. Training data were developed using 1024 tumors of known origin obtained either from publicly available Gene Expression Omnibus datasets or from the Department of Genome Biology at Kindai University, and a tenth of the data was randomly set aside for use as a validation cohort to determine the proportion of true classifications. This procedure was repeated 10 times, resulting in an average 78.6% cross validated estimate of accuracy. However, this microarray classifier has not been validated on additional datasets. In the trial, 130 patients were randomized, with gene expression profiling successfully performed for all of them. Twenty nine patients did not proceed to study treatment, 19 of whom had a primary site subsequently identified, leaving 101 patients for efficacy analysis. Of these, 26 (26%) had pancreas, 23 (23%) had gastric, 11 (11%) had lymphoma, eight (8%) had urothelial, seven (7%) had cervical, and six (6%) had ovarian sites predicted.
The one year survival rate was 44.0% in the site specific arm and 54.9% in the empiric carboplatin-paclitaxel arm (P=0.264). At a median follow-up time of 9.7 (range 1.2-83.9) months for the site specific arm and 12.5 (0.9-66.5) months for the empiric carboplatin-paclitaxel arm, the median overall survival was 9.8 (95% confidence interval 5.7 to 13.8) months for the site specific arm and 12.5 (8.9 to 16.1) months for the empiric chemotherapy arm (stratified log rank P=0.896), with a stratified Cox hazard ratio of 1.03 (0.68 to 1.56).55 Several limitations of the study must be considered, however, before concluding that site specific therapy does not improve outcomes. Firstly, this study did not enroll a sufficient number of patients in whom efficacy could be evaluated to satisfy its initial statistical analysis, which called for a total sample size of 114 patients, possibly resulting in an underpowered study. More importantly, the molecular assay that was the integral biomarker for this study was not robustly validated with independent samples, so whether this microarray based assay has the requisite sensitivity and specificity for clinical use remains unclear, and we can thus not be sure whether the data from this study can be generalized to other molecular classifiers. Finally, the patient mix was quite different from the previous retrospective studies, and the site specific therapies may not be optimal for the most common disease types. Given that the prognosis and treatment paradigms differ markedly with lymphoma, any patients suspected to have lymphoma should optimally be treated under a different paradigm and generally would not be considered for empiric carboplatin-paclitaxel.
Although the Japanese study has several key flaws, the lack of a significant survival improvement from site of origin testing was also seen with the presentation of the results of the GEFCAPI-04 randomized trial at the ESMO congress in 2019. This trial also showed no significant survival benefit of using molecular classifiers to provide site specific therapy compared with empiric chemotherapy. In this study, patients with CUP were randomized 1:1 to receive either empiric gemcitabine-cisplatin or to have gene expression testing followed by site specific treatment. The classifiers used were extensively studied in previous publications, including the 92 gene CancerTYPE ID classifier (n=222) and the Tissue Of Origin Pathwork test (n=21). The primary endpoint was progression-free survival. Patients were enrolled from March 2012 through February 2018, and 243 patients were randomized. The most commonly reported primary tumor types from molecular classifiers included pancreaticobiliary cancer in 19%, squamous cell carcinoma in 11%, kidney cancer in 8%, and lung cancer in 8%. Tailored site specific treatment was given in 91/123 patients randomized to site specific treatment. No significant difference was seen in median progression-free survival, with a hazard ratio of 0.95 (0.72 to 1.25) on central radiologic review and 0.80 (0.60 to 1.06) on local radiologic review. The secondary endpoint of overall survival was also similar in the overall population, with a hazard ratio of 0.92 (0.69 to 1.23) and a median of 10 months versus 10.7 months.56
The results of the Japanese trial and GEFCAPI-04 indicate that site specific therapy does not improve outcomes such as median progression-free or overall survival compared with empiric chemotherapy. However, treatment paradigms are markedly different and non-chemotherapy centered in some cancer types, such as renal cell carcinoma. In these subsets of patients, it is plausible that treatment with empiric chemotherapy could result in inferior outcomes, so current clinical practice and future trials should focus on identifying patients in whom identifying a primary disease site would markedly affect treatment paradigms.
Next generation sequencing
Carcinomas of unknown primary commonly have potentially targetable genomic alterations and mutations. Identifying genomic aberrations for which aberration specific therapies are available results in clinically meaningful changes in outcomes for patients with a variety of solid tumors, including non-small cell lung cancer, colorectal cancer, and melanoma. Available data suggest that carcinomas of unknown primary similarly harbor clinically meaningful rates of actionable mutations.
A retrospective study of 200 carcinomas of unknown primary that had next generation sequencing (NGS) performed by Foundation Medicine found that 192 (96%) had at least one alteration identified, with 169 (85%) having at least one clinically relevant alteration, and found 26 alterations for which targeted therapies approved in a known tumor type are available.57 These included six (3%) with EGFR substitution, six (3%) with ERBB2 amplification, 11 (6%) with BRAF substitution, and two (1%) with ALK substitution.57 Importantly, cases were also described of patients treated with targeted therapies, such as crizotinib for MET amplification or ALK fusion, in which patients had significant responses,57 suggesting that identifying actionable alterations is clinically meaningful.
A second retrospective study of 150 patients who had NGS performed at Memorial Sloan Kettering Cancer Center found that 137 (91%) had at least one alteration detected and 45 (30%) had potentially targetable alterations. These included six (4%) with BRAF V600E mutation, seven (5%) with ERBB2 amplification, and four (3%) with FGFR2/3 fusion. Of these patients, 15 (10%) received targeted therapies, with time to treatment failure ranging from less than one month to 14 months; the longest durations of treatment (of five to more than 14 months) were seen in patients with genomic rearrangements, including ALK fusion, RET fusion, FGFR2 fusion, and NTRK1 fusion.58
Use of circulating tumor DNA (ctDNA) to facilitate assessment of genomic alterations via a liquid biopsy is an emerging platform. Analyses of ctDNA in patients with CUP have shown rates of actionable mutations comparable to the two previous series using tumor tissue. In a study from the University of California San Diego, 442 patients with CUP had ctDNA testing using the Guardant assay between 2014 and 2016, during which time 54-70 genes were tested using NGS. Of the samples from 442 patients, 353 (80%) had ctDNA alterations detected and 290 (66%) had at least one characterized, likely pathogenic alteration. Targetable alterations included nine (2%) patients with BRAF V600E, four (1%) with EGFR L858R, and three (1%) with RET fusion. A total of 282/442 (64%) had an alteration theoretically actionable by an agent approved by the US Food and Drug Administration.59 This may thus serve as a potential novel platform for testing.
When interpreting these retrospective reports of genomic tumor testing, it is important to note that the definition of a “potentially actionable” alteration is a matter of debate. For example, although the ctDNA study reported that 64% of patients were found to have a potentially actionable mutation, many of those are truly only theoretically actionable, such as the 164 patients with a TP53 mutation. These studies also do not clearly describe the number of patients who have therapy changed as a consequence of results of testing, and nor do they robustly show clinical benefit of mutation specific testing. Finally, even if a “potentially actionable” alteration is detected, attaining access to targeted therapies may be challenging as no targeted agents have regulatory approval in CUP. Given this, enrollment in clinical trials is recommended, as this will allow patients access to targeted therapies while generating data to determine the efficacy and safety of these therapies.
However, no difference exists in the rationale for performing large, multigene NGS in CUP compared with other advanced solid tumors. Recent regulatory approvals of biomarker selected therapies agnostic of primary tumor site, including treatments for cancers with NTRK fusions or microsatellite instability, show that sequencing results may affect standard therapy choices, even in carcinoma of unknown primary. However, additional prospective clinical trials would ideally be conducted to show a clinically meaningful benefit in other biomarker defined CUP populations, although such studies are unlikely to be conducted. The benefit is likely limited to the small proportion of patients who are found to have a clearly validated alteration that serves as a predictive biomarker for a targeted therapy.
Potential utility of NGS results to identify tissue of origin
Many genomic aberrations or mutation patterns that are detected by NGS are strongly associated with certain primary tumor sites and thus may prove useful to identify a primary tumor site. For example, APC loss of function mutations are highly associated with colorectal cancers, and mutational signatures consistent with ultraviolet light exposure are highly associated with cutaneous melanomas. These patterns may be discernible using machine learning to result in enhanced ability to determine a primary tumor type. Recently, NGS data from Memorial Sloan Kettering were used to develop a genomics based classifier, using 7791 samples in 22 cancer types as a training set, and the model then underwent cross validation on training data along with an independent test set. On cross validation, 5748/7791 (73.8%) tumor types were successfully identified, and the classifier successfully identified tumor type in 8623/11 644 (74.1%) cases in an independent test set. The probability of each prediction varied in different tumor types—it exceeded 95% in 43.5% of cases, whereas most of the inaccurately classified samples had a probability below 50%. This classifier successfully identified a likely tissue of origin with a probability greater than 50% in 95/141 (67%) patients who clinically had CUP and occasionally resulted in changes in clinical management of patients.60 With increasingly widespread use of NGS in metastatic carcinomas, these classifiers may provide additional opportunities to identify tumor type in CUP. Caris has developed an analysis based on a machine learning algorithm trained and validated on more than 40 000 tumor samples with NGS data to identify primary tumor site, with tumor lineage classifier accuracy ranging from 82% to 96%.61 This assay, dubbed the MI Genomic Profiling Similarity (GPS) score, is now commercially available. Overall, these early results seem to be comparable to the accuracy of previous tissue of origin tests, although whether these result in clinically meaningful improvements in outcome remains unknown.
Potential immunologic biomarkers
An appreciable proportion of CUPs have biomarkers suggestive of likely response to immune checkpoint inhibitors, including microsatellite instability, tumor mutation burden, and high programmed death-ligand 1 (PD-L1) expression. A retrospective analysis of 389 cases of CUP used NGS to elucidate total mutational load (TML) and microsatellite instability and immunohistochemistry to determine PD-L1 expression. High microsatellite instability was found in 7/389 (1.8%) cases and was confirmed by immunohistochemistry, showing loss of a mismatch repair protein in six cases. High TML (≥17 mutations/Mb) was found in 46/389 (11.8%) samples, and PD-L1 expression (using SP142 antibody) on at least 5% of tumor cells was found in 82/365 (22.5%) samples.62 These findings present the hypothesis that a subpopulation of CUPs may respond to immune checkpoint inhibitors, most notably the group with high microsatellite instability tumors, who have the potential for a profound difference in their prognosis with the use of checkpoint inhibitors. Further assessment of the optimal biomarker to use for selection of patients for checkpoint immunotherapy is important for all diseases, including CUP.
Prospective clinical trials of mutation testing in CUP
Given the clearly demonstrated benefit of treating patients whose cancer is found to harbor one of several genomic aberrations that can be detected in a range of malignancies, clinical trials are now prospectively assessing genomic aberrations and assigning patients to a therapy option based on the underlying aberration. High profile trials such as NCI-MATCH and TAPUR enroll patients with a variety of cancers including CUP. The CUPISCO study (clinicaltrials.gov identifier NCT03498521) is a randomized phase II trial focused on CUP that enrolls previously untreated patients with adenocarcinoma or poorly differentiated carcinoma of unknown primary. CUPISCO compares molecularly directed therapy based on detected genomic aberrations with platinum based chemotherapy. In this study, patients receive three cycles of induction chemotherapy with the investigator’s choice of carboplatin-paclitaxel, carboplatin-gemcitabine, or cisplatin-gemcitabine. Archival tissue is tested using hybrid capture based comprehensive genomic profiling, and microsatellite instability, TML, and genomic loss of heterozygosity are calculated. Additionally, tumor expression of PD-L1 is assessed by immunohistochemistry using DAKO 22C3 antibody. Patients with complete or partial response or stable disease following induction chemotherapy are then randomized 3:1 to receive molecularly directed therapy based on the counsel of a molecular tumor board and the final decision of the treating investigator, or otherwise to continue with three additional cycles of chemotherapy. Patients with progressive disease will receive molecularly directed therapy as per the molecular tumor board counsel. The primary endpoint of this study is investigator assessed progression-free survival, with accrual planned for 790 patients.63
The study is still open and enrolling, and interim feasibility results were presented at the 2019 ESMO congress. Of 303 patients whose data were examined retrospectively, 96 (32%) would have been matched to a molecularly directed therapy arm. Notably, key genomic alterations included HER2 (7%), PIK3CA (6%), NF1 (6%), NF2 (5%), BRAF (4%), PTEN (4%), FGFR2 (4%), EGFR (4%), and MET (4%). Gene fusions were also observed, including ALK (1%), RET (1%), and ROS1 (1%). High TML (≥20 mut/Mb) was found in 9%, and 1% had microsatellite instability. Additionally, 14% had high PD-L1 with a tumor proportion score of 50% or higher.64
Additional studies of novel molecularly targeted systemic therapies in CUP are ongoing. For example, the CUPem study is enrolling patients with CUP who have received at least one previous regimen of chemotherapy to receive the PD-1 antibody pembrolizumab. Correlative studies, including archival tissue at baseline to document PD-L1 expression, are planned, but there is no intrinsic biomarker selection for the study (NCT03752333). The results of the NivoCUP study were presented at the 2020 ASCO virtual meeting and showed an overall response rate of 10/45 (22%, 95% confidence interval 11% to 37%) among patients who had received previous treatment and 2/11 (18%, 2% to 52%) among previously untreated patients with CUP.65
Emerging assays
Most of the published studies using tissue of origin testing with commercially available tests use relatively older technologies to assess gene expression profiles. The proliferation of cancer genome data published has allowed for additional analyses of transcriptomics, genomics, and epigenomics to identify tissue of origin, which may offer more accurate results. A 154 gene expression signature correctly classified 97.1% of 9626 primary tumor samples and 92% of 1248 poorly differentiated or metastatic samples.66 An ensemble of neural networks was trained using data from the Cancer Genome Atlas and used to retrospectively determine primary tumor type from RNAseq data from 201 metastatic tumors of known tumor types; it had an overall mean accuracy of 86%.67 A microarray DNA methylation signature, dubbed EPICUP, was also trained to identify 38 tumor types and in a validation study showed 97.7% (96.1% to 99.2%) sensitivity and 99.6% (99.5% to 99.7%) specificity.68
Guidelines
Guidelines from NCCN, ESMO, and NICE for management of cancers of unknown or occult primary are highlighted throughout this review. The Spanish society of medical oncology (Sociedad Espanola de Oncologia Medica; SEOM) published clinical guidelines for cancer of unknown primary in 2017,69 which are generally concordant with NCCN and ESMO guidelines. NCCN guidelines specify that “tumor sequencing and gene signature profiling for tissue of origin is not recommended for standard management at this time,” noting that “there may be diagnostic benefit, though not necessarily clinical benefit.” However, this recommendation was category 3, indicating that major disagreement exists within the NCCN as to whether the intervention is appropriate.17 SEOM guidelines also state that “the impact on clinical benefit of targeted treatment based on molecular studies remains controversial and the level of evidence and degree of recommendation is low.” 69 ESMO guidelines similarly state that tissue of origin tests “may aid in the diagnosis of the putative primary tumour site in some patients… However, their impact on patient outcome via administration of primary site-specific therapy remains questionable and unproven in randomized trials.” 18 Finally, NICE guidelines also direct: “Do not use gene-expression-based profiling to identify primary tumours in patients with provisional CUP.” 19 All guidelines thus stress the need for randomized phase II and III studies to show the utility of tissue of origin testing. Future versions of the guidelines will likely incorporate the final results of GEFCAPI-04 and CUPISCO, and final publication will be eagerly awaited.
Conclusion
Modern oncologic treatment increasingly stresses personalization of therapy, as novel predictive biomarkers and targeted therapies are identified. Current management of unfavorable subtypes of CUP focuses on empiric chemotherapy. Studies to date have not shown disease specific chemotherapy to result in a significant improvement in outcomes, with the exception of the small number of clinical scenarios in which therapy with curative intent is feasible. Just as is the case with patients in whom the site of origin is known, future management of CUPs will emphasize identification of subgroups of patients who have diseases that are susceptible to targeted therapies and immunotherapies on the basis of a predictive biomarker.
Research questions
Do additional subsets of patients exist in whom identification of tissue of origin will result in a significant improvement in progression-free and overall survival, particularly given the modern use of immune checkpoint inhibitors as a standard of care in many tumor types?
Are new methods or assays available that can identify tissue of origin with improved sensitivity and specificity?
Will identifying actionable somatic or germline mutations in patients with cancer of unknown primary improve progression-free and overall survival in these subgroups?
Is the use of tissue of origin testing for cancers of unknown primary cost effective?
Footnotes
Series explanation: State of the Art Reviews are commissioned on the basis of their relevance to academics and specialists in the US and internationally. For this reason they are written predominantly by US authors
Contributors: Both authors contributed to the planning, conduct, and reporting of the work in the article, and both are responsible for the overall content as guarantors.
Competing interests: We have read and understood the BMJ policy on declaration of interests and declare the following interests: MSL has received consulting fees from AstraZeneca, Bayer, and Genentech/Roche and research grants from Genentech/Roche, Pfizer, Amgen, Bristol-Myers Squibb, EMD Serono, and Exelixis; HKS has received research grants from Bayer, Merck, and Precision Biologics.
Patient involvement: No patients were asked for input in the creation of this article.
Provenance and peer review: Commissioned; externally peer reviewed.