Assessing the utility of molecular diagnostic classification for cancers of unknown primary

Abstract Background Roughly 5% of metastatic cancers present with uncertain origin, for which molecular classification could influence subsequent management; however, prior studies of molecular diagnostic classifiers have reported mixed results with regard to clinical impact. In this retrospective study, we evaluated the utility of a novel molecular diagnostic classifier by assessing theoretical changes in treatment and additional testing recommendations from oncologists before and after the review of classifier predictions. Methods We retrospectively analyzed de‐identified records from 289 patients with a consensus diagnosis of cancer of uncertain/unknown primary (CUP). Two (or three, if adjudication was required) independent oncologists separately reviewed patient clinical information to determine the course of treatment before they reviewed results from the molecular diagnostic classifier and subsequently evaluated whether the predicted diagnosis would alter their treatment plan. Results Results from the molecular diagnostic classifier changed the consensus oncologist‐reported treatment recommendations for 235 out of 289 patients (81.3%). At the level of individual oncologist reviews (n = 414), 64.7% (n = 268) of treatment recommendations were based on CUP guidelines prior to review of results from the molecular diagnostic classifier. After seeing classifier results, 98.1% (n = 207) of the reviews, where treatment was specified (n = 211), were guided by the tissue of origin‐specific guidelines. Overall, 89.9% of the 414 total reviews either expressed strong agreement (n = 242) or agreement (n = 130) that the molecular diagnostic classifier result increased confidence in selecting the most appropriate treatment regimen. Conclusions A retrospective review of CUP cases demonstrates that a novel molecular diagnostic classifier could affect treatment in the majority of patients, supporting its clinical utility. Further studies are needed to prospectively evaluate whether the use of molecular diagnostic classifiers improves clinical outcomes in CUP patients.


| INTRODUCTION
Cancer of unknown primary (CUP) describes a heterogeneous group of metastatic tumors that are defined by the absence of a clinically known tissue of primary origin.Patients diagnosed with CUP represent 2%-5% of all cancers, 1,2 and treatment options available to CUP patients are limited. 3,4Compared to patients with known primary tumors, the prognosis for CUP patients is generally poor, with median overall survival rates reported as <12 months. 2,5,6n the absence of a known primary site, clinicians gather information from a range of sources-including imaging and endoscopy as well as histopathological, molecular, and serum biomarkers-to aid in treatment selection. 6,7[10] The Tempus Tumor Origin (TO) test uses RNA expression data to select from 68 possible histological subtypes and showed a 91% classification accuracy when applied to an independent validation set of samples with known primary. 11 limited number of studies have evaluated the clinical utility of tissue-of-origin diagnosis via molecular diagnostic classifier in the setting of CUP owing to methodological challenges with studying this population. 12Several randomized studies have found no significant difference in CUP patient outcomes according to whether they received tissue-specific or empiric therapies, 13,14 while another showed that patients receiving tissue-specific therapies have improved outcomes. 159][20] Despite limitations, results from these studies indicate potential clinical benefit of having a diagnosis assigned by a molecular diagnostic classifier.
Current NCCN treatment guidelines focus on the tissue of origin for recommendation (i.e., breast cancer), 21 suggesting that a molecular-guided diagnosis for CUP patients could alter therapeutic management.Toward this end, a prospective study from 2018 found that 47% of patients had an altered treatment plan following review of tissue-of-origin predictions from a 92-gene assay. 22However, this study was limited by the inclusion of a considerable number of patients with a suspected site of origin and few predicted tumor (sub-)types in the population; in addition, since 2018, treatment paradigms based on genomic profiling have changed dramatically.Here, we present a decision impact study where oncologists retrospectively reviewed redacted clinical records from 289 confirmed CUP patients.From these reviews, we assessed treatment and testing recommendations both prior to and after review of molecular diagnostic classifier results, as well as reported confidence that the classifier results aided in selecting the most appropriate treatment regimen.

| Study population
An initial set of 336 de-identified records from patients with a CUP diagnosis whose clinician ordered Tempus xT (CAP/CLIA validated DNA-seq of 648 genes) 23,24 and the Tempus Tumor Origin (TO) test (v.1.0for reports before 10/28/2021, v.1.1 for subsequent orders) were selected by simple random sampling (each case was randomly assigned an integer and a random number generator selected a subset of integers) from the full set of 519 patients meeting all eligibility criteria in the Tempus clinicalgenomic database.These patients came from 169 different clinical sites across the United States.At least 285 patients were required based on statistical power requirements of 80% power at an alpha of 0.05 to test the null hypothesis of treatment change exceeding 40% and a non-inferiority difference of 0.07.A total of 336 patients were included based on the expectation that approximately 10% of patients would be screened out of the study due to not having a confirmed CUP diagnosis.
The classification of CUP in the Tempus clinicalgenomic database is established by Tempus pathologists who review pathology reports and prognosis notes as part of routine sample accessioning and report signout (cancer diagnoses are used to guide therapy and clinical trial matching on the xT report).When a definitive diagnosis is not provided in the associated reports the patient is classified as CUP (i.e., Tempus pathology supporting its clinical utility.Further studies are needed to prospectively evaluate whether the use of molecular diagnostic classifiers improves clinical outcomes in CUP patients. does not assign speculative diagnoses when the clinical care team is considering multiple diagnoses).Each eligible patient had a TO result issued over a period in 2021, a predicted probability for one or more diagnoses generated by the TO test, an ECOG score ≤2, and at least one pathology report (inclusive of any available diagnostic IHC, microscopic findings, any relevant addendums, etc.) and one progress note (inclusive of any available relevant patient and familial history, lab results [especially blood tumor markers], radiology, etc.) from the time of testing on file.

| Review of clinical records and treatment assignment
Each patient record was randomly assigned to be initially reviewed by two independent board-certified oncologists from an NCI-designated Comprehensive Cancer Center, with sub-specialties including gastrointestinal cancers, genitourinary cancers, neuroendocrine cancers, and cancers of unknown primary.The initial set of two reviews for each patient were performed by seven different oncologists in total.We utilized a case assignment method that aimed to achieve equal overlap between each reviewer and their counterparts, where a given reviewer shared an equal number of cases with each of the other reviewers.In the event of disagreement about whether the patient was CUP or on the primary endpoint (a yes or no question asking whether the results from the molecular diagnostic classifier would change treatment recommendations), a third oncologist reviewer adjudicated the discordance.These reviews were randomly assigned to one of three oncologists, and the adjudication reviewer never adjudicated a set of reviews that they initially reviewed.
For the determination of CUP status and inclusion in the study, reviewing oncologists used clinical and pathologic data available in progress notes from when the molecular diagnostic classifier was ordered, pathology reports, and NGS results (including tumor mutational burden, microsatellite instability status, and DNA mutational status from any of the available 648 genes on the panel including fusions).The reviewing oncologists provided their responses in a standardized questionnaire template (see Data S1 for a full list of all questions that reviewers were asked).Specifically for confirmation of CUP status, reviewing oncologists were asked "Does the patient's cancer appear to be of uncertain/unknown primary?" and required to pick either: (1) Yes, (2) No, or (3) Not enough information.In the event that the available clinicopathologic information and NGS results met criteria for assigning a primary diagnosis, the reviewer designated the case as non-CUP, and the patient was excluded from the study.Only patients with a consensus answer of "yes"indicating they are definitionally CUP-were included in the study.
For assessment of the primary endpoint, the reviewing oncologists specified the patient's treatment regimen before and after receiving the molecular diagnostic classifier result in free text, while having the option to leave responses blank if they felt there was not enough information.For reviews where additional testing was not needed, they were then asked "Did the treatment regimen change?" and had to pick one of the following answer options: (1) Yes, the entire regimen was updated, (2) Yes, I added an agent, (3) Yes, I removed an agent, (4) No, (5) N/A-Not enough information to select a treatment, or (6)  Other (opens free text).If additional testing was needed, the reviewer answered "How would the results of additional organ-specific testing change treatment?"and selected: (A) Treatment would change regardless of the results of testing, (B).Treatment change is dependent on result of testing, (C)Treatment would not change, or (D) Other (opens free text).The answers that counted as "yes" toward the primary endpoint of whether treatment would change included selection of either 1, 2, 3, A, or B. If there was disagreement regarding whether treatment would have changed between the initial two reviewing oncologists, a third oncologist adjudicated the result to allow the primary endpoint to be assessed at the patient level.
In addition to providing treatment recommendations pre-and post-review of the molecular diagnostic classifier results, oncologists were also asked several follow-up questions regarding the need for additional diagnostic testing, the evidence that guided treatment recommendations, and confidence that the molecular diagnostic classifier results aided in selecting the most appropriate treatment regimen.Any differences in secondary endpoints or follow-up questions (including specific treatment recommendations at any stage of review) were not adjudicated.

| Data collection and statistical analyses
Standardized patient study assessment forms were filled out electronically by oncologists before and after reviewing the results of the molecular diagnostic classifier.Structured fields and free text results associated with medications and regimens were curated and aggregated for study analysis.Descriptive study analysis was performed according to SAP v1.0 using SAS version 9.4 (SAS Institute Inc.).
Patient characteristics were summarized as counts (%) for categorical variables and mean (standard deviation) and median (25th, 75th percentiles) for continuous variables.The primary endpoint was presented as a binomial proportion with 95% confidence intervals calculated using the Wilson score method without continuity correction.Results of the oncologist surveys were summarized on the review or patient level according to whether the question was adjudicated.

| Molecular diagnostic classifier description
The Tempus TO test is a CAP/CLIA validated molecular diagnostic classifier that uses whole-exome capture RNAseq data as input to provide diagnostic predictions from 68 possible subtypes.The model was trained on 43,726 tumor samples of known origin and has an overall accuracy of 91% based on an independent validation cohort consisting of 9210 tumor samples. 11For individual patient records, results are reported with one or more predicted diagnoses including the probability that the sample belongs to the given label.Samples where the highest-probability label is lower than the 30% (TO v1.0) or 35% (TO v1.1) threshold are reported with an "indeterminate" result.The differences between the v1.0 and v1.1 versions of the TO test are minor and not expected to impact the results of the study.

| Study overview and clinical characteristics
We collected de-identified records from patients with confirmed cancers of unknown primary-defined here as a patient with two or more possible sites of tumor origin.Treatment recommendations and confirmation of CUP status were made by oncologist reviewers based on review of all available clinical information (including molecular NGS reports) before viewing predictions from the molecular diagnostic classifier.In the event that the reviewer felt there was not enough information available to select an initial treatment, they were not required to select one.After reviewing the classifier results, oncologists again reported treatment recommendations.Each patient record was independently assessed by two oncologists, with any differences in the eligibility (whether the patient had a CUP) or primary endpoint (a yes/no response regarding whether results from the molecular diagnostic classifier produced a change in treatment recommendations) adjudicated by a third oncologist.Data S1 lists the full set of questions that oncologists were asked to fill out for each patient.
Following review of available clinical information, 47 patients with a cancer classified as "tumor of unknown origin" did not meet the study inclusion criteria.For 31 of these patients, both initial reviewers agreed that the cancer was not CUP, and for 16 patients, a third reviewer confirmed that the cancer was not CUP. Figure 1  The most common predicted subtypes reported by the molecular diagnostic classifier were: cholangiocarcinoma (n = 58, 20.1%), lung adenocarcinoma (n = 41, 14.2%), pancreatic adenocarcinoma (n = 22, 7.6%), lung squamous cell carcinoma (n = 19, 6.6%), colorectal adenocarcinoma (n = 18, 6.2%), and gastroesophageal adenocarcinoma (n = 18, 6.2%).Table 1  characteristics including biopsy site and demographics, as well as a summary of results from the molecular diagnostic classifier.

| Treatment recommendations before and after review of molecular diagnostic classifier results
Of the 289 patients analyzed in this study, the reviewing oncologists indicated that 81.3% (n = 235, 95% confidence interval = [76.4,85.4]) of patients would have some form of change in their treatment plan based on the results from the molecular diagnostic classifier-Either specifying a treatment where there was previously not enough information available to do so, addition/removal of a treatment agent, complete update of the treatment regimen, or based on additional biomarker testing.Figure 2 shows that for the 125 patients that did not require adjudication of the primary endpoint (i.e., there was consensus from the two reviewers as to whether the treatment would/would not change following review of classifier results as well as on the initial CUP diagnosis), reviewers agreed that treatment recommendations changed for 92.0%(n = 115) and did not change for 8.0% (n = 10) of patients.For the 164 patients that required adjudication from a third oncologist, the adjudicating reviewer indicated that treatment changed for 73.2% (n = 120) and did not change for 26.8% (n = 44) of patients.Considering only the subset of patients where the assessed review/s contained a selected treatment prior to review of the molecular diagnostic classifier (n = 222), reviewing oncologists indicated that 75.7% (n = 168) would have some form of a change in their treatment plan.Finally, to assess possible reviewer heterogeneity, we looked at data from individual reviewers (8 reviewers, each reviewing between 9 and 90 patient records) and found that the percentage of cases where reviewers stated that treatment recommendations would change following review of classifier results ranged from 73%-97%.Consensus patient-level results were used for the aforementioned primary endpoint because responses to this line of inquiry were adjudicated.By contrast, subsequent analyses with regard to questions that were not adjudicated were performed on the reviews from individual oncologists.In total, there were thus 125 patient records where the initial two reviewers agreed on both the CUP diagnosis as well as the primary endpoint (resulting in 250 reviews).There were 164 patient records confirmed CUP through adjudication where the first two reviewers disagreed on either CUP status (n = 111) or primary endpoint (n = 53), necessitating adjudication by a third reviewer who summarized and resolved conflicting information from the first two reviewers.For these adjudicated patients, the third review was regarded as the consensus review and was subsequently assessed in review-level analyses.Thus, the 289 patient records were associated with 414 reviews (250 + 164).Figure 3 represents the assessment flow and the number of reviews included in each subanalysis that follows.
Prior to review of the classifier results, oncologists characterized the type of treatment regimen (n = 339, 81.9%) or indicated there was not enough information available to select a treatment (n = 75, 18.1%; Table 2).Treatment recommendations from these individual reviews (n = 414) were most commonly based on CUP guidelines (n = 268, 64.7% of reviews) and a presumed primary site (n = 55, 13.3% of reviews).Following review of the classifier results, the oncologist could again specify a treatment or indicate that additional testing would be needed to specify a treatment.Of the 211 reviews where treatment was specified post-classifier review, 98.1% (n = 207) were guided by the tissue of origin-specific guidelines.
Of the 163 reviews where review of the molecular diagnostic classifier results led to an altered or new treatment plan without additional testing required (n = 135 reviews where a previous treatment was specified and n = 28 where no previous treatment was specified), the treatment recommendations most commonly added were gemcitabine (n = 69, 42.3%) and cisplatin (n = 62, 38.0%; Table 3).Other notable therapeutic changes following review of the classifier results were that checkpoint inhibitor therapy was added to 16.0% of reviews (n = 26) and a TKI was added to 8.0% of reviews (n = 13).
In 135 reviews, a treatment regimen was specified prior to molecular diagnostic classifier review (i.e., there was sufficient information to determine a treatment plan based on other available clinical information) but subsequently updated post-classifier review (with no additional testing needed).These reviews represent potentially the most impactful shifts in treatment recommendations in response to the results of the molecular diagnostic classifier, given that enough information was present to specify a treatment prior to review of the classifier results but that treatment regimen was subsequently changed without the need for additional testing.In these reviews, the treatments that were most commonly removed from recommended patient treatment plans following the review of results from the molecular diagnostic classifier were: fluorouracil (n = 46, 34.1%), oxaliplatin (n = 42, 31.1%),carboplatin (n = 41, 30.4%), and paclitaxel (n = 37, 27.4%; Table 3).When characterizing the types of changes made to therapy recommendations, 64.4% of changes (n = 87) were from one chemotherapy regimen to another, and 5.9% of changes (n = 8) were due to the removal of all chemotherapy agents and switch to immune checkpoint inhibitors only (CPIs).Further, 3.0% of patients (n = 4) were switched from systemic to local therapies.

| Subtype-specific biomarker and other additional testing recommendations
For the 49.0% of reviews (n = 203) where some form of additional testing was recommended, oncologists reported that treatment would change as a result of these tests in 92.1% (n = 187) of reviews (Figure 3).A majority of the additional testing recommendations were for assessment of biomarkers that are required for pre-treatment evaluation in the guidelines corresponding to the cancer type predicted by the molecular diagnostic classifier.The types of additional testing most often specified as being needed to make a treatment decision were PD-L1 (n = 118, 58.1%), additional imaging (n = 60, 29.6%), ER/PR/HER2 (n = 31, 15.3%), and HRD (n = 19, 9.4%), among others.The predicted subtypes with the greatest proportion of reviews requiring additional testing (with a minimum of 5 reviews) were gastroesophageal adenocarcinoma (20/24 reviews, 83.3%), gastroesophageal squamous cell carcinoma (5/6 reviews, 83.3%), head and neck squamous cell carcinoma (9/11 reviews, 81.8%), and breast carcinoma (12/15 reviews, 80.0%).
F I G U R E 3 Overview of individual reviews split according to responses to clinical management questions.The number of reviews analyzed for certain secondary endpoints-including characterizing therapy changes and the evidence guiding treatment regimen determination-is dependent on answers to these questions.Of the questions shown, only "Change in treatment?"was adjudicated in the event of disagreement by the initial reviewers.
We separately assessed the 64 reviews from patients where consensus-level recommendations indicated that there would be no treatment change following review of the classifier results.In this set, additional testing would have been ordered in 25.0% (n = 16) of reviews including PD-L1 (n = 7) and HRD (n = 6).
Additionally, we found that confirmatory germline testing was recommended for 25.4% (n = 105) of the full set of 414 reviews, of which 49.5% (n = 52) were reported as being directly based on the results from the molecular diagnostic classifier.

| Evaluating confidence in molecular diagnostic classification
While the Tempus TO test and other molecular diagnostic classifiers can provide predicted primary sites for individual patients, a critical component of these predictions is the degree to which clinicians find them diagnostically valuable or believable.To assess this, we asked reviewing oncologists their agreement with the following statement "The TO result increased my confidence in selecting the most appropriate treatment regimen" for each review.Of the 414 reviews, the response breakdown was: Strongly Agree (n = 242, 58.5%), Agree (n = 130, 31.4%),Neutral (n = 31, 7.5%), Disagree (n = 10, 2.4%), and Strongly Disagree (n = 1, 0.2%; Table 2).
For the 64 reviews from patients where consensus-level recommendations noted no treatment change following review of the classifier results, 85.9% of reviews from oncologists either strongly agreed (n = 36, 56.2%) or agreed (n = 19, 29.7%) that the molecular diagnostic classifier results increased their confidence in selecting the most appropriate treatment regimen.Thus, while the classifier results did not lead to a treatment change in these reviews, the results added confidence to the presumed diagnosis and treatment plan.

| DISCUSSION
This retrospective study was designed to assess the potential clinical utility of a molecular diagnostic classifier in informing clinical management decisions for patients who have cancer with an uncertain primary site.For 81.3% of the 289 patients in the study, reviewing oncologists stated that the predicted diagnosis from the molecular diagnostic classifier would result in a change in treatment recommendations.
The results of this study illustrate the potential clinical benefit of performing molecular diagnostic classification in patients where standard-of-care diagnostic workups have been unable to render a definitive diagnosis.One explanation for the large rate of change in treatment recommendations that was observed is that treatment regimens for metastatic cancer patients are increasingly based on guidelines (such as the NCCN), which are in turn based on tissue site, histology, and precision medicine biomarkers.Prior studies have assessed the clinical impact of molecular diagnostic classifiers and demonstrated variable rates of treatment impact for CUP patients. 22Our study was unique in its attempt to understand the overall impact on treatment management in addition to describing specific types of treatment changes and the need for further biomarker testing, which provides additional details on how molecular diagnostic classifiers can impact clinical management.
When evaluating recommended treatment changes from individual clinical reviews, we observed the addition and removal of essentially all classes of systemic therapies including chemotherapy, tyrosine kinase inhibitors, and CPIs, highlighting the broad spectrum of treatment modalities used in histologic and tissuespecific regimens.CPIs specifically represent an interesting treatment subset to highlight given the recency of the indications and specificity of guidelines, which are based on histology, site of diagnosis, and PD-L1 status. 25 We observed that 19% of reviews would have added CPIs to the patient's treatment regimen, and 6% of reviews would switch the patient from chemotherapy to a CPI alone as their first systemic treatment recommendation.The importance of CPIs was also captured in reviewlevel assessment with the need for additional biomarker testing: PD-L1 testing was recommended in the majority (n = 118, 58%) of the 203 reviews that specified a requirement for additional testing.
While the molecular diagnostic classifier in the study uses transcript expression to predict tumor subtypes, pairing this test with an NGS sequencing panel allowed us to investigate the prevalence of mutations in specific clinically relevant genes and their association with predicted subtypes.Figure 4  289 patient cohort.Notably, the identified genes highlight additional clinical scenarios where having an accurate diagnosis is important for clinical management decisions.For instance, PARP inhibitors have approval for use in patients with BRCA1/2 alterations in ovarian, prostate, breast, and pancreatic cancer; however, there is no pancancer BRCA1/2 approval for these agents. 26Additionally, alterations to clinically relevant genes are frequently used as biomarkers for inclusion in clinical trials where knowledge of a tissue-specific diagnosis may be required for enrollment, such as FIGHT-101 and BEACON.It should be noted that genomic alterations can also be used to inform the tumor diagnosis such as IDH or FGFR fusions in patients with liver lesions, indicating a high likelihood of cholangiocarcinoma. 27Identification of highly specific genes such as these in specific clinical contexts could potentially obviate the need for the addition of a molecular diagnostic classifier to render or confirm a diagnosis.
Importantly, in the majority of reviews, the molecular diagnostic classifier increased confidence in the therapeutic plan regardless of whether the test results altered treatment recommendations.This confidence is particularly important for cases in the clinical care setting when diagnosis and histology are uncertain.In addition to influencing treatment selection, this confidence is critical in cases with adverse subsequent outcomes.Under these circumstances, the acceptance of subsequent supportive care could be delayed or declined by patients seeking additional opinions and evaluations.
There are several limitations to this analysis.While the primary endpoint of our study was assessed at the patient level and received adjudication by a third oncologist when there was discordance, secondary analyses were performed on review-level data and did not undergo adjudication.The lack of adjudication for these secondary endpoints is a limiting factor for interpretability of some sub-analyses given the substantial amount of reviewer discordance observed in these nuanced scenarios of clinical management.Discordance in some of these scenarios is to be expected, as clinical practice heterogeneity may be amplified in the setting of advanced-stage cancer.In this study, the reviewing oncologists represented a spectrum of oncology specializations and years in practice, which is expected to lead to some degree of clinical management heterogeneity.A further limitation of this survey is that we did not require oncologists to specify a treatment prior to review of the diagnostic classifier results.In clinical practice, even if the ultimate diagnosis is unknown, some form of treatment would be given.
Finally, given that our study is retrospective and deidentified, we could not evaluate if or how actual clinical management was impacted by the molecular diagnostic classifier results, and instead all of the recommendations that we report are theoretical.Consequently, our study design also does not allow us to determine whether patient outcomes were improved.Future studies aimed at prospectively assessing treatments and outcomes in cohorts of patients diagnosed with CUP are critical to better understand the impact of tissue-of-origin classification.
Currently, molecular diagnostic classifiers can support rendering an accurate and actionable cancer diagnosis but are not supported by the NCCN for patients with CUP as indicated in the Occult guidelines. 6This position may evolve as precision medicine approaches increasingly depend on tissue and histology diagnoses as observed elsewhere in the NCCN guidelines. 21Future prospective and outcome-based trials are necessary to further demonstrate the clinical utility of molecular diagnostic classifiers for CUP patients.
shows the flow of patients with regard to screening and eligibility.Our study thus assessed 289 patients with a consensus diagnosis of CUP, and of these, 111 (38.4%) required adjudication based on initial disagreement about the CUP diagnosis.An additional set of patients (n = 53, 18.3% of the assessed patient population) required adjudication due to disagreement regarding whether treatment would change following review of the molecular diagnostic classifier results.

F I G U R E 1
Overview of patient inclusion criteria and reviewer confirmation of cancer of uncertain/ unknown primary (CUP) status.

F I G U R E 2
Overview of patientlevel findings with regard to the primary study endpoint (oncologist assessment of treatment change based on molecular diagnostic classifier results).
highlights the 9 most common subtypes (with the remaining subtypes combined in the "Other" category for the purposes of visualization) and 17 biologically relevant genes corresponding to those subtypes across our F I G U R E 4 Oncoplot of pathogenic, clinically relevant gene alterations in the study cohort.
Patient demographics and molecular diagnostic classifier results.
T A B L E 1 Change in treatment regimen and feedback after review of molecular diagnostic classifier results (reviewer level).
T A B L E 2