Abstract
Liquid biopsy (LBx) for tumor profiling is increasingly used, but concerns remain regarding negative results. A lack of results may truly reflect tumor genomics, or it may be a false negative that would be clarified by tissue testing. A method of distinguishing between these scenarios could help clarify when follow-on tissue testing is valuable.
Here we evaluate circulating tumor DNA (ctDNA) tumor fraction (TF), a quantification of ctDNA in LBx samples, for utility in identifying true negative results. We assessed concordance between LBx and tissue-based results, stratified by ctDNA TF, in a real-world genomic dataset of paired samples across multiple disease types. We also evaluated the frequency of tissue results identifying driver alterations in patients with lung cancer after negative LBx in a real-world clinicogenomic database.
The positive percent agreement and negative predictive value between liquid and tissue samples for driver alterations increased from 63% and 66% for all samples to 98% and 97% in samples with ctDNA TF ≥1%. Among 505 patients with lung cancer with no targetable driver alterations found by LBx who had subsequent tissue-based profiling, 37% had a driver, all of which had ctDNA TF <1%.
Patients with lung cancer with negative LBx and ctDNA TF ≥1% are unlikely to have a driver detected on confirmatory tissue testing; such informative negative results may benefit instead from prompt treatment initiation. Conversely, negative LBx with ctDNA TF <1% will commonly have a driver identified by follow-up tissue testing and should be prioritized for reflex testing.
This work illustrates the utility of circulating tumor DNA tumor fraction to distinguish true negative liquid biopsy results from indeterminate results that may benefit from follow-up tissue testing. This will support clinicians in the decision of whether to act on negative liquid biopsy results and begin non-targeted therapy or to prioritize obtaining tissue for confirmatory comprehensive genomic profiling.
Introduction
Negative results from liquid biopsy (LBx) cancer molecular profiling can be challenging to interpret. An absence of targetable alterations may accurately reflect the tumor genotype (i.e., a true negative). Alternatively, it may represent a false negative due to insufficient circulating tumor DNA (ctDNA) shed, concealing a targetable oncogenic driver. Reflex to tissue biopsy (TBx) profiling is thus advised, per FDA label, for negative LBx results across tumor types (1). Confidence in a true negative LBx could be helpful to inform timely treatment selection.
Non–small cell lung carcinoma (NSCLC) is the disease most commonly tested using liquid biopsy (2). Targetable alterations are well defined and mutually exclusive (3). Availability of molecular genotyping results before first-line (1L) therapy has been found to be associated with significantly better survival (4). Thus, a positive result for any National Comprehensive Cancer Network–defined driver could invite prompt therapy initiation. In the absence of a molecular driver alteration, however, the clinical decision to initiate immunotherapy with or without chemotherapy is urgent. Therapy selection in colorectal cancer similarly requires confidence in negative results in RAS genes, as well as BRAF, as do various other tumor types such as breast and prostate (5–8). Frustratingly, ctDNA shed is variable, making negative results via LBx insufficient for ruling out driver mutations (1). Confirmation of the presence of ctDNA in a sample would increase confidence in negative results. Conversely, knowledge that a LBx did not adequately sample a patient's underlying tumor via ctDNA could prompt tissue molecular testing. One key challenge is that detection of germline and clonal hematopoiesis (CH) alterations in LBx profiling can appear to represent tumor alterations but may obscure an otherwise non-shedding tumor (9–11).
A potential marker of the presence of genuine tumor signal is ctDNA tumor fraction (TF), a biomarker which has been studied across a range of LBx assays (12). We here define ctDNA TF as the proportion of total cell-free DNA in a sample which is tumor derived (ctDNA). We developed an improved method of calling ctDNA TF from LBx comprehensive genomic profiling (CGP) incorporating aneuploidy, variant allele frequency, and canonical alterations. This new ctDNA TF biomarker uses a new technique to avoid confounding signal from germline and CH variants. Rather than incorporating routine peripheral blood mononuclear cell (PBMC) sequencing (13), which can be challenging to scale across diverse global populations, we incorporated algorithmic removal of germline based on variant allele frequency and removal of CH signal based on multiomic patterns identified by PBMC sequencing. Here we investigate whether algorithmic quantification of ctDNA by TF in an LBx sample can increase confidence in negative LBx results, reducing need for confirmatory TBx testing.
Materials and Methods
Study cohorts
We interrogated an institutional database of 3,854 patients with both LBx (FoundationOneLiquidCDx, F1LCDx) and TBx (FoundationOneCDx, F1CDx) to calculate positive percent agreement (PPA; sensitivity), negative predictive value (NPV), and negative percent agreement (NPA; specificity) of LBx compared with TBx, provided that blood collection for liquid CGP was after tissue collection: NSCLC (n = 1,753), breast carcinoma (n = 835), colorectal cancer (n = 859), and pancreatic cancer (n = 407; Supplementary Fig. S1). Approval for this study, including a waiver of written informed consent and Health Insurance Portability and Accountability Act waiver of authorization, was obtained from the Western Institutional Review Board (protocol 20152817). Conduct was performed in accordance with the principles of research ethics as set forth in the Belmont Report.
In parallel, we queried the nationwide (U.S.-based) deidentified Flatiron Health–Foundation Medicine (FH-FMI) NSCLC clinico-genomic database (CGDB) from approximately 280 U.S. cancer clinics (∼800 sites of care). Retrospective longitudinal clinical data were derived from electronic health record data, comprising patient-level structured and unstructured data, curated via technology-enabled abstraction, and were linked to genomic data derived from FMI CGP tests in the FH-FMI CGDB by deidentified, deterministic matching (14). The study included 6,810 patients diagnosed with NSCLC from January 2011 to December 2022 receiving any multigene LBx [Guardant360, FoundationOneLiquid (F1L, a laboratory developed test prototype of F1LCDx), or F1LCDx]. Patients were considered biomarker positive if one of seven common targetable NSCLC driver alterations was detected: KRAS, EGFR, MET exon 14, BRAF V600E mutations or ALK, ROS1, and RET fusions. Patients with a birth year of 1937 or earlier may have an adjusted birth year in Flatiron datasets due to patient deidentification requirements. The data in the CGDB are deidentified and subject to obligations to prevent reidentification and protect patient confidentiality.
CGP
CGP was performed in a laboratory that is Clinical Laboratory Improvement Amendment certified, College of American Pathologists accredited, and New York State approved (Foundation Medicine).
LBx samples were profiled using a validated, FDA-approved next-generation sequencing panel assay F1LCDx (15). Circulating cell-free DNA was extracted from whole blood. CGP was performed using hybrid-capture, adaptor ligation–based libraries. F1LCDx reports single-nucleotide variants, insertions/deletions, genomic rearrangements, copy-number amplifications, and losses in 324 cancer-related genes. F1LCDx also reports genomic signatures including ctDNA TF.
TBx were analyzed using F1CDx as described previously (16, 17). Briefly, the pathologic diagnosis of TBx was confirmed on routine hematoxylin and eosin–stained slides. Samples with a minimum of 20% tumor nuclei underwent DNA extraction, and hybrid capture-based sequencing was performed on the same 324 cancer-related genes interrogated by F1LCDx.
TF quantification
ctDNA TF was quantified by combining multiple methods. For samples with significant aneuploidy, the purity assessment from a robust copy-number model that accounts for both the observed coverage variation and allele frequencies of genome-wide SNP allele frequencies is used to determine the ctDNA TF estimate (2). When significant aneuploidy is not present, the allele frequencies of short variants and rearrangements deemed very likely to be somatic are used to estimate ctDNA TF. The somatic status determination for short variants and rearrangements may be made in one of two ways: (i) presence on a white list of specific variants and categories of variants that are highly biased toward being somatic (and not due to CH), or (ii) evidence of a statistically significant difference in quantitative metrics from fragments harboring the mutated allele. These approaches do not identify all true somatic variants found for a given sample, but the substantial majority have at least one or a few variants that can be used. The ctDNA TF estimate may include the highest VAF for a short variant or rearrangement deemed to be somatic. The specific set of short variants and rearrangements deemed very likely to be somatic was determined on the basis of assessment of rates of occurrence of variants in clinical plasma samples compared with the corresponding rates in PBMC samples. Examples of variant categories included are loss-of-function variants in PTEN and known activators of PIK3CA. In addition, an assessment of cfDNA fragment sizes is used to exclude CH-derived aneuploidy from copy-number modeling.
Real-world clinical outcomes determination
Real-world progression (rwP), real-world progression-free survival (rwPFS), and real-world overall survival (rwOS) events were captured using a clinician-anchored abstraction approach, such that the date of cancer progression was identified as the date of the first source evidence for progression referenced by the clinician (e.g., radiology report, pathology, or exam; ref. 18, 19). rwPFS was defined as time from start of systemic therapy to either the date of first rwP 14 days following start of therapy or death. Patients were censored at the date of their last clinic visit. rwOS was defined as time from start of systemic therapy to date of death. Patients were censored at the later of their last documented visit date, their last documented medication administration, or their last abstracted oral therapies order.
Statistical analysis
A total of 95% confidence intervals for PPA, NPV, and NPA (Supplementary Fig. S1C) were calculated using the Wilson score interval. For a subset of patients in the FH-FMI CGDB, rwOS and rwPFS from start of 1L treatment were estimated using Kaplan–Meier analysis. For rwOS estimates, risk set adjustment was applied to account for left truncation. Patients entered the at-risk population at the latter of either their second visit to the Flatiron Health network or the date of FMI report. Statistics, computation, and plotting were carried out using Python 2.7 (Python Software Foundation, RRID: 008394) packages matplotlib (RRID:SCR_008624), statsmodels (RRID:SCR_016074), and Pandas (RRID:SCR_018214), or R 4.2.1 (Posit, RRID:SCR_001905) packages ggplot2 (RRID:SCR_014601), survminer (RRID:SCR_021094), survival (RRID:SCR_021137), and tidyverse (RRID:SCR_019186).
Data availability
All relevant data are provided within the article and its accompanying Supplementary Data. Because of Health Insurance Portability and Accountability Act requirements, we are not consented to share individualized patient genomic data, which contains potentially identifying or sensitive patient information. Foundation Medicine is committed to collaborative data analysis, and we have well established and widely used mechanisms by which investigators can query our core genomic database of >600,000 deidentified sequenced cancers to obtain aggregated datasets. More information and mechanisms for data access can be obtained by contacting the corresponding authors or the Foundation Medicine Data Governance Council at [email protected].
The CGDB data that support the findings of this study have been originated by Flatiron Health, Inc. and Foundation Medicine, Inc. The data are deidentified and access is restricted to ensure appropriate use in compliance with all of the conditions imposed by 45 CFR Part 164.514(b; ref. 1), including but not limited to prohibitions against attempts to reidentify. Foundation Medicine and Flatiron Health are committed to collaborative data analysis and have well established and widely used mechanisms by which qualified researchers can pursue scientific research with the data. Requests for data sharing by license or by permission for the specific purpose of replicating results in this article can be submitted to [email protected] and [email protected].
Results
ctDNA TF is the key determinant of LBx concordance for TBx-detected driver alterations
In the five most prevalent diseases represented in the Foundation Medicine genomic database, PPA, NPV, and NPA for detection of common oncogenic alterations were assessed for patients with tissue biopsy followed by liquid biopsy with a median time between TBx specimen collection and LBx specimen collection of 357 days (range: 0–3,914; interquartile range: 37.3–580; Supplementary Fig. S1A). In total, 3,854 patients (NSCLC: 1,753; breast: 835; colorectal cancer 859; pancreatic: 407; Supplementary Fig. S1B) were included.
Overall sensitivity for drivers was variable across diseases, ranging from 49% for KRAS alterations in pancreatic cancer to 76% for ESR1 mutations in breast cancer (Fig. 1A). However, when limiting to patients with elevated ctDNA TF (≥1%), sensitivity for all alterations in total was 97.3% indicating ctDNA TF is a key measure for informing assay sensitivity (Fig. 1B). Conversely, in patients with low or undetectable ctDNA TF, sensitivity was consistently low (Fig. 1C). For select driver variants in NSCLC and colorectal cancer, NPV was calculated across the full cohort, in ctDNA TF-elevated specimens and in ctDNA TF-low/undetectable specimens. NPV was 66% for all specimens, consistent with the need for confirmatory tissue testing. Yet when considering ctDNA TF, NPV was 97% for those with elevated ctDNA TF and less than 50% for those with low/undetectable ctDNA TF (Fig. 1B and C). We additionally analyzed samples collected within 0–30, 0–60, and 0–90 days and found numerically higher PPA and NPV results for all cases with ctDNA TF≥1%, though with broader confidence intervals due to lower sample pair counts (Supplementary Figs. S2A–S2C, S3A–S3C, S4A–S4C).
We further analyzed PPA and NPV at different ranges of TF. PPA and NPV remained >95% for all ranges studied above 1% (Supplementary Fig. S5A and S5B). For the 257 pairs with ctDNA TF between 1% and 2%, PPA remained 98% and NPV was 97%.
Like an informative negative, confidence in a biomarker positive liquid CGP test is critical for informing guideline recommended choice of therapy. We assessed specificity (i.e., the true negative rate, which informs the risk of false-positive results) across the same oncogenic biomarkers and found that regardless of the level of ctDNA TF, specificity was generally high for oncogenic biomarkers (Supplementary Fig. S6A–S6C). Notably, LBx found ESR1 alterations in patients with breast cancer not found via TBx, resulting in a lower specificity. This was found among all patients (609 LBx negative/736 TBx negative, 82.7%) and in patients with ctDNA TF ≥1% (308/408, 75.5%). Similarly, LBx found RAS alterations in colorectal cancer not found via TBx. Because these mutations emerge as the result of selective pressure from therapy and are often subclonal (20–22), it is expected that liquid CGP may detect these mutations when previous tissue CGP testing did not. High specificity provides confidence in a biomarker positive result even if ctDNA TF is low.
Reflex to tissue CGP identifies drivers missed by LBx in patients with advanced NSCLC
To better understand how ctDNA TF could inform the need for confirmatory reflex testing on TBx in the real world, we studied a retrospective database of patients with advanced NSCLC (aNSCLC) who had LBx-based CGP results prior to 1L therapy followed by TBx-based CGP results at any point (Fig. 2). Of 2,265 LBx sent prior to 1L therapy in patients with aNSCLC, 1,355 (60%) were negative for all of the seven common driver alterations (Fig. 3A). Among patients with negative LBx, 505 (412 evaluated by Guardant360, 93 F1L or F1LCDx) received follow-up TBx testing and 37% had a targetable driver detected on reflex to TBx. For patients with driver-positive TBx after a negative LBx CGP, KRAS (46%), EGFR (22%), and ERBB2 mutations (10%) along with ALK rearrangements (6%) were most common NCCN driver mutations detected (Fig. 3B).
ctDNA TF informs the relative benefit from reflex to TBx CGP
For the 505 patients who received follow-up TBx after a negative LBx result, the median time to treatment initiation was 3.0 weeks. Unsurprisingly, patients who went on to receive TBx CGP prior to 1L therapy had a longer time to treatment start (median 4.7 weeks) than those who initiated therapy without waiting for confirmatory TBx CGP (median 1.8 weeks). We then focused on the 80 patients with a ctDNA TF estimate (10 of 13 F1L results, 70 of 70 F1LCDx results) and used TF to stratify patients receiving confirmatory TBx CGP to elucidate the value of TBx reflex testing after a negative LBx result. In patients with TF <1%, driver alterations were detected in 52% (29/56) of TBx reflex testing, whereas confirmatory TBx testing was negative for all patients (24/24) who had ctDNA TF ≥ 1%. (Fig. 4). These results are consistent with the high NPV for driver alterations in ctDNA TF-elevated LBx samples.
Timely tissue reflex testing enables use of targeted therapy and is associated with prolonged benefit for patients
We then studied the clinical value of confirmatory tissue testing for those with indeterminate liquid results (ctDNA TF <1%). Subsequent TBx CGP after LBx prior to 1L therapy was associated with higher likelihood of receiving matched targeted therapy (43/254, 17% vs. 9/220, 4%; Fig. 5A). Targeted therapy use was limited in patients whose tissue CGP detected KRAS and MET alterations (2/16 and 1/7, respectively), likely due to lack of options (non-G12C KRAS, Supplementary Fig. S7) or testing occurring prior to approval of appropriate targeted therapies. When excluding non-G12C KRAS patients along with driver negative patients, 130 patients had a targetable driver on confirmatory TBx testing after a negative LBx. Among 81 with TBx results prior to 1L therapy, 53% (43) received targeted therapy compared with 18% (9/49) of patients with testing results after start of 1L treatment (Fig. 5B; Supplementary Table S1). For patients treated with targeted therapy after timely tissue reflex CGP, both median rwPFS (15.9 months; Fig. 5C) and median rwOS (45.8 months; Fig. 5D) are favorable, consistent with prolonged benefit from the selected therapy.
Discussion
LBx is a pragmatic option for CGP in multiple disease settings where tissue is not readily available, but false negatives are an established limitation. Despite the frequent ability to detect ctDNA of advanced tumors in peripheral blood (2), a sample identifying no driver alterations is an uncertain result due to variable sensitivity dependent upon ctDNA shed. Thus, the FDA label for approved LBx recommends confirmation of negative results with a reflex to tissue testing (2). This confirmatory tissue testing risks creating a diagnostic burden for patients and physicians, which could be reduced if we knew which patients would benefit the most from follow-up tissue testing after a negative LBx result.
We find that the ctDNA TF biomarker described here can identify samples containing adequate ctDNA content to enable increased concordance with tissue testing. Consistent with previous findings (23), we find that 37% of patients with negative CGP results on LBx have on-guideline biomarkers identifiable by tissue testing. We examined the changes of sensitivity and specificity of LBx-based CGP at low or high TF, which incorporates multiple features to minimize variability of individual variant allele fractions while optimizing sensitivity. We found that the PPA of LBx CGP dramatically increases to >90% for NCCN-designated drivers across major tumor types for samples with high ctDNA TF. Importantly, this increased sensitivity yields a high NPV (>95%). High ctDNA TF thus increases confidence in negative LBx results, informing the decision to reflex to TBx-based CGP or start therapy based on LBx results alone (Fig. 6).
For driver-negative samples with low ctDNA TF, concordance is reduced and confirmatory tissue testing is a priority as it may reveal actionable alterations. Timely reflex testing via TBx after negative LBx CGP is associated with a 3-fold increase in targeted therapy use and approximately 50% reduction in chemotherapy use, as well as potentially improved rwPFS and rwOS. However, the clinical condition and personal preferences of a patient may strongly contraindicate a tissue biopsy. A low TF may provide further information to aid in the decision of collecting tissue for confirmatory testing. Interestingly, we have separately identified that low ctDNA TF is associated with a more favorable prognosis (24), potentially indicating that these low ctDNA TF cases may have a more indolent course amenable to delaying therapy to complete tissue testing.
For driver-negative samples with high ctDNA TF, high concordance with tissue means the value of reflex to tissue testing is reduced. When there is time urgency, prompt initiation of treatment may be beneficial. The 24 patients with high ctDNA TF and no driver alteration in our dataset might have avoided reflex to confirmatory TBx and associated delay in treatment initiation. Identification of patients likely to have samples with high ctDNA TF would be of great interest. At this time, the authors are unaware of any clinical factors that are predictive of ctDNA TF other than known markers of disease burden such as Eastern Cooperative Oncology Group performance status, disease grade, and various lab values (25, 26).
We acknowledge certain limitations to the interpretation of this work. First, these data focus on short variants and rearrangements due to their role as targetable drivers in the tumor types studied, which can be detected at high sensitivity at TF of 1% or greater. ctDNA TF may require additional calibration for application to copy-number amplifications and losses which commonly require greater tumor content for reliable detection (15). Though we find improved sensitivity for copy-number alterations with increasing TF, additional investigation into copy-number amplifications and losses will be needed to establish similarly high PPA and NPV at relatively low (1%–5%) ctDNA TF. A second limitation is that these samples were not collected in a prospective, randomized manner. Even still, the data presented here fulfill most of the Bradford Hill Criteria for causation from epidemiologic data, specifically strength (patient n), consistency (27–29), temporality (Fig. 4; ref. 30), biological gradient (Supplementary Fig. S5), plausibility (27), and analogy (Fig. 6). Finally, these data are specific to this biomarker performed on a large CGP panel incorporating multiomic methods to avoid CH signal, and thus may not be applicable to ctDNA TF estimations from other panels.
Conclusion
Here we present evidence that ctDNA TF is a sensitive and specific method for determining whether a blood LBx sample has a sufficient quantity of ctDNA for detecting alterations in solid tumors. Taken together, these results demonstrate that ctDNA TF could inform the decision between whether to pursue follow-up tissue molecular testing versus timely initiation of systemic therapy, reserving additional diagnostic testing for patients where it may yield informative results and adding confidence to others for a prompt initiation of therapy.
Authors' Disclosures
C.D. Rolfo reports personal fees from AstraZeneca, Roche, MSD, Inivata, Archer, MD Serono, Boston Pharmaceuticals, Invitae, Regeneron, BostonGene, Novartis, Bayer, and Imagene; grants from LCRF-Pfizer and NCRF; and non-financial support from GuardantHealth and Foundation Medicine outside the submitted work. R.W. Madison reports personal fees from Foundation Medicine, Inc. and Roche Holding AG during the conduct of the study; in addition, R.W. Madison has a patent for WO2022271159 pending and a patent for 63/466541 pending. L.W. Pasquina reports other support from Foundation Medicine during the conduct of the study, as well as other support from Foundation Medicine and Roche Holdings outside the submitted work. D.W. Brown reports other support from Foundation Medicine during the conduct of the study, as well as other support from Foundation Medicine outside the submitted work. Y. Huang reports personal fees from Foundation Medicine, Inc and Roche Holding AG during the conduct of the study; in addition, Y. Huang has a patent for WO2024015973 pending and a patent for 63/466541 pending. J.D. Hughes reports a patent for WO2024015973 pending and a patent for 63/466541 pending. R.P. Graf reports other support from Foundation Medicine, Inc during the conduct of the study. G.R. Oxnard reports personal fees from Foundation Medicine and Roche during the conduct of the study, as well as personal fees from Eli Lilly outside the submitted work. H. Husain reports personal fees from Foundation Medicine, Neogenomics, AstraZeneca, Janssen, Mirati, EMD Serono, Amgen, Sanofi, Roche, Regeneron, Merck, and Daiichi Sankyo; grants and personal fees from BillionToOne; and grants from Bristol Myers Squibb and Lilly outside the submitted work. No other disclosures were reported.
Authors' Contributions
C.D. Rolfo: Conceptualization, supervision, investigation, writing–review and editing. R.W. Madison: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. L.W. Pasquina: Formal analysis, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. D.W. Brown: Data curation, software, formal analysis, validation, investigation, methodology, writing–review and editing. Y. Huang: Software, validation, methodology, writing–review and editing. J.D. Hughes: Software, methodology, writing–original draft, writing–review and editing. R.P. Graf: Conceptualization, investigation, methodology, writing–review and editing. G.R. Oxnard: Conceptualization, resources, supervision, funding acquisition, investigation, visualization, methodology, writing–original draft, writing–review and editing. H. Husain: Software, supervision, investigation, writing–review and editing.
Acknowledgments
Foundation Medicine is a wholly owned subsidiary of Roche. Foundation Medicine is a producer of FDA-regulated molecular diagnostics. Authors employed by Foundation Medicine were involved in the design and conduct of the study, analysis, interpretation of the data, preparation, review, and approval of the article.
This study was supported by Foundation Medicine, Inc. (R.W. Madison, L.W. Pasquina, D.W. Brown, Y. Huang, J.D. Hughes, R.P. Graf, G.R. Oxnard).
Note: Supplementary data for this article are available at Clinical Cancer Research Online (http://clincancerres.aacrjournals.org/).