Clinical and biological determinants of circulating tumor DNA detection and prognostication using a next-generation sequencing panel assay

ABSTRACT

Circulating tumor DNA (ctDNA) is utilized for molecular profiling of cancers, and is under investigation for a growing number of applications based on the assumption that ctDNA levels faithfully reflect disease burden. Our objective was to investigate whether patient and tumor characteristics may impact ctDNA detection or levels and the prognostic significance of ctDNA levels or mutations. We performed a retrospective cohort analysis of a comprehensively annotated cohort of 561 patients at a National Cancer Institute–designated comprehensive cancer center with advanced solid cancers who underwent ctDNA testing using a commercial targeted next-generation sequencing assay. ctDNA detection in advanced cancers was associated with older age, non-obese body mass index, and diabetes, but not with tumor diameter, volume, lesion number, or other pathological features. Regression models indicate that no more than 14.3% of the variance in ctDNA levels between patients was explained by known clinical factors and disease burden. Even after adjusting for established prognostic factors and tumor burden, ctDNA levels were associated with worse survival among patients without prior systemic therapy, while ctDNA mutations were associated with survival among patients who previously received systemic treatment. These findings uncover clinical factors that affect ctDNA detection in patients with advanced cancers and challenge the convention that ctDNA is a surrogate for tumor burden. Our study also indicates that the prognostic value of ctDNA levels and mutations are independent of tumor burden and dependent on treatment context.

KEYWORDS:

• Ctdna
• cell free dna
• circulating tumor dna
• obesity
• prognostic biomarkers

Introduction

Cell-free DNA (cfDNA) shed from cancer cells termed circulating tumor DNA (ctDNA) are detectable in liquid biopsies from patients with early and advanced-stage cancers using various technologies.^1–4 Early observations demonstrating that ctDNA declines after surgical treatment and are detectable at a higher frequency among advanced-stage cancers than early-stage cancers supported a close association between ctDNA and tumor burden, with some studies even indicating a linear relationship.^1,2,5–12 Accordingly, ctDNA has been employed as a surrogate for tumor burden and liquid biopsies are currently under investigation as early detection, pharmacodynamic, and minimal residual disease markers.^8,13–19 However, large-scale studies across many cancer types show that 10–40% of late-stage cancers have no detectable ctDNA.^{1,5,6,12,20–23} Different methodologies for ctDNA detection used in these studies and evidence that ctDNA can capture the genetic heterogeneity of metastatic cancers including subclonal mutations not apparent from single-site tissue biopsies suggest that undetectable ctDNA levels cannot be wholly attributed to sampling bias or pervasive deficiencies in any one ctDNA assay.^12,24–26 Although ctDNA biology remains poorly understood, the abundance of ctDNA is determined by mechanisms governing its production, release into circulation, stability in blood, and clearance.²⁷ Thus, other factors beyond tumor mass or cell divisions are likely to be important determinants of ctDNA detectability. This is supported by evidence that cfDNA concentrations are influenced by various aspects of human biology and disease including demographics, chronic comorbidities, inflammatory states, and tissue injury.^28–32

We hypothesized that patient and tumor characteristics may influence ctDNA detection across cancer types. Elucidation of such determinants would be critical to understanding the limitations of liquid biopsies for individual patients and guide the interpretation of ctDNA analyses for an increasing number of clinical applications. Herein, we examined a comprehensively annotated cohort of 561 patients with advanced cancers who underwent liquid biopsies analyzed by targeted next-generation sequencing.

Patients and Methods

Patient data and samples

This retrospective cohort analysis included consecutive eligible patients seen at the University of Texas Southwestern Harold C. Simmons Comprehensive Cancer Center and Parkland Health & Hospital System, who had liquid biopsies analyzed using the Guardant360 CDx assay. Eligible patients included all individuals with either unresectable or metastatic solid cancers. Collection of liquid biopsies was performed as part of routine clinical care. As such, liquid biopsies were generally performed in place of molecular profiling of tumor tissue for most patients. Hematology and blood chemistry test values were only abstracted if laboratory tests were drawn the same day as the liquid biopsy. Acute phase reactants and serum tumor markers were abstracted if laboratory tests were drawn within 2 weeks of the liquid biopsy. Characteristics of tumor pathology were abstracted from pre-systemic treatment biopsy or resection specimens. Disease site and tumor burden data were abstracted from radiographic images including computed tomography and magnetic resonance imaging performed within 4 weeks of the liquid biopsy. The sum of all diameters was the aggregate of the longest diameter on either the axial, sagittal, or coronal plane of all lesions. To calculate tumor volumes, 3-deminsional measurements for each lesion were used to calculate the volume of an ellipsoid (4/3 x π x width x length x height) which has been shown to approximate tumor volumes.^33,34 This study was performed in accordance with Good Clinical Practices and the Declaration of Helsinki and approved by the University of Texas Southwestern Institutional Review Board.

Liquid biopsy and ctDNA sequencing

Liquid biopsies were collected according to the Guardant360 Clinical Blood Collection Kit instructions (Guardant Health, Inc.).^9,35 As previously described, the Guardant360 CDx assay is a targeted high throughput hybridization-based capture technology for the detection of single nucleotide variants, insertions, and deletions in 73 genes by paired-end synthesis-sequencing using the NextSeq 500 and/or HiSeq 2500 platforms (Illumina, Inc.).^9,35 The complete list of genes and variants analyzed by the assay is viewable under the GTR Test ID GTR000527948.6. Putative germline mutations including variants identified by allele fractions between 40% and 60%, prior annotation as germline mutations, and manual review as indicated in clinical reports were excluded from our analysis.

Statistical analysis

Odds ratio for undetectable ctDNA was assessed by univariable logistic regression for all variables and multivariable logistic regression for variables with significant associations in univariable analysis. Linear regression using the ordinary least squares method was used to assess the coefficient of determination of univariable and multivariable models. Overall survival was determined from the date of the initial liquid biopsy to the time of death or most recent follow-up. Kaplan–Meier survival curves were compared using log-rank tests. Unadjusted and adjusted hazard ratios were determined using Cox regression models. The proportional hazard assumption was tested by including a time-dependent covariate in regression models and results were only reported if the covariate was not significantly associated with the outcome, indicating that the assumption was not violated. SPSS Statistics 23 (IBM) and GraphPad Prism 8 were used for statistical analysis.

Results

Identifying factors associated with undetectable ctDNA

Among 561 patients with unresectable or metastatic solid cancers who underwent liquid biopsies and NGS of ctDNA using the Guardant360 CDx assay, the median age was 68 years (interquartile range 60–75), 54.2% were male, and 42.2% had newly diagnosed disease (Table 1). We defined undetectable ctDNA as the absence of any somatic mutation by liquid biopsy testing, and found that 70 patients (12% of the entire cohort) had undetectable ctDNA, which is comparable to rates reported in prior studies.

Table 1. Characteristics of patients with ctDNA testing

Characteristics
Age, mean (SD)		66.5 (12.1)
Sex, No. (%)	Male	304 (54.2)
	Female	257 (45.8)
Race, No. (%)	White	199 (35.5)
	Black	40 (7.1)
	Asian	15 (2.7)
	Unknown	307 (54.7)
Cancer type, No. (%)	Non-small cell lung	229 (40.8)
	Pancreatic	69 (12.3)
	Prostate	54 (9.6)
	Colon	51 (9.1)
	Breast	38 (6.8)
	Hepatocellular	29 (5.2)
	Bile duct	26 (4.6)
	Head and neck	13 (2.3)
	Neuroendocrine	10 (1.8)
	Other	42 (7.5)
ECOG, No. (%)	0	165 (29.4)
	1	250 (44.6)
	2	88 (15.7)
	3 or greater	58 (10.3)
Stage, No. (%)	Locally advanced	40 (7.1)
	Metastatic	521 (92.9)
Lines of therapy, No. (%)	0	237 (42.2)
	1	113 (20.1)
	2	94 (16.8)
	3 to 5	99 (17.6)
	6 or more	18 (3.2)

For each patient, a comprehensive panel of clinical, tumor, and treatment characteristics including demographics, comorbidities, hematology and chemistry lab tests, pathological characteristics, disease sites, tumor measurements, and treatment history was curated (Figure 1). Variables were systematically tested in univariable logistic regression analyses, which showed that age, history of smoking, diabetes, metastatic disease, bone disease, number of organs involved, diameter of the largest and all lesions, and progression of disease at the time of liquid biopsy were associated with ctDNA detection (Figure 1). In contrast, obesity was significantly associated with undetectable ctDNA (odds ratio, 3.46; p < .01). A multivariable analysis including cancer type as a covariate revealed that only clinical factors, namely age, obesity, and diabetes, remained statistically significant predictors of ctDNA status. No multicollinearity was detected among the aforementioned variables related to tumor burden including the number of organs involved, diameter of the largest lesion, diameters of all lesions, and number of lesions based on the determination of variance inflation factors using linear regression models (all factors being less than 4).³⁶ Identical univariable and multivariable analyses were performed after excluding synonymous mutations or variants of unknown significance, which confirmed that age was associated with ctDNA detection, while obesity was associated with undetectable ctDNA. Given the possibility that prior treatment may impact ctDNA detection, we conducted logistic regression analysis of patients who had ctDNA collected prior to any systemic therapy, which similarly showed that obesity was negatively associated with ctDNA detectability (odds ratio, 2.52; p = .03), while age or tumor burden parameters were not significantly associated in univariable or multivariable analyses.

Figure 1. Patient and tumor characteristics associated with ctDNA detection. Variables relating to clinical factors, tumor biology, and treatment history were abstracted for 561 patients with advanced stage cancers to assess for predictors of ctDNA detection using the Guardant360 CDx assay. The odd ratios (OR) of undetectable ctDNA were determined using univariable and multivariable logistic regression analyses

Among the most frequently mutated genes linked to clonal hematopoiesis of the indeterminate potential (CHIP) including DNMT3A, TET2, TP53, and CEBPA, only TP53 is assessed in the Guardant360 CDx assay. We repeated our regression analysis after excluding 20 patients who only had ctDNA mutations in TP53 and no other genes. Multivariable analysis showed that age (odds ratio, 0.96; p < .01), obesity (odds ratio, 3.61; p < .01), and diabetes (odds ratio, 0.41; p = .03) remained significant predictors of undetectable ctDNA.

Examining the relationship between ctDNA levels and tumor burden

The lack of an association between tumor characteristics and ctDNA detection in our cohort led us to evaluate the relationship between ctDNA levels and tumor characteristics as pathologic factors including disease burden have been associated with ctDNA abundance.^{1,5,6,25–27,37} We quantitated ctDNA levels using three methods that have been applied in past studies including the maximal variant allele frequency (VAF) of any somatic mutation, the mean VAF of all somatic mutations, and the sum of VAF of somatic mutations.^6,38–43 We assessed the relationship between ctDNA levels and either the sum of the largest diameters for all lesions, the standard measurement for gauging responses in clinical practice, and the volume of all lesions, a metric which provides a more faithful reflection of disease burden. While there was a positive correlation between maximal VAF levels with the diameters (Pearson r, 0.38; p < .01) or volume (Pearson r, 0.30; p < .01) of all lesions, this relationship was associated with a large degree of scatter as observed in bivariate plots (Figure 2a). Consistent with a limited association between ctDNA levels and tumor burden, univariable linear regression analyses demonstrated that the explained variance of ctDNA levels by the diameter or volume of all lesions was only 14.3% and 9.3%, respectively. The explained variance of ctDNA levels calculated as the mean VAF or sum of VAF by the diameters or volume of all lesions was also poor (Supplementary Figure 1). In addition, the association between ctDNA levels and tumor burden was limited across cancer types with the coefficient of determinations ranging from 6.5% for lung cancers to 21.9% for pancreatic cancers (Supplementary Figure 2). Notably, the explained variance of ctDNA levels was even lower, ranging from 0% to 7% for other parameters of disease burden including the number of lesions, number of nodal metastases, and the diameter and volume of the dominant lesion (Supplementary Figure 3).

Figure 2. Pathologic characteristics and tumor burden are poorly associated with ctDNA levels. (a) Scatter plots depict the relationship between ctDNA levels and tumor burden defined as the diameters of all lesions or the volume of all lesions. The gray line depicts the best-fitting line. The inset shows the same plots with axes in log scale to demonstrate the degree of scattering across ctDNA values and tumor burden measurements. (b) Multivariable linear regression analyses adjusting for cancer type were used to assess the effects of variables with significant associations in univariable linear regressions on ctDNA abundance defined as maximal VAF, mean VAF, or sum of VAFs. Standardized coefficients denoting the fraction of variance explained by each model is shown the bar chart

We used multivariable linear regression analysis to assess independent associations between ctDNA level and different tumor burden parameters that were significant predictors in univariable analyses including the number of organs involved, the volume of all lesions, the number of lesions, and the presence of peritoneal disease (Figure 2b). Given the collinearity detected between the diameters and volume of all lesions, we chose to include only the volume of all lesions as a covariate in multivariable models as it is a more accurate estimate of tumor mass. Among predictors of ctDNA detection, age but not BMI or diabetes was significantly associated with ctDNA levels in univariable analyses, and therefore age was also included in the final model. Multivariable regression analysis adjusting for cancer type demonstrated that the number of organs involved and volume of all lesions remained positively associated with ctDNA levels, while the opposite was observed for age (Figure 2b). Similar results were seen when ctDNA was calculated as the mean or maximal VAF (Figure 2b). However, the coefficient of determination for any model never exceeded 14.5%, indicating that the variance of ctDNA levels across patients is poorly explained by known clinical factors or tumor biology (Figure 2b). There were also no statistically significant correlations detected between ctDNA levels and acute phase reactants and tumor markers with the exception of prostate-specific antigen (r = 0.29, p = .04)(Supplementary Figure 4).

In our cohort, six patients had three serial liquid biopsies obtained throughout their treatments, which allowed us to assess intraindividual variation in ctDNA levels and tumor burden longitudinally. In three patients, ctDNA kinetics largely paralleled changes in tumor burden (Figure 3 and Supplementary Fig. 5). In contrast, three patients demonstrated discordant trends in ctDNA levels and tumor burden (Figure 3 and Supplementary Fig. 5).

Figure 3. Longitudinal analysis of intraindividual relationships between ctDNA and tumor burden. Kinetics in ctDNA levels and tumor volumes for patients who had serial liquid biopsies. Representative axial slices from computed tomography scans are shown

ctDNA levels are independent prognostic biomarkers

There are conflicting data regarding the prognostic relevance of ctDNA detectability, and whether the abundance or number of ctDNA mutations is independently associated with survival.^44–48 In addition, prior studies have not assessed whether ctDNA is a prognostic biomarker after accounting for tumor burden. Using the Kaplan–Meier analysis and log-rank test, we found that patients with undetectable ctDNA had a median overall survival of 68.4 months versus 15.6 months for patients with detectable ctDNA (chi-square, 11.1; p < .001) (Figure 4a). When patients with detectable ctDNA were segregated by ctDNA levels greater or less than the median maximal VAF value, patients with ctDNA levels less than the median were noted to have longer overall survival (21.7 versus 10.8 months; chi square, 24.5; p < .001) (Supplementary Fig. 6A). Patients with mean VAF or the sum of all VAFs less than the median were similarly associated with longer overall survival (Supplementary Fig. 6B and 6 C). We also segregated patients into quartiles by ctDNA levels, which showed that overall survival was shortest for patients in the third (11 months) and fourth quartiles (10.5 months), followed by patients in the second quartile (18.6 months), and then patients in the first quartile (35.4 months) (Figure 4b).

Figure 4. ctDNA is an independent prognostic biomarker in advanced cancers. (a) Kaplan-Meier survival curves for patients with detectable and undetectable ctDNA. The median overall survival for patients with detectable and undetectable ctDNA was 15.6 and 68.4 months, respectively. p < .001 by the Log-rank test. (b) Kaplan-Meier survival curves for patients stratified by ctDNA level quartiles. The median overall survival for patients in the first, second, third, and fourth quartiles was 11, 10.5, 18.6 and 35.4 months, respectively. p < .001 by the Log-rank test. (c) Kaplan-Meier survival curves for patients stratified by the number of gene mutations. The median overall survival for patients with 1, 2, 3–5, and >5 mutations was 48.2, 20.8, 11.1 and 11.1 months, respectively. p < .001 by the Log-rank test. (d) Kaplan-Meier survival curves for patients stratified by tumor volume quartiles. The median overall survival for patients in the first, second, third, and fourth quartiles was 11, 10.5, 18.6 and 35.4 months, respectively. p < .001 by the Log-rank test. (e) Adjusted hazard ratios (HRs) of death with 95% confidence intervals for patients stratified by systemic treatment exposure was determined by multivariable Cox regression models adjusted for cancer type. Models include all variables listed as factors

Among patients with detectable ctDNA, a higher ctDNA mutation count (stratified by the median or quartiles) was associated with worse survival (Figure 4c and Supplementary Fig. 6D).

Greater tumor volumes (stratified by the median or quartiles) were also associated with worse survival (Figure 4d and Supplementary Fig. 6E). We used multivariable Cox regression models to determine whether ctDNA levels, mutation number, and tumor volumes were independent prognostic factors. Patients were initially stratified by whether they had received prior systemic treatment to assess whether the timing of ctDNA collection at baseline or after treatment exposure would impact its prognostic value. Additional variables included in multivariable Cox regression models included factors significantly associated with an increased risk of death in univariable analyses including age, male sex, ECOG score, Charlson Comorbidity Index, number of organs involved, and number of lesions as these. In multivariable analyses adjusting for cancer type among systemic treatment naïve patients, increasing age, ECOG score, ctDNA levels, number of ctDNA mutations, and number of organs involved were significantly associated with an increased risk of death (Figure 4e). Among patients previously treated with systemic therapies, only ECOG score, and number of ctDNA mutations, but not ctDNA levels, were associated with overall survival (Figure 4e). While small sample sizes limited the statistical power of analyses of individual cancer types, we note that similar findings were observed in multivariable Cox regression analyses of patients with non-small cell lung cancer and other cancer types (Supplementary Fig. 7).

Discussion

Elucidating clinical and pathologic factors that impact ctDNA is crucial to assessing the clinical validity and utility of ctDNA as a biomarker. Herein, we systematically interrogated a large and comprehensively annotated cohort of patients with advanced cancers who underwent ctDNA analysis using Guardant360 CDx, a highly sensitive and clinically validated assay.³⁵

Our analysis indicates that patient characteristics including age, obesity, and diabetes rather than tumor biology are key determinants of ctDNA detectability. While the precise mechanisms of these associations remain to be elucidated, host factors may impact the distribution, stability, and metabolism of ctDNA.²⁷ This is supported by prior findings demonstrating the impact of maternal health on fetal cfDNA detection and variation in cfDNA levels as a result of changes in physiology and diseased states.^{28–32,49–54} In fact, multiple studies have shown that maternal weight and age are negatively associated with fetal cfDNA detection, and these prior findings are akin to our results.^52–54 Our analysis does not exclude technical causes of undetectable ctDNA given constraints on sensitivity for any ctDNA assay including error rates of DNA polymerase for polymerase chain reaction-based methods and sequencing errors in NGS.⁵⁵ In addition, it is unclear whether patients with undetectable ctDNA using Guardant360 CDx includes those with ctDNA levels below the threshold of detection or those with a complete absence of ctDNA.

The limited association between ctDNA levels and parameters encompassing different facets of cancer growth in our patient cohort calls into question the widely held assumption that ctDNA reflects tumor burden.^{2,6,7,10,19,56} This discrepancy may be in part attributed to the smaller sample sizes of prior studies and methodological differences in ctDNA quantification, but also alludes to additional undefined or poorly appreciated patient or tumor characteristics that regulate ctDNA. Despite our efforts to capture major characteristics of tumor biology, not every aspect could be accounted for in our dataset. For instance, while the generation of ctDNA has been largely attributed to cell turnover, ctDNA may be shed by non-passive processes including release through exosomes, which are not readily detected by the Guardant360 CDx assay.^27,55,57 It is also unknown whether ctDNA is derived from all subclones, or if certain cells within the tumor hierarchy or within geographic locales such as invasive edges may contribute disproportionately to ctDNA levels.²⁷ This brings into the question whether subgroups of cancers with inflammatory phenotypes such as a high number of infiltrating lymphocytes or the tumor microenvironment may influence ctDNA levels. In addition, the abundance of specific ctDNA fragments may be influenced by molecular characteristics of cancer cells as the representation of the genome by cfDNA is dictated by nucleosome occupation which is in turn regulated by transcriptional and epigenetic processes.^58,59 Our longitudinal analysis of ctDNA in six patients with at least three serial measurements to allow correlative analyses with tumor measurements should only be considered exploratory, but nonetheless supports our primary findings that ctDNA does not strictly reflect tumor burden.

Our study indicates that the prognostic values of ctDNA levels and mutations are independent of tumor burden and context dependent. Specifically, ctDNA levels are only prognostic when collected at baseline, while ctDNA mutation counts are prognostic even after prior exposure to systemic treatment. The mechanism underlying the association of ctDNA mutation counts with survival in treated patients remains unclear, but may be explained by the fact that treatment induced selection of subclones may lead to greater genetic heterogeneity. Inconsistent results in past studies describing the prognostic value of ctDNA or lack thereof may be attributed to varying sample sizes, adjustment for a limited number of confounding variables, and other methodological differences.^{23,26,37,38,44–48} It is unclear if the association between ctDNA and survival is correlative or causative, but it is possible that ctDNA has functional effects based on evidence that cfDNA from various sources are ligands for toll-like receptors and may be internalized by cells.^60–62 Given that tumor mutation burden may be associated with cancer prognosis and impact ctDNA, it remains to be investigated whether tumor mutation burden contributes to the prognostic value of ctDNA mutations and abundance. Whether this type of treatment also affects the prognostic value of ctDNA is also unknown, although prior studies including patients who received chemotherapies, targeted therapies, or immunotherapies similarly showed that baseline ctDNA levels are prognostic.^15,37,63

Strengths of our analysis include its large sample size, which exceeds that of many prior studies, comprehensive clinical annotations, and the inclusion of major solid cancer types. In addition, we provide a thorough characterization of tumor biology including precise volumetric measurements rather than just lesion diameters as assessed in past studies. While our study incorporates a large panel of clinical and biological factors, missing data for a few variables such as race may limit the power to detect significant associations. Other limitations of our study includes the lack of a comparison of ctDNA to matched molecular profiling of tissue specimens, which is due to the fact that in nearly all instances liquid biopsies were performed in place of tissue biopsies. Thus, a lack of detectable ctDNA could be attributed to scarce mutations in the tumor itself or mutations not captured in the gene panel analyzed by the Guardant360 CDx assay. However, several studies including a recent rapid autopsy study comparing multi-site tissue and ctDNA sequencing indicate that cfDNA analysis may sensitively capture all if not most clonal mutations in patients and may even be superior to single biopsies in the detection of driver gene mutations.⁶⁴

Given the association between age and CHIP, it is expected that a non-trivial fraction of patients in our cohort who are predominantly older aged will have CHIP and CHIP mutations may thus confound ctDNA analyses. However, a couple of factors mitigate the contribution of CHIP in our analyses including the fact that the most common genes associated with CHIP including DNMT3A, TET2, ASXL1, PPM1D, SF3B1, ASXL2, GNB1, CBL, and SRSF2 are not assessed in the Guardant360 CDx assay.^65,66 Nonetheless, several genes with mutations associated with CHIP including TP53, JAK2, IDH1/2, and GNAS are assessed in the Guardant360 CDx assay. A sensitivity analysis was conducted by excluding cases who only had mutations in TP53 detected in ctDNA because there were no cases where JAK2, IDH1/2, and GNAS mutations were the only mutations detected in ctDNA, and we found no differences in the subsequent results from our primary analysis. It should also be noted that a pan-cancer analysis of 10,719 patients across 30 solid cancer types in The Cancer Genome Atlas shows a substantially increased frequency of TP53, IDH1/2, and GNAS mutations in solid cancers than in CHIP, corroborating the likely nominal contribution of CHIP mutations to our analyses.^65,67

Another limitation of this study is the number of genes analyzed by the Guardant360 CDx assay, and it remains to be investigated whether a larger gene panel with equivalent depth of sequencing coverage may lead to similar results. In addition, somatic mutations in nonmalignant cells have now been observed in an increasing number of tissue types, and whether these may be another source of mutations in cfDNA remains to be clarified.⁶⁸

Author contributions

D.H. conceived the project, collected data, performed data analyses, and wrote the manuscript. D.E.G., and M.S.B contributed patient data. M.E., D.E.G., and M.S.B participated in data interpretation and contributed to the final editing and approval of the manuscript.

Acknowledgments

We thank Dr. Hao Zhu for the discussions and review of the manuscript.

Disclosure statement

D.E.G. reports advisory or consulting roles for Bristol-Myers Squibb, Karyopharm, and Catalyst and receives research funding from AstraZeneca, BerGenBio, Karyopharm, and 3V Biosciences. D.H., M.E., and M.S.B. declare no competing interests.

Supplementary material

Supplemental data for this article can be accessed on the publisher’s website

Additional information

Funding

This work was supported by a Cancer Prevention and Research Institute of Texas Early Clinical Investigator Award (to DH) and NIH grants 2UM1CA186644-06 (to MSB), U01 AI156189, K24CA201543-01 (to DEG).