Dynamic Mutational Landscape of Cerebrospinal Fluid Circulating Tumor DNA and Predictors of Survival after Proton Craniospinal Irradiation for Leptomeningeal Metastases

Leptomeningeal metastasis (LM) is a dreaded complication of solid tumor malignancy with poor prognosis and limited treatment options. Conventional treatments result in median survival of 3 to 6 months, depending on the histology, functional status, and response to systemic therapy (1). Recently, proton craniospinal irradiation (pCSI) has emerged as a potentially efficacious treatment for patients with solid tumor LM with promising survival outcomes (2, 3).

Irradiation of mammalian cells results in direct and indirect DNA damage that can lead to genetic alterations, including mutations, insertions, deletions, and copy-number aberrations (CNA; ref. 4). Because tumor cells often carry highly abnormal and heterogeneous genomes, the genetic effects of ionizing radiation on tumor cells may differ greatly, even among cells of the same histology (5). Thus, the efficacy of radiotherapy may depend in part on the complement of genetic alterations present in tumor cells before and after treatment. Analysis of detectable genetic aberration in LM can reveal clonal evolution of metastatic cancer cells, and the emergence of specific clones after pCSI may predict specific responses to radiotherapy (6).

Though there are currently no established biomarkers of response to radiotherapy, liquid biopsies are an emerging noninvasive method to diagnose tumors (7), characterize their natural history (8), and understand their response to therapy (9). Cerebrospinal fluid (CSF) can be obtained through lumbar puncture, a minimally invasive procedure with a low complication rate (10). We have previously shown that baseline levels and dynamic changes in the number of CSF circulating tumor cells (CTCs) in patients with LM could be used to stratify response to pCSI (11). Like CTCs, circulating tumor DNA (ctDNA) in the CSF (12, 13) and plasma (14) of patients with LM have been shown to reflect tumor evolution and are of potential use as a biomarker of treatment response. When compared with plasma ctDNA, CSF ctDNA demonstrates higher sensitivity to detect LM compared with traditional cytology (15). Mutational analysis of ctDNA of the CSF reveals clinically relevant genomic alterations in patients with primary central nervous system (CNS) tumors and CNS metastases with higher allelic fractions compared with plasma-detected mutations (16, 17).

Even when detecting genetic variation within a tumor or ctDNA, most detected variants have unclear significance and clinical applications pose a significant challenge (18, 19). Several methods have been proposed to derive clinically relevant genetic alteration summary scores, with varying application. In patients with primary brain tumors or CNS metastases, variant allele frequencies (VAF) have been shown to correlate with tumor burden and decrease with therapy (i.e., surgery and systemic therapy) (12). However, the predictive value of ctDNA alterations and mutational summary measures in patients with LM treated with pCSI remains unknown.

We analyzed the ctDNA in patients from our phase I trial of pCSI for solid tumor LM (NCT03520504). We assessed the changes in the mutational landscape that are specific to the selective pressure of pCSI. We also explored whether ctDNA summary measures, like mean VAF and mean cancer cell fraction (CCF), which incorporates the inferred clonal architecture, can predict clinical outcomes in patients with LM treated with pCSI.

For the patients on protocol receiving pCSI for LM, ctDNA from the CSF and plasma were collected at baseline, and at 1 month, 3 months, and 6 months post-pCSI (2). Fourteen patients had both baseline CSF ctDNA measured and at least one follow-up timepoint measurement. Systemic therapy was held 7 days prior to initiation of and during pCSI, with the exception of endocrine therapy and trastuzumab for patients with breast cancer, and patients resumed their previous systemic therapy or initiated a new systemic therapy 4 weeks after completing pCSI. Per study exclusion criteria, patients with extensive systemic disease without reasonable systemic therapy options and with multiple, serious neurologic deficits were excluded. Patients were followed until death or censoring on December 31, 2020.

For all ctDNA samples, we performed targeted exome sequencing using Memorial Sloan Kettering-Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT), an in-house panel assaying 505 genes (20). All samples were compared with their germline tissue to determine somatic mutations, except for 1 patient who was missing germline tissue and was compared with a pooled normal tissue reference. The sequencing depth for each sample is reported in Supplementary Table S1. Mutations were independently called without any prior knowledge using total depth ≥ 20, variant allele depth ≥ 8 and VAF ≥ 2% for known mutational hotspots and VAF ≥ 5% for novel mutations (ref. 6; Supplementary Table S2). On manual verification of all called variants, 14 variants were included that had a total read depth < 20 but ≥ 10, because they occurred in multiple samples from the same patient or had a VAF > 25%. All variants had at least 3 supporting reads which is the absolute minimum recommended allele depth to be considered a mutation (21). For all mutations called in either the plasma or the CSF, a secondary analysis was performed for all other samples collected from the same patient, in which less stringent criteria were applied to detect the full compilation of alterations (called as present if variant allele depth ≥ 1). If these criteria were not met, a mutation was marked as not present and considered to be limited to the plasma or CSF. For this secondary mutation analysis, we report only the frequency of mutations that are common and unique to plasma and CSF as well as the overall agreement between the plasma and CSF pre- and post-pCSI using Cohen's kappa statistic. Mutations pre- and post-pCSI were compared using a paired Wilcoxon signed rank test.

The mean VAF was calculated as the average proportion of sequencing reads showing any alteration (22), and the mean VAF pre-CSI (VAF_pre) and post-CSI (VAF_post) were determined for all patients. The mean change in VAF pre- and post-pCSI (Δ_VAF) was derived by calculating the average difference in VAF for each variant and using a value of zero for undetected variants (i.e., a mutation that is not detectable or was only detectable post-pCSI).

Allele-specific copy-number variation and the clonality of mutations were determined using FACETS (version 0.3.9; ref. 23). A two-pass implementation with a low-sensitivity run (critical value = 100) was used to determine the purity and sample ploidy, and a high sensitivity run (critical value = 50) was used to determine gene-level segmentation and integer copy-number calls. A threshold of ±0.5 was used to call an amplification or deletion of a chromosomal segment. Samples with a purity that could not be estimated (typically ≤ 0.30) were excluded from the clonality analysis. Estimated purity of plasma and CSF ctDNA samples was compared using an unpaired Wilcoxon signed rank test. For 11 analyzable samples, preexisting MSK-IMPACT mutation calls using an automated clinical pipeline were obtained for either their primary tumor or a metastatic tumor from prior biopsies. The clonal evolution that took place from initial biopsy (i.e., either primary tumor or metastasis) to ctDNA before pCSI, and ctDNA after pCSI was estimated using CCF. The FACETS algorithm is an exploratory approach, and with the default settings, mutations with a high estimated CCF are determined to be clonal whereas lower CCF estimates (i.e., ≤ 80%) indicate a subclonal mutation. For new subclonal mutations that were only detected after pCSI, we calculated the mean CCF (CCF_SUBCLONAL). CCF_SUBCLONAL was only calculated for all samples with a high enough purity to determine clonality using the default FACETS algorithm settings.

We separately analyzed all genes with alterations present in tumor and metastases, genes with alterations occurring in ctDNA of the CSF and plasma, genes with alterations that are unique to ctDNA of the CSF and plasma before pCSI and genes with alterations in ctDNA after pCSI. For these gene lists, we used gene set enrichment analysis (GSEA) to interrogate any significant overlap with Reactome pathways genes. We report all pathways with significant overlap at an FDR q value ≤ 0.05.

We identified an independent cohort of patients with CNS metastases, including brain metastases or LM, who had CSF ctDNA MSK-IMPACT testing at the time of diagnosis. Genomic variants had already been determined according to a clinically validated pipeline (20). These patients were further stratified into those who had CNS-directed radiotherapy, defined as whole brain radiotherapy, spine radiotherapy, or both, as a component of their initial CNS metastatic disease treatment but did not have pCSI, and those who did not have radiotherapy. For these samples, the mean VAF at the time of diagnosis prior to intervention was calculated as specified previously.

We estimated overall survival (OS), defined as the time from the start of pCSI to death or last follow-up using the Kaplan–Meier method. Analysis was performed to find associations between VAF, CCF, and OS. CNS response and progression of disease was recorded per protocol using MRI and clinical assessment, and we also analyzed CNS progression-free survival (PFS), defined as the time from start of pCSI to CNS progression, death, or last follow-up (2). Variables were either dichotomized by the median value or, for measures of change pre- and post-pCSI, a value close to zero was chosen to include patients with decreased and essentially no change. For the comparison cohort, OS was calculated as the time from initial therapy after diagnosis to death or last follow-up. Survival estimates are shown with 95% confidence intervals (CI). Unless otherwise specified, significance was assessed at alpha = 0.05.

Written informed consent to use patient samples was obtained from all participants in adherence with the Memorial Sloan Kettering Cancer Center (MSKCC) Institutional Review Board. This study was conducted in accordance with the ethical guidelines outlined in the Belmont Report.

Raw data for this study were generated at the MSKCC Integrated Genomic Operation. Derived data supporting the findings of this study are available from the corresponding author upon request. Other data generated in this study are available within the article and its Supplementary Data files.

The patient characteristics are shown in Table 1. The median patient age was 53 years. The majority of patients (64%) had non–small cell lung cancer (NSCLC). All patients received at least 1 systemic therapy since cancer diagnosis, and most patients were heavily pretreated with the median number of lines of pre-CSI chemotherapy as 4 (range 1–9). At the study end, 79% (n = 11) of patients had died and 79% (n = 11) had CNS progression of disease. No patients had CNS progression within 3 months of pCSI, and though 6 patients had systemic progression before CNS progression, only 3 were within 3 months of the pCSI. The CNS and systemic disease status is shown in Supplementary Table S3, and most patients had stable systemic disease limited to bone metastases, pulmonary nodules, or no detectable lesions. The median OS for the cohort was 9 months [interquartile range (IQR), 5–21 months, Supplementary Fig. S1].

Of the 14 patients with pre-pCSI CSF ctDNA measured, 13 had a post-CSI CSF ctDNA measurement taken at 1-month post-pCSI, and 1 patient had a measurement taken 3 months post-pCSI. We observed changes in genetic aberrations occurring in ctDNA in the CSF after pCSI compared with the baseline (Fig. 1A), but there was no significant change in the overall number of observed mutations after pCSI (paired Wilcoxon signed rank P = 0.66; Fig. 1B). A nonsignificant increase in the number of patients with altered GNAS, PIK3C2G, and NOTCH2 was observed after pCSI (Fig. 1C). In patients with available data, we observed variants in ctDNA pre- and post-pCSI that can be classified as those that were: (i) present in the primary tumor or metastatic tumor (Supplementary Fig. S2A), (ii) present in ctDNA but were not observed in the primary or metastases (Supplementary Fig. S2B), (iii) present before pCSI but not detectable afterwards (Supplementary Fig. S2C), or (iv) only detectable after pCSI (Supplementary Fig. S2D). Using GSEA analysis of ctDNA, we found that many altered genes were involved in transcription regulation such as the RUNX1 pathway, DNA damage repair, and cell-cycle regulation, and there appears to be enrichment for altered transcription pre- and post-pCSI (Supplementary Table S4). GNAS, encoding the G-protein alpha subunit, was the only recurrently mutated gene after pCSI, occurring in 2 patients.

We next compared mutations detected in the plasma and CSF ctDNA in pre-pCSI and post-pCSI samples. The purity of the ctDNA samples, determined using FACETS, from the plasma was significantly different than that of the CSF and typically lower (Wilcoxon rank sum P = 0.001; Supplementary Fig. S3). The kappa statistic pre-pCSI comparing mutations found in the CSF and plasma was 0.255 (P < 0.001) indicating fair agreement. Conversely, the kappa statistic post-pCSI comparing mutations found in the CSF and plasma was 0.005 (P = 0.92) indicating no agreement. There were no differences in the distribution of clonal and subclonal alterations in CSF ctDNA pre- and post-pCSI, Wilcoxon paired signed rank tests for clonal mutations, P = 0.432, indeterminate mutations, P = 0.779, and subclonal, P = 0.922 (Supplementary Fig. S4). While some patients showed a large number of shared mutations, there were a considerable number of mutations that were unique to the CSF ctDNA (Supplementary Fig. S5). An increased number of mutations unique to the CSF at 1-month post-pCSI was associated with worse OS [HR, 1.24 per mutation (95% CI, 1.05–1.45); P = 0.009].

We then examined whether the mean VAF before and after pCSI was associated with survival outcomes. Having a higher VAF_pre when dichotomized by the median (VAF_pre = 0.32) was associated with worse OS [median 6 months (95% CI, 5 months–not estimated (NE) for VAF_pre ≥ 0.32 vs. 9 months (95% CI, 9 months–NE) for VAF_pre < 0.32; P = 0.05; Figure 2A]. However, VAF_post (when dichotomized by the median, VAF_post = 0.35) was not associated with differences in OS [median 8 months (95% CI, 5 months–NE) for VAF_post ≥ 0.35 vs. 9 months (95% CI, 6 months–NE) for VAF_post < 0.35; P = 0.9]. We instead considered whether Δ_VAF, which increased for 5 patients and decreased or remained stable for 9 patients (Fig. 2B) could predict survival outcomes. Patients with an increased Δ_VAF after pCSI had significantly worse OS [median 6 months (95% CI, 5 months–NE) for Δ_VAF ≥ 0.01 vs. 18 months (95% CI, 9 months–NE) for Δ_VAF < 0.01; P = 0.008; Fig. 2C]. Similarly, allele fractions incorporating clonality also predicted response to pCSI in LM patients: CCF_SUBCLONAL greater than the median was associated with worse OS [median 8 months (95% CI, 5 months–NE) for CCF_SUBCLONAL ≥ 0.61 vs. 17 months (95% CI, 6 months–NE) for CCF_SUBCLONAL < 0.61; P = 0.05; Fig. 2D]. We did not find significant associations with CNS PFS.

We present the putative framework for clonal evolution observed in ctDNA (Fig. 3A) and for three patients, clonal evolution from a solid tumor to LM is shown along with treatment history in Fig. 3B–D. For patient 10, who had a relatively poor survival (OS = 8.5 months), we detected the emergence of multiple alterations after pCSI (Fig. 3B). In contrast, for patient 5, who had extended survival (OS = 28 months), we did not appreciate a substantial increase in detected alterations after pCSI (Fig. 3C). For patients with analyzable ctDNA in both the plasma and the CSF, we examined clonal evolution and show an example of a patient with poor survival (OS = 3.6 months) in Fig. 3D, to demonstrate that alterations can be shared between ctDNA in the plasma and CSF or unique to each compartment. Through this approach, it is possible to test whether pCSI provides a selection pressure for clonal evolution. While there are some alterations unique to each compartment, we observe that only the CSF demonstrated an increase in observed alterations, some of which are clonal mutations (SESN3, BRD4, CCNE1, GNAS) and CNA with loss of EGFR and FGFR.

We identified 50 patients with CSF MSK-IMPACT performed at CNS metastasis diagnosis. Within this cohort, 60% (n = 30) received non-pCSI CNS radiotherapy; 40% (n = 20) did not receive radiotherapy. Like our study patient population, the comparison cohort consisted primarily of lung (42%, n = 21) and breast cancers (42%, n = 21). There was no difference in median OS for the patients who received radiotherapy versus those who did not (P = 0.70). Among patients who received radiotherapy, lower CSF VAF at baseline was associated with improved OS [median 8 months (95% CI, 4 months–15 months) vs. 9 months (95% CI, 7 months–NE) in with lower CSF VAF at baseline; P = 0.05], while no difference in OS by CSF VAF was seen in patients who did not receive RT (P = 0.10; Fig. 4). Therefore, pretreatment CSF VAF may be a predictor of survival in patients with solid tumor CNS metastases who receive radiotherapy.

In this study, we examined the genetic variation of ctDNA before and after pCSI in patients with LM in a subset of a patients who were treated on a phase I trial of pCSI for solid tumor LM. We found that the pre- and post-pCSI VAF and the change in VAF may have utility as a clinical biomarker. We also provided evidence that pCSI may exert a selective pressure in the CSF compartment potentially inducing genomic changes that are specific to LM treated with pCSI. To our knowledge, this is the first study to evaluate CSF ctDNA alterations associated with pCSI for LM. Though we detected associations with CSF ctDNA VAF_pre, Δ_VAF, CCF_SUBCLONAL, and OS, we did not find any association between these measures and CNS PFS. We have previously shown that pCSI is associated with improved CNS PFS and OS in patients with LM (3), and we are actively pursuing biomarkers that may be associated with both outcomes (11). It is possible that evaluating genetic changes detected in the CSF at 1 month for most patients was too early for detectable associations with CNS PFS, and a larger study with additional, regularly obtained ctDNA measurements may be needed.

For patients with CNS metastases, it is not currently standard of care to incorporate detectable genetic aberration, especially for variants of unknown significance, into clinical decision making for CNS-directed radiotherapy. In this study, we showed that the VAF is a simple measurement that can be applied clinically to stratify patient response to pCSI and target high-risk patients for earlier intervention after pCSI. The ability to calculate VAF likely reflects both the tumor burden and the degree of tumor heterogeneity, and we showed the higher pre-pCSI VAF as well as an increasing change in VAF after pCSI are associated with worse OS. Our results are consistent with reports demonstrating that a higher tumor burden in the CNS is associated with worse outcomes in LM (24).

Because the VAF does not take into account the clonality of mutations in ctDNA or the purity of the cfDNA sample, we also demonstrated that CCF calculated only using subclonal mutations is a significant predictor of survival after pCSI for LM in a cohort of patients with multiple types of cancer, supporting the robustness of ctDNA mutational analysis in this context. Given our limited sample size and heterogeneous cohort, we were not powered to discriminate between small differences in the CCF of clonal mutations, which generally only had values greater than 80%. Future studies will likely be able to determine whether there is biologic significance to altered clonal mutations before and after pCSI, and it is possible that this effect will be different by histology, requiring a stratified analysis. In addition, we compared our findings with a cohort of patients with CNS metastases who did not receive pCSI and found that VAF of CSF ctDNA was associated with OS after radiotherapy, indicating potential broad application for response in patients with CNS metastasis undergoing radiotherapy, although larger validation cohorts with statistical analysis controlling for potential confounders are needed to understand the relationship between CSF ctDNA VAF and OS after radiotherapy for CNS metastasis.

Targeted exome sequencing revealed low tumor purity and clonal architecture that could not be assessed in plasma ctDNA samples compared with CSF ctDNA. Although alternative methods have been demonstrated to have better efficacy in sequencing low purity plasma ctDNA using a hybridization capture of a smaller panel of genes and deep sequencing using techniques to suppress background sequencing errors (25), the high purity of ctDNA in the CSF may overcome the known limitations of bulk sequencing of heterogeneous solid tumors or competing cfDNA in the plasma (16). In a cohort of patients with CNS tumors of which 3 had LM, CSF ctDNA was shown to have a greater mutational diversity compared with plasma ctDNA (26). Similarly, in patients with brain metastases, mutational diversity of CSF ctDNA was observed to be distinct from tumor primaries and plasma ctDNA (12, 25, 27). In this context, CSF ctDNA may be an ideal, highly sensitive method to detect subclonal mutations. We show that using the exploratory CCF estimates of subclonal mutations is also associated with OS.

We detected substantial genetic mutational and copy-number diversity in the CSF of patients with LM, even in a limited cohort. By tracing the presence of mutations from primary/metastatic tissue to ctDNA (Fig. 3A), we found that many of the observed mutations in ctDNA were present in preceding solid tumor tissue, and we showed that the presence of more alterations before pCSI was associated with worse OS. Interestingly, while we observe that some variants are not detectable after pCSI, some patients have an increased genetic diversity after treatment, and larger increases in detected alterations are associated with poorer survival after pCSI (Fig. 3B and C). We also demonstrated that there are genetic variants that are unique to the CSF ctDNA, consistent with prior findings in patients with LM showing CNS-limited single nucleotide variants (refs. 16, 28; Fig. 3D). Overall, our findings suggest that there could be mutations that occur in response to pCSI, and radiation is known to lead to genetic variation through several mechanisms (29). While alterations to ctDNA after pCSI are expected, it remains unclear why some patients would develop new clonal and subclonal mutations and others would have decreased alterations after pCSI. This difference may be related to defective DNA damage response (DDR) pathways in some tumors, which may also be a risk factor for poor patient survival. Many of the detected mutations after pCSI involve genes implicated in DDR, cell-cycle control, and radiosensitivity, such as ARID1B (30), EPHB2 (31), BRD4 (32), CCNE1 (33), NKX2–2 (34), RTEL1 (35), and CDKN2A (36).

There are several possible sources of observed genetic variation that may correlate with outcomes independently of pCSI. Because patients in this study, and patients with LM in general, have often had multiple systemic therapies, it is likely that the observed mutations reflect prior selective pressures. While the baseline ctDNA would reflect the selective pressures of prior treatment, which we would also expect to vary by histology, this cohort represents a heterogeneous sample of patients with LM. A larger, stratified cohort could help distinguish the contribution of prior systemic therapy from the unique metastatic and leptomeningeal genetic signals. Given that the bulk of our VAF measurements were pre-pCSI compared with 1 month post-pCSI, the changes have a higher probability of being related to the pCSI than to the ongoing systemic therapy, which was held until 1month post-pCSI. In addition, the observed ctDNA in the plasma and CSF may reflect systemic and CNS disease, respectively, but the possibility of ctDNA crossover remains. Though the blood brain barrier is thought to minimize plasma ctDNA transfer to the CSF (37), CNS confined disease was demonstrated to have some ctDNA circulating in the periphery at relatively low levels (38). In the current study, only 3 patients had systemic disease progression before CNS progression within 3 months of pCSI; 2 had a decrease in CNS Δ_VAF and 1 showed an increase, indicating that the presence of systemic disease is not likely biasing our findings.

There are several limitations to this study. This study is exploratory in nature with a small cohort, so we did not perform correction for multiple comparisons; thus, all specific cut points, thresholds, and biomarkers may be unique to this cohort and, therefore, require further study. Though only a subset of the trial patients underwent both pre- and post-pCSI ctDNA measurements, the median OS was similar between this cohort and the full phase I trial cohort (2), and a larger cohort (8) of patients treated with pCSI for LM. The sample set was too small to detect many common mutations between samples; we overcame this limitation by focusing on summary measures and pathway level findings. We also chose mutation calling parameters that were more permissive, while still requiring sufficient evidence for mutation calls, to increase our mutation calling sensitivity. This allowed us to broaden our analysis of a limited number of patients at the expense of specificity, and we performed manual validation of called variants to increase our confidence in mutation calls. However, as a result, the presence or absence of an alteration cannot exclude false positives or lack of coverage, respectively. Lastly, we reported the CCF as determined using FACETS, but CCF is still an exploratory method, and the clinical utility and interpretation of CCF requires additional study.

Proton CSI is an emerging and promising therapy for patients with LM. Analysis of ctDNA from patients with LM before and after pCSI shows that pCSI efficacy may depend on genomic characteristics before treatment and may exert genomic selection after treatment. We found that patients with improved outcomes are associated with lower frequency of mutations, and in particular, a lower mean VAF. We showed that the quantification of these mutations has important implications for patient stratification in CNS-directed radiotherapy.

M.F. Berger reports personal fees from Eli Lilly and AstraZeneca outside the submitted work. D.S. Higginson reports grants from SQZ Biotechnologies, Natera, AstraZeneca, and NCI outside the submitted work. L. Schaff reports personal fees from DebioPharm outside the submitted work. S.N. Powell reports grants from Varian Medical Systems and Philps outside the submitted work; and personal fees from Rain Therapeutics. A. Boire reports grants from NCI and Pew Charitable Trusts during the conduct of the study and grants from Keck Foundation, Center for Experimental Therapeutics, Pershing Square Sohn Cancer Research Foundation, and NCI outside the submitted work; in addition, A. Boire has a patent for United States Provisional Application No.: 62/258,044. November 20, 2015. pending to Sloan Kettering Institute and a patent for United States Provisional Application No.: 63/052,139 pending to Sloan Kettering Institute. J.T. Yang reports grants from Cycle for Survival Equinox Innovation Initiative Award during the conduct of the study; personal fees from Nanocan Therapeutics, Debiopharm Group, and Galera Therapeutics; personal fees and other support from AstraZeneca; and other support from Kazia Therapeutics and X-RAD Therapeutics outside the submitted work. No disclosures were reported by the other authors.

N.A. Wijetunga: Conceptualization, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. A.G. Goglia: Data curation, formal analysis. N. Weinhold: Supervision. M.F. Berger: Data curation, supervision, writing–review and editing. M. Cislo: Data curation, project administration. D.S. Higginson: Supervision, writing–review and editing. K. Chabot: Data curation, project administration. A.M. Osman: Data curation, formal analysis, writing–review and editing. L. Schaff: Writing–review and editing. E. Pentsova: Conceptualization, investigation, writing–review and editing. A.M. Miller: Supervision, investigation, writing–review and editing. S.N. Powell: Conceptualization, supervision, investigation. A. Boire: Conceptualization, formal analysis, supervision, funding acquisition, investigation, methodology, writing–original draft, writing–review and editing. J.T. Yang: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing.

This work was supported by grants from Cycle for Survival Equinox Innovation Initiative Award (to J.T. Yang) and the NIH/NCI (P30-CA008748; to Sloan Kettering Institute, and R01 CA245499; to A. Boire).

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.