Sequential tumor molecular profiling identifies likely germline variants

Abstract

Purpose:

To identify likely germline DNA variants from sequential tumor profiling data from hematopoietic malignancies (HMs).

Methods:

The coefficient of variance was calculated from variant allele frequency of next generation sequencing assays. Variants’ likelihood of being germline was ranked on a 1-5 scale. Outcomes were examined in patients with such variants.

Results:

In a pilot set of 33 genes, 89% of grade 1, 77% of grade 2, 62% of grade 3, 52% of grade 4, and 21% of grade 5 variants were confirmed to be germline. Among those, 22% were pathogenic or likely pathogenic in genes recognized as conferring hereditary HM (HHM) risk, including BRCA1/2, CHEK2, CSF3R, and DDX41. To determine if this approach identified genes with known autosomal dominant inheritance, we analyzed sequential data from 1336 genes in 1135 HM patients. Among unique variants, 16% occurred in HHM genes, and 15% were deleterious. Patients with grade 1/2 alleles had decreased survival two years following initial molecular testing (78% versus 88%, p=0.0037) and increased all-cause mortality compared to those without (HR 2.02, 95%CI 1.18–3.46, p=0.01).

Conclusion:

Variant germline status may be predicted using sequential tumor profiling and patients with likely germline variants experience inferior outcomes compared to those without.

INTRODUCTION

Clinical next generation sequencing (NGS) is standard of care for risk stratification and surveillance of hematopoietic malignancies (HMs), yet maximizing utility of these data remains challenging¹. NGS data from malignant cells contribute to disease diagnosis and prognosis, while also defining clonal sub-populations. However, most clinical applications rely on data from a single time point². Determining whether a DNA variant detected by tumor profiling is germline or somatic can be difficult³. Variants allele frequencies (VAFs) between 0.3-0.7 that persist over time suggest germline status, especially when a patient fluctuates in and out of remission³. We sought to develop a quantitative method of tracking VAFs in sequential samples to identify those most likely to be germline to allow prioritization of patients for genetic counseling and testing.

MATERIALS AND METHODS

We analyzed data from OncoPlus, a targeted-capture NGS assay comprising 1336 genes⁴, collected from March 2016 to December 2021, and analyzed on an IRB-approved study (Supplemental Methods). DNA variants were quantified using the coefficient of variation (CV) to characterize VAF changes in sequential samples⁵. CV was computed by calculating the standard deviation of the VAFs, divided by the mean VAF over time:

$${\sigma }^2=\frac{\sum {(\chi -\mu )}^2}{N}$$

$$CV=\frac{\sigma }{\mu }$$

where $\sigma$ is the standard deviation of VAFs within an individual patient; $\chi$ is the individual VAF; and $\mu$ is the mean of the VAFs. Low CV numbers reflect less, and high CV numbers reflect more variability between sequential tests. Using colony stimulating factor 3 receptor (CSF3R) variants, we hypothesized that CV≤0.1 suggested germline status (Supplemental Figure 1A-B)⁶. We then applied this CV cutoff to variants in 33 genes associated with inherited HM risk and tested germline status with DNA derived from cultured skin fibroblasts (Supplemental Table 1)⁷. We optimized this prediction algorithm by comparing variants within patients who had at least one variant with low CV and another variant with CV>0.1. We filtered out common polymorphisms by excluding those with population frequencies ≥ 0.0001 in the gnomAD v2.1.1 database⁸. We then analyzed OncoPlus data from malignant cells, curating variants according to American College of Medical Genetics and Genomics/Association of Molecular Pathology standards using ClinVar and Varsome⁹. Time-to-event data were analyzed using Kaplan-Meier and Cox proportional hazards methods. Data analysis was performed using R (v4.2.1), with statistical significance, p<0.05.

RESULTS

We hypothesized that some variants with VAFs 0.3–0.7 on recurrent NGS tumor profiling are germline and could be identified by calculating the CV (Figure 1A-B)⁶. To test this, we examined 1364 HM patients in whom 2132 unique variants in 33 genes associated with hereditary HMs (HHMs) were identified (Supplemental Table 1). Unique recurrent variants (n=708; 33% of all variants detected) on more than one test were identified in 26% (352/1364) of patients. Using CV≤0.1, VAF≥0.3, and population frequency ≤0.0001, we identified 19% (265/1364) of patients with 1000 (0.06% of total, 374 unique) recurrent variants (Figure 1C and Supplemental Table 2). The germline nature of each was determined by augmented exome sequencing of DNA from cultured skin fibroblasts in 143 (54%) individuals (Figure 1C and Supplemental Table 2). To rank the accuracy of prediction, we graded each variant using a 1-5 scale (Figure 1D and Supplemental Methods, Supplemental Figure 2A), with 1 being most and 5 least likely to be germline. Overall, 89% (48/54) of grade 1, 77% (50/65) of grade 2, 62% (23/37) of grade 3, 52% (12/23) of grade 4, and 21% (18/87) of grade 5 variants were confirmed to be germline alleles (Figure 1D, Supplemental Table 2, and Supplemental Figure 2A). A receiver operator curve, comparing variant grades to confirmed germline status, demonstrated an area under the curve of 0.82 (95% CI 0.77-0.87) (Supplemental Figure 2B). Among unique confirmed germline variants, 22% (33/151) were pathogenic (P) or likely pathogenic (LP; Figure 1E and Supplemental Table 2).

Figure 1. CV identifies gene variants likely to be of germline origin.

(A) Variant allele frequency over time for an individual patient. A 66-year-old woman with myelodysplastic syndrome presented with fatigue and was found to have a white blood cell count of 1400/μL, Hb 9.3 g/dL, and platelets 112,000/μL. A bone marrow biopsy demonstrated a hypercellular bone marrow with 29% blasts, consistent with transformation to AML. Time point (1) represents NGS results at AML diagnosis. The patient received chemotherapy with CPX-351 (liposomal encapsulation of cytarabine and daunorubicin). At time point (2), the patient was noted to have recurrent disease and started treatment with decitabine and venetoclax. The patient’s course was complicated by delayed count recovery after each cycle. Time point (3) represents progressive disease, and the patient started treatment with ivosidenib due to the acquisition of the isocitrate dehydrogenase 1 (IDH1) c.395G>A p.(Arg132His) variant at VAF 6%. At time point (4), the patient had progressive disease and elected for hospice care. (B) Formulas to calculate variance and coefficient of variation. Calculated values are shown for the CSF3R c.1213G>A p.(Glu405Lys) allele germline variant compared to the IDH1 c.395G>A p.(R132H) somatic variant. (C) Filtering algorithm for germline validation showing patients with unique potential germline variants. (D) Normalized bar graph of variant etiology by grade: 48/54 (89%) were grade 1, 50/65 (77%) grade 2, 23/37 (62%) grade 3, 12/23 (52%) grade 4, and 18/87 (21%) grade 5. (E) Waterfall plot of the number of DNA variants in each gene within the validation cohort. Colors correspond to ClinVar or Varsome predicted pathogenicity: red, P/LP; green, variant of uncertain significance; and purple, benign or likely benign. Abbreviations used: CV, coefficient of variation; μ, mean of variant allele frequency for an individual patient; σ, standard deviation of variant allele frequency for an individual patient.

To test if this unsupervised grading scale enriched for HHM gene variants, we analyzed NGS data from HM cells in 1336 genes. We found that 23% (260/1135) of patients had grade 1 (14.9%) or 2 (12.0%) variants (Supplemental Table 3). Among unique variants, 16% (137/840) occurred in HHM genes^3,10, including TET methylcytosine dioxygenase 2 (TET2), Fanconi Anemia complementation group A (FANCA), DEAD-box helicase 41 (DDX41), and ATM serine/threonine kinase (ATM) (Figure 2A and Supplemental Figure 3A). Additionally, we identified variants in genes not currently associated with HHMs, such as FAT atypical cadherin 3 (FAT3), Patched 1 (PTCH1), TSC complex subunit 2 (TSC2), ALF transcription elongation factor 3 (AFF3), ERB-B2 receptor tyrosine kinase 3 (ERBB3), and ERB-B2 receptor tyrosine kinase 4 (ERBB4) (Figure 2A). Among unique grade 1/2 variants in autosomal dominant (AD) HHM genes, 15% (20/137) were classified as P/LP (Supplemental Table 3).

Figure 2: Outcome analysis of patients with potential germline variants.

(A) Waterfall plot of grade 1/2 variants for each gene analyzed. Colors indicate the recognized association of grade 1/2 variants with germline cancer inheritance¹⁰ : red, HHM risk; blue, HHM risk and somatic variants; green, somatic variants; and purple, unknown association with HMs. (B) Cumulative hazards modeling of patient age by variant grade. Likelihood of germline status was graded as: grade 1/2 (most likely to be germline; green line) versus grade 1/2 variants in genes known to confer AD cancer risk (red line) versus grade 3/4/5 variants (dark blue line). HR = 2.02 (95% CI 1.18–3.46, p = 0.01). Abbreviations used: CI, confidence interval; HR, hazard ratio.

Finally, to determine if there are outcome differences in patients with likely germline variants, we compared overall survival of patients with grade 1/2 HHM gene variants, patients with grade 1/2 variants not in HHM genes, and patients with only grade 3/4/5 variants (Supplemental Table 3)¹⁰. The two-year overall survival was 78% (95% CI 68.2%–88.6%) for patients with grade 1/2 HHM gene variants, 88% (95% CI: 81.9%–95.2%) with grade 1/2 variants not in HHM genes, and 92% (95% CI: 89.9%-94.3%) with only grade 3/4/5 variants (Supplemental Figure 3B). Patients with deleterious grade 1/2 HHM variants trended toward inferior outcomes (Supplemental Figure 3C). Finally, we observed increased risk of all-cause mortality in patients with a grade 1/2 HHM gene variant compared to those without (HR 2.02, 95% CI 1.18–3.46, p = 0.01) (Figure 2B).

DISCUSSION

The identification of deleterious germline variants can be challenging due to the lack of systematic germline testing and the need to recognize genes associated with germline predisposition^11,12. We created an unsupervised approach to identify potential germline alleles using the CV for individual variants over multiple tests³. This method identifies individuals who should be prioritized for germline genetic counseling and testing of deleterious variants and identifies novel genes that might confer AD germline cancer predisposition. Many potential germline variants identified were classified as variants of uncertain significance, highlighting the need for research efforts to determine their functionality and contribution to disease pathogenesis¹².

Routinely revisiting prior patient NGS data throughout clinical care helps to suggest germline etiology beyond that collected at static time points. In this study, we performed our analysis using a retrospective cohort. However, in clinical practice we recalculate the CV for each variant on an iterative basis as new NGS data become available for each patient, thus, actively comparing newly acquired data to a pool of previously reported variants and flagging patients with likely germline variants^7,13. Pathology laboratories could apply longitudinal data analysis systematically by incorporating the CV and/or germline grading algorithm for each variant into clinical bioinformatic variant callers.

Our methodology has some important limitations. First, our method requires at least two NGS tests separated in time, limiting the ability to identify potentially unknown germline variants on initial diagnostic profiling. The area under the curve using CV of an individual variant was limited in its ability to predict germline etiology reliably, so we included a comparison of other variants within an individual patient to improve our prediction accuracy creating the germline grading scale incorporating information from both the variant of interest and other variants in patient samples (Supplemental Methods). We used CV values of 0.1, 0.03, and 0.05 to define our grading scale based on our single center, derivation cohort. Although lower CV values correlated with germline status, specific cutoffs used to create a grading scale were ultimately arbitrary but necessary to incorporate information from multiple variants within a single individual. Variants with grade ≥ 3 could be of germline etiology, mandating individual review (Supplemental Table 2). Additionally, variants associated with clonal hematopoiesis, variants in genes common to chronic myeloid neoplasms such as Janus kinase 2 (JAK2), and variants associated with targeted therapies may have similar VAFs when NGS tests are repeated at short intervals, yielding a low CV and germline grade (Supplemental Figure 3D)^14,15. Pathogenic germline variants in genes such as sterile alpha motif domain containing 9 (SAMD9) and sterile alpha motif domain containing 9 like (SAMD9L) may be present below our VAF cutoff of 0.3 and subsequently missed by our algorithm. This method relies on stable copy number variations within tumor cells. Among the 12 patients (18 variants), with grade 5 predictions, all had complex karyotypes with additions, deletions, or loss of heterozygosity in regions of the variant and/or received an allogeneic hematopoietic cell transplant (HCT) during the analysis period (Supplemental Table 4). Finally, 11% (6/54) grade 1 variants were somatic, highlighting the importance of confirmatory testing (Figure 1D).

The prognostic implication in patients with AD germline P/LP alleles should be interpreted based on the gene identified^10,16-19. Our univariate, Kaplan-Meier analysis of OS suggests that P/LP, AD germline alleles may be associated with decreased overall survival among patients with HMs, raising the question if such individuals should be treated differently than those without such alleles. However, patient comorbidities, myeloid versus lymphoid disease-specific factors, therapy toxicity, and potential radiation exposure may confound these data, and we were unable to perform a multivariate analysis due to lack to inclusion in the clinical database. Further, although the population treated at the University of Chicago Comprehensive Cancer Center reflects a wide range of individuals from varied genetic ancestry, we were unable to account for polymorphisms that are associated with groups of patients experiencing disparity in the US healthcare system. Additionally, we used the time of first NGS test as Day 0 suggesting that this is a surrogate for patient presentation at our tertiary center. However, patients may have been treated prior to presentation at other institutions and dates of disease diagnosis were not available in our database. Like the observation that patients with germline P/LP DDX41 variants transplanted with wild-type donors may have increased incidents of severe graft-versus-host disease, these data suggest that outcomes might differ, contributing to hypothesis generation, but do not provide mechanistic or further insight into pathogenicity²⁰. In fact, there remains a paucity of data regarding isolated germline contribution towards disease, risk-adapted upfront therapy in patients with germline predisposition, optimal timing of allogeneic HCT, and risk for therapy-induced complications. Subsequent prospective, targeted studies are needed to answer these questions.

DATA AVAILABILITY

R scripts have been uploaded to the Godley Laboratory github repository (https://github.com/LucyGodley/SequentialTumorProfiling). Clinical sequencing data are available upon request.