Introduction

Newborn screening is one of the most successful public health initiatives in modern medicine and has led to early diagnosis and improved outcomes for many serious pediatric conditions. Current research efforts are focused on incorporating primary DNA sequencing into newborn screening, allowing detection of disorders not identified through traditional biochemical approaches1,2,3,4,5,6. Cancer predisposition syndromes, caused by deleterious germline (or germline mosaic) variants in cancer-associated genes, are present in at least 10% of children with malignancy7,8,9. Genomic newborn screening could identify these at-risk children, which may translate into patient benefit through syndrome-specific cancer surveillance protocols10,11. For example, approximately 40% of children with retinoblastoma have a germline heterozygous deleterious variant in the RB1 gene, and over 90% of these children develop bilateral retinal tumors by age three12. In known RB1 carriers, early ophthalmologic surveillance promotes detection of small subclinical tumors that have favorable visual and survival outcomes after treatment with localized therapies13. However, most children with this genetic cancer predisposition syndrome go undetected until they develop symptomatic intraocular tumors; these (usually bilateral) tumors require aggressive treatment, including ocular enucleation, chemotherapy and/or radiation. Survivors often have significant visual impairment and are at risk for acute and long-term treatment-related toxicity14. Similarly, in children with other genetic cancer predisposition syndromes, such as those associated with pathogenic germline variants in WT1, TP53, and RET, presymptomatic clinical care is focused on prevention or early cancer detection through application of syndrome-specific guidelines focused on improving survival and reducing exposure to treatment-related morbidity10,11,15.

Prospective genomic newborn screening research is underway worldwide1,2,3, but would require very large population-based studies and extended follow-up to prospectively study the impact of the inclusion of cancer genes on outcomes. In the absence of prospective data, we previously developed a simulation model which projected an 8% reduction in overall childhood cancer mortality through adoption of an 11-gene cancer predisposition panel into population-based newborn screening. The 11 genes were selected for their association with increased risk of specific pediatric cancers, the availability of gene-specific cancer surveillance guidelines, and the potential benefit of early detection (See Table 1)16.

Table 1 Genetic Predisposition Syndromes Included in Study, Their Association with Early Onset Childhood Cancers, Surveillance Recommendations, and Potential Impact of Early Detection
Full size table

In the current study, we took advantage of linkage between Michigan’s Cancer Surveillance and Newborn Screening Programs to further examine the role that genomic newborn screening might have in detecting infants at risk for early- onset cancers. We identified, from Michigan births between 1987−2020 the children who developed a solid or central nervous system malignancy by age 8 years whose archived newborn dried blood spots were available for research. We used targeted next-generation sequencing to estimate the prevalence, among these children with early-onset cancers, of pathogenic or likely-pathogenic germline variants in 11 autosomal dominant cancer risk genes. We compared the prevalence of deleterious variants to their prevalence both in internal control groups and external cohorts. Using cancer registry data, we evaluated clinical features of cancers in germline carriers which might be impacted by early detection.

Here, we show that targeted sequencing of a panel of cancer predisposition genes would identify approximately 1/27000 infants who will develop cancer by age 8 in the setting of a genetic cancer predisposition syndrome. Our data suggest that 80% of children who will develop bilateral retinoblastoma could be identified at birth using this technology.

Results

Cohort development and sequencing analysis

Overall, 1948 subjects met criteria for inclusion in the study cohort. Clinical and demographic characteristics of the overall cohort are shown in Table 2. The distribution of cancer types in the overall cohort is shown in Fig. 1a. Archived newborn dried blood spot samples were obtained from all subjects, with adequate DNA extracted from 1943 of 1948 (99.7%) samples. All 1943 samples met quality standards for variant and CNV analysis.

Fig. 1: Cancer cohorts by diagnosis and detected germline variant counts.
Fig. 1: Cancer cohorts by diagnosis and detected germline variant counts.
Full size image

Panels A, B depict the distribution of cancer types in the overall cohort (A, n = 1948) and in the cases with a detected pathogenic or likely-pathogenic germline variant in a targeted gene (B, n = 132). Panel C shows the number of pathogenic or likely-pathogenic variants detected in each targeted gene (except for ALK, which had no detected pathogenic or likely-pathogenic variants detected). Bar colors represent cancer type. Sarcoma cases include rhabdomyosarcoma, osteosarcoma, and other sarcomas; renal tumor cases include Wilms and other renal tumors; CNS tumors include pineoblastoma, medulloblastoma, choroid plexus carcinoma, and other brain tumors. (CNS Central Nervous System, tNGS targeted Next Generation Sequencing). Source data are available as a source data file.

Table 2 Characteristics of Cancer Cohort (N = 1948)
Full size table

Germline variant detection and characteristics of P/LP carriers

A pathogenic or likely-pathogenic germline variant in one of the targeted genes was detected in 132 (6.8%) subjects. The distribution of cancer diagnoses among the 132 germline variant carriers is shown in Fig. 1b. The number of pathogenic or likely-pathogenic variants detected in each targeted gene, and associated diagnoses, is shown in Fig. 1c. Variants were most often identified in RB1 (n = 69), TP53 (24), SMARCB1 (8), and WT1 (7). No pathogenic or likely-pathogenic germline variants were detected in the ALK gene. Variants included 17 large deletions and 115 single nucleotide variants, small insertions, or deletions. Notably, 130 of 132 variant carriers developed a tumor type with a closely established association with the targeted gene.

The prevalence of pathogenic or likely-pathogenic variants in specific cancer diagnostic groups is shown in Fig. 2. The 11-gene panel identified germline pathogenic or likely-pathogenic variants in at least 10% of cases for seven diagnoses: medullary thyroid carcinoma (100%), retinoblastoma (40%), choroid plexus carcinoma (30%), adrenocortical carcinoma (20%), pineoblastoma (17%), lung/pleural malignancy (11%) and medulloblastoma (11%), Based upon the 3,554,464 births in Michigan contributing to the cohort during the study period, approximately 1 in 27,000 newborns went on to develop an early-onset malignant solid or brain cancer in the presence of detected variant in a cancer predisposition gene.

Fig. 2: Prevalence of detected pathogenic or likely-pathogenic variants by cancer diagnosis with affected gene color-coded.
Fig. 2: Prevalence of detected pathogenic or likely-pathogenic variants by cancer diagnosis with affected gene color-coded.
Full size image

N = 1948 cancer cases overall and 132 cases with detected P/LP variants. Note change in range of Y-axis after first three diagnoses. (NOS not otherwise specified, P/LP pathogenic or likely-pathogenic). Source data are available as a source data file.

Genetic, demographic, and clinical findings in the 132 children with pathogenic or likely-pathogenic variants are depicted in Fig. 3 and detailed in Supplementary Dataset 1 and 2. The median age at cancer diagnosis in carriers was 14 months (IQR 5.0-26.5) compared with 32 months (IQR: 13.0–55.0) in those without detected germline variants (p (Wilcoxon) < 0.001). Among the 132 children with detected cancer predisposition syndromes, there was a slight male excess (ratio M:F of ~1.3); race/ethnicity was classified as White in 68.2%, Black in 6.1%, Hispanic in 3.8%, and unknown/suppressed in 22.0% of cases. Second cancer data were available from 126/132 (95%) of P/LP variant carriers. Eleven of these 126 carriers (8.7%) developed second cancers, compared to 85/1755 (4.8%) in non-carriers with available data. The odds of second cancer among P/LP variant carriers were 1.88 times that in non-carriers (OR = 1.88 (95% CI (asymptotic): 0.98, 3.62), Wald χ²₁ = 3.67, p = 0.06). Second cancers in carriers occurred in individuals with P/LP variants in RB1 (n = 5), TP53 (n = 4), APC (n = 1), and DICER1 (n = 1).

Fig. 3: Clinical and genomic variables in 132 cases detected.
Fig. 3: Clinical and genomic variables in 132 cases detected.
Full size image

Oncotype plot48 depicting the clinical and genomic characteristics of cases with detected cancer predisposition syndromes (n = 132). Data are sorted by age of cancer diagnosis and by gene. Each column represents an individual child, with corresponding clinical data at the top of the figure (cancer diagnosis, age at cancer diagnosis, vital status, and second cancer occurrence). Each row of the bottom part of the graph depicts the affected gene, with variant type for each genetic change color-coded. Percentage distribution of variant types for each gene is illustrated in a pie chart adjacent to the row for each gene. Sarcoma cases include rhabdomyosarcoma, osteosarcoma, and other sarcomas, and renal tumor cases include Wilms tumors and other renal tumors. (unk: unknown) Source data are available as a source data file.

RB1 and retinoblastoma predisposition

Pathogenic or likely-pathogenic variants in the RB1 gene were detected in 68/168 (40%) of children who developed retinoblastoma, and in one child who developed metastatic pineoblastoma, a rare pediatric brain tumor associated with germline RB1 pathogenic variants. There were no detected RB1 variants in the internal control group of 1774 infants who did not develop retinoblastoma or pineoblastoma (p (exact) <0.0001) (Supplementary Table 1).

Germline RB1 pathogenic and likely-pathogenic variants were detected in 41/51 (80%) of known bilateral retinoblastoma cases and 27/117 (23%) of the remaining retinoblastoma cases (with unilateral or unrecorded laterality). RB1 variant carriers developed retinoblastoma at a younger age compared to those without germline variants (9 months (IQR: 2-17) vs. 23 months (IQR: 11-35.5) (p (Wilcoxon) < 0.001)). No differences in stage distribution or primary therapy (data not shown) or overall survival (Supplementary Fig. 1a) were noted in the RB1 carriers compared to non-carriers. Among the 68 retinoblastoma cases with germline RB1 variants, treatment included chemotherapy (32), surgery (likely ocular enucleation) (40), or radiation therapy (3). Among individuals with retinoblastoma (excluding 11 with missing data regarding secondary cancers), secondary cancers occurred in 5/64 (7.8%) with germline P/LP RB1 variants compared with 4/93 (4.3%) without RB1 variants (OR = 1.89 (95% CI (exact): 0.49, 7.31), p (exact) = 0.488). Among the 100 retinoblastoma cases without identified germline RB1 variants, there were four deaths from retinoblastoma and no secondary cancer deaths. In the 68 cases with germline RB1 variants, there were six cancer-related deaths, three due to retinoblastoma and three after secondary cancers (bone tumor (2) and malignant melanoma (1)).

Other pediatric cancer predisposition syndromes

Among the 1780 children with cancers other than retinoblastoma, P/LP germline variants were detected in 64 (3.6%) cases. Median age at diagnosis was 22.5 months in these carriers (IQR 12.5-45.5 months) compared with 32 months (IQR: 13–56 months) in non-carriers (p (Wilcoxon) = 0.072). Regional or metastatic disease was recorded in 21 carriers (33%), and radiation was included in primary treatment in 12 (19%) carriers. No statistically significant differences in stage distribution or primary treatment type were noted between children with or without detected variants. In the 59 carriers for whom follow-up data were available, 21 were known to have died, 19 due to malignancy. Overall survival was lower in these variant carriers compared to non-carriers (p (log rank) = 0.0324) (Supplementary Fig. S1b).

Six children in the cohort developed medullary thyroid carcinoma; all six had germline RET pathogenic variants. There were no detected RET variants in the 1942 newborns who did not develop medullary thyroid cancer (p (exact) <0.0001). WT1 variants were detected in 7 infants who later developed renal tumors, and there were no WT1 variants detected in cases without renal tumors (p (exact) <0.0001) (Supplementary Table 1).

Pathogenic or likely-pathogenic germline variants in TP53, the gene associated with Li-Fraumeni Syndrome, were noted in 24 newborns, with cancer diagnoses occurring at a median age of 35 months (IQR: 19.5-48 months). Cancer diagnoses included sarcomas (10) and CNS tumors (9), as well as adrenal (3), renal (1), and liver (1) malignancies. Four second cancers and nine deaths due to malignancy were reported in the children with TP53 variants.

Germline variants in PTCH1 and SUFU, genes associated with Nevoid Basal Cell Carcinoma Syndrome, were detected in nine newborns who developed medulloblastoma and one with rhabdomyosarcoma. Germline variants in SMARCB1, associated with Rhabdoid Tumor Predisposition Syndrome, were detected in eight newborns, one with a neonatal kidney tumor and seven with brain tumors, occurring at a median age of three months (IQR: 0.0−8.5).

Comparison cohorts

We observed strong gene–tumor specificity, with variants in RB1, RET, and WT1 confined to subjects with associated tumors (p (exact) <0.001) (Supplementary Table 1). In total, only two variant carriers were detected in the internal control groups: one PHOX2B carrier (among 1498 subjects without a history of neuroblastoma, the tumor expected in PHOX2B predisposition syndromes) and one APC carrier (among 1873 subjects without a history of hepatoblastoma, the tumor expected in young children with APC-associated predisposition).

Data from 38,448 infants reported in four published population-based genomic newborn-screening studies2,4,5,6 identified two detected cancer predisposition-associated germline variants, one in RB12 and one in DICER15 (Supplementary Table 2). The variant prevalence data in our childhood cancer cohort were compared with data from the Genome Aggregation Database v.4.0 and showed that ClinVar pathogenic or likely-pathogenic variants in the 11 pediatric cancer-predisposition genes were exceptionally rare or absent in the general population (Supplementary Table 3 and Supplementary Fig. S2). Overall comparisons of the frequency of germline variants in our cancer cohort to the comparison cohorts are summarized in Supplementary Table 4.

Discussion

We performed a population-based investigation of newborn genomic sequencing focused on cancer predisposition syndromes. Using archived dried blood spots from newborns who later developed early-onset cancers, we demonstrated that sequencing a limited number of selected cancer-risk genes would identify approximately 7% of children who will develop a solid or brain malignancy by age 8 years. Overall, this screening would identify 1 in 27,000 newborns with a cancer predisposition syndrome and early-onset cancer, a prevalence consistent with, or even higher than, other diseases currently included in newborn screening, such as Pompe Disease (1/18,000 births), Severe Combined Immunodeficiency (SCID) (1/59,000 births), and Maple Syrup Urine Disease (1/200,000 births)17.

Our study has several notable strengths. First, we demonstrate the technical capability of population-based genomic screening using DNA extracted from a single dried blood spot punch, achieving successful library preparation and targeted sequencing in 99.7% of samples. Recent evaluations of DNA isolation protocols from dried blood spot punches have demonstrated reliable DNA recovery suitable for downstream next-generation sequencing across multiple platforms18. Our use of a high-throughput variant classification workflow further supports the technical robustness of our approach under conditions compatible with routine newborn screening programs. Second, by taking a phenotype-first approach—linking a statewide cancer registry to a neonatal dried blood spot biobank—we were able to directly assess the proportion of children whose cancer risk could have been predicted at birth and to evaluate the potential clinical impact of this detection: children with germline variants developed tumors at very young ages, underscoring the potential clinical value of early recognition of cancer predisposition. Third, we were able to assess the specificity of our approach, demonstrating the absence of off-target variants within our own cohort and the rarity of pathogenic and likely-pathogenic variants, in both published newborn (non-cancer) cohorts and in large genomic datasets derived from adult populations. Finally, unlike most studies reporting the prevalence of germline cancer predisposition in hospital-based pediatric cancer patients, our analysis uniquely provides population-based estimates of genetic risk among children with early-onset malignancy.

Our study has several limitations that also highlight priorities for future research. First, the ideal control group—a very large cohort of newborns known to remain cancer-free through early childhood, with biospecimens analyzed using identical genomic and analytic methods—was not feasible to assemble. Although emerging data from large newborn genomic screening studies demonstrate the rarity of pathogenic variants in cancer predisposition genes, these studies employ heterogeneous sequencing platforms, gene panels, and variant interpretation approaches, and study populations with different race and ethnicity2,4,5,6. Extrapolation from the Michigan birth cohort to populations with different ancestry, demographic structure, and/or potential founder variants should be made cautiously. Second, because clinical data available through the birth and cancer registries are limited, we were unable to assess the proportion of cases with clinically known cancer predisposition syndromes or a family history of cancer. Certain long-term outcomes, such as visual impairment following retinoblastoma or renal dysfunction after Wilms tumor, were also not measured. Nonetheless, earlier identification of at-risk infants would be expected to improve these outcomes through enhanced surveillance and earlier intervention. Third, analysis of cancer risk by race and ethnicity was limited by privacy-related suppression of demographic data and small subgroup sample sizes. Future studies with more complete demographic linkage and larger case numbers will be required to robustly evaluate demographic variation in cancer risk and variant penetrance. Finally, by excluding autosomal recessive cancer predisposition syndromes and leukemia predisposition genes, as well as restricting our analysis to 11 solid tumor genes, we likely underestimated the overall number of genetically at-risk newborns. Leukemia predisposition genes were excluded because the benefits of presymptomatic detection are less well established than for solid and brain malignancies. Future studies evaluating expanded panels will help define which cancer genes warrant consideration in newborn screening.

The economic implications of newborn screening for cancer risk were not evaluated in the current study. In our previous modeling study of universal newborn genetic screening for cancer risk16, we projected reductions in pediatric cancer mortality after implementation of universal newborn screening for cancer risk, with cost-effectiveness strongly influenced by sequencing costs. Given that sequencing costs continue to decline, and that genomic technologies will be increasingly incorporated into newborn screening programs, the yield and gene–tumor specificity we observe in the current study provide important data for further estimates of the economic and public health implications of genomic screening for early cancer risk. While surveillance strategies for individuals with cancer predisposition syndromes have been shown to enable early tumor detection, their cost-effectiveness likely varies by syndrome, by cancer risk and by screening modality. Potential risks of early identification should also be considered, including psychological distress for families, the burden of surveillance in young children, and the possibility of overdiagnosis or unnecessary interventions arising from indeterminate findings. Although surveillance is recommended for many cancer predisposition syndromes and can enable earlier tumor detection, implementation into clinical practice is variable, and at-risk infants may remain undiagnosed without health care systems supporting this intervention19.

Despite the identification of RB1 more than four decades ago, most children with bilateral retinoblastoma continue to present with locally advanced, vision-compromising disease. Similarly, most children with genetic cancer predisposition syndromes remain undiagnosed until symptomatic malignancies develop. Together, these observations underscore the potential clinical value of earlier identification through genomic newborn screening. Realizing this potential would require collaboration among oncologists, geneticists, newborn screening experts, and public health laboratories, along with careful attention to ethical implications. Beyond immediate clinical impact, newborn genomic screening for cancer risk would offer an unprecedented opportunity to advance research in pediatric cancer pathogenesis and prevention.

Methods

Human subjects research approval

This study was approved by the Dana-Farber/Harvard Cancer Center Institutional Review Board (Study #22-354), Michigan Department of Health and Human Services Institutional Review Board, the Michigan BioTrust for Health and Michigan Cancer Surveillance Program Scientific Review processes (Michigan Department of Health and Human Services Study #20220907-EA)20. The BioTrust allows research use of newborn dried blood spots under a waiver of consent for samples collected before 2010; for samples collected after 2010, broad parental consent for research on stored samples was required. There were 4,328,248 births between 1987 and 2020 in Michigan, from which 3,554,464 blood spots (82.1%) were available for research.

Study population

Cancer cases were identified through the Michigan Cancer Surveillance Program. Eligibility included: born in Michigan between 1987 and 2020, diagnosis of a malignant solid or brain tumor by age 8 years, and availability of a newborn dried blood spot for research. Children with diagnoses of germ cell or skin cancers were excluded due to the lack of association of malignant germ cell tumors with known autosomal dominant pediatric cancer predisposition syndromes and the difficulty of complete ascertainment and proper categorization of malignant skin cancers in childhood, respectively. Primary malignant tumors of the lung and pleura were analyzed together to account for evolving histopathologic classification over the study period, including recognition of pleuropulmonary blastoma as a distinct entity associated with DICER1. Eight years of age at cancer diagnosis was chosen as a cut-off, given the age of onset expected in the predisposition syndromes included in our analysis and given a focus on the potential of our data to inform newborn screening. Cancer outcomes were ascertained through linkage with the cancer registry for cases occurring between 1987 and 2020. As a result, children born in more recent years had shorter follow-up time in which to develop cancers. Data available for each case included birth year within a 2-year range, diagnosis, age, stage, primary treatment type, occurrence of second cancer, and demographic data. Some data were suppressed before release to the study investigators to assure anonymity of children with very rare cancers. We identified 1948 cases, including 977 males (50.2%), 889 females (45.6%) and 82 (4.2%) with sex not-recorded or suppressed. Some cancer registry diagnosis codes were grouped by study investigators to minimize the need for suppression of data (see Supplementary Fig. S3).

Newborn screening panel for cancer predisposition syndrome

Eleven genes (RB1, RET, TP53, SMARCB1, SUFU, PTCH1, WT1, DICER1, APC, ALK, PHOX2B) were chosen for analysis based upon their association with autosomal dominant early onset cancer predisposition syndromes for which there are recommended preventive cancer surveillance guidelines (Table 1). A targeted next-generation sequencing panel, which includes these genes, was designed and optimized for detection of single-nucleotide variants, small insertions and deletions of up to 20 bp, and copy number variants in coding and adjacent untranslated regions (including splice junctions).

Processing of dried blood spots, laboratory analysis, and variant classification

Genomic DNA was isolated for each case from a single 3.2 mm dried blood spot (DBS) punch using the Chemagic 360 instrument following the manufacturer’s protocol. Isolated DNA was fragmented enzymatically, and enriched by target capture21. DNA quality metrics included OD_260/280 and OD_260/230 ratios initially across the first 96 samples, and subsequently using Qubit; fragment size was evaluated using TapeStation PCR performance using library test fragments, and coefficient of variation across replicates. Paired-end (2 × 150 base-pairs) sequencing was performed using a NovaSeq 6000 Illumina Sequencer (San Diego, CA). Sequencing data were considered acceptable if libraries achieved >250× average coverage and >99% of targets at ≥20× coverage.

Raw sequencing data were aligned to human genome assembly GRCh38 and processed using the Fabric Genomics Sentieon pipeline for SNV and small indel detection. High-confidence variants were retained using standard thresholds, including minimum read depth ≥20×, mapping quality ≥20, and heterozygous allele balance 0.3–0.7, with exceptions reviewed in IGV. CNVs were called using CNVkit with exon-level parameters and required at least one exon with log2 ratio exceeding thresholds and read support, confirmed visually in IGV.

The Artificial-Intelligence driven Classification Engine (ACE) software application22 (Fabric Genomics; Oakland, CA) was used to classify single nucleotide variants, small insertions and deletions. Copy number changes were identified based upon loss of genetic material in or near the targeted gene, and then further reviewed for a deletion event involving >20 bp in at least one exon of the gene of interest. Variants were classified as pathogenic, likely-pathogenic, benign/likely-benign, or of uncertain significance using criteria from the American College of Medical Genetics and Association of Molecular Pathology23. ACE-designated variants of uncertain significance were secondarily filtered for review for possible pathogenicity if categorized by ClinVar as pathogenic or likely-pathogenic, or if the variant was associated with predicted loss-of-function (due to a gain of a stop codon, loss of a start codon, splice site donor/acceptor variant, or frameshift). A study geneticist reviewed all pathogenic and likely-pathogenic ACE-identified variants and deletions, as well as all secondarily filtered variants of uncertain significance to determine final classification using criteria from the American College of Medical Genetics and Association of Molecular Pathology23. To enhance transparency, the variant detection and adjudication workflow used in this study are summarized in Supplementary Fig. S4.

Statistics & reproducibility

As a population-based study, sample size was determined by the size of the population meeting inclusion criteria (see above). No statistical method was used to predetermine sample size; excluded cases and suppressed data are described above under “Study Population”. Descriptive statistics of the demographic and clinical features of the cohort, as well as prevalence of pathogenic and likely-pathogenic variants, are reported. Continuous variables are presented as mean ± standard deviation (SD) or median (IQR) for skewed distributions (skewness > |0.8 | ). Observed variant frequencies in this cancer cohort were not adjusted for ancestry, so any effects due to population structure or bias cannot be fully excluded. Differences by variant status were tested using Wilcoxon rank-sum tests for continuous variables and chi-square tests (or Fisher’s exact tests for cells with counts <5) for categorical variables; all tests were two-sided with statistical significance set at p < 0.05. Using Fisher’s exact test, we compared the proportion of cases with pathogenic/likely-pathogenic germline variants within selected cancer predisposition syndrome-specific diagnoses to internal controls (for example, the proportion of WT1 variants among infants who developed renal tumors compared to that among the remaining cases). In addition, four genomic newborn screening (non-cancer) studies (total N = 38,448 infants), chosen for large sample size, public availability of variant data, and inclusion of at least one of our study’s targeted genes in the study’s gene list, are described for comparison2,4,5,6. To estimate variant frequency in the general population, we analyzed ClinVar pathogenic/likely-pathogenic allele frequencies in the 11 cancer predisposition genes using the Genome Aggregation Database v4.0, a genomic dataset which includes 807,162 individuals24,25. Details of comparison analyses are provided in the Supplementary Methods. Overall survival was visualized using the Kaplan-Meier estimate of the survival function with differences between groups tested using the log-rank test. Analyses were conducted using SAS software, Version 9.4 (SAS Institute Inc., Cary, NC, USA), and data were additionally visualized and analyzed using R26 and visualized using cBioPortal27.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.