Therapy-Related Mutational Signatures in Subsequent Neoplasms among Survivors of Childhood Cancer

Analysis of the genomes of 200 subsequent breast, meningioma, and thyroid neoplasms from childhood cancer survivors reveals the potential mutagenic impact of radiation, platinum, and nitrogen mustard treatment.

Abstract

Childhood cancer survivors have a heightened risk of developing subsequent neoplasms (SN) related to therapy. We analyzed whole-genome, exome, and RNA sequencing of 200 breast, meningioma, and thyroid SNs, which developed a median of 26.4 years after childhood cancer, among 160 survivors. Meningioma and thyroid SNs were enriched for driver gene rearrangements compared with de novo tumors, including NF2-disrupting alterations and kinase fusions potentially induced by radiation. Radiation correlated with increased insertion–deletion signature ID5. Nitrogen mustard treatment correlated with elevated “flat” signature SBS5 in breast and meningioma SNs; in vitro, these agents caused an unresolved flat signature associated with multiple flat signatures from the Catalogue of Somatic Mutations in Cancer. In meningioma, platinum therapy correlated with NF2 splice-site variants. Analysis of 19 multisample survivors revealed intrapatient heterogeneity in meningioma, including clonally independent tumors. These results demonstrate the long-term impact of childhood cancer treatment on the genomes of SNs developing in adulthood, which may guide SN treatment and prevention.

Significance:

This represents the most comprehensive genomic characterization of SNs from childhood cancer survivors to date, revealing the mutagenic impact of multiple therapies on the SN genome, including the potential impact of nitrogen mustards such as cyclophosphamide. These results may guide the optimization of future cancer treatment regimens to prevent SN development.

See related commentary by Bertrums and van Boxtel, p. 1483

Introduction

Although >85% of children diagnosed with cancer will become >5-year survivors (1, 2), it is well established that these survivors have over a fivefold increased risk of developing a malignant subsequent neoplasm (SN) compared with cancer risk in the general population (3). In fact, malignant SNs are the most common nonrelapse cause of late mortality in childhood cancer survivors (4, 5). Other than approximately 7% of malignant SNs that are associated with germline cancer predisposition (6), the majority of SNs, which occur at ages much younger than in the general population, are therapy-related (3). After nonmelanoma skin cancer, breast and thyroid cancers are the most common therapy-related malignant SNs in childhood cancer survivors (7). Although thyroid malignancies are often responsive to therapy, 20% of female survivors who develop a subsequent breast cancer will not survive (8). Survivors are also at increased risk of developing nonmalignant SNs such as meningioma, which, though typically histologically benign, can be associated with tremendous neurologic comorbidity (9). New approaches are needed to prevent and treat SNs in childhood cancer survivors.

Prior exposure to DNA-damaging therapies increases the risk of developing SNs, including chest-directed irradiation for breast SNs (10), cranial radiotherapy for meningioma SNs (9), neck/thyroid-directed radiotherapy for thyroid SNs (11), alkylating agents (12), anthracyclines (12), and epipodophyllotoxins (13–15). Though these risk factors are well established, the genomic landscapes of the most common SNs among pediatric cancer survivors, including breast, thyroid, and meningioma SNs, are incompletely understood. Exome and targeted sequencing were previously used to study the genomic landscapes of 31 meningiomas, most of which were SNs from childhood cancer survivors (16); however, an analysis of therapy-induced mutational signatures, indicative of treatment exposure, was not possible due to the limited number of exonic mutations. To our knowledge, therapy-induced mutational signatures in breast, meningioma, and thyroid SNs among survivors of childhood cancer have not been explored. Understanding how prior exposures shape the genomic landscapes of these common SNs using whole-genome sequencing (WGS) to enable identification of all classes of genomic alterations and mutational signatures may inform underlying mechanisms of SN development, provide new therapeutic targets, and suggest alternative primary therapies to reduce SN risk.

To this end, we performed WGS, exome, and RNA sequencing (RNA-seq) on 200 breast, meningioma, and thyroid SN samples from 160 childhood cancer survivors. The samples were obtained through the Childhood Cancer Survivor Study (CCSS), a multi-institutional cohort of 5-year survivors of childhood cancer. The CCSS has collected and archived formalin-fixed, paraffin-embedded (FFPE) SNs, pathologically verified by centralized review, since 1994. By comparing the mutational profile of SNs from this cohort with those of de novo cancers characterized by The Cancer Genome Atlas (TCGA; refs. 17, 18) and other studies (19, 20), we were able to assess the long-term impact of treatment exposure on SN tumorigenesis in this unique cohort.

Results

Cohort Composition and Sequencing

We originally performed WGS, exome, and RNA-seq on 306 breast, meningioma, and thyroid SNs from FFPE tissue obtained through the CCSS and matched germline tissue for WGS and exome. We excluded 106 of these from analysis due to low DNA sequencing quality (n = 59), low tumor purity as evidenced by a lack of somatic alterations (n = 34), or evidence of transplant or tumor/normal tissue mismatch (n = 13; Supplementary Fig. S1A; see “Methods”).

The remaining 200 high-quality samples were derived from 160 individual childhood cancer survivors, including 63 with breast cancer, 57 with meningioma (98% of which were nonmalignant, including 72% grade I and 26% grade II), and 42 with thyroid cancer (including 90% papillary and 10% follicular) SNs (Supplementary Table S1). Among these 160 individuals, two had multiple SNs of different histologies. All 200 samples had DNA sequencing from WGS and/or exome, including 154 with WGS (median 46× and 36× coverage for tumor and germline, respectively) and 182 with exome sequencing (median 174× and 46×; Supplementary Fig. S1B). Of these 200 DNA-characterized samples, we obtained high-quality RNA-seq data on 128 (Supplementary Fig. S1C). No statistically significant biases were observed with respect to histology, stage, grade, subtype, or age when comparing the 200 high-quality samples with the 106 excluded samples (Supplementary Tables S1 and S2). For example, among the good-quality breast cancers, 78 were ductal (96%) and 3 were lobular (4%), whereas among the excluded breast cancer samples, 40 were ductal (93%) and 3 were lobular (7%; P = 0.42 by the two-sided Fisher exact test; Supplementary Table S2).

The median age at childhood cancer diagnosis was 10.7 years (range, 0.5–20.7), and SNs occurred at a median age of 37.8 years (range, 13–54.4; Fig. 1A). The median time between childhood cancer diagnosis and SN occurrence was 26.4 years (range, 9.9–43.8). Radiation was the most frequent DNA-damaging therapy given for childhood cancer among the 160 survivors with SNs (87% of individuals), followed by vinca alkaloids (61%), nitrogen mustard alkylators (46%), and anthracyclines (38%), with varying frequencies by SN type (Fig. 1B; Supplementary Table S3). The most frequent childhood cancer diagnoses were leukemia (36%), Hodgkin lymphoma (26%), central nervous system (CNS) tumors (13%), and bone tumors (10%; Supplementary Table S4). Patients with meningioma SN were most commonly survivors of childhood leukemia who had received prophylactic and therapeutic radiation for CNS involvement historically (21, 22), whereas breast and thyroid SNs were most commonly diagnosed in Hodgkin lymphoma survivors [Fig. 1B (right)] due to chest and neck radiation exposures (23).

Figure 1.

SN cohort and mutation burden. A, The x-axis shows the age at diagnosis of childhood cancer (diamond) and later SN (circle) in each patient. Childhood and SN cancer types are indicated in the legend at the bottom. B, Left, Percentage of patients with SN receiving each class of chemotherapy or radiation treatment. Right, Stacked bar plot showing the percentage of patients with SN with each childhood cancer diagnosis. 1° indicates primary (childhood) cancer. C, Top, Number of coding-region SNVs per Mb in each SN type (this study) and in control cohorts from other studies. The median is shown as a horizontal bar. All data are based on exome sequencing. N at the bottom indicates the number of patients. D, Age at sample collection or diagnosis in each cohort, with the median indicated as a horizontal bar. Patients with germline cancer predisposition mutations were excluded. E, As in C, except that indels rather than SNVs are shown. P values in all panels are from multivariate linear regression incorporating clinical variables, including age, sex (when relevant), and grade/stage (“Methods”).

Single-Nucleotide Variant Mutation Burden in SNs

To assess the impact of therapy on the mutation burden of SNs, we analyzed the somatic single-nucleotide variant (SNV) burden in each SN type; insertions and deletions (indel) were analyzed separately as their therapeutic etiologies may differ (24–26). In patients with multiple samples, one sample per patient was used to assess mutation burden. As genomic sequencing of FFPE samples leads to artifacts (27), we assessed potential artifacts using exome sequencing from 33 available pediatric tumor samples spanning 18 tumor types (Supplementary Fig. S2A; Supplementary Table S5), in which we profiled matched fresh-frozen and FFPE tissue from the same sample, along with matched normal exome sequencing. Filtering SNVs using criteria established to remove variants likely representing FFPE artifacts resulted in better correlation (Pearson r = 0.897) than unfiltered SNVs (Pearson r = 0.811) between fresh-frozen and FFPE SNV burdens (Supplementary Fig. S2B; see “Methods”). Likewise, filtering indels yielded similar indel burdens between fresh-frozen and FFPE samples (0.026 vs. 0.034 indels per Mb, respectively; P = 0.16 by the two-sided Wilcoxon rank-sum test), whereas unfiltered indel burdens were significantly higher in FFPE (0.043 vs. 0.087 indels per Mb; P = 0.001; Supplementary Fig. S2C; see “Methods”). Thus, we used these filtering criteria for our FFPE exome data and comparison cohorts in assessing mutation burden, as described below. As further technical validation, using the abovementioned filtering criteria, SNV burdens for exome and WGS data were significantly correlated (Supplementary Fig. S3).

We used exome data for SNV and indel burden comparisons with other studies that relied on exome data, from which we downloaded raw data and called mutations consistently, including a random selection of 100 TCGA breast cancers and 100 TCGA thyroid cancers as de novo tumor comparisons. As a technical validation, we confirmed that our SNV and indel burdens strongly correlated with the TCGA calls from these 200 TCGA breast and thyroid cancers (Supplementary Fig. S4A and S4B; Pearson r = 0.999 for SNV and r = 0.970 for indels) and that these randomly selected cases had representative SNV and indel burdens relative to the entire cohort of TCGA breast and thyroid cancer exomes (Supplementary Fig. S4C and S4D).

Breast SNs had a median of 0.32 SNVs per megabase (Mb), lower than de novo breast cancers from TCGA (0.67 per Mb; Fig. 1C; ref. 17), though this difference was not statistically significant upon multivariate linear regression analysis, including age, tumor purity, stage, histology (ductal vs. lobular), estrogen receptor (ER) status, and HER2 status (Supplementary Tables S6 and S7). De novo patients with breast cancer were significantly older (median 60 years) than those with breast SN (median 41; Fig. 1D) by multivariate analysis including histology, breast cancer stage, ER status, and HER2 status (Supplementary Table S8; patients with germline cancer predisposition mutations were excluded from the age comparison in Fig. 1D, as these are associated with younger age at diagnosis; ref. 28). This age difference may explain the lack of statistical significance of the decreased SNV burden in breast SNs compared with de novo breast tumors. A similarly and significantly reduced age was observed in meningioma and thyroid SNs relative to de novo cancers by multivariate analysis, including histology, grade, and sex (Fig. 1D; Supplementary Table S8).

Meningioma SNs had a median of 0.35 SNVs per Mb, significantly higher than de novo meningiomas (0.15), which were originally reported by Clark and colleagues (Fig. 1C; ref. 20) using multivariate linear regression that included age, grade, tumor purity, and sex as covariates (Supplementary Table S7). We also analyzed, for comparison, the SNV burden in an exome-based study of radiation-induced meningiomas (RIM; Agnihotri and colleagues; ref. 16), in which 15 of 16 samples were also SNs from childhood cancer survivors. This RIM cohort also had a significantly higher SNV burden (median 0.34 SNVs per Mb) than de novo meningioma, whereas our meningioma SNs’ median of 0.35 was not significantly higher than that of the RIM cohort (Fig. 1C). Thyroid SNs also had a significantly higher SNV burden (median 0.26 per Mb) than de novo thyroid cancers from TCGA (0.18; ref. 18) by multivariate linear regression incorporating age, sex, tumor purity, and histology (Fig. 1C; Supplementary Table S7).

Increased exome coverage could not account for the increased SNV burdens in meningioma SNs, as the de novo meningiomas were profiled at significantly higher coverage than meningioma SNs or the RIM cohort (Supplementary Fig. S5A). Furthermore, inclusion of only samples with lower coverage (≤100×; Supplementary Fig. S5B) did not substantially change the SNV burdens in breast or thyroid SNs relative to de novo breast and thyroid cancers (whereas this analysis could not be done for meningioma as all de novo tumors were above 100×), and the increased SNV burden in thyroid SNs compared with de novo thyroid cancers remained statistically significant by multivariable analysis when including only these lower-coverage samples (P = 0.0075; Supplementary Table S9; Supplementary Fig. S5C). Furthermore, downsampling of 10 thyroid SNs to match de novo TCGA thyroid cancer coverage did not significantly change SNV burdens (Supplementary Fig. S5C–S5E).

To test whether germline alterations affected the SNV burden in SNs, we first identified germline cancer–predisposing variants from WGS and exome data. In total, 14 of 160 (8.8%) patients with SN had germline cancer–predisposing variants, including 8 patients with breast cancer, 3 patients with meningioma, and 3 patients with thyroid SN (Supplementary Table S10), similar to the frequency reported in pediatric cancer (28). These included five BRCA1 and BRCA2 variants (none of which were associated with biallelic inactivation/LOH in SNs; Supplementary Table S10), three CHEK2 variants, and two TP53 inactivating variants. However, no significant association was found between SNV burdens in germline carriers versus others in breast, meningioma, or thyroid SNs (P > 0.05 by the two-sided Wilcoxon rank-sum test). These data suggest that germline alterations are not the primary cause of the increased SNV burden in SNs.

Indel Mutation Burden in SNs

The somatic indel burden was similar in breast SNs (median 0.058 indels per Mb) versus de novo TCGA breast cancer (also 0.058; Fig. 1E; Supplementary Table S11) by multivariable linear regression using the same covariates described above for SNV burden analysis (Supplementary Table S12). In meningioma, the indel burden was significantly increased in SNs (median 0.058 indels per Mb) and the RIM cohort (also 0.058) versus de novo meningioma (0.0 per Mb; Fig. 1E, in which ladder appearance is due to integer values where 1 indel is 0.029 per Mb; ref. 16) by multivariable linear regression (Supplementary Table S12). Likewise, thyroid SNs had a significantly elevated indel burden (median 0.029 per Mb) versus de novo TCGA thyroid cancer (median 0; Fig. 1E) by multivariable linear regression (Supplementary Table S12), though the effect size was small, representing a median difference of 1 indel per genome.

Together, these data indicate that meningioma and thyroid SNs have higher SNV and indel burdens relative to de novo tumors. We next explored whether this may be therapy-related using mutational signature analysis.

SNV Signatures and Associations with Nitrogen Mustard Exposure

We assessed the impact of therapy on SN genomes using SNV mutational signature analysis, given the elevated SNV burden in meningioma and thyroid SNs. We used a reported SNV signature framework (29), including known signatures curated by the Catalogue of Somatic Mutations in Cancer (COSMIC; ref. 30), and analyzed WGS samples with sufficient and uniform coverage to assess SNV genome-wide, including 21 breast, 44 meningioma, and 15 patients with thyroid SN (“Methods”). We also analyzed a published dataset of matched fresh-frozen and FFPE WGS from eight ovarian cancers (31) and observed statistically significant correlations between SNV (Pearson r = 0.91; P = 0.002) and indel (r = 0.84; P = 0.008) burdens, indicating FFPE suitability for signature analysis for these mutation types (Supplementary Fig. S6A and S6B). However, structural variants (SV) did not significantly correlate between fresh-frozen and FFPE samples, with frequent dropout in FFPE (Supplementary Fig. S6C; r = 0.47; P = 0.24), indicating that SV signatures could not be assessed with our WGS data. The same applied to copy-number signatures, due to frequent oversegmentation in FFPE data.

De novo signature extraction (Supplementary Fig. S7) identified six SNV signatures, which were deconvoluted into known COSMIC (30) single-base substitution (SBS) signatures, including ubiquitous age-related signatures SBS1 (induced by 5-methylcytosine deamination) and SBS5 (unknown etiology; ref. 32); SBS3 (homologous recombination deficiency) and SBS40a (unknown etiology) in one tumor each; apolipoprotein B mRNA editing catalytic polypeptide-like enzyme-induced signatures SBS2 and SBS13 in 6 of 21 breast SNs; and platinum treatment-associated signatures SBS31 and SBS35 in 4 SNs (Fig. 2A; Supplementary Table S13) of the 8 SNs with prior platinum treatment. SBS5 accounted for the majority (>80%) of SNVs in each SN type. No novel signatures were identified. Seventy-five of eighty (94%) tumors could be reconstructed with cosine similarity ≥0.85 using the tested COSMIC SBS signatures, and 69 of 80 (86%) with cosine similarity ≥0.90 (Supplementary Table S13). The 11 tumors with cosine scores below 0.9 were enriched for low-SNV burden samples (mean 374 SNVs vs. 1,022 in other samples; P = 1.8 × 10⁻⁷ by the two-sided t test), suggesting technical variation related to low mutation burdens. These data indicate that the tested COSMIC SBS signatures effectively reconstruct most tumors in our cohort.

Figure 2.

SNV mutational signatures in SNs. A, Strength of each COSMIC SNV mutational signature in each patient based on WGS data. Patients are ordered along the x-axis by SNV burden, and the y-axis indicates the number of SNVs attributed to each signature (top) or the percentage of SNVs (bottom). The rightmost bar in each SN cancer type indicates the mean across all samples in that SN type. Colors indicate specific mutational signatures as shown at right. B and C, Strength of COSMIC signature SBS5 in breast (B) or meningioma SNs (C), with patients separated by prior treatment with the indicated drug or drug class. Red indicates patients that were treated with the indicated drug or class for their childhood cancer (within 5 years of childhood cancer diagnosis) prior to SN development (binary; yes/no); and blue indicates patients not treated. Only drug classes with at least five treated patients and five untreated patients were analyzed, and within these classes, drugs with at least two treated patients were analyzed. Box, IQR (25th–75th percentile); middle bar, median. Whiskers are described in R boxplot documentation (a 1.5× IQR rule is used). P values are from multivariate linear regression with covariates described in “Methods.” D, Schematic showing experimental treatment of MCF10A cells with drugs or control for 5 weeks, followed by isolation of three single-cell clones per treatment group and WGS of these clones (and of bulk pretreatment control to subtract preexisting SNVs). E, SNV burden in MCF10A single-cell clones (y-axis) in each treatment group after 5 weeks of treatment, subtracting out SNVs already present in bulk pretreatment control. P values are from a two-sided t test assuming equal variance. F, Cosine similarities between experimental drug signatures from the treatment of MCF10A cells (mean of three clones per drug or drug combination) versus COSMIC SNV signatures detected in SNs, with the control background signature subtracted out from MCF10A treatments. G, Ninety-six-channel SNV mutational spectrum of MCF10A cells treated with 4-OH-cyclophosphamide or 4-OH-cyclophosphamide plus cytarabine (top two plots) or three COSMIC SNV signatures with the highest similarity to the 4-OH-cyclophosphamide signature (cosine similarities in F).

Given the increased SNV burden in meningioma and thyroid SNs (Fig. 1C) and the lack of novel SNV signatures, we hypothesized that therapy may contribute to known signatures. We focused on SBS5, given its abundance, and asked whether prior childhood cancer treatment with various DNA-damaging therapies was associated with SBS5 levels (Fig. 2B and C; Supplementary Fig. S8A; Supplementary Tables S14 and S15). SBS5 is a ubiquitous signature increasing with age (32), may result from generalized mutation processes, and is a “flat” signature affecting the 96 SNV types with only moderately varying frequency (29). In breast and meningioma SNs, prior treatment with a nitrogen mustard–alkylating agent (cyclophosphamide, ifosfamide, or mechlorethamine) was associated with significantly increased SBS5 as a drug class (Fig. 2B and C; Supplementary Tables S14 and S15; in meningioma, cyclophosphamide encompassed the entire nitrogen mustard class, as other mustards were not used), with breast SNs having a median of 1,319 versus 585 SBS5 SNVs in treated versus untreated, respectively (Fig. 2B; P = 3 × 10⁻³ by multivariate linear regression including age, breast cancer stage, tumor purity, and ER/HER2 status as covariates after multiple hypothesis correction; Supplementary Table S14), whereas meningioma SNs had a median of 990 versus 709 SBS5 SNVs in treated versus untreated, respectively (Fig. 2C; P = 4.1 × 10⁻³ by multivariate linear regression including age, sex, grade, and tumor purity as covariates after multiple hypothesis correction; Supplementary Table S15; “Methods”).

No significant treatment-SBS5 associations were observed in thyroid SNs, perhaps partly due to lower sample numbers (n = 15; Supplementary Fig. S8A). Furthermore, cyclophosphamide-induced mutagenesis may be suppressed in thyroid SNs due to higher expression of cyclophosphamide-inactivating ALDH1A1 (33) compared with breast and meningioma SNs, with a similar pattern in these SNs’ normal tissue of origin based on GTEx (Supplementary Fig. S8B; ref. 34).

As nitrogen mustards were associated with SBS5 (Fig. 2B and C), we asked whether there was a dose–response relationship, focusing on meningioma in which a sufficient number of patients received cyclophosphamide (n = 18), whereas breast cancer had only three cyclophosphamide-treated patients and three treated with other nitrogen mustards. Cyclophosphamide dose was not significantly associated with SBS5 in meningioma SNs, either when including all samples or treated samples only, by multivariable linear regression including age, sex, grade, and tumor purity (Supplementary Fig. S8C; Supplementary Table S16), despite increased SBS5 when comparing cyclophosphamide-treated versus untreated as a binary variable (Fig. 2C). The lack of dose–response correlation may be due to widely differing drug combinations coadministered with cyclophosphamide (Supplementary Table S17); as shown experimentally below, drug coadministration may suppress cyclophosphamide-induced mutagenesis.

To test whether nitrogen mustards can induce SBS5, we treated MCF10A breast epithelial cells with an active cyclophosphamide metabolite, 4-hydroxy-cyclophosphamide (4-OH-cyclophosphamide), which is generated in the liver after cyclophosphamide treatment (35). Cells were treated for 5 weeks with 4-OH-cyclophosphamide (1 μmol/L), cisplatin (0.5 μmol/L) as a positive control with a known signature (25), a control culture without drug, cytarabine (25 nmol/L) as a negative control not associated with SBS5 (Fig. 2C), or the combination of 4-OH-cyclophosphamide (1 μmol/L) and cytarabine (10 nmol/L) to assess how drug combinations affect mutagenesis. Dosages were selected based on dose-response curves in MCF10A cells (Supplementary Fig. S9), choosing doses that inhibited proliferation only moderately (10%–40% inhibition) to avoid the selection of resistant clones, as we did previously with the discovery of another therapy-induced mutational signature (36). Cells were treated in biological triplicate for each drug, followed by single-cell cloning (n = 3 clones total per drug or drug combination) and WGS (Fig. 2D). WGS of pretreatment bulk cells was used to subtract preexisting SNVs (Fig. 2D). 4-OH-cyclophosphamide significantly increased the SNV burden in MCF10A clones compared with control (Fig. 2E; mean 2010 SNVs vs. 616 SNVs, respectively; P value shown is from the two-sided t test), whereas cisplatin also increased the SNV burden as expected and cytarabine did not. Interestingly, combination treatment (4-OH-cyclophosphamide plus cytarabine) yielded 1,514 SNVs, significantly above control but significantly below 4-OH-cyclophosphamide alone (36% fewer SNVs relative to 4-OH-cyclophosphamide alone; Fig. 2E). This may be due to the suppression of proliferation, which may facilitate 4-OH-cyclophosphamide–induced mutagenesis through DNA replication. Notably, 92% of acquired SNVs detected in this experiment (across all MCF10A clones sequenced) were unique to one clone, indicating that mutations were not primarily due to selection. As a further technical validation, cisplatin treatment in MCF10A cells yielded a signature resembling known platinum-associated COSMIC signatures SBS31 (cosine similarity 0.95) and SBS35 (0.85; Fig. 2F; Supplementary Fig. S10).

Among the eight SNV signatures detected in SNs, the MCF10A-derived 4-OH-cyclophosphamide signature (after subtracting the control untreated SNV signature; Supplementary Fig. S11A–S11C) most closely resembled “flat” signatures SBS3, SBS5, and SBS40a, with moderate cosine similarities of 0.84, 0.77, and 0.83, respectively (Fig. 2F and G), consistent with a recent study showing that treatment with the nitrogen mustard maphosphamide induces an SBS5-like signature in hematopoietic stem and progenitor cells (HSPC; ref. 37). The results from that study’s experimental maphosphamide signature also showed the highest similarity to flat signatures, including SBS3, SBS5, and SBS40a across COSMIC signatures, with cosine similarities of 0.77, 0.76, and 0.78, respectively (Supplementary Fig. S11D and S11E). The study also reported increased flat signature abundance, including SBS5, after a cyclophosphamide-containing regimen in neonatal HPSCs, consistent with our results. SBS5, SBS3, and SBS40a, along with the 4-OH-cyclophosphamide signature, are all broad-spectrum, relatively featureless signatures (Fig. 2G), potentially making them difficult to disambiguate computationally. When comparing the 4-OH-cyclophosphamide signature to all 86 COSMIC SNV signatures (instead of just the 8 we detected in SNs), the top 4 signatures with the highest cosine similarity were SBS3, SBS5, SBS40a, and additionally SBS40c (0.80), the last of which was not detected in SNs. This indicates the uniqueness of the similarity of the 4-OH-cyclophosphamide signature to flat COSMIC signatures, despite having absolute cosine scores below 0.9.

We performed the following computational simulation to test whether mutagenesis caused by cyclophosphamide may partly contribute to the abundance of flat COSMIC signatures. Specifically, the in silico addition of the MCF10A 4-OH-cyclophosphamide to WGS-based mutational spectra of nitrogen mustard–untreated breast and meningioma SNs (to simulate 4-OH-cyclophosphamide treatment) increased SBS5 in most tumors, with an SBS40a increase in a smaller subset (Supplementary Fig. S12A–S12C). SBS5 represented 60.2% of the overall signature increase upon simulated 4-OH-cyclophosphamide addition in breast SNs and 89.4% in meningioma SNs, whereas SBS40a represented 33.5% and 9.4%, respectively (Supplementary Fig. S12D and S12E). These simulated treatment profiles were reconstructed with the eight COSMIC signatures found across SNs (Fig. 2A) with virtually the same cosine similarity as the original profiles of patients with SN [0.93 cosine similarity for both simulated and original profiles in meningioma and in breast SNs, 0.92 for simulated and 0.93 for original profiles; Supplementary Fig. S12B and S12C (right)]. The simulated profiles were generated by adding 734 or 281 SNVs from the 4-OH-cyclophosphamide signature to breast or meningioma SNs, respectively, corresponding to the median SBS5 increase seen in patient tumors treated with nitrogen mustards (Fig. 2B and C). These data may explain the lack of prior recognition of cyclophosphamide-induced mutagenesis in patient tumors compared with more clearly defined drug signatures such as those of cisplatin (29).

Together, these data indicate that cyclophosphamide treatment induces an unresolved flat signature with moderate similarity to multiple flat COSMIC signatures and that the increased SNV burden in some SNs (Fig. 1C) may be due to alkylators such as nitrogen mustards.

Indel Signatures and Associations with Radiotherapy

We next performed indel signature analysis based on a published framework that relies on indel size, context, and repeat pattern to classify indels, resulting in 83 indel types (29). De novo extraction [Supplementary Fig. S7 (bottom)] yielded three signatures, which were deconvoluted into four COSMIC indel signatures, with no novel signatures identified. The four COSMIC indel (ID) signatures identified in SNs (Fig. 3A; Supplementary Table S18) included age-related signature ID5 as the most abundant (mean >40% of indels in each cancer type); ID1 and ID2, which result in T indels, respectively, at poly-T tracts; and ID3, which is enriched in (though not specific to) patients who smoke. No chemotherapy was significantly associated with indel signature abundance. Prior radiation dose to relevant organs (chest dosage for breast SNs, brain dosage for meningioma, and thyroid dose for thyroid SNs) showed a positive association with ID5 by multivariate analysis in meningioma (P = 0.01; β = 0.022 indels per cGy of radiation) but not in breast or thyroid SNs (Fig. 3B), with age, stage, tumor purity, sex, grade, histology, ER status, HER2 status, and tumor purity as covariates, as relevant (Supplementary Table S19; “Methods”). When combining all three cancer types, the strength of the association between ID5 and the organ-relevant radiation dose was more significant (P = 0.001; β = 0.017) by multivariate linear regression incorporating age, tumor purity, sex, and cancer type as covariates (Fig. 3C; Supplementary Table S20), consistent with Chernobyl-associated radiation correlating with ID5 in thyroid cancer (21). ID5 was the only indel signature that was significantly associated with radiation. Together, these findings suggest that radiation may induce indels in SNs by accelerating ID5. ID5 is primarily associated with small deletions, not insertions (29), consistent with experimental data showing that radiation primarily induces deletions rather than insertions (38).

Figure 3.

Indel mutational signatures in SNs. A, Strength of each COSMIC indel mutational signature in each patient based on WGS data. Patients are arranged along the x-axis, and the y-axis indicates the number of somatic indels attributed to each signature (top) or the percentage of indels (bottom). The rightmost bar in each SN cancer type indicates the mean across all samples in that SN type. Colors indicate specific mutational signatures as shown on the right. B and C, Correlation between prior radiation dose to the relevant organ (within 5 years of childhood cancer diagnosis; x-axis) and ID5 strength (y-axis) in SNs as determined from WGS, with (B) showing each cancer type individually and (C) showing the three cancer types combined. P and β (the latter representing indels per cGy) values are from multivariate linear regression incorporating age and other clinical features (see “Methods”), and each point represents one SN sample. Dotted lines demarcate the 95% confidence interval of the linear regression line.

Doublet Signatures

We also performed mutational signature analysis on somatic doublet (e.g., AA > TT) mutations using WGS data. We limited the analysis to 28 samples with at least 15 doublet variants due to the paucity of this variant type. De novo signature extraction yielded three signatures (Supplementary Fig. S13A), none of which resembled COSMIC doublet signatures (cosine similarity <0.80). However, our signature 3 did resemble aspects of COSMIC DBS5 (platinum-associated), including CT > AA and CT > AC enrichment (Supplementary Fig. S13B). Indeed, this signature was only found in three platinum-treated tumors (Supplementary Fig. S13C), suggesting it may have been platinum-induced.

Radiation-Associated Driver Rearrangements Enriched in Meningioma and Thyroid SNs

We next analyzed the genomic landscapes of each of the three SN histologies, focusing first on meningioma and thyroid SNs, which had a common theme of driver rearrangement enrichment. Among 57 meningioma SNs with WGS and/or exome data, we identified somatic NF2 alterations in 61% of patients and loss of chromosome 22, in which NF2 resides, in 92% (Fig. 4A; Supplementary Fig. S14), similar to the RIM cohort (16). NF2 alterations included SVs in 17 (30%) of patients with meningioma SN, which disrupt the NF2 reading frame through inter- or intrachromosomal rearrangements (Fig. 4A; Supplementary Table S21), as in the RIM study (16). These NF2 SVs may have been therapy-related given their absence in de novo meningiomas [0% of cases; Fig. 4A (right)], which represents a statistically significant depletion by multivariable logistic regression with age, sex, and grade as covariates (P = 2.5 × 10⁻⁸; Supplementary Table S22). Among the 16 NF2 SV-bearing patients with known treatment history, 15 had received prior cranial radiotherapy, including 3 without chemotherapy.

Figure 4.

Meningioma SN genomic landscape. A, Oncoprint showing the presence/absence of alterations in each gene or chromosomal alteration in patients with meningioma SN. Mutation types are color-coded as indicated in the legend at the left; deletion indicates focal deletion of all or part of the gene. SV disrupting NF2 due to rearrangement. Bar plots at right indicate percentage of patients with alterations in each gene in this cohort (SN; n = 57) or other indicated cohorts (n = 31 for Agnihotri; n = 300 for Clark). Top indicates the proportion of indels attributed to each indel signature, followed by SNV signatures below. Sequencing platforms and selected treatments (given for the childhood cancer prior to SN development) are also indicated at the top. * indicates two samples with platinum mutational signatures and NF2 splice SNVs. B, Percentage of meningioma SNs with splice site NF2 SNVs in patients without (“−”) or with (“+”) platinum SNV signatures SBS31 or SBS35. P value by two-sided Fisher exact test. C, Schematic showing locations of NF2 splice site SNVs in two patients with platinum SNV signatures. D, Functional analysis of NF2 splice-site SNVs using a plasmid encoding exons 11–14 of NF2 with native intron 12 between exons 12 and 13, followed by GFP in the same open reading frame. The two NF2 splice site SNVs shown in C were also introduced into this construct, followed by the transfection of 293T cells with each construct to assess fluorescence microscopically (left, brightfield; right, green fluorescence; scale bar indicates 200 μm), where fluorescence indicates successful splicing due to preservation of the reading frame. RNA-seq was also performed on transfected cells (right) to assess the percentage of exon 12 to 13 splice junction-spanning reads that were properly spliced (gray) versus unspliced (red). E, Bar plots showing the 96-SNV mutational spectrum of the meningioma SN with an NF2 exon 12 splice variant versus SBS35 found in the patient (top) and of the meningioma SN with an NF2 exon 13 splice variant versus SBS31 found in this patient (bottom). Bar plots at the right indicate the proportion of SNVs in each sample attributed to each SNV signature and the probability that each NF2 splice site SNV was induced by each signature.

Among 44 mutational signature-characterized meningioma SNs, 4 had received prior platinum treatment, and 3 of these bore platinum signatures [SBS31/SBS35; Fig. 4A (top), where white indicates lack of high-quality WGS data for the sample, leading to an inability to assess mutational signatures]. Interestingly, platinum signatures and NF2 splice-site SNVs significantly co-occurred (Fig. 4B; P = 0.009 by the two-sided Fisher’s exact test), with two of three platinum signature-positive patients bearing NF2 splice-site SNVs (67%), compared with 1 of 41 signature-negative patients (2%). These SNVs occurred at exon 12 or 13 splice donor/acceptor sites [Fig. 4A (asterisks) and C] and likely inactivate NF2 as they are associated with neurofibromatosis type 2 (39, 40). Given the lack of RNA-seq in these SNs, we tested whether these SNVs affect NF2 splicing by transfecting 293T cells with GFP constructs preceded by NF2 exon 12, intron 12, and exon 13, with or without the splice mutations [Fig. 4D (left)]. Successful splicing would produce an in-frame, fluorescent NF2-GFP chimera, whereas unsuccessful splicing leads to an out-of-frame, nonfluorescent protein, similar to a reported approach (41). Wild-type (WT) NF2-GFP had GFP fluorescence consistent with splicing, whereas the exon 12 and exon 13 splice mutants abrogated fluorescence (Fig. 4D) and splicing based on RNA-seq [Fig. 4D (right); Supplementary Fig. S15A and S15B]. Thus, the NF2 exon 12 and 13 splice-site SNVs cause splice defects and NF2 inactivation.

The two NF2 splice SNVs were at platinum signature hotspot contexts (Fig. 4E), suggesting they may have been platinum-induced [A(C > A)C for the exon 12 variant; C(C > T)T for the exon 13 variant]. Using each signature’s abundance in these two samples and each signature’s SNV predilection, we calculated an 86.9% probability that the exon 12 NF2 SNV was SBS35-induced and a 79.7% probability that the exon 13 SNV was SBS31-induced [Fig. 4E (right); Supplementary Fig. S16], using an approach we employed previously (42, 43). Together, these data suggest that platinum treatment may induce NF2 splice-site SNVs, which contribute to meningioma development.

We also analyzed the landscape of 42 thyroid SNs with WGS and/or exome data (Fig. 5A), focusing on known thyroid cancer drivers (18, 26), as no other genes were significantly mutated. Thyroid SNs were separated into four subgroups defined by driver alterations and transcriptional profiles (Fig. 5A–C), including (i) PPARG fusion-positive patients (n = 3 or 7% of patients); (ii) a kinase fusion-driven group, including 14 patients (33%) with RET, NTRK3, NTRK1, or ALK fusions (Supplementary Table S23); (iii) a Ras-mutant group, including 9 patients (21%) with SNVs or indels in BRAF, HRAS, NRAS, or NF1, which clustered transcriptionally with the kinase fusion group (Fig. 5C); and (iv) an “aneuploid/deletion-enriched” group of 16 patients (38%) lacking driver fusions or SNVs/indels and instead harboring copy-number alterations such as chromosome-level copy alterations and focal deletions [PTEN and CDKN2A; Fig. 5A and B (right)]. We compared the thyroid SN driver frequency with radiation-associated thyroid cancers from the Chernobyl accident (26) in addition to de novo TCGA thyroid cancers (18).

Figure 5.

Thyroid SN genomic landscape. A, Oncoprint showing the presence/absence of fusions (detected by WGS and/or RNA-seq) or gene-specific mutations in patients with thyroid SN. Mutation types are color-coded as indicated in the legend at the left; deletion indicates focal deletion of all or part of the gene. Bar plots at the right indicate the percentage of patients with alterations in each gene in this cohort (SN; n = 42) or other indicated cohorts (n = 305 for Morton; n = 484 for TCGA). Top indicates the proportion of indels attributed to each indel signature, followed by SNV signatures below. Sequencing platforms, radiation treatment (given for childhood cancer prior to SN development), and transcriptional subtype (as defined in C) are also shown at the top. B, Copy-number profiles of each thyroid SN for each chromosome, ordered by chromosome position from top to bottom, with patients (x-direction) ordered the same as in A. Integer DNA copy level is indicated by color (legend at left), with gray indicating unknown copy status (primarily at centromeres). C, Transcriptional clustering using t-SNE of all three SN cancer types using RNA-seq data from this study. Top-left, thyroid SNs cluster into three subgroups, which are also annotated in A. D, Percentage of patients in our SN cohort versus other cohorts with various classes of alterations. Subgroups are defined as in A, with the “other” group including patients lacking PPARG, kinase fusions, and Ras pathway mutations.

Furthermore, thyroid SNs and Chernobyl-associated thyroid cancers were significantly enriched for kinase fusions compared with de novo TCGA thyroid cancers [33%, 36%, and 12% of samples, respectively; Fig. 5D; P = 0.042 for SNs vs. TCGA, and P = 0.008 for Chernobyl vs. TCGA by multivariable logistic correction with age, sex, and histology as covariates, with multiple hypothesis correction (Supplementary Table S24; Fig. 5A (right)); ref. 18]. This fusion enrichment may result from radiation-induced DNA rearrangements, as 83% of thyroid SNs and all Chernobyl-associated thyroid cancers were radiation-exposed. Ras pathway mutations were less frequent in thyroid SNs compared with Chernobyl-associated and de novo thyroid cancers, particularly BRAF mutations [10%, 47%, and 62%, respectively; Fig. 5A (right) and D; Supplementary Table S25; P < 0.05 for all pairwise comparisons between the three cohorts by multivariable logistic regression after multiple hypothesis correction]. Furthermore, thyroid SNs were enriched for cases lacking driver fusions or RAS pathway mutations (aneuploid/deletion-enriched), with 38% of samples in this group, compared with 4% for Chernobyl-associated and 14% for de novo thyroid cancers (Fig. 5D; P < 0.001 for all pairwise comparisons). Together, these data suggest that radiation exposure, whether therapeutic (SNs) or environmental (Chernobyl-associated), enriches for fusions, potentially due to radiation-induced rearrangements. Furthermore, the enrichment of the aneuploid-enriched subgroup in thyroid SNs suggests that therapy or age may also enrich for this subtype.

Breast SN Genomic Landscape Resembles That of De Novo Breast Tumors

We analyzed the driver alterations found in 63 breast SNs, focusing on known breast cancer drivers (19), as no other genes were significantly mutated (Fig. 6). In total, 14% of breast SNs harbored focal ERBB2 copy gains, 19% bore PIK3CA mutations, 21% had TP53 mutations, and 10% each harbored CDH1 or GATA3 mutations (Fig. 6). We compared these frequencies to 560 de novo breast cancers with driver alteration annotation [Fig. 6 (right); ref. 19], revealing a similar ERBB2 copy gain frequency in de novo (13%) versus SNs (14%), an increased PIK3CA mutation frequency in de novo (28% vs. 19%), and an increased TP53 frequency in de novo (39% vs. 21%). The latter was perhaps due to the enrichment of ER-negative cases in the de novo cohort (19). Other genes mutated in ≤10% of de novo breast cancers were also altered at low frequency in SNs (Fig. 6). Notably, no driver gene was altered at a significantly different frequency in breast SNs versus de novo breast cancers by multivariate logistic regression incorporating age, stage, ER status, HER2 status, and histology (Supplementary Table S26). Overall, these results indicate that breast SNs have a comparable driver gene profile to de novo breast cancers.

Figure 6.

Breast SN genomic landscape. Oncoprint showing the presence/absence of alterations in each gene in patients with breast SN. Mutation types are color-coded as indicated in the legend at the left; deletion indicates focal deletion of all or part of the gene. ERBB2 copy alterations are separated into focal high-level amplification (eight or more copies) or focal gain (5–7 copies). Bar plots at right indicate the percentage of patients with alterations in each gene in this cohort (SN; n = 63) or de novo breast cancers from Nik-Zainal (n = 560). The top indicates the proportion of indels attributed to each indel signature, followed by SNV signatures below. Sequencing platforms, radiation treatment (given for childhood cancer prior to SN development), and clinical HER2 and ER status are also shown at the top.

Nineteen of sixty-three breast SNs (30%) lacked alterations in the 15 driver genes shown in Fig. 6, though these samples had characteristic breast cancer aneuploidies (Supplementary Fig. S17). Thirteen percent of de novo breast cancers (19) also lacked alterations in these 15 driver genes, lower than the 30% in SNs, a difference that was driven largely by TP53 enrichment in de novo (as 17% of de novo cases had TP53 as the only driver alteration).

Clonal Heterogeneity in Multisample Patients

We analyzed clonal heterogeneity in 20 multisample patients (2–4 samples per patient; median 2) with exome sequencing, including 10 patients with breast cancer, 5 patients with meningioma, and 5 patients with thyroid SN (Supplementary Figs. S18–S21). For all but two patients, samples were collected temporally proximate (usually on the same day; always within 2 months) and spatially proximate (same laterality for breast and thyroid; same brain/skull region for meningioma; spatial details were unknown for two patients with thyroid; Supplementary Table S1). For two patients with meningioma (“A” and “B” in Supplementary Fig. S18A), samples were collected 9 to 19 years apart from multiple locations.

In each patient, we classified SNVs as ubiquitous (detected in all samples, indicating early mutations), shared (multiple but not all samples), and private SNVs (one sample, indicating later mutations; Supplementary Fig. S18A). Meningioma had the greatest heterogeneity, with only 22% of SNVs per patient being ubiquitous on average, compared with 84% in breast and 73% in thyroid SNs [Supplementary Fig. S18A (right)].

In 19 of 20 patients, we confidently reconstructed clonal evolution structures by clustering SNVs into truncal SNVs (present in all clones) or later evolutionary clusters using mutation variant allele frequency (VAF) and copy status, as we have done previously (Supplementary Fig. S18B–S18D; “Methods”; refs. 36, 42, 44, 45). For breast SNs, in eight of nine patients, driver alterations were truncal, including in PIK3CA, TP53, MAP3K1, and CDH1, as were key copy gains such as 1q, 8q, and 17q (Supplementary Fig. S18B, where branch length is proportional to SNV number, and clone sizes are proportional to area in ovals). By contrast, driver alterations were exclusively truncal only in two of five patients with meningioma [Supplementary Fig. S18C (left)], with three of five having heterogeneity in NF2 alterations, including one patient with multiple independent NF2 alterations [Supplementary Fig. S18C (right)]. Two of five patients with meningioma lacked truncal SNVs [Supplementary Fig. S18C (right)], suggesting genetically independent tumors consistent with multifocal disease (46). This heterogeneity was not driven solely by the inclusion of two temporally distinct patients (“A” and “B,” Supplementary Fig. S18A), as one of these two (“A,” patient 003099) had evidence of two clonally independent tumors acquired only 3 months apart [D4 and D2, Supplementary Fig. S18C (middle)]. Thyroid SNs showed less heterogeneity, with truncal drivers in four of five patients [Supplementary Fig. S18D (left)], although one patient had two thyroid tumors with different drivers (RET fusion and hyperdiploidy) and different transcriptional clustering despite sharing most SNVs [Supplementary Fig. S18D (right)].

Discussion

To our knowledge, we provide results from the most comprehensive genomic study to date of matched germline and tumor tissue for 200 SNs among survivors of childhood cancer exposed to radiation and chemotherapy, providing new insights into the mechanisms of SN development, the leading cause of late mortality in childhood cancer survivors. Overall, Fig. 7 synthesizes the effects of DNA-damaging therapy on the risk for cancer development, based on this and prior studies.

Figure 7.

Impact of DNA-damaging therapies on SN genomes and possible modifiers. Schematic of the impact of various treatments on SN genomes, including variant types, signatures, and driver alterations. SN tissue types affected by each signature and mutation type are indicated by gray shapes shown below the “ID5” text, from left to right: breast, meningioma, and thyroid SNs.

Radiation primarily induces indels (24, 26, 47), consistent with an association with ID5 in our study, as seen with Chernobyl-associated thyroid cancers (26), and structural rearrangements (24, 48, 49), consonant with the enrichment of NF2 rearrangements and kinase fusions in SNs (Fig. 7). However, radiation does not seem to be the major driver of SNVs in SNs (26, 47). Based on our analysis of intrapatient heterogeneity in meningioma SNs (Supplementary Fig. S18), brain irradiation may lead to multiple independent malignancies rather than disease arising from a single clone, suggesting the need to optimize radiation dosage and localization when feasible. One limitation of this study is the use of FFPE data, making it difficult to assess radiation-associated SV signatures due to high SV dropout (Supplementary Fig. S6C) and copy-number signatures due to oversegmentation of FFPE WGS profiles. Future studies of SNs focused on fresh-frozen tissue would be valuable in elucidating these signatures.

In contrast to radiation, alkylating agents, including nitrogen mustards and platinums, seem to primarily induce SNVs in SNs (Fig. 7). Surprisingly, no nitrogen mustard-associated SNV signature has been previously discovered in relapsed cancers treated with such agents, including cyclophosphamide (36, 50), despite evidence that cyclophosphamide induces SNVs in human iPS cells based on a study by Kucab and colleagues (51). Our 4-OH-cyclophosphamide signature differed from the cyclophosphamide signature of Kucab and colleagues (Supplementary Fig. S11E), potentially due to (i) the short treatment duration in their study (3 hours), lasting less than a full cell cycle, which may mitigate a drug’s full mutagenic effects (vs. 5 weeks in our study); (ii) variable metabolism of the cyclophosphamide prodrug by rodent liver (vs. active metabolite in our case); and (iii) the use of iPS cells, which have enhanced DNA repair compared with differentiated cells. These factors may explain the absence of their signature in cyclophosphamide-treated relapsed cancers. Furthermore, our findings are consistent with a recent study analyzing the effects of chemotherapy treatment in pregnant patients with cancer, in which the cyclophosphamide analog maphosphamide induced an SBS5-like signature in HSPC experimentally (37). They are also consistent with a study of therapy-related myeloid neoplasms, showing the enrichment of age-related mutational signatures in these neoplasms (52), which our results suggest may be promoted by therapy.

Whereas the platinum SNV signature shows clear enrichment for specific trinucleotide contexts and mutation classes (29), the nitrogen mustard signature may be computationally difficult to distinguish from endogenous mutational processes (such as SBS5 and other flat signatures), leading to previous underappreciation of its impact on SNs and relapsed cancers. As cyclophosphamide is a known risk factor for SN development (53), these findings suggest potential further optimization of cyclophosphamide treatment regimens to prevent SNs. The convergence of exogenous (cyclophosphamide) and endogenous mutational processes is analogous to the prior finding that 5-fluorouracil–induced mutagenesis resembles COSMIC signatures SBS17a/b, which are associated with oxidative DNA damage (54). Furthermore, we note that the experimental 4-OH-cyclophosphamide signature had only moderate cosine similarity to flat COSMIC signatures, perhaps due to differences in clock-like mutational processes in cultured cells, which instead may have model-specific background mutational signatures as seen by us [i.e., control cells in Supplementary Fig. S11A (top)] and others (51). The enrichment for SBS5 in patient data with simulated 4-OH-cyclophosphamide treatment, relative to other flat signatures (Supplementary Fig. S12), may reflect a preference for SBS5 in the computational deconvolution of admixed signatures.

The lack of a dose–response relationship between SBS5 and cyclophosphamide dose in SNs (Supplementary Fig. S8C) may be due to the variable combinations given with cyclophosphamide (Supplementary Table S17), as the addition of cytarabine significantly decreased 4-OH-cyclophosphamide–induced mutagenesis relative to 4-OH-cyclophosphamide alone in MCF10A cells (Fig. 2E). For example, among 18 cyclophosphamide-treated meningiomas with high-quality WGS, we were able to ascertain the chemotherapies coadministered at the same time as cyclophosphamide in 10 patients (with the other 8 on treatment protocols with a known cyclophosphamide dose but variable or uncertain timing of the drug). Among these 10 patients, cyclophosphamide was coadministered with seven unique drug combinations, with five of these combinations given only to a single patient (e.g., cyclophosphamide plus doxorubicin and cyclophosphamide plus methotrexate plus mercaptopurine; Supplementary Table S17).

We also note the lack of a procarbazine-associated mutational signature in SNs, despite the inclusion of survivors with prior procarbazine treatment (e.g., five patients with thyroid SN; Supplementary Table S1). We and others previously identified the procarbazine-associated COSMIC signature SBS25 in normal blood (55, 56). Procarbazine is associated with second cancer development, particularly gastrointestinal malignancies (57) and leukemia (58). We speculate that procarbazine, as an oral drug, may thus have the highest exposure in the gastrointestinal tract and blood, whereas its exposure in solid tissue may not be sufficient to generate a strong mutational signature. Relatedly, further work is needed to determine the tissue-specific effects of various chemotherapies, which some researchers have investigated with regard to platinum treatment (bioRxiv 2025.10.12.681852).

Platinum treatment is associated with SN development in childhood cancer survivors (59), including meningioma (60). Our findings suggest that platinums may contribute to meningioma development through direct induction of NF2 splice-site SNVs (Fig. 7). Platinums may be more likely to induce NF2-inactivating splice-site SNVs compared with nonsense inactivating SNVs, due to their predilection to mutate at splice donor (GT) and acceptor (AG) motifs (Fig. 4C–E), with lesser capability of generating new stop codons. NF2 is a large gene with 16 exons, providing a substantial number of splice sites for platinum-induced mutagenesis. This phenomenon, combined with the observation of radiation-associated NF2 rearrangements, indicates that individual tissues retain the need for specific genes to be altered for malignancy development, independent of the mutagenic insult (whether endogenous or environmental).

Indeed, oncologists treating patients with SN should expect a similar spectrum of driver alterations compared with de novo cancers, though with some deviations in their frequencies (Figs. 4 and 5), potentially due to therapy-related mutational processes. Together, these results may guide future efforts to treat or prevent SNs and provide a resource for understanding the etiology of these diseases.

Methods

Patient Samples and Genomic Sequencing

SN tumor samples were obtained as FFPE scrolls or slides through the CCSS Biopathology Center at Nationwide Children’s Hospital. Samples selected for sequencing had ≥30% pathologically assessed tumor purity and ≥75% non-necrotic tissue in tumor regions. Pathology reports were reviewed to determine if multiple specimens from an individual were likely to represent relapse or an independent SN. Germline and tumor WGS and tumor-exome sequencing were performed primarily by the HudsonAlpha Institute for Biotechnology, with library preparation performed using either the FFPE Exome Library KAPA Hyper Prep protocol, Whole Genome Library KAPA Hyper Prep protocol, Swift 1S protocol, or FFPE Exome Whole Genome Library NEBNext Ultra II protocol, followed by sequencing performed using an Illumina NovaSeq 6000 instrument (RRID: SCR_016387). Germline WGS was obtained in some cases through the St. Jude Lifetime Cohort study. Germline exome data were obtained from dbGaP accession phs001327.v2.p1. RNA-seq was performed by the Computational Biology Genomics Laboratory at St. Jude Children’s Research Hospital using Illumina Stranded Total RNA Prep, Ligation with Ribo-Zero Plus, and sequencing on an Illumina NovaSeq 6000. Written informed consent to use tissue for research was obtained from patients and/or legal guardians for children, and all patients were reconsented at the age of majority. Research was reviewed and approved by the Institutional Review Board of St. Jude Children’s Research Hospital. This study complies with the Declaration of Helsinki and all other relevant ethical regulations.

Matched fresh-frozen and FFPE exome data from 33 pediatric cancers, along with matched germline exome data, were used to assess FFPE artifacts. These samples were obtained through the St. Jude Children’s Research Hospital Genomes for Kids study (61). Written informed consent to use these samples was obtained from patients and/or legal guardians under approval from the Institutional Review Board of St. Jude Children’s Research Hospital, in compliance with the Declaration of Helsinki and all other relevant ethical regulations.

Matched fresh-frozen and FFPE WGS data from patients with ovarian cancer were obtained through dbGaP accession phs003036.

Treatment History

Prior treatments analyzed consisted of those given within 5 years of childhood cancer diagnosis. Although treatment information for the SNs themselves was unknown, minimal radiation or chemotherapy would have been given before most SNs’ sample acquisition, as 85% of SN samples were collected at SN diagnosis or less than 1 month after SN diagnosis, among samples with known diagnosis and acquisition dates.

SNV and Indel Mutation Identification and Burden

Somatic SNVs and indels were identified by first aligning exome and WGS data to GRCh37 using BWA (sampe; ref. 62) version 0.7.15.

SNVs were identified using three callers: Bambino (63) version 1.6, Mutect2 (GATK version 4.1.8.0; ref. 64), and Strelka (65) version 2.9.10. Exome- and WGS-detected SNVs required Bambino plus one other caller to be included as valid. SNVs and indels in repetitive regions of the genome were excluded from analysis.

Indels were identified after BWA alignment as described for SNVs, followed by indel calling with four callers: Bambino (63) version 1.6, Mutect (GATK version 4.1.8.0; ref. 64), Strelka (65) version 2.9.10, and SvABA (66) version 1.1.3. Exome-detected indels had to be detected by at least two callers (including Bambino as one of them) to be considered valid. WGS-detected indels had to be detected by at least two callers.

FFPE tissue preparation is known to cause artifactual mutations (27). To filter out FFPE artifacts, we excluded SNVs and indels with VAF below 0.2 or fewer than six mutant reads (except for indels in exome data, for which we only filtered out indels with VAF below 0.1). These thresholds were effective in removing FFPE-related SNV and indel artifacts based on datasets of matched fresh-frozen versus FFPE exome data (n = 33; Supplementary Fig. S2) and matched fresh-frozen versus FFPE WGS data (Supplementary Fig. S6). A lower VAF threshold (0.1) was used for cross-cohort exome indel burden comparison, as a higher threshold would have resulted in a majority of samples with 0 coding indels in several cohorts (e.g., the Clark cohort; ref. 20). Given that FFPE SN samples analyzed here did not exhibit indel superabundance (evidence of extensive FFPE indel artifacts) and given the high concordance of the indel burdens from matched fresh-frozen and FFPE exomes (Supplementary Fig. S2C), a lower indel VAF threshold was justified for generating meaningful results for the cross-cohort comparison of indel burdens in coding regions (Fig. 1E), without compromising data quality.

SNV and indel burdens were determined from exome data for our SN samples and for the comparison cohorts by downloading raw exome data from TCGA (dbGaP accession phs000178; a random selection of 100 breast and 100 thyroid cancers was used; refs. 17, 18), de novo meningioma (kindly provided by the authors; ref. 20), and RIM (EGA accession EGAS00001002317; ref. 16) cohorts. Comparisons with TCGA mutation calls were based on SNV and indel files downloaded from the GDC Data Portal. For SNV and indel burden calculations, we included only variants in coding regions of exons (including silent SNVs) as determined from UCSC’s hg19 refGene.txt annotation, while excluding variants in introns and untranslated regions. SNV and indel burdens were normalized to the length of the coding regions of the genome (SNVs per Mb or indels per Mb).

MutSigCV (67) and dNdScv (68) were used to identify significantly mutated genes based on SNVs and indels, though no novel genes were identified.

The effects of downsampling on SNV burden were assessed by running Picard version 2.9.4 using the DownSampleSam function on aligned BAM files, followed by SNV calling using the abovementioned callers.

Doublet Mutation Identification

Doublet variants were identified after BWA alignment as described for SNVs, followed by variant calling with three callers: Bambino (63) version 1.6, Strelka (65) version 2.9.10, and FreeBayes (arXiv 1207.3907v2) version 1.2.0. High-quality variants called by at least two callers were included in the analysis.

Copy-Number Variation Analysis

Somatic copy-number variation analysis was performed on WGS and exome data using CNVKit (69) version 0.9.10, following alignment with BWA (sampe) version 0.7.15 to GRCh37. For WGS data, the CNVKit option “-m wgs” was used. Each sample’s copy profile was manually reviewed to correctly center the two-copy (diploid) regions at the copy value of 2, if needed, based on allelic imbalance information and copy level, as done previously (70). Focal copy-number gains and losses in driver genes were identified as those spanning 5 Mb or less.

SV Identification

SV analysis was performed on WGS data to identify NF2 disruptions, driver fusions, and SV burdens. Raw WGS data were aligned to hg19 using BWA (mem) version 0.7.12. Somatic SVs were identified using Manta (71) version 1.6.0 or by searching for split reads within known thyroid cancer fusion genes (18). SVs were manually reviewed in BamViewer (63). For the comparison of SV burdens between matched fresh-frozen and FFPE ovarian cancer data (Supplementary Fig. S6C), SVs were filtered to include only those with at least 10 supporting split reads.

Fusion Identification

To identify fusions from RNA-seq data, raw reads were first mapped to GRCh37 using StrongArm (72). Fusions were then identified using Cicero (73) and reviewed in GenomePaint (74) to verify that orientation and exonic structure matched those of known functional driver fusions in thyroid cancer (18) or were disruptive of NF2. We also identified fusions from RNA-seq using Fuzzion2 (75) version 1.3.0, which identifies driver fusions directly from FASTQ input by searching for split reads matching patterns of known driver fusions.

Germline Cancer–Predisposing Variant Identification

Germline cancer–predisposing variants were prioritized by running MedalCeremony (28, 76, 77) on SNVs and indels called from WGS or exome sequencing of germline samples and corroborated in tumor sequencing to determine loss of heterozygosity (LOH) status. Gold and silver medal variants were further annotated using VEP (78) and ClinVar. The candidate variants underwent manual inspection and internal expert curation to generate the final list of pathogenic/likely pathogenic variants.

Transcriptional Clustering and Analysis

Transcriptional clustering to generate t-distributed stochastic neighbor embedding (t-SNE) plots was performed by first obtaining gene expression counts from raw RNA-seq data with the Rsubread package in R, using the UCSC hg19 genome and annotation. The resulting gene expression count matrix (excluding genes on sex chromosomes) was transformed using varianceStabilizingTransformation in R. t-SNE was performed on the 1,000 genes with the top mean absolute deviation using Rtsne, with a perplexity of 20 and 10 initial dimensions. Only samples with at least 20% of exons covered at 20× coverage or more in RNA-seq were included. PAM50 subtypes were determined using the genefu (79) package in R.

Mutational Signature Analysis

SNV, indel, and doublet signature analyses were performed using SigProfiler version 2.3.1 (for de novo signature extraction) or SigProfilerSingleSample version 1.3 (to determine signature presence and strength in individual samples), with SBS1 and SBS5 allowed in all samples regardless of rules or sparsity, and SBS2 and SBS13 as connected signatures for SNV signature analysis. Only WGS samples with ≥80% of exons covered at 20× or higher were analyzed [as a surrogate for sufficient and even coverage throughout the genome; this corresponded approximately to ≥85% of the genome covered at 20× or higher (referred to as coverage breadth; Supplementary Fig. S22A)], to exclude samples with uneven coverage leading to a lack of effective detection of genome-wide SNVs and indels. This filtering eliminated the significant correlation between sequencing breadth that was present before this filtering step was done (Supplementary Fig. S22B and S22C). Furthermore, we excluded samples with less than 50% tumor purity (measured by copy-number and mutation VAF density; refs. 44, 45), as the high VAF and mutant read count filtering strategy to remove FFPE artifacts (VAF of ≥0.2 and ≥6 mutant reads in tumor samples) required high tumor purity to properly assess mutational signatures quantitatively. COSMIC version 3.4 signatures were used. De novo signature extraction resulted in extracted signatures that were mixtures of COSMIC SBS and ID signatures for SNV and indel signatures, respectively, for which presence we tested in patient samples using SigProfilerSingleSample. Doublet signatures did not resemble COSMIC signatures (cosine similarity <0.80), and so the three extracted signatures were analyzed using SigProfilerSingleSample. The probability that NF2 SNVs were induced by SBS signatures was calculated as we described previously (42, 43). Cosine similarities were calculated using the lsa package in R.

In silico addition of the MCF10A 4-OH-cyclophosphamide spectrum to patient SN profiles, to simulate 4-OH-cyclophosphamide treatment (Supplementary Fig. S12), was performed as follows: The mean MCF10A 4-OH-cyclophosphamide profile (after subtracting the control background signature, as shown in Fig. 2G) was normalized to 1. This 96-channel profile was then multiplied by 734 (for breast SNs) or 281 (for meningioma SNs), as these values represent the median increase in SBS5 in each of these cancer types in samples treated with nitrogen mustards (Fig. 2B and C); each value in the resulting 96-channel vector was rounded to an integer. Then, this vector was added to the original somatic SNV spectra of each nitrogen mustard-untreated breast or meningioma SN (having high-quality WGS) to simulate 4-OH-cyclophosphamide treatment. SigProfilerSingleSample was then run on these simulated profiles to measure the abundance of the eight COSMIC SNV signatures (shown in Fig. 2A) for the simulated profiles, in the same way as it was performed for the original SNV spectra from patients with SNs. The cosine similarities of the original and simulated profiles reconstructed using the eight COSMIC signatures are shown in Supplementary Fig. S12B and S12C (boxplots at right). For each SN sample, the change in each SNV signature’s abundance (quantified as the number of SNVs) was also quantified in the simulated versus original SN profiles to identify cases in which a signature increased. The sum of all of these increases for each signature across breast and meningioma SNs is represented as pie charts in Supplementary Fig. S12D and S12E to indicate which signatures increased most upon simulated 4-OH-cyclophosphamide treatment.

Experimental 4-OH-cyclophosphamide Mutational Signature

MCF10A cells were plated and treated for 5 weeks with 1 μmol/L 4-OH-cyclophosphamide (4-Hydroxy Cyclophosphamide Preparation Kit, Toronto Research Chemicals, cat. #H926301), 25 nmol/L cytarabine (Selleck Chemicals, cat. #S1648), 0.5 μmol/L cisplatin, 1 μmol/L 4-OH-cyclophosphamide combined with 10 nmol/L cytarabine, or untreated control. MCF10A cells (ATCC cat. #CRL-10317, RRID: CVCL_0598) were grown in Mammary Epithelial Growth Medium (Lonza Bioscience, CC-3150) with cholera toxin (100 ng/mL). Three biological replicates (wells of six-well plates) were treated for each drug or drug combination. After 5 weeks of treatment, one single-cell clone per well (3 per drug or drug combination) was isolated, followed by genomic DNA isolation and WGS. The pretreatment bulk control MCF10A cell line was also analyzed by WGS and used as a “germline” to subtract out preexisting SNVs. Somatic SNVs (found only in treated clones and not in pretreatment bulk) were called and filtered as described for SN samples. To subtract background mutagenesis present during normal cell culture, the mean of the three control clones’ SNV spectra was subtracted from those of the drug-treated clones (with negative values set to zero), and the resulting profiles were used for the cosine comparisons in Fig. 2F and for the plots in Fig. 2G and Supplementary Fig. S10.

The above drug doses were selected based on dose–response assays in MCF10A cells (Supplementary Fig. S9), which were performed using the MTT assay for 4-OH-cyclophosphamide and cisplatin or CellTiter-Glo for cytarabine. For MTT assays, MCF10A cells were seeded in 96-well plates in the medium mentioned above, and the drug was added in the medium the following day with three technical replicates per dose. Four days later, the MTT assay was performed by adding 50 μL per well of 5 mg/mL MTT, sterile-filtered in PBS. After 1 to 2 hours of incubation at 37°C, the medium was removed, and 100 μL per well of DMSO was added. Plates were mixed to dissolve formazan crystals, followed by reading the absorbance at 560 nm. Viability was normalized to vehicle-treated cells. For the CellTiter-Glo assay (cytarabine), cells were plated and treated as above with six technical replicates per dose, in 96-well white-walled plates. After 4 days of drug treatment, CellTiter-Glo was added to each well, and the plates were mixed for 2 minutes with gentle shaking. Luminescence was then read using a 1-second integrated reading, and viability was normalized to vehicle-treated cells.

Clonal Heterogeneity Analysis

Clonal heterogeneity of multisample patients was analyzed, as we have done previously (36, 42, 44, 45), by clustering SNVs into truncal SNVs (present in all clones) or later evolutionary clusters. Somatic SNVs in each patient were first clustered by their presence or absence in each sample (at VAF 0.05 or above) and further subdivided into additional clusters based on the presence of multiple VAF density peaks within the cluster (Supplementary Figs. S19–S21). A lower VAF threshold was used in these multisample patients (0.05, as opposed to the threshold of 0.2 used for mutation burden analysis) due to the lack of mutation-superabundant samples in this subcohort (which superabundance is indicative of FFPE artifacts) and high concordance between exome and WGS SNVs among samples with matched exome and WGS data among these multisample patients (in which 93% of exome-detected SNVs were also detected in WGS among SNVs with at least 95% likelihood of being detected in at least 1 read in WGS, based on beta-binomial predictive modeling), indicating minimal FFPE-related SNV artifacts. Clusters with only one SNV (or with 1–3 SNVs that contradicted clonal deconvolution ascertained from higher-SNV clusters) were generally excluded from analysis. For 15 of 19 patients, we constructed evolutionary trees based solely on somatic SNVs found in 2-copy regions in all samples within the patient, as there were a sufficient number of these to analyze in these 15 patients. In these patients, VAFs were adjusted for tumor purity [VAFcorrected = VAF/(tumor purity)], and CCFs of each SNV cluster (in a given sample) were determined by multiplying the median corrected VAF for that cluster (in a given sample) by 2. CCFs of subclones were normalized to the CCF of the truncal variants, with the latter being set to 1. In four patients, analysis was expanded beyond SNVs in 2-copy regions due to widespread aneuploidy leading to few 2-copy-region SNVs (these patients lack VAF density plots and have asterisks in Supplementary Figs. S19–S21). In these patients, the multiplicity of SNVs was uncertain as SNVs in, for example, a 4-copy region could be found on 1, 2, 3, or 4 copies. Therefore, the CCFs of clones in these samples were estimates, as denoted by asterisks in Supplementary Figs. S18–S21.

Functional Analysis of NF2 Splice Variants

NF2 exons 11 to 12, followed by intron 12 and then exons 13 to 14 (NM_000268), were cloned immediately 5′ of GFP (mGFP specifically) in the pLenti-C-Myc-DDK-P2A-Puro-mGFP plasmid (PS100123, obtained from OriGene). NF2 exon 11 begins with an in-frame ATG, forming the new start codon for the resulting chimeric NF2-GFP open reading frame, which is in frame when intron 12 is properly spliced out but out of frame otherwise. This construct was further modified to introduce the NF2 exon 12 splice site variant (chr22, position 30069476, G > T in GRCh37 coordinates, or +1 splice donor position in exon 12) found in meningioma 416414_D1, or the NF2 exon 13 splice site variant (chr22, position 30070824, G > A in GRCh37, or −1 splice acceptor position in exon 13) found in meningioma 413349_D1. 293T cells were plated in six-well plates and transfected with WT NF2-GFP and the two splice mutants using LipoD293 (SignaGen Laboratories). After 2 days, brightfield and green fluorescence were visualized using an EVOS FL microscope (Life Technologies). We also isolated RNA from these cells, and RNA-seq was performed using the Illumina Total Stranded RNA prep.

Statistics and Reproducibility

Genomic sequencing was performed on each sample one time with no technical replicates. Two-sided tests were used for all statistical tests, if applicable. P < 0.05 was considered statistically significant. The statistical test used in each case is described in the main text or figure legends.

Multivariate linear regression was performed using the “lm” function in R for age and mutation burden comparisons in Fig. 1 and signature associations in Figs. 2 and 3. For the age comparisons in Fig. 1D, covariates for breast cancer included stage, histology (ductal vs. lobular), ER status, and HER2 status; for meningioma, covariates were grade and sex; and for thyroid cancer, covariates were histology (papillary vs. follicular) and sex (Supplementary Table S8). For the SNV and indel mutation burden comparisons in Fig. 1C and E, covariates for breast cancer included age, stage, tumor purity, ER status, HER2 status, and histology; for meningioma, covariates were age, grade, tumor purity, and sex; and for thyroid cancer, covariates were age, histology, tumor purity, and sex (Supplementary Tables S7 and S12). Multivariate linear regression to compare SBS5 burdens in Fig. 2 used, as covariates, in breast cancer: age, breast cancer stage, ER status, HER2 status, and tumor purity (Supplementary Table S14) and in meningioma: age, sex, grade, and tumor purity (Supplementary Table S15). Individual drugs (or drug classes if drugs belonged to a drug class such as anthracyclines with multiple drugs in the class used in the cohort) and prior treatment status (binary; yes or no) were included in the model if at least five SNs in the cancer type had received the treatment. The resulting P values for each drug or drug class were multiplied by the number of drugs (or drug classes) assessed within each cancer type (Bonferroni correction) for multiple hypothesis correction, and these corrected P values are shown in Fig. 2 and Supplementary Tables S14 and S15. For four breast SNs lacking known clinical ER and HER2 status, these were inferred using PAM50 subtypes, with luminal A considered ER+/HER2−, basal considered ER−/HER2−, and HER2-enriched considered ER−/HER2+ (for the SBS5-drug correlations in breast SNs and ID5-radiation correlations in breast SNs). For the radiation dose correlation with indel signatures, covariates for breast cancer were age, breast cancer stage, ER status, HER2 status, and tumor purity; in meningioma: age, sex, grade, and tumor purity; and in thyroid cancer: age, tumor purity, sex, and histology (Fig. 3B; Supplementary Table S19). For the combined analysis including all three cancers (Fig. 3C; Supplementary Table S20), covariates included age, tumor purity, sex, and cancer type. For the correlation between cyclophosphamide dose and SBS5 in meningioma (Supplementary Fig. S8C), covariates included age, sex, grade, and tumor purity (Supplementary Table S16).

To compare the frequency of driver alterations across cohorts in Figs. 4A, 5A, D, and 6, multivariate logistic regression was used via the “logistf” function in R. For meningioma driver alteration frequency, covariates included age, sex, and grade, and variants occurring in at least 5% of samples were analyzed (Supplementary Table S22). For thyroid cancer driver alteration frequency, covariates included age, sex, and histology (Supplementary Tables S24 and S25). For breast cancer driver alteration frequency, covariates included age, breast cancer stage, ER status, HER2 status, and histology (Supplementary Table S26). A Bonferroni correction was performed for each P value by multiplying the P value by the number of alterations analyzed, as indicated in the aforementioned Supplementary Tables.

Supplementary Material

Supplementary Tables S1-S26

Supplementary Table S1 shows SN sample clinical variables and metadata. Supplementary Table S2 shows a comparison of variables between good-quality vs. excluded samples. Supplementary Table S3 shows a summary of prior treatments across SN patients. Supplementary Table S4 shows a summary of original childhood cancer diagnoses for SN patients. Supplementary Table S5 shows metadata for 33 pediatric cancers with matched FFPE and fresh-frozen exome data. Supplementary Table S6 shows a list of coding-region somatic SNVs used for mutation burden analysis. Supplementary Table S7 shows multivariable analysis comparing SNV burdens between cohorts. Supplementary Table S8 shows multivariable analysis comparing age between cohorts. Supplementary Table S9 shows multivariable analysis comparing SNV burdens between thyroid cancer cohorts among samples with coverage below 100x. Supplementary Table S10 shows cancer-predisposing germline alterations among SN patients. Supplementary Table S11 shows a list of coding-region somatic indels used for mutation burden analysis. Supplementary Table S12 shows multivariable analysis comparing indel burdens between cohorts. Supplementary Table S13 shows SNV signature levels among SNs. Supplementary Table S14 shows multivariable analysis comparing SBS5 burdens in breast SNs stratified by prior treatment. Supplementary Table S15 shows multivariable analysis comparing SBS5 burdens in meningioma SNs stratified by prior treatment. Supplementary Table S16 shows multivariable analysis testing SBS5-cyclophosphamide dose-response relationship in meningioma SNs. Supplementary Table S17 shows variation in cyclophosphamide-containing regimens among meningioma SNs. Supplementary Table S18 shows indel signature levels among SNs. Supplementary Table S19 shows multivariable analysis testing ID5-radiation dose-response relationship in each SN type. Supplementary Table S20 shows multivariable analysis testing ID5-radiation dose-response relationship with all 3 SN types combined. Supplementary Table S21 shows NF2-disrupting structural variants detected in meningioma SNs. Supplementary Table S22 shows multivariable analysis comparing driver alteration frequency between meningioma cohorts. Supplementary Table S23 shows driver fusions in thyroid SNs. Supplementary Table S24 shows multivariable analysis comparing mutation group frequency between thyroid cancer cohorts. Supplementary Table S25 shows multivariable analysis comparing driver alteration frequency between thyroid cancer cohorts. Supplementary Table S26 shows multivariable analysis comparing driver alteration frequency between breast cancer cohorts

Supplementary Figures S1-S22

Supplementary Figure S1 shows SN sample exclusion, sequencing coverage, and sequencing types. Supplementary Figure S2 shows filtering approaches for exome FFPE artefacts in 33 pediatric tumors with matched FFPE vs. fresh-frozen data. Supplementary Figure S3 shows the correlation between SNV burden in SN exome vs. WGS data. Supplementary Figure S4 shows TCGA exome SNV and indel burden analysis for de novo cancer cohorts. Supplementary Figure S5 shows coverage data and absence of significant coverage effects on mutation burdens. Supplementary Figure S6 shows a comparison of SNV, indel, and SV burdens in 8 ovarian tumors with matched FFPE vs. fresh-frozen data. Supplementary Figure S7 shows de novo signature extraction data for SNV and indel signatures. Supplementary Figure S8 shows SNV signature data for thyroid SNs and dose-response analysis in meningioma SNs. Supplementary Figure S9 shows drug dose response curves in MCF10A cells. Supplementary Figure S10 shows the cisplatin signature in MCF10A cells. Supplementary Figure S11 shows additional details of experimental drug signatures and comparison signatures from other experimental studies. Supplementary Figure S12 shows the effects of simulated 4-OH-cyclophosphamide treatment on SNV signatures in patient tumors. Supplementary Figure S13 shows doublet signatures in SNs. Supplementary Figure S14 shows meningioma SN copy profiles. Supplementary Figure S15 shows details of functional NF2 splicing assays. Supplementary Figure S16 shows how probabilities were predicted for NF2 variant induction by platinum signatures. Supplementary Figure S17 shows breast SN copy profiles. Supplementary Figure S18 shows clonal evolution trees for breast, meningioma, and thyroid SNs. Supplementary Figure S19 shows details of clonal evolution in breast SNs. Supplementary Figure S20 shows details of clonal evolution in meningioma SNs. Supplementary Figure S21 shows details of clonal evolution in thyroid SNs. Supplementary Figure S22 shows selection of high-quality WGS samples for mutational signature analysis based on coverage breadth

Data Availability

The CCSS is a US NCI-funded resource (U24 CA55727) to promote and facilitate research among long-term survivors of cancer diagnosed during childhood and adolescence. Raw sequencing data for SNs and matched germline samples are available through the dbGaP repository phs001327. In addition, utilization of the CCSS data that leverages the expertise of CCSS Statistical and Survivorship research and resources will be considered on a case-by-case basis. For this utilization, a Research Application of Intent followed by an Analysis Concept Proposal must be submitted for evaluation by the CCSS Publications Committee. Users interested in utilizing this resource are encouraged to visit http://ccss.stjude.org. Matched fresh-frozen and FFPE exome raw data for 33 pediatric tumors and associated germline samples are available on St. Jude Cloud (76), with relevant sample identifiers indicated in Supplementary Table S5.

Authors’ Disclosures

S.W. Brady reports other support from American Lebanese Syrian Associated Charities (ALSAC) of St. Jude Children’s Research Hospital and grants from NCI during the conduct of the study. M.A. Khan reports other support from ALSAC during the conduct of the study. M.N. Edmonson reports grants from ALSAC during the conduct of the study. N.V. Terekhanova reports grants from NCI Cancer Center and other support from ALSAC during the conduct of the study. C. Im reports grants from the National Institutes of Health (NIH) outside the submitted work. R.M. Howell reports grants from the NCI during the conduct of the study, as well as grants from the NCI outside the submitted work. W.M. Leisenring reports grants from the NIH during the conduct of the study. G.T. Armstrong reports grants from the NIH during the conduct of the study. No disclosures were reported by the other authors.

Authors’ Contributions

S.W. Brady: Data curation, formal analysis, investigation, visualization, methodology, writing–original draft, writing–review and editing. M.A. Arnold: Data curation, formal analysis, investigation. M. Wang: Data curation, formal analysis, investigation. R. Alsallaq: Data curation, formal analysis, investigation. L. Dong: Formal analysis, validation, investigation, methodology. M.A. Khan: Investigation, methodology. W. Yang: Formal analysis. K.L. Stratton: Data curation, investigation. W. Liu: Formal analysis, investigation. Y. Chen: Formal analysis, investigation. E. Plyler: Investigation. J.A. Steele: Formal analysis, investigation. B.B. Powers: Formal analysis, investigation. D. Rosenfeld: Formal analysis, investigation. M.N. Edmonson: Formal analysis, investigation. Y. Feng: Formal analysis. N.V. Terekhanova: Formal analysis. K. Hagiwara: Formal analysis, methodology. S. Arunachalam: Formal analysis, investigation. H.L. Mulder: Investigation, methodology. D.K. Srivastava: Data curation, formal analysis, investigation. M. Rusch: Resources, software, formal analysis. V.G. Nolan: Resources, formal analysis. A. McDonald: Resources, data curation. Y. Sapkota: Resources, investigation. M.M. Gramatges: Resources, investigation. L.M. Turcotte: Data curation, methodology. C. Im: Data curation, methodology. R.M. Howell: Resources, data curation. J. Easton: Investigation, methodology. X. Ma: Formal analysis. Z. Wang: Formal analysis, investigation. W.M. Leisenring: Resources, data curation. M. Conces: Resources, methodology. J.P. Neglia: Resources, investigation. Y. Yasui: Investigation, methodology. S. Bhatia: Resources, methodology. D.W. Ellison: Resources, data curation, supervision, project administration, writing–review and editing. J. Zhang: Conceptualization, resources, data curation, formal analysis, supervision, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. G.T. Armstrong: Conceptualization, resources, data curation, supervision, writing–original draft, project administration, writing–review and editing.