Early Cancer Detection in Li–Fraumeni Syndrome with Cell-Free DNA

Li–Fraumeni syndrome (LFS; OMIM #151623) is a highly penetrant hereditary cancer disorder caused by a pathogenic germline variant in the tumor suppressor gene TP53 (1, 2). Individuals with germline TP53 mutations have an estimated lifetime risk of ∼75% in males and ∼100% in females of developing at least one cancer (3, 4). The most common cancers in TP53-mutation (TP53m)-carriers are soft-tissue sarcoma, osteosarcoma, brain tumors, breast cancer, and adrenocortical carcinoma (5, 6), although there are a spectrum of additional rare tumors (7). Approximately 50% of TP53m-carriers will develop multiple cancers (3) with a younger age of diagnosis for the first cancer (8) and previous treatment with radiotherapy (9) as factors contributing to the development of subsequent cancers.

Due to their high risk of developing cancer, intensive surveillance with frequent diagnostic imaging, physical examination, and blood tests are recommended for all TP53m-carriers; these have been shown to detect cancers earlier and improve patient outcomes (10, 11). Despite this, ∼20% of TP53m-carriers are often noncompliant, citing several negative contributing psychosocial factors such as inconvenience, cost, logistical barriers, anxiety, and physical and emotional exhaustion (12, 13). Cell-free DNA (cfDNA) analysis, or “liquid biopsy”, is an emerging technology that may help to alleviate logistical barriers and complement (or even replace) current screening modalities. Liquid biopsy relies on the identification of tumor-associated genetic alterations or signatures [circulating tumor DNA (ctDNA)] reflected in DNA fragments released into the blood plasma (14).

With the development of cfDNA sequencing and analysis techniques, studies have demonstrated limits of detection as low as 0.1% for mutations and 3% for copy-number alterations (15–17). More recent studies have shown that ctDNA is often shorter compared with cfDNA from noncancerous cells (18) and can provide complementary information agnostic of genetic alterations (19). Another field of advancement in liquid biopsy is the detection of cancer-associated methylation signatures using cell-free methylated DNA immunoprecipitation sequencing (cfMeDIP-seq) which enriches for 5-methylcytosine, an essential component of tumor and normal cell identity (20–22). The wide scope of analyses, noninvasive collection, and lack of need for specialty medical instruments make liquid biopsy an attractive tool for the monitoring of patients with high-risk hereditary cancer syndromes (HCS) such as LFS. However, cfDNA studies have primarily utilized sporadic cancer patients resulting in a dearth of data in HCS.

Here, we present a cohort of 193 blood samples collected from 89 TP53m-carriers. Using a unified library preparation protocol (∼40 ng cfDNA) that incorporates targeted panel sequencing (TS), shallow whole-genome sequencing (sWGS), and cfMeDIP-seq, we profiled the cfDNA landscape and developed a customized multimodal approach using genome, fragmentome, and methylome analyses to detect cancer-associated signal across a wide spectrum of cancers (Fig. 1A). Analyses were first assessed individually for their performance in detecting ctDNA followed by multimodal and longitudinal analysis.

A retrospective cohort of 193 blood samples was collected from 89 TP53m-carriers (pediatric: individuals = 26, samples = 70, adult: individuals = 63, samples = 123; age = 1–67) recruited from clinics at the Princess Margaret Cancer Centre (n = 57) and the Hospital for Sick Children (n = 32). Fifty-two of these TP53m-carriers belong to 21 LFS families (samples = 102; Supplementary Fig. S1A). The remaining 37 TP53m-carriers were sole representatives within these families. Male TP53m-carriers (n = 26, samples = 50) were found to be disproportionately less represented within the adult population compared with females (n = 63, samples = 143; Supplementary Fig. S1B). TP53m-carriers were classified into three categories: (i) TP53m-carriers who have never had cancer [“LFS Healthy” (LFS-H), individuals = 42], (ii) cancer-free TP53m-carriers with a history of previous cancer [“LFS Past Cancer” (LFS-PC), individuals = 21], and (iii) TP53m-carriers with active cancer [“LFS Active Cancer” (LFS-AC), individuals = 26; Fig. 1B]. Samples from LFS-H and LFS-PC were classified as cancer-negative. All individuals were confirmed to have been diagnosed with LFS through clinical germline testing of TP53 (Supplementary Fig. S1C). A median of three serial blood samples (range = 2–11) were collected from 42 TP53m-carriers (14 pediatric and 28 adult), with the remaining TP53m-carriers (n = 47) only having one blood sample collected (Supplementary Fig. S1D). Eighteen TP53m-carriers, termed “phenoconverters,” transitioned from cancer-negative to cancer-positive (forward; n = 6), cancer-positive to cancer-negative (reverse; n = 9), or developed multiple cancers (n = 3; Fig. 1C). A total of 154 cancer-negative (from cancer-free TP53m-carriers) and 39 cancer-positive (from TP53m-carriers with an active cancer diagnosis) samples were collected. Cancer-positive samples were collected from TP53m-carriers with a wide spectrum of cancers that were initially detected through standard surveillance methods according to the “Toronto Protocol” with follow-up tumor-directed imaging and/or biopsy for diagnostic confirmation (Fig. 1D; ref. 11). Plasma samples were also collected from 30 healthy TP53-wild-type individuals (“Healthy Controls”). Sample descriptions can be found in Supplementary Table S1. Clinical follow-up was available for a mean of 48.1 months (range, 2.3–101.2, median = 46.5) from the first blood collection date and a mean of 35.8 months (range, 2.3–60.5, median = 40.3) from the last blood collection date (Fig. 1E).

DNA extraction yielded a median of 7.94 ng of DNA/mL of plasma (SD = 6.16) with no differences detected between healthy controls (median = 10.0, SD = 5.57) and TP53m-carriers (median = 7.44, SD = 6.19), pediatric (median = 8.47, SD = 5.68) and adult (median = 6.67, SD = 6.49) TP53m-carriers, male (median = 8.37, SD = 4.50) and female (median = 7.23, SD = 6.68) TP53m-carriers, or cancer-negative (median = 7.17, SD = 5.89) and cancer-positive (median = 8.33, SD = 7.20) TP53m-carriers (Supplementary Fig. S1E). However, due to the lower volume of blood collected from pediatric individuals, total DNA yields were much lower in pediatric individuals (median = 14.07 ng, SD = 10.84) compared with adults (median = 67.60 ng, SD = 69.98). In total, 23 samples (pediatric = 10, adult = 13; 1 cancer-positive, 22 cancer-negative) did not yield >10 ng of DNA and did not proceed with sequencing. This resulted in a loss of one forward and one reverse phenoconverter. The remaining 170 samples from 82 TP53m-carriers and 30 healthy control samples were submitted for a combined TS, sWGS, and cfMeDIP sequencing protocol (Fig. 1F; Supplementary Fig. S1F and S1G).

TS was performed on 118 plasma samples from 69 TP53m-carriers (25 individuals with serial samples). A total of 52/170 plasma samples did not have >40 ng of DNA and were excluded from TS (43 pediatric and 9 adult) and 1 sample failed sequencing. Using germline variant identification, we identified 109 germline TP53 variants [single-nucleotide variant (SNV)/indel]. In samples where a germline SNV/indel variant was not detected (n = 8), we performed targeted panel copy-number analysis using PanelCNmops (23) and VisCap (24), which identified a germline TP53 copy-number variant (CNV) in 7 patients (Supplementary Fig. S2A). Comparison between the two panel-based copy-number methods resulted in high concordance (R² = 0.53; Supplementary Fig. S2B). In one patient with a lymphoma (LFS90), we were unable to detect their germline TP53 exon 2–9 duplication using our targeted panel copy-number analysis applied to cfDNA, likely due to the contribution of a somatic TP53 deletion event given that a shallow deletion event was detected in exons 1 and 11. Reflexive sequencing of the patient's matched buffy coat identified the exon 2–9 duplication. In total, we identified germline TP53 variants in 116/117 plasma samples, all of which were concordant with clinical germline testing.

Among the 31 cancer-positive samples from 22 LFS-AC, 20 were from late-stage cancers (stage III/IV), 5 from early-stage cancers (stage 0/I/II), and 6 from tumors with no or unknown stage. TS identified 16 somatic TP53 alterations in 13 samples (41.4%) from 12 individuals: 6 from late-stage and 3 from early-stage cancer patients, resulting in a detection rate of 30.0% (6/20) and 60.0% (3/5), respectively. Additionally, two BRCA2, one PALB2, one MSH6, one PMS2, and three APC somatic pathogenic variants were identified. In cancer-negative samples (n = 86) from 60 cancer-free TP53m-carriers, we identified 20 somatic TP53 variants in 16 samples from 11 individuals and an additional three BRCA1, five BRCA2, two PALB2, three MLH1, four MSH2, four MSH6, and two APC somatic pathogenic variants (Fig. 2A; Supplementary Table S2).

Copy-number analysis of sWGS using ichorCNA (15) in 170 samples (60 pediatric and 110 adult) from 82 TP53m-carriers (23 pediatric and 59 adult) identified 22 samples with detectable tumor fractions (TF; TF >0.03), 7/131 cancer-negative and 15/38 cancer-positive. One sample failed sequencing. By restricting analysis to short fragments (90–150 bp) to enrich for ctDNA (25), ichorCNA identified an additional four TF-positive samples (1 cancer-negative, 3 cancer-positive; Fig. 2B; Supplementary Table S3). Comparing the two ichorCNA methodologies (all fragments vs. short fragments), we observed a strong correlation between the predicted TF from each analysis (R² = 0.91; Supplementary Fig. S2C). Unlike TS, sWGS analysis was not able to detect TP53 CNVs, likely due to the large (1 Mb) bin size used by ichorCNA (Supplementary Fig. S2B). Detection of CNVs was higher in samples from patients with late-stage (13/24, 54.2%) compared with early-stage (1/8, 12.5%) cancers (Supplementary Fig. S3). In one individual, also diagnosed with Turner syndrome (LFS88), tumor fraction prediction was not confounded by the loss of chromosome X (Supplementary Fig. S4A). In contrast, copy-number alterations associated with clonal hematopoiesis confounded tumor fraction prediction in one patient (LFS64, excluded from the study; Supplementary Fig. S4B).

In total, 118 samples had both TS and sWGS data available: 86 cancer-negative and 32 cancer-positive. Within the cancer-positive samples, combined TS and sWGS profiling was able to identify somatic alterations in 20/32 samples (6 TS only, 7 sWGS only, and 7 both), a detection rate of 62.5% (Fig. 2C). The lack of detection in the remaining samples may be due to the relatively high TF required for robust copy-number analysis (3%–5% TF), lack of adequate library diversity for TS, or low shedding of ctDNA.

Studies have shown that ctDNA is enriched for shorter DNA fragments, and analysis of fragment length can be useful for the detection of cancer-associated genetic alterations (18, 25, 26). To further validate somatic variants and to rule out other false positives such as clonal hematopoiesis and sequencing errors, we calculated fragment length–weighted variant scores using a reference set of tumor mutation–informed cancer- and healthy control–associated fragment lengths (26). Comparing the difference in mean fragment length between fragments with variants and wild-type fragments in our LFS cohort, we observed a wider distribution in somatic variants compared with heterozygous germline variants (P = 0.018; Supplementary Fig. S5A). Variant scores for germline variants showed a linear relationship between wild-type and mutant scores (R² = 0.85) whereas somatic variants showed decreased correlation (R² = 0.23) due to increased variability in the mutant variant scores (P = 0.017; Supplementary Fig. S5B). Fragments with germline TP53 variants were found to have a similar size distribution compared with wild-type fragments (P = 0.187), whereas fragments with somatic TP53 mutations were found to be shorter (P < 0.001; Supplementary Fig. S5C). Using this approach, we observed decreased fragment size in one patient (excluded) with a TP53 p.Y220C mutation [variant allele frequency (VAF) = 45.6%] that was classified as germline but was later found to be a somatic variant associated with their leukemia (LFS64; Supplementary Fig. S5D). Conversely, in one patient with a known clonal hematopoiesis of indeterminate potential (CHIP)-associated TP53 p.R175H variant (LFS47; VAF = 5.75%), we observed similar fragment sizes compared with wild-type fragments (P = 0.566; Supplementary Fig. S5D). These observations are consistent with sporadic cancers and provide further confidence that the somatic variants identified in our analysis are likely cancer-associated (18, 27).

Several studies have demonstrated that analysis of fragment size can detect cancer-agnostic genomic alterations (19, 28). We first determined whether the fragmentation score metric (described above) could be used to detect cancer signals in an external cohort of 494 plasma samples across 7 different cancer types (261 healthy controls; ref. 29). Fragmentation scores showed strong correlation to the proportion of short (>150 bp) fragments (R² = 0.94; Supplementary Fig. S6A) but weak correlation to copy-number–informed predicted tumor fraction (R² = 0.14; Supplementary Fig. S6B). Fragmentation score was able to achieve a specificity of 39.9% at 95% sensitivity across all cancer types (range, 12.5%–65.4%) and stages (range, 31.8%–48.6%; Supplementary Fig. S6C and S6D). Next, using the fragmentation score metric in our cohort, we calculated a gene-level TP53 fragmentation score using only reads that mapped to TP53 in our TS and a genome-wide fragmentation score using all reads from sWGS. In both metrics, we observed increased fragmentation scores in TP53m-carriers independent of cancer status compared with healthy controls (Supplementary Fig. S6E). Therefore, we used the 90th percentile of LFS Healthy patients as our detection cutoff. Bootstrap analysis (100 iterations of downsampling) of fragmentation scores in TP53m-carriers, comparing LFS-AC to LFS-H, showed sensitivities comparable to those observed in external validation cohorts (genome: mean = 36.8%, range = 23.3%–48.2%; TP53: mean = 24.1%, range = 9.1%–34.6%; Supplementary Fig. S6F). Gene-level TP53 and genome-wide fragmentation scores showed strong correlation (R² = 0.67; Supplementary Fig. S6G). However, the correlation was not strong when comparing copy-number–informed predicted tumor fraction and fragmentation scores (TP53 R² = 0.09, genome R² = 0.07) and TP53 somatic mutation VAFs and TP53-level fragmentation scores (R² = 0.05; Supplementary Fig. S6H). Using fragmentation scores, we were able to detect an additional 4 cancer-positive samples (2 genome, 1 TP53, 1 both) where genomic alterations were not detected (Fig. 2).

Due to the increased fragmentation scores observed in TP53m-carriers compared with healthy controls, we investigated further by calculating the proportion of short (<150 bp) cfDNA fragments genome-wide. TP53m-carriers had the typical peak at ∼167 bp (LFS-H: mean = 170.9, median = 170, mode = 168; LFS-AC: mean = 169.6, median = 169, mode = 168) similar to healthy controls (mean = 171.1, median = 170, mode = 168). However, compared with healthy controls, we observed an increased proportion of short fragments in TP53m-carriers, independent of cancer status (cancer-negative P = 0.003, cancer-positive P = 0.0005; Supplementary Fig. S7A). This was consistent in TP53m-carriers who have never had cancer (LFS-H; P = 0.027; Supplementary Fig. S7B) and not attributable to a prior history of cancer (P = 0.085) or age (R² = 0.003; Supplementary Fig. S7C). Cancer-positive samples additionally, had increased proportions of short fragments compared with cancer-negative samples (P = 0.037), suggesting that the mechanism of altered fragmentation in cancer cells is conserved between LFS-associated cancers and sporadic cancers. In healthy controls and TP53m-carriers, we observed a strong correlation between the proportion of short fragments and genome-wide fragmentation scores (R² = 0.88–0.96; Supplementary Fig. S7D) and a weak correlation between the proportion of short fragments and copy-number–informed predicted tumor fraction (R² = 0.06; Supplementary Fig. S7E).

To investigate differences in the distribution of fragment lengths in our cohort, we compared the median distribution of samples from healthy controls (n = 30), LFS-H (n = 58), and LFS-AC (n = 38) across four fragment length compartments (10–89, 90–150, 151–220, and 221–320 bp; Fig. 3 and Supplementary Fig. S7F). Random forest classification achieved high performance between cancer-negative TP53m-carriers and healthy controls (AUC = 0.810), and LFS-H and healthy controls (AUC = 0.817; Supplementary Fig. S7G). Compared with healthy controls, LFS-H samples had an increased proportion of fragments within the 10–89, 90–150, and 221–320 bp compartments and decreased proportions in the 151–220 bp compartments. This pattern suggests increased fragmentation within the nucleosome resulting in the decreased size of mononucleosome fragments and increased release of dinucleosome fragments in TP53m-carriers. In contrast, samples from LFS-AC exhibited an increased proportion of fragments in the 90–150 bp and front half of the 221–320 bp compartments, and a decreased proportion of fragments in the 151–220 bp compartments and later half of the 221–330 bp compartments compared with samples from LFS-H. This pattern of differential fragmentation suggests that the fragment length distribution of TP53m-carriers with active cancer shifts toward shorter mono- and dinucleosome-associated fragments, consistent with sporadic cancers (18).

To investigate cancer-associated DNA methylation patterns in TP53m-carriers, we performed cfMeDIP-seq on 169 LFS samples (38 cancer-positive and 131 cancer-negative) from 82 TP53m-carriers (59 adult and 23 pediatric) and 28 healthy controls. Two LFS samples failed library construction and/or sequencing. The remaining samples showed enrichment of methylated Arabidopsis spike-in controls (81.4%–99.5%) and CpG enrichment efficiency (relH = 2.48–3.90, GoGE = 1.65–2.16; Supplementary Fig. S8). Similar to our sWGS fragment analysis, TP53m-carriers, independent of cancer status, showed an increased proportion of short fragments in our cfMeDIP-seq libraries (Supplementary Fig. S9A and S9B) which was correlated with the proportion of short fragments in matching sWGS libraries (R² = 0.58–0.80; Supplementary Fig. S9C). As blood cells from TP53m-carriers have been reported to exhibit methylation patterns distinct from noncarriers (30), we performed differential methylation analysis comparing LFS Healthy patients and healthy controls. This comparison identified 300 differentially methylated regions (DMR; hypermethylated; Supplementary Table S4) which we subsequently excluded from downstream analyses.

TP53m-carriers are at a high risk to develop a wide spectrum of different tumor types. However, due to the challenges associated with curating sufficiently large training data sets, the rarity of some LFS-associated tumor types, and the multiple one-off tumor types in our cohort, we first explored a pan-cancer approach. Using a universal cancer marker set identified using microarray analysis by Vrba and colleagues, we built a pan-cancer signature by mapping the hypermethylated CpGs to the corresponding genome coordinates followed by the removal of immune-cell–specific methylation sites, resulting in 105 DMRs (Supplementary Table S5; ref. 31). The pan-cancer signature was then validated in an external data set (samples = 388, healthy = 86, cancer types = 7; ref. 21) and achieved a sensitivity of 47.3% (range, 10%–71.4%) at 95% specificity (Supplementary Fig. S10A). A similar approach using differentially methylated CpGs from 27 cancer types from The Cancer Genome Atlas (TCGA) was also validated but did not perform as well (sensitivity = 42.7%; range = 5.0%–60.9%; Supplementary Fig. S10B; ref. 32). Using the 95th percentile of pan-cancer methylation scores from healthy controls as the detection cutoff, we detected pan-cancer–associated methy­lation in 65 LFS samples, 19 cancer-positive and 46 cancer-negative (Fig. 4A). We were able to detect 15/23 (65.2%) late-stage (stage III/IV) and 3/7 (42.9%) early-stage (stage 0/I/II) cancers using our pan-cancer approach.

The most prevalent cancer type among TP53m-carriers is breast cancer; a finding that is also reflected in our cohort (n = 14, 36.8%). Therefore, we built a breast cancer–specific methylation classifier by identifying the hypermethylated DMRs in plasma samples from TP53m-carriers with active breast cancer (n = 14) against samples from TP53m-carriers with active non-breast cancer (n = 23), and healthy controls (n = 28). This resulted in a signature of 24 DMRs (Supplementary Table S6). In cross-validation, we achieved a sensitivity of 92.9% and 40.1% at 95% specificity against healthy controls and non-breast cancers, respectively (Supplementary Fig. S10C). Using this breast cancer signature and the 95th percentile of healthy controls as the detection cutoff, 57 LFS samples scored positive (39 cancer-negative and 18 cancer-positive). Of the 18 cancer-positive samples, 12 were breast cancers (8 late-stage, 2 early-stage, 2 unknown stage; Fig. 4B). Four of these 12 breast cancer samples were not detected by the pan-cancer classifier. Six additional non-breast cancers also scored highly using this signature, suggesting that there may be some cancer-type crossover. The higher sensitivity of our breast cancer signature suggests that given robust enough training data, future studies may be able to incorporate a wide array of cancer type–specific signatures.

Individually, we demonstrated that each of our investigations was able to detect cancer-associated signals, but often lacked sensitivity. To determine if an integrated approach improved our sensitivity, we compared the findings from genome, fragmentome, and methylome analyses in 38 cancer-positive samples (TS = 31, sWGS = 38, cfMeDIP = 37, all = 31). Variant calling identified somatic TP53 variants in 13 (41.9%), copy-number analysis identified CNVs in 18 (47.4%), fragment size analysis identified enriched short fragments in 8 (TP53, 25.8%) and 13 (genome, 34.2%), and methylation analysis identified breast cancer–associated methylation in 18 (47.4%) and pan-cancer methylation in 19 (50.0%) samples (Fig. 5A). Combined, we were able to detect a cancer-associated signal in 81.6% (31/38) samples from TP53m-carriers with active cancer, a 31.6%–55.8% improvement over each individual analysis (Fig. 5B; Supplementary Fig. S11A). In late-stage (III/IV) cancers, our detection rate increased to 22/24 (91.7%; Supplementary Fig. S11B) while in early-stage (0/I/II) cancers, our detection rate was 4/8 (50.0%; Supplementary Fig. S11C). Of the 7 cancer-positive samples with negative ctDNA, 3 were from low-grade intracranial gliomas (LFS2 and LFS5), 1 from a localized Gleeson score 6 prostate adenocarcinoma (LFS81), 2 from slow-growing lung metastases from a prior leiomyosarcoma (LFS40 and LFS59), and 1 from a breast cancer recurrence (LFS58). Within samples from individuals with active breast cancer (n = 14), methylation profiling identified 12 (85.7%) and was the only positive detection method in 2 samples (14.3%; Supplementary Fig. S11D).

In all of our analyses, several samples from cancer-free TP53m-carriers showed evidence of ctDNA which may be indicative of false positives or clinically undetected cancer. Of 131 cancer-free samples from 73 TP53m-carriers, we identified 74 (56.5%) samples from 48 (65.8%) individuals with any combination of somatic TP53 variants (n = 16), CNVs (n = 7), increased fragmentation (TP53 n = 13; genome n = 20), or positive breast (n = 39) or pan-cancer (n = 46) methylation scores (Fig. 5C; Supplementary Fig. S11E). Cancer-associated signal was detected in only one analysis in 32/74 (43.2%) cancer-negative samples compared with 3/38 (7.9%) in cancer-positive samples. Following an in-depth review of clinical data and clinical follow-up, we identified a later cancer diagnosis or suspicious finding by imaging in 26/48 (54.2%) cancer-free TP53m-carriers with a positive ctDNA signal (50/74 samples, 67.6%; Supplementary Fig. S11F and Supplementary Table S7). These findings suggest that our false-positive rate is 18.3% (24/131) within samples and 30.1% (22/73) within individuals. To assess negative ctDNA time points and patients, a true negative was deemed validated if the patient did not develop cancer within 1 year of the negative ctDNA result. Using this metric, within ctDNA-negative TP53m-carriers (patients = 43, samples = 57) only two samples from two patients were deemed as false-negatives. In the first patient (LFS3), ctDNA was negative 3 months prior to their astrocytoma diagnosis. In the second patient (LFS45), ctDNA was negative 11 months prior to a lung cancer diagnosis. In both cases, we detected positive ctDNA closer to their cancer diagnosis (LFS3: 1 month after diagnosis at the time of surgical resection; LFS45: 5 months before diagnosis). Together, within cancer-free TP53m-carriers, our positive predictive value was 54.2% (26/48) and our negative predictive value was 95.4% (41/43; Fig. 5D), an increase of 16.6%–34.6% (NPV) over uni-modal analysis (Supplementary Fig. S11G).

To explore longitudinal ctDNA trends, we compared ctDNA analyses within phenoconverters (forward = 5, reverse = 8, multiple = 3; Supplementary Fig. S12A) and individuals with serial cancer-negative time points (n = 16; Supplementary Fig. S12B). We first evaluated an integrated cancer status score (negative vs. positive) by applying a logistic regression model to our cfDNA analysis metrics (mutation, copy-number, fragmentation scores, methylation scores). Probability scores from this model correlated well with the clinical timelines of 6/16 phenoconverters (LFS15, LFS78, LFS2, LFS35, LFS57, and LFS88). However, in 6 individuals (LFS3, LFS5, LFS40, LFS59, LFS37, and LFS89), clinically relevant signal in one analysis was obscured upon integration. Within cancer-free TP53m-carriers, individuals LFS45 (lung cancer T2-T3) and LFS68 (schwannoma T3-T4) developed and were treated for cancers in-between two cancer-negative time points. In both patients, we observed an increase in short fragments preceding their diagnosis followed by a decrease following surgical resection.

To determine whether the ctDNA signal was apparent prior to cancer detection by conventional screening protocols using imaging modalities, we reconstructed the clinical and molecular timelines of select individuals (Fig. 6).

In LFS3, we observed an increase in fragmentation scores and pan-cancer methylation scores contemporary with and 1 month following the patient's diagnosis of a low-grade astrocytoma as seen on MRI (left). Using cfDNA, we observed an increasing pan-cancer methylation signal starting at 20 months prior to detection of the individual's osteosarcoma in the right zygomatic arch (right). An MRI 10 months prior to their osteosarcoma diagnosis did not show any evidence of cancer (middle).

For LFS15, we observed increasing copy-number–informed predicted tumor fraction and pan-cancer methylation signal starting at 6 and 18 months, respectively, prior to their leukemia diagnosis which was detected by complete blood count following abnormal signal intensity in the bone marrow and cortical parenchyma of the kidneys on MRI (middle, right). Interestingly, MRI 1 year prior (left) and routine bloodwork performed ∼6 weeks prior to their diagnosis showed no evidence of leukemia as lactic dehydrogenase, a nonspecific blood marker, was within normal limits.

In LFS5, we observed increased genome-wide fragmentation signal 6 months prior to their melanoma diagnosis and contemporary to their osteosarcoma diagnosis. Similarly, pan-cancer methylation was also detected 6 months and 7 months prior to melanoma and osteosarcoma diagnoses, respectively. Genome-wide fragmentation decreased following the treatment of both tumors whereas pan-cancer methylation remained elevated following osteosarcoma treatment.

For LFS78, we detected a ctDNA signal in both methylation analyses preceding a negative screening whole-body MRI (left). Elevated ctDNA signal continued to increase on multiple analyses up until the diagnosis of metastatic breast cancer (liver) detected using contrast-enhanced CT (right). In all cases, multimodal analysis of cfDNA outperformed or was contemporary with current clinical screening modalities. The high sensitivity and identification of matched tumor alterations by our multimodal cfDNA methodology suggest that incorporation of cfDNA surveillance can improve and complement the current management of TP53m-carriers.

Lastly, to validate that the ctDNA signals detected in the plasma originated from the matched tumor, we compared our plasma findings to genomic findings from whole-genome sequencing for 3 tumors (LFS3, osteosarcoma; LFS5, osteosarcoma; LFS15, leukemia; ref. 33). The frameshift p.R280fs TP53 mutation identified in the plasma of LFS3 (VAF = 0.60%) was not identified in the matching osteosarcoma which had a copy-number loss of the wild-type TP53 allele. However, in LFS15, TS detected a focal copy-number loss of TP53 in the plasma, which was also detected in their matched leukemia sample. CNVs identified in all 3 tumors showed stronger correlation and concordance to the copy-number profiles from the plasma closest to diagnosis (R² = 0.17–0.93) compared with matched lymphocytes (R² = −0.04 to 0.08; Supplementary Fig. S13) and suggests that the genomic alterations observed in ctDNA reflect those of the tumor.

Here, we present an integrated analysis of 170 liquid biopsy samples from 82 TP53m-carriers profiled using genomic, fragmentomic, and epigenomic methods. Our approach provides proof-of-principle and supports the complementary value of a combined multiomic and multimodal ctDNA assay for the early detection of cancer in HCS, such as LFS. Each assay (TS, sWGS, cfMeDIP-seq) enables analyses that measure independent biological signals that, when evaluated together and in conjunction with other clinical screening modalities, can improve the overall sensitivity, specificity, and robustness of predictions. The high negative predictive value (95.4%) but relatively low positive predictive value (54.2%) within clinically cancer-free patients suggest that liquid biopsy may serve as a prediagnostic screening test complementary to current annual/biannual imaging surveillance that can be performed at a higher frequency.

In this study, analysis of TP53 somatic mutations and genome-wide CNVs resulted in a detection rate of 62.5% in cancer-positive patients, suggesting that relying solely on targeted sets of genomic alterations is not sufficiently sensitive. As our study is retrospective, some pediatric samples did not have sufficient residual cfDNA for TS profiling as some were banked up to 9 years prior and partially used for other studies. Future prospective studies can address this limitation by ensuring the collection of sufficient blood (∼10 mL) in Streck tubes from all patients. Obtaining this volume consistently from pediatric LFS individuals has been shown to be feasible in the implementation of the clinical surveillance protocol termed the “Toronto Protocol” (11). Another avenue of alleviating cfDNA input requirements is through the improvement of methylation analyses which require much lower cfDNA input amounts (∼5–10 ng). Cell-free methylation studies have demonstrated high performance of cancer type–specific detection in sporadic cancers (refs. 21, 32; medRxiv 2023.01.30.23285027). However, these studies have been in the context of common cancer types with vast amounts of reference data, which are required to build robust classifiers from patients with a confirmed clinical diagnosis. As many LFS-associated cancers are less common in the general population, much more epigenetic profiling of these tumor types, ideally using cfDNA from LFS carriers with active cancer, will be required to enable specific cancer-type methylation analyses. This aspect further highlights the need for concerted profiling efforts and data sharing in the context of rare diseases.

Previous studies assessing fragment size in sporadic cancers have often utilized patients with late-stage or metastatic disease and shown high performance (18). In contrast, studies that include early-stage cancers show comparable levels of sensitivity with our study (34). When fragment size was assessed longitudinally within individuals in our cohort, we observed a correlation between fragmentation scores to the individual's clinical cancer status. Given that there are differing baseline levels of fragmentation between patients based on genetic and lifestyle factors (35), establishing patient-specific fragmentomic baselines and detection thresholds may help to increase sensitivity and account for interpatient variability. In the context of LFS, this would be feasible as these individuals already receive routine bloodwork, at minimum, annually, and in pediatrics, often much more frequently. The need for these baselines is especially useful in TP53m-carriers due to the inherently altered cfDNA fragmentation observed in this cohort. This observation may also suggest that TP53m-carriers exhibit baseline differences in their fragmentation landscape, potentially from altered chromatin architecture and cellular homeostasis (30, 36, 37), compared with TP53-wild-type individuals and thus also require the establishment of a TP53m-carrier baseline. These differences may also extend to other HCSs, thus patient and/or HCS-specific baselines and controls may be required for successful implementation of genome-agnostic cfDNA analyses such as fragmentation and methylation.

The increased sensitivity observed in our study through a combination of multimodal analyses suggests that future clinical tests aimed toward early detection will likely need to utilize several analysis types that leverage different biological measurements. This was evident in our study where the ctDNA signal was often only detected by one analytic modality in cancer-free individuals prior to a clinical cancer diagnosis. In this study, we were able to detect a cancer-associated signal in 35.6% (26/73) of patients during a clinically cancer-free timepoint, wherein the participant subsequently developed a cancer or suspicious finding on imaging at follow-up. The sensitivity gained from multimodal analyses should also be balanced with false-positive rates which can lead to unnecessary and costly follow-up procedures. However, this may be tolerated in high-risk individuals such as those with LFS, who already undergo routine imaging modalities that can also have comparable rates of false positivity or low sensitivity, depending on the imaging method used (38–40). Healthcare providers are enthusiastic about the potential for liquid biopsy in HCS but remain cognizant of false-positive rates (41) and patients with HCS also believe that the value of current rigorous surveillance methods outweighs the associated psychosocial costs (12, 42, 43). Given the relative lower invasiveness and increased accessibility of liquid biopsy, compared with current surveillance methods, future studies should assess the effects of liquid biopsy on the psychosocial impact of current surveillance programs.

In this study, 30.1% (22/73) of cancer-free TP53m-carriers showed a cancer-associated signal by ctDNA analysis but no subsequent clinical finding. However, due to the high prevalence of cancer in this population, further targeted and specific follow-up should be considered to determine if these cases are genuine false positives. Combination of imaging surveillance and integrated ctDNA analysis may be complementary in both increasing sensitivity and decreasing the rate of false positives of each screening method, respectively.

The methods for TS and sWGS have been published widely and are routinely used clinically by several academic and com­mercial laboratories. Although cell-free fragmentation and methylation analyses are relatively new and not as established clinically, this study highlights the added benefit of including genome-agnostic methodologies in the setting of early detection. The highest barrier of entry in our suite of techniques is cfMeDIP-seq which is not yet widely adopted and uses a specialized antibody pulldown method. However, the landscape of cell-free methylation assays is expanding with several methodologies having emerged that are more similar to TS and WGS, such as enzymatic methyl-seq (44) and panel-based methylation sequencing (20). One study has even shown the ability to infer methylation status through analysis of cell-free fragmentation surrounding methylation sites using plasma whole-genome sequencing, which would eliminate the need for a standalone methylation assay (45).

Many ctDNA studies are currently focused on singular or bimodal approaches to ctDNA detection in sporadic cancer (20, 46). However, our study suggests that multiomic, multimodal analyses should be adopted and explored further to assess the relative benefits of including additional assays, analyses, and biological information. We demonstrate that integration of multiple plasma-based analyses provides a more holistic view of the ctDNA landscape in low tumor burden settings, such as early detection; and show the relative benefit and complementary nature of liquid biopsy for the management of individuals with LFS. Lastly, these findings can be extended to all HCS with the potential for improved specificity and sensitivity, especially in syndromes with a more homogenous spectrum of associated tumors, and those affected by high-incidence tumor types.

Written informed consent was obtained from all participants of this study. This study was approved by the UHN institutional review board (REB# 19-6239) in accordance with recognized ethical guidelines (e.g., Declaration of Helsinki, Council for International Organizations of Medical Sciences, Belmont Report, U.S. Common Rule). All patients underwent routine clinical care by board-certified clinicians as per the standard-of-care.

Venous blood samples were collected in EDTA or Streck tubes (Streck). EDTA collection was processed within 2 hours. Whole blood samples were centrifuged at 4°C (1,900 × g, 10 minutes). Peripheral blood mononuclear cells (PBMCs) were separated from plasma and stored at −80°C. Isolated plasma was centrifuged a second time at 4°C (16,000 × g, 10 minutes) to remove residual cells and debris. Purified plasma was stored at −80°C until cfDNA extraction. The full protocol can be found at https://charmconsortium.ca/protocols-database/.

DNA from blood was extracted using the QIAGEN QIAamp Circulating Nucleic Acid Kit (Qiagen). Genomic DNA from PBMCs was extracted using the DNeasy Blood and Tissue Kit (Qiagen). After extraction, genomic DNA was sheared to similar sizes as cfDNA using an ultrasonicator (LE220, Covaris). Library preparation for TS, sWGS, and cfMeDIP-seq was performed only for samples with >10 ng DNA yield. TS was excluded for samples with <40 ng DNA yield. Samples with <10 ng DNA yield were not processed. The full protocol can be found at https://charmconsortium.ca/protocols-database/.

cfMeDIP-seq for each sample was prepared as previously described (22) using 10 ng of unique molecular identifier (UMI) ligated cfDNA library with the following deviations from the protocol. Briefly, 0.1 ng of A. thaliana DNA methylation control package containing one methylated and one unmethylated spike-in control BAC (Diagenode) was spiked into 10 ng of cfDNA library. 5% of the sample was aliquoted as a control library. The remainder underwent immunoprecipitation using monoclonal antibody targeting 5-methylcytosine (5-mc; Diagenode, clone #33D3, cat. # C15200081-100, RRID:AB_2572207). Immunoprecipitated libraries and control libraries were amplified, indexed, and pooled.

Precapture libraries were prepared using KAPA Hyper Prep Kit (Kapa Biosystems) and xGen Duplex Seq Adapter-Tech Access [Integrated DNA Technologies (IDT)]. UMI-ligated cfDNA libraries were then split for each assay (TS, sWGS, cfMeDIP-seq). To generate TS libraries, hybrid capture using a custom 100 kb panel of 829 probes targeting TP53, BRCA1, BRCA2, PALB2, MLH1, MSH2, MSH6, PMS2, EPCAM, and APC (Supplementary Table S8) was performed on dual indexed precapture libraries and then PCR-amplified. sWGS and TS libraries were indexed, pooled, and sequenced on the NextSeq 500 using 150-bp paired-end sequencing reads (2 × 150 bp; Illumina). sWGS libraries were sequenced to a target 1×, and TS libraries were sequenced to a target 20,000×. cfMeDIP-seq libraries were sequenced to a target 60 million clusters on the MiSeq Nano using 150-bp paired-end sequencing reads (2 × 150 bp; Illumina). UMI extraction and sequencing alignment to human genome reference GRCh38 was performed using Burrows-Wheelers Aligner version 0.7.12 (RRID:SCR_010910; ref. 35) and deduplicated using Samtools version 1.9 (RRID:SCR_002105) according to the following workflow: https://github.com/oicr-gsi/bwa. All sequencing coverage can be found in Supplementary Table S9.

Aligned reads were error corrected and amalgamated using ConsensusCruncher (https://github.com/pughlab/ConsensusCruncher; ref. 47) according to the following workflow: https://github.com/oicr-gsi/consensusCruncherWorkflow. Germline variant calling was performed using The Genome Analysis Toolkit version 3.8 (RRID:SCR_001876) HaplotypeCaller (Broad Institute; ref. 48). Somatic variant calling was performed using MuTect2 version 3.8 (Broad Institute; bioRxiv 2021.06.30.861054). Variant calling was performed individually on single-strand consensus sequences, duplex consensus sequences, and all unique ConsensusCruncher outputs then merged. VAF was annotated using read support from the all unique ConsensusCruncher output. Single-nucleotide polymorphisms considered benign were removed from further analysis. Candidate pathogenic variants were manually reviewed using Integrative Genomics Viewer version 2.8.13 (Illumina, RRID:SCR_011793). Germline variants were classified according to the 2015 American College of Medical Genetics and Genomics guidelines (49).

TP53 copy-number variants were detected using PanelCNmops (v 1.14.0; ref. 23) and VisCap (https://github.com/pughlab/VisCap; ref. 24). GC-correction, read count normalization, copy-number prediction, and tumor burden prediction were performed using the ichorCNA tool version 0.2.0 (Broad Institute; https://github.com/broadinstitute/ichorCNA; ref. 15) using the recommended default settings. An in-house panel of normals was generated using in-house healthy blood controls (n = 31). To enrich for ctDNA, ichorCNA was performed once on all fragments and again using only short fragments (90–150 bp). This was performed due to increased detection of ctDNA reported in previous studies (25). IchorCNA outputs were then manually reviewed. The higher predicted tumor fraction (all vs. short) was used as the ichorCNA output. For all ichorCNA analyses, only predicted TFs >0.03 were considered positive (based upon the ichorCNA limit of detection).

Global fragment size distributions were calculated using Picard (RRID:SCR_006525) CollectInsertSizeMetrics (v4.0.1.2). Fragment sizes of mutations detected using targeted panels were performed using a custom script and Samtools (v1.9). Briefly, aligned reads overlapping with a mutation of interest were pulled from the bam files and then binned into wild-type and mutant reads. Reads that had alterations at the site of interest not concordant with either mutation or wild-type were discarded. Fragmentation and variant scores were calculated according to Vessies and colleagues (26). Briefly, for the reference set, each fragment length was assigned a weighted value (Log₂Difference) based upon the difference in relative frequency between the distribution of tumor-informed mutant cfDNA fragments and healthy control cfDNA fragments. Then, for each sample, fragments that overlapped with a variant, the TP53 gene, or all fragments genome-wide were assigned a fragment length–weighted value based on the reference values described above. Fragmentation and variant scores were calculated as the mean of the weighted values.

Bootstrap analysis within TP53m-carriers was performed comparing LFS-AC versus cancer-negative TP53m-carriers and LFS-AC versus LFS-H. Briefly, balanced downsampling was performed followed by random selection (80%). Sensitivity at 90% specificity was assessed using genome-wide and TP53-level fragmentation scores. Downsampling was iterated over 100 times to yield a performance distribution.

Fragment size classification of TP53-wild-type controls versus cancer-negative TP53m-carriers and TP53-wild-type controls versus LFS-H were performed using a random forest algorithm. Fragment length frequency distributions (0–600 bp) were used as input for classification. Briefly, for each comparison, balanced downsampling was performed followed by stratified 10-fold cross-validation (90% train, 10% test). Downsampling was iterated over 100 times, and ROC-AUC analysis was performed on the aggregated test scores.

The cfMeDIP-seq FASTQ files were analyzed using an integrated pipeline known as MedRemix (https://github.com/pughlab/cfMeDIP-seq-analysis-pipeline). First, extract_barcodes.py from ConsensusCruncher (https://github.com/pughlab/ConsensusCruncher, May 19, 2021 commit) was used to extract UMI barcodes and remove spacers from unzipped FASTQ files (47). Then, alignment was performed to the human genome (genome assembly GRCh38/hg38) by BWA-MEM v0.7.17, sorted and indexed by Samtools v1.14. Next, the depth of aligned fragments generated from paired reads of cfMeDIP-seq libraries were counted within nonoverlapping 300 bp windows, along with the number of CpGs within each window. The coverage depth was modeled as a two-component mixture, with the two components representing methylated and nonmethylated bins, accounting for CpG density and GC content by negative binomial regression. The mean coverage of nonmethylated reads was inferred from bins with zero CpGs as a function of GC content and treated as a fixed mixture component. The mixing coefficient of methylation status and regression coefficients between the mean coverage of methylated bins and CpG density were jointly inferred using an expectation-maximization process and iterated until convergence. In this process, a posterior probability of each bin being methylated was computed.

The pan-cancer signatures were obtained from Vrba and colleagues (31). The authors obtained 1,250 hypermethylated CpGs by comparing the Illumina HumanMethylation450 (450K) data of 23 types of cancers with their matched normal adjacent samples in TCGA. After mapping these CpGs to their chromosome coordinates (and then methylation bins), removing CpGs mapped to the sex chromosomes, and filtering out CpGs recurrently and highly methylated in blood cells, we obtained a 105 CpG signature (Vrba signature). Blood-cell methylation filtering was performed to remove CpGs in regions methylated in peripheral blood leukocytes, the greatest normal tissue contributor of cfDNA. First, CpGs were restricted to those overlapping the PBL-depleted regions derived from prior PBL MeDIP-seq. To further suppress peripheral immune signal, we used data from tissue-specific 450K methylomes downloaded from the Gene-Expression Omnibus (GEO; GSE122126) and applied a stringent filter to only include CpGs with mean PBL beta value below 0.1 and each individual cell type beta value below 0.15. Finally, we filtered out any CpGs with a mean methylation probability of >0.01, 90th percentile methylation probability of >0.1, or maximum methylation probability >0.5 among cfMeDIP-seq data from normal controls. In addition to the Vrba-derived pan-cancer signature, we also explored a similar method using 450K methylation array data from the TCGA PanCanAtlas. 27 cancer types were included, with the criteria that the data set contain at least 50 independent primary tumors profiled by 450K methylation array. From each cohort, 50 tumors were sampled at random. Differentially methylated CpGs (DMC) were identified between every pair of cancer types using limma R package v3.48.3, using an adjusted P value threshold of <0.05. DMCs were classified as those hypermethylated at minimum 2-fold over at least two thirds of all other TCGA cohorts. Hypermethylated DMCs were further filtered through our blood-cell filter described above. This ultimately yielded a pan-cancer methylation signature consisting of a 243-CpG signature (TCGA signature).

To identify LFS-specific and LFS breast cancer–specific signatures, we first removed bins that overlapped with regions in the ENCODE blacklist. Bins were further filtered by removing regions frequently methylated in healthy control samples, which were bins with average posterior probabilities >0.1 in healthy control samples.

The LFS-specific signatures were identified by comparing the 59 LFS Healthy samples with 28 healthy non-LFS control samples. To generate a balanced data set, we performed downsampling by randomly choosing 28 samples from the LFS Healthy group and combining them with the 28 healthy control samples. Stratified 10-fold cross-validation was then performed on this subset of samples. In each iteration of the cross-validation, 10% were regarded as testing data, whereas 90% were used to identify DMRs hypermethylated in the LFS Healthy samples by limma R package v3.48.3 (RRID:SCR_010943). The top 300 DMRs were selected as features, and a random forest classifier (caret R package, v6.0, RRID:SCR_022524) was trained using training samples in the 9 groups to predict the probabilities of testing samples being LFS. The entire downsampling and cross-validation process was repeated for 100 times. In total, 1,000 differential methylation analysis was performed, and a DMR was determined as LFS-specific only if it was selected for more than 250 times in this process. Finally, 300 DMRs were identified as LFS-specific.

The breast cancer DMRs were identified based on 2 groups of hypermethylated DMRs in our LFS breast cancer samples: breast cancer vs. non-breast cancer and breast cancer vs. healthy control. Similarly, we used 100 10-fold cross-validation to identify DMRs. In each cross-validation, control samples were first downsampled (n = 14) to generate a balanced subset. Then, 300 hypermethylated DMRs in the LFS breast cancer samples versus control samples were identified by limma from the cross-validation training data. Meanwhile, another 300 DMRs were identified by comparing the LFS breast cancer samples to LFS non-breast cancer samples. The intersection of these two sets of DMRs was used and regions that were selected for more than 250 times resulted in a total of 24 breast cancer–specific DMRs.

The methylation score (cancer score) of a sample is defined as the cumulation posterior probabilities of the methylation bins that correspond to the DMRs. The higher the score, the more probable that the cancer–specific DMRs are methylated in the corresponding sample, which indicate the patient is more likely to be cancer-positive.

Integration was performed using all ctDNA metrics (somatic TP53 mutations, copy-number tumor fraction, TP53 fragmentation score, genome-wide fragmentation score, pan-cancer methylation score, breast cancer methylation score). Somatic TP53 mutations were converted to 0 for not mutation present and 1 for mutation present. Training was performed using a logistic regression model and the caret R package. Briefly, cancer-negative samples (n = 131) were downsampled to match cancer-positive (n = 38) samples to create a training data set. Training was performed by stratified 10-fold cross-validation with 90% used for training, and 10% used for validation. This process was repeated 100 times. Integrated ctDNA scores were calculated using the mean of prediction probabilities across all validation sets.

Healthy blood controls were consented to and recruited with institutional approval (REB#: 19-6239) and underwent TS (n = 24), sWGS (n = 29), and cfMeDIP-seq (n = 28). All healthy control data were aligned to GRCh38 as described above and processed according to GATK best practices. Healthy control plasma analyses were performed the same as described above. Heat map visualizations were created using the R package Complexheatmap (v2.9.4; ref. 50) and statistical analyses were performed using R (v4.03, RRID:SCR_001905) in RStudio (1.4.1103, RRID:SCR_000432).

All sequencing data files (TS, sWGS, cfMeDIP) are deposited in the European Genome-Phenome Archive (EGA) under the accession number EGAS00001006539. Code and processed files to reproduce all analyses and figures are available at https://github.com/pughlab/LFS-early-detection-ctdna.

J.A. Sobotka reports grants from the Canadian Institutes of Health Research during the conduct of the study. C. Man reports grants from the Terry Fox Foundation during the conduct of the study. P. Veit-Haibach reports grants from Siemens Healthineers, nonfinancial support from Siemens Healthineers, and personal fees from Wellcome Trust, Ontario Association of Radiologists, JCA Medical Seminars PET/CT course, Springer Nature, Taiwan Oncology Society, and German Cancer Center outside the submitted work. R.H. Kim reports grants from the Princess Margaret Cancer Foundation, the Canadian Institutes for Health Research, TD Ready Challenge, and McLaughlin Centre for Molecular Medicine during the conduct of the study. T.J. Pugh reports grants from Terry Fox Research Institute, Canadian Institutes for Health Research, TD Ready Challenge, and MacLaughlin Centre at the University of Toronto during the conduct of the study; grants and personal fees from AstraZeneca, personal fees from Chrysalis Biomedical Advisors, Merck, SAGA Diagnostics, and grants from Roche/Genentech outside the submitted work. No disclosures were reported by the other authors.

D. Wong: Conceptualization, resources, data curation, software, formal analysis, supervision, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. P. Luo: Conceptualization, resources, data curation, software, formal analysis, validation, investigation, visuali­zation, methodology, writing–review and editing. L.E. Oldfield: Resources, data curation, investigation, project administration. H. Gong: Resources, data curation, investigation, project adminis­tration. L. Brunga: Resources, data curation, investigation, project administration. R. Rabinowicz: Data curation, investigation. V. Subasri: Data curation, investigation, writing–review and editing. C. Chan: Data curation, investigation. T. Downs: Investigation. K.M. Farncombe: Resources, data curation, investigation. B. Luu: Data curation, investigation. M. Norman: Data curation, investigation. J.A. Sobotka: Data curation, investigation, project administration. P. Uju: Data curation, investigation. J. Eagles: Resources. S. Pedersen: Resources, project administration. J. Wellum: Resources. A. Danesh: Software. S.D. Prokopec: Software. E.Y. Stutheit-Zhao: Software, investigation. N. Znassi: Software, investigation. L.E. Heisler: Resources, data curation, software, project administration. R. Jovelin: Data curation, software. B. Lam: Resources, data curation, project administration. B.E. Lujan Toro: Resources, data curation. K. Marsh: Methodology. Y. Sundaravadanam: Methodology. D. Torti: Investigation, methodology, project administration. C. Man: Data curation, investigation, methodology. A. Goldenberg: Supervision. W. Xu: Validation, investigation, methodology. P. Veit-Haibach: Resources, data curation, investigation, visualization, methodology, writing–review and editing. A.S. Doria: Resources, data curation, supervision, validation, visualization, methodology, writing–review and editing. D. Malkin: Conceptualization, resources, formal analysis, supervision, funding acquisition, validation, investigation, methodology, writing–original draft, writing–review and editing. R.H. Kim: Conceptualization, resources, supervision, funding acquisition, validation, investigation, writing–original draft, writing–review and editing. T.J. Pugh: Conceptualization, resources, data curation, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing.

This work would not have been possible without the patients and their generous participation in this study. This work was supported by grants from the Terry Fox Research Institute (grant #1081) with funds from the Terry Fox Foundation (D. Malkin), Canadian Institutes for Health Research (D. Malkin; CIHR-159453), TD Ready Challenge (R.H. Kim and T.J. Pugh), and the McLaughlin Centre at the University of Toronto (R.H. Kim and T.J. Pugh). This study was a collaboration with the CHARM consortium (https://charmconsortium.ca) and is performed under the auspice of the LIBERATE study (NCT 03702309), which is an institutional liquid biopsy program at the University Health Network supported by the BMO Financial Group Chair in Precision Cancer Genomics (Chair held by Dr. Lillian Siu). This work was supported in part by the Canadian Institutes for Health Research Foundation Scheme Grant #143234 (D. Malkin) and the Garron Family Cancer Centre through funds from SickKids Foundation (D. Malkin, R. Rabinowicz, V. Subasri). Additional funding support for this project was made possible by the Shar Foundation, FDC Foundation, Soccer for Hope, Princess Margaret Cancer Foundation, and the Ontario Institute for Cancer Research (OICR). D. Wong is supported by a Princess Margaret Cancer Center Fellowship, a Princess Margaret Cancer Digital Intelligence SPARK Award, a Canadian Institutes of Health Research Fellowship, and a Children's Tumor Foundation Young Investigator Award. R. Rabinowicz is supported in part by a Fellowship from the Israel Cancer Research Fund and the Garron Family Cancer Centre. R.H. Kim is supported by The Bhalwani Family Charitable Foundation. Goldie R. Feldman, Karen Green and George Fischer Genomics and Genetics Fund, Lindy Green Family Foundation, the Devine/Sucharda Charitable Foundation, and Hal Jackman Foundation. T.J. Pugh holds the Canada Research Chair in Translational Genomics and is supported by a Senior Investigator Award from the OICR and the Gattuso-Slaight Personalized Cancer Medicine Fund. D. Malkin holds the Canadian Imperial Bank of Commerce Children's Foundation Chair in Child Health Research. We thank the staff of the OICR Genomics Program (https://genomics.oicr.on.ca) for their expertise in generating the sequencing data used in this study. OICR is supported by funding provided by the Government of Ontario. We thank members of the Pugh Lab, Kim Lab, and Malkin Lab for their extensive peer editing and feedback to improve the manuscript. Lastly, we particularly thank all of the Li–Fraumeni syndrome family members who contributed samples for the study.

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