利用配对种系基因组预测儿童癌症的发病年龄

Background: Germline events play a role in influencing the age of onset of pediatric cancers, but their specific contribution remains incompletely understood. Approximately 18% of children with cancer harbor a germline gene mutation associated with a Cancer Predisposition Syndrome (CPS). However, this fraction is believed to be an underestimation, given that a substantially larger proportion of childhood cancers is thought to have a genetic basis due to the limited environmental exposure in children compared with adults. Carriers of cancer predisposition gene mutations typically undergo clinical surveillance for early tumor detection. However, many children carry germline risk that would not be recognized by their cancer or family cancer phenotypes and are therefore not identified for surveillance, reflecting a major gap in leveraging whole genomic data for individual prediction. One notable example of a monogenic CPS is Li-Fraumeni Syndrome (LFS), caused by germline mutations in the TP53 tumor suppressor gene. Our lab has demonstrated that additional (epi)genetic events that modify the LFS phenotype and enable more precise prediction of tumor onset, underscoring the need for a better understanding of the contribution of germline events to age of onset across pediatric cancers. In this study we analyze a broad range of pediatric cancer types, using their paired germline genomes to develop machine learning (ML) models that predict age of onset, uncover genomic predispositions, and support earlier tumour detection. Methods: Whole genome sequencing data from the SickKids Cancer Sequencing (KiCS) program's initial cohort of poor-prognosis childhood cancer patients (n=333), ~18% of whom meet CPS criteria, were analyzed. After grouping germline variants by gene and pathogenicity according to ACMG guidelines, these variants were used as input features for UMAP and random forest models to classify the age of onset. Generative AI was used solely to modify code, refine written content, and identify relevant journal articles. All AI-assisted content was verified. Results and Conclusion: Preliminary models achieved an average AUROC of ~0.63 in classifying onset before versus after various selected ages (1-17). These findings suggest that genomic profiles are dependent on developmental stages. The UMAP performed on the same variants led to the observation that one cluster contained a higher proportion of pediatric-enriched cancer subtypes (75%) compared to the other clusters (58%). Preliminary findings suggest that genomic profiles that are more consistent with hereditary cancer can be separated from genomic profiles that are more consistent with sporadic cancer. Integrating genomic data with predictive modelling may improve our understanding of how germline events collectively influence age of onset across pediatric cancer types, which could inform risk stratification in clinical surveillance protocols.