Abstract A038: Unmasking hidden mutational footprints of therapy in pediatric tumors

Abstract The substantial improvement in childhood cancer survival rates has been primarily achieved through therapy intensification, often at the cost of life-long adverse effects. Therapies can be associated with characteristic patterns of mutations found in the tumor genome, known as mutational signatures. These signatures, therefore, can act as markers of resistance and late effects of therapy. We employed a large-scale analysis of 611 whole-genome sequenced pediatric tumors with harmonized therapy data, integrating de novo mutational signature extraction with ensemble machine learning (ML) models trained on comprehensive genomic features (single- and double-base substitutions, indels, structural variants, and copy number alterations) to identify therapy-associated patterns. De novo extraction across single-nucleotide variants, indels, structural variants, and copy number alterations revealed 95 distinct mutational signatures, including 31 novel ones. Platinum therapy was associated with the highest mutagenic burden, with canonical signatures (e.g., SBS31, SBS35, DBS5) detected in 41% of exposed tumors. By aligning with treatment timelines, we defined a median latency of 91 days and a burden threshold of ∼1500 mutations for platinum signatures to emerge—appearing in all osteosarcomas (most common primary bone cancer) and two-thirds of neuroblastomas (most common extracranial solid tumor in children). Despite these insights, the reasons why some tumors do not exhibit these signatures, and how this relates to their response to therapy, remain poorly understood. To uncover subtler therapy-associated signals beyond known COSMIC signatures, we trained ensemble ML models on various types of genomic alterations. Our ML models achieved high predictive performance (F1-score = 0.89) in classifying platinum-exposed tumors. Importantly, predictive accuracy persisted even after masking features corresponding to canonical platinum signatures, indicating the presence of non-canonical mutational patterns. Feature analysis identified biologically plausible non-canonical genomic features enriched in platinum-exposed tumors across pediatric cancers, suggesting previously unannotated mutational footprints of therapy. Notably, ML analysis revealed distinct non-canonical genomic features associated with anthracyclines, and select antimetabolites and alkylating agents, shedding light on therapies previously lacking characterized mutational signatures. These novel genomic signals were also validated across independent pediatric and adult cancer cohorts. This work highlights the power of ML to detect therapy-induced mutational patterns beyond known mutational signatures. As cancer remains the leading cause of disease-related death among children in many countries, this integrated ML-driven approach offers a powerful strategy for improved therapy monitoring, early resistance prediction, and ultimately, the development of more personalized and effective treatment strategies (e.g., stratification and de-escalation) for children with cancer. Citation Format: Mehdi Layeghifard, Marcos Díaz-Gay, Pedro L. Ballester, Elli Papaemmanuil, Mark Cowley, Anita Villani, Ludmil B. Alexandrov, Adam Shlien. Unmasking hidden mutational footprints of therapy in pediatric tumors [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A038.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.