Abstract 5506: An ensemble machine learning approach to ALT detection revals therapeutic vulnerabilities in pediatric cancers lacking actionable drug targets.

Abstract Alternative lengthening of telomeres (ALT) maintains telomeres in 15-20% of cancers through a recombination-based mechanism associated with poor outcomes. ALT is frequently driven by ATRX/DAXX loss-of-function mutations and is enriched in pediatric malignancies including osteosarcoma, neuroblastoma, and soft tissue sarcomas—cancers with limited therapeutic options. Current ALT detection methods are low-throughput and not scalable to large cohorts, impeding systematic discovery of ALT-specific vulnerabilities. We developed ALTitude, a whole-genome sequencing (WGS)-based ensemble machine learning framework to predict ALT status. Using telomere-relevant genomic features extracted from >1,000 cancer cell lines in the Pediatric Cancer Dependencies Accelerator and Cancer Dependency Map (DepMap) resources. We trained and validated the model against orthogonal ALT detection methods. We integrated ALTitude predictions with genome-wide CRISPR loss-of-function screening data in DepMap to identify ALT-specific dependencies. ALTitude achieved high accuracy in ALT classification across diverse cancer types, providing the first scalable map of ALT in a densely functionally characterized cell line panel. Integration with CRISPR essentiality screens revealed SMARCAL1 as a top selective dependency in ATRX/DAXX mutant ALT - positive osteosarcoma, soft tissue sarcomas, and neuroblastoma. SMARCAL1 harbors a helicase domain distantly paralogous to ATRX and resolves stalled replication forks at telomeres during DNA damage response, suggesting that increased replication stress at ALT telomeres creates essentiality for SMARCAL1 function. This work establishes a scalable method to infer ALT status from WGS data and systematically connects ALT biology to therapeutic vulnerabilities in cancer cell line models. By identifying SMARCAL1 as selective dependency in ALT-positive cell lines, we nominate novel therapeutic targets for challenging pediatric cancer types. Our approach demonstrates how integrating features derived from genomic data with functional screening and other 'omics data can reveal novel insights into drivers across aggressive cancer subtypes. Citation Format: Declan Bennett, Monika Weirdl, Lillian M. Guenther, Paul Geeleher, . An ensemble machine learning approach to ALT detection revals therapeutic vulnerabilities in pediatric cancers lacking actionable drug targets [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 5506.