A machine learning framework leveraging off-target clinical sequencing reads accurately predicted alternative lengthening of telomeres status across 78,704 tumors with a mean ROC-AUC of 0.84.
A novel machine learning framework can accurately infer alternative lengthening of telomeres (ALT) status from routine clinical sequencing data, enabling scalable pan-cancer ALT detection.
Absolute Event Rate: 0% vs 0%
Abstract Replicative immortality is a hallmark of cancer, achieved by activating telomere maintenance mechanisms (TMMs), which prevent telomere shortening and senescence. While the majority of tumors achieve this via the reactivation of telomerase, a significant subset (10-15%) relies on alternative lengthening of telomeres (ALT), a recombination-driven mechanism of telomere maintenance. ALT is highly tumor type-specific, most frequently seen in mesenchymal tumors. ALT is strongly associated with inactivating mutations in chromatin remodeling genes ATRX and DAXX, however, many ALT tumors lack alterations in these genes. Additionally, ATRX loss alone in in vitro models is insufficient for triggering ALT. Together, these underscore that the underlying mechanisms of ALT remain incompletely understood. A critical gap has been incomplete characterization of the genetic landscape of ALT, as laborious detection assays restrict studies to specific histologies or to tumors with limited genetic profiling. To address this, we sought to use the MSK-IMPACT clinical cohort. MSK-IMPACT is an FDA-approved targeted sequencing panel used for routine genetic profiling of tumors, with 100,000 patients profiled across 80+ tumor types. We developed a machine learning framework to infer ALT status by leveraging typically discarded off-target reads. These reads contained telomeric sequences, enabling us to quantify telomere content and repeat composition. We then assessed the presence of ALT experimentally using the C-circle assay (CCA) for 700 patient samples. ATRX/DAXX truncations were strongly associated with ALT but did not show complete correspondence. Using telomeric features from 300 tumors with high-confidence CCA results, we trained an ensemble of Random Forest classifiers with stratified five-fold cross-validation to predict ALT status, achieving robust performance (mean ROC-AUC = 0.84; PRC-AUC = 0.76). Models were then calibrated and ensembled to generate predictions for 78,704 patient tumors in MSK-IMPACT. Model predictions corresponded well with known ALT patterns. Highest prevalence was seen in sarcomas, gliomas, and neuroendocrine tumors, particularly in ATRX/DAXX mutant tumors, with low or no prevalence in TERT-altered tumors. Interestingly, the model predicted low but notable ALT prevalence in ATRX/DAXX wild-type tumors within ALT-relevant pathologies. MSK-IMPACT annotations enabled identification of ALT-associated genetic factors beyond ATRX/DAXX in these tumors. Our framework enables scalable ALT detection from routine clinical sequencing data and provides an updated picture of ALT prevalence across a wider range of tumor types than previously possible. This resource can be used to identify novel genetic associations, assess differential treatment response and prognosis, and guide therapeutic decisions targeting TMMs. Citation Format: Harshit Sahay, Bill H. Diplas, Oluchi C. Ezekwenna, Divya Koyyalagunta, Simran Chhabria, Madison Darmofal, Quaid Morris, Agnel Sfeir. Pan-cancer landscape of alternative lengthening of telomeres revealed by machine learning analysis of clinical sequencing data 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 1307.
Sahay et al. (Fri,) reported a other. A machine learning framework leveraging off-target clinical sequencing reads accurately predicted alternative lengthening of telomeres status across 78,704 tumors with a mean ROC-AUC of 0.84.