Abstract B077: M&M: An RNA-seq based pan-cancer classifier for pediatric tumors

Abstract With many documented tumor entities, acquiring the correct diagnosis is a challenging but crucial process in pediatric oncology. Notably, rare tumors present a unique challenge given their infrequency and relative unfamiliarity among pathologists. As a result, these tumor entities tend to be affected by higher misclassification rates. Here we present M&M, a pan-cancer ensemble-based machine-learning algorithm specifically tailored towards inclusion of rare pediatric tumor (sub)types. The RNA-seq based algorithm can classify 52 different tumor types, plus the underlying 96 tumor subtypes. Furthermore, M&M encompasses samples from all tumor stages, treatment statuses and from several non-neoplastic tissues. To facilitate infrequently occurring tumor (sub)type classifications, two different classifiers were created and integrated: a Minority classifier tailored towards correct classification of rare tumor (sub)types, and a Majority classifier with more predictive power for high frequency tumor entities. Each classifier was created using the same four steps of feature selection, feature reduction, down-sampling, and classification algorithm selection, using different focusing methods. Classification took place on the tumor subtype level, from which the tumor type could be extrapolated.M&M could correctly classify the tumor type for 94.5% of the samples within the reference cohort, and the underlying tumor subtype for 86.3%. When filtering on high-confidence classifications, M&M could reach a precision of ∼99% for ∼80% of the tumor type, and a precision of ∼96% for 70% of the tumor subtype classifications. For the low-confidence classifications, the correct tumor classification was often included in the three highest-scoring labels, leading to an overall accuracy of 98% within the top 3. For the tumor subtype classifications, this score was 95%. More than two-third of the samples from infrequently occuring tumor types (3-5 samples) received a high-confidence classification, accompanied by a precision of ∼94%. For classes covered within the classifier, M&M’s performance is comparable to existing class-restricted classifiers like the DKFZ methylation classifier for central nervous system tumors. An independent test cohort confirmed the robustness of M&M's performance.Machine-learning algorithms for both adult and childhood cancer are increasingly used in the clinic, contributing towards increased patient survival. However, many tumor entities are currently missing from existing classifiers. Developing and introducing an extensive agnostic pan-cancer classifier in diagnostics has the potential to increase the diagnostic accuracy for many pediatric cancer cases, thereby contributing towards optimal patient survival and quality of life. Citation Format: Fleur S.A. Wallis, John L. Baker-Hernandez, Marc van Tuil, Claudia van Hamersveld, Marco J. Koudijs, Eugène T.P. Verwiel, Alex Janse, Laura S. Hiemcke-Jiwa, Ronald R. de Krijger, Mariëtte E.G. Kranendonk, Marijn A. Vermeulen, Pieter Wesseling, Uta E. Flucke, Valérie de Haas, Maaike Luesink, Jayne Y. Hehir-Kwa, Bastiaan B.J. Tops, Patrick Kemmeren, Lennart A. Kester. M&M: An RNA-seq based pan-cancer classifier for pediatric tumors [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Advances in Pediatric Cancer Research; 2024 Sep 5-8; Toronto, Ontario, Canada. Philadelphia (PA): AACR; Cancer Res 2024;84(17 Suppl):Abstract nr B077.

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.