Abstract 5039: Machine learning enhances histology-agnostic detection of PTEN deletions in panel sequencing of advanced cancers

Abstract Targeted panel sequencing of known genetic driver alterations is an established part of the initial clinical diagnostic routine for many tumors. However, exploratory sequencing for advanced and rare cancers remains experimental. MASTERsg is a prospective clinical broad panel sequencing program that utilizes the TSO500 and TST170 DNA/RNA sequencing assays to identify therapeutically relevant alterations in patients with advanced cancers. The TSO500 pipeline detects copy number variations (CNVs) by dividing the CNV caller’s raw data into gene-specific bins, each consisting of 50-250 base pairs. This approach enables the detection of both amplifications and deletions through bin count averaging. However, it may miss focal CNVs that do not meet the standard caller’s threshold fold change. This study aimed to enhance the CNV calling capabilities of the TSO500 pipeline to include focal CNVs. To achieve this, we developed a graphical representation of each gene's bins through plotting, enabling the visual identification of focal CNVs. Additionally, we sought to automate this identification process using machine learning, exemplified by detecting focal deletions of Phosphatase and Tensin homolog (PTEN), a tumor suppressor gene whose functional loss often serves as the basis for recommending mTOR/Akt inhibition in molecular tumor boards. For this purpose, we implemented an Extreme Gradient Boosting (XGBoost) model utilizing normalized PTEN bin counts, enhanced with feature transformations such as polynomial line fitting and derivative to capture read depth variations. Our dataset included 59 patients representing 20 advanced solid tumor entities: 29 PTEN deletions identified by the TSO500 pipeline with a gene-specific fold change cutoff of <0.65 (FC range: 0.05-0.64), five focal PTEN deletions detected only by visual bin plot analysis (FC range: 0.66-1.051), and 25 wildtype cases (FC range: 0.65-1.096). A balanced test set of 16 patients was selected for evaluation. To verify the results, all sequenced tumor samples underwent PTEN immunohistochemical (IHC) staining, with staining results quantified using the H-Score, which ranges from 0 (completely negative) to 300 (strongly positive).Our best-performing model achieved a Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.97, accurately identifying all five focal PTEN deletions missed by the TSO500 pipeline. The biological relevance of these deletions was confirmed by IHC, with H-Scores of 0 in 4/5 cases and 35 in 1/5 cases. This study demonstrates that machine learning can enhance PTEN deletion detection in TSO500 panel sequencing, with potential applications for other genes. Prospective validation on additional tumor samples is currently ongoing. Citation Format: Irina A. Kerle, Oleksandr Husak, Thomas Gross, Andreas Hartig, Daniela Richter, Sascha Brückmann, Andreas Mock, Anja Kögler, Rebecca Prause, Alexander A. Wurm, Maximilian Werner, Dorothea Hanf, Lino Möhrmann, Doreen William, Daniela E. Aust, Evelin Schröck, Christoph Heining, Hanno Glimm. Machine learning enhances histology-agnostic detection of PTEN deletions in panel sequencing of advanced cancers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 5039.