Abstract 1142: Non-invasive cell free DNA (cfDNA) methylation profiling for accurate proportional quantification of lung cancer subtypes

Abstract Background: Neoplastic progression in lung cancer often involves tumors with heterogeneous histologies, harboring multiple subclones of distinct histological subtypes - such as lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LSCC), and small cell lung carcinoma (SCLC) - within a single tumor. Determining histological subtype of the tumor typically requires invasive lung tissue biopsy, which precludes serial monitoring both during and after therapy. Treatment regimens targeting the predominant histology may inadvertently select for drug-resistant subclones of alternative subtypes, necessitating timely adjustments in treatment strategies. Comprehensive cfDNA methylation profiling across thousands of cancer-associated regulatory regions offers an approach for accurate quantification of lung cancer histological subtypes, facilitating enhanced noninvasive disease monitoring and more effective therapeutic strategies. Methods: We developed a novel cfDNA methylation-based deconvolution model to identify and quantify the proportional contributions of LUAD, LSCC and SCLC in patient blood samples. The model leverages hypermethylation signals across over 9000 of cancer-associated genomic regions to estimate the contribution of each histological subtype while simultaneously predicting patient-specific subtype hypermethylation signatures and non-cancer component signatures across regions. We modeled each sample as a linear combination of its three histological subtype and non-cancer components weighted by their proportional contributions. The model was trained using stochastic gradient descent to minimize both reconstruction and proportional contribution losses. Using samples from cancer-free individuals (N=6, 142), and lung cancer patients (LUAD N=2, 721, LSCC N=663, and SCLC N=243), the model was trained using 90% of the data and tested on the remaining 10%. For each test sample, the model was used to predict proportions of three histological subtypes summing to 100%. Model performance was evaluated by comparing the predicted histological subtypes to the known histological subtype annotations of lung cancer samples. Results: The overall accuracy on the test set was 85.1% down to 0.1% tumor fraction, with the highest accuracy for LUAD (90.4%), followed by LSCC (72.4%) and SCLC (63.5%). Importantly, the model demonstrated capability to detect minor subclonal contributions as low as 0.1%. Conclusions: Our methylation-based subtype deconvolution model provides an accurate method for quantifying proportions of lung cancer histological subtypes in patient blood samples, with the unique ability to detect minor subclonal populations. This approach enables the early non-invasive detection of treatment-resistant lung cancer subtypes, thereby holding significant potential to improve patient outcomes. Citation Format: Anton Valouev, Wei Tian, Kunwar Singh, Sheila Solomon, Justin Odegaard, Darya Chudova, Helmy Eltoukhy. Non-invasive cell free DNA (cfDNA) methylation profiling for accurate proportional quantification of lung cancer subtypes [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 1142.