Abstract 1911: Platform agnostic lung cancer plasma signature identified using multidimensional proteomics and explainable machine learning

Introduction: Lung cancer is the leading cause of cancer mortality worldwide with 70% diagnosed at the late stage. Low-dose CT (LDCT) is available for screening but its uptake remains low. A blood test providing a molecular signature that identifies those at highest risk of lung cancer can integrate with existing screening pathways to increase uptake and enable earlier detection. The blood proteome harbors promising information for minimally invasive biomarker discovery especially for early stage and low tumor-burden lung cancer with limited circulating tumor DNAs. Methods: We combined data-independent acquisition mass spectrometry (MS) as an unbiased approach with a targeted proximity extension assay (PEA, Olink) for biomarker discovery in 614 patients: 490 lung cancer (180 stage I/II, 310 stage III/IV, all major histologies) and 124 matched control (from a LDCT screening program) plasma samples collected under standardized protocol (PMCC, Toronto, 2008-2019). To identify a lung cancer molecular signature, we used a 80/20% discovery/holdout testing approach and incorporated explainability in a tailored machine learning (ML) platform (iScience, 2023) to provide ranking and rationale for feature selection decision making. Finally, aptamer based proteomic profiling (SomaScan) validated a focused molecular signature panel. Results: In total 5,403 distinct protein groups were identified, 3,655 by MS, 2,884 by PEA, with 1,136 overlapping. Both approaches identified proteins significantly enriched for inflammation, chemotaxis/migration, humoral immune response, and complement activation in lung cancer. MS identified additional proteins distinctly enriched for epithelial cell migration and ROS metabolism. ML directed feature selection identified focused panels of 17 and 20 proteins from MS and PEA datasets, each achieving 92% and 98% area under the ROC curve (AUC) respectively in hold-out validation, demonstrating sensitivity/specificity for overall lung cancer detection at 88/90% and 94/90% using ≤20 biomarkers. Importantly, each panel accurately detects early stage, late stage, small cell, adenocarcinoma, and squamous cell carcinoma of the lung, with AUCs of 88%-98% (MS) and 95%-99% (PEA) for each subtype. Further evaluation of the signatures using an aptamer platform confirmed a biomarker panel with strong concordance across the three distinct proteomic approaches, creating a platform agnostic lung cancer plasma signature for disease detection. Conclusion: We identified a platform agnostic plasma molecular signature for lung cancer using three distinct proteomic approaches and explainable ML. This molecular signature can be highly scalable across different protein detection platforms, with prospective studies underway to assess its impact enhancing existing screening programs, enabling earlier detection, and improving survivorship. Citation Format: Peter Jianrui Liu, Harriet R. Ferguson, Luke Hankey, Honglei Huang, Junetha Syed, Iliyana Kaneva, Ella Mi, Daniel A. Szulc, Luna Jia Zhan, Devalben Patel, Ming-Sound Tsao, Roman Fischer, Benedikt M. Kessler, Geoffrey Liu, Nikola I. Gushterov, Andreas J. Halner. Platform agnostic lung cancer plasma signature identified using multidimensional proteomics and explainable machine learning [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 1911.