Abstract 6273: AI-based serum proteomics models predict survival outcomes in NSCLC patients undergoing immunotherapy

Abstract Background: Clinical responses to immune checkpoint inhibitors (ICIs) are highly variable. Pre-treatment serum proteomic biomarkers offer a promising tool to predict response outcomes by capturing the biological mechanisms underlying response variability. Methods: This retrospective cohort study analyzed pre-treatment serum samples from 158 consecutive locally advanced or metastatic (palliative intent treatment) NSCLC patients treated with ICI monotherapy or ICI-chemotherapy combination. Deep proteomic profiling via mass spectrometry identified protein interaction patterns, and three machine learning models—XGBoost, logistic regression, and multilayer perceptron (MLP)—were evaluated using 5-fold cross-validation. The primary outcomes were progression-free survival (PFS) and overall survival (OS) at multiple time points (3, 6, 9, 12, 36, and 60 months), with performance assessed using the area under the receiver operating characteristic curve (AUC). Results: The study group had a median age of 70.8 years, with 51.9% female and 67.7% with adenocarcinoma. Smoking history showed 15.8% never smoked, 69.6% were former smokers, and 13.9% were current smokers. Performance status (PS) was 0 in 19.0%, 1 in 54.4%, 2 in 19.6%, and 3 in 5.7%. The median tumor mutational burden (TMB) was 5.3 mut/Mb, with PD-L1 <1% in 31.6%, >1% in 55.7%. Treatment regimens included mono-ICI (38.0%), dual ICI (13.9), and ICI + chemotherapy (48.1%). Median OS was 19.5 months, PFS was 6.6 months. Proteomic-based models showed varying performance across outcomes and time points. For PFS, XGBoost achieved the highest performance at 90 days (AUC=0.67), maintaining moderate prediction at 270 days (AUC=0.65) and 1 year (AUC=0.65). For OS, performance improved over time, with the best results at 1 year (AUC=0.61). Using 4, 000 highly variable proteomic mass spectrometry features, risk stratification at 1 year OS showed significant differences between groups (log-rank p=0.05) in PFS, with median PFS of 187 and 283 days for high- and low-risk groups, respectively (HR=1.38, 95% CI: 0.95-2.02, log-rank p=0.09). Conclusion: Serum proteomics-based AI model can predict PFS and OS in NSCLC patients treated with ICIs, achieving AUCs up to 0.67 for PFS and 0.61 for OS. Larger studies are needed to validate our findings. Citation Format: Junho Song, Leeseul Kim, Donghoon Shin, Jongwoo Kim, Taegyu Um, Allen Cho, Jaeyoung Bae, Liam IL-YOUNG Chung, Jeeyeon Lee, Young Kwang Chae. AI-based serum proteomics models predict survival outcomes in NSCLC patients undergoing immunotherapy [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 6273.