Abstract 1098: Combining random projection and stacking learning for NSCLC subtype classification based on transcriptomic data

Abstract Lung cancer is the leading cause of cancer-related death and the second most diagnosed cancer worldwide. Non-small cell lung carcinoma (NSCLC), the predominant type of lung cancer, presents diverse pathological features and cellular heterogeneity. As the two major NSCLC subtypes, lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) differ in their cells of origin and exhibit distinct disease progression, which, however, are difficult to distinguish solely based on conventional experimental methods like histological examination. Previous studies indicated that predicting subtypes based on transcriptomic profiles is a feasible and promising method. However, existing computational approaches may not fully capture information from transcriptomics data, and they are not effective in dealing with high dimensionality data. To address these concerns, we present a novel method called RPSLearner that combines Random Projection (RP) with Stacking Learning, to efficiently aggregate diverse information from base models for accurate identification of NSCLC subtypes. Specifically, we applied multiple RPs to reduce the dimensionality of transcriptomic profiles, which were then combined to capture diverse feature information. RP could efficiently reduce the dimensionality of the high-dimensional omics data while preserving distances between data points. Subsequently, the dimension-reduced feature vectors were classified by ten machine learning base models, which were further integrated by a stacking meta-learner. The stacking strategy combines the prediction results from various base models using a multi-layer perceptron (MLP) to boost the prediction performance. Experimental results based on 1,333 lung cancer patient samples from The Cancer Genome Atlas Program (TCGA) database suggested that RPSLearner significantly outperformed existing state-of-the-art methods. We further benchmarked RPSLearner against different cases including using only base classifiers, ensemble random projection with classifiers, and score ensemble methods, and demonstrated that RPSLearner achieved the best performance across multiple metrics including accuracy, AUC and F1-score. We believe that RPSLearner has direct positive impacts on downstream lung cancer risk stratification and tailored treatment design. We also expect that RPSLearner can be customized and extensible to identify subtypes of other types of cancer. Citation Format: Xinchao Wu, Jieqiong Wang, Shibiao Wan. Combining random projection and stacking learning for NSCLC subtype classification based on transcriptomic data [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 1098.