A Novel Deep Learning-Based Multi-Model Ensemble Approach for the Prediction of Non-Small Cell Lung Cancer (NSCLC) Metastasis via Integration of Multi-omics Data

Lung cancer is the second most prevalent cancer globally, with 2.2 million new cases and 1.8 million associated deaths annually. This research focuses on Non-Small Cell Lung Cancer (NSCLC), which is around 80-85% of all Lung cancer cases. While early detection of metastasis is crucial, current methods such as imaging face limitations as they lack a comprehensive understanding of the molecular causes of metastasis, causing 90% of all cancer-related deaths. This research aims to predict NSCLC metastasis by integrating high-throughput multi-omics data such as copy number variation, RNA sequencing, reverse phospho-protein arrays, and mutation data for 220 patients into a deep-learning model. A denoising autoencoder, an unsupervised method for efficient data representation by eliminating noise, was used. The class imbalance was addressed via the Synthetic Minority Over-sampling Technique. Least Absolute Shrinkage and Selection Operator reduced features to the top 100 components for dimensionality reduction. The pipeline was extended to other subclasses within the multi-omics dataset, utilizing a supervised approach with Random Forest Classification models. Utilizing a modified Multi-Model Ensemble Learning approach with a Stacking Classifier, this study achieved a 94.32% accuracy, surpassing models proposed in previous studies at 85%, and obtaining a Cohen's Kappa value of 0.804, highlighting the reliability of this model's predictions. This study identified thirty genes highly associated with NSCLC metastasis, laying the foundation for refining patient stratification and identifying potential therapeutic targets. This study recommends expanding the model pipeline for more omics types, paving the way for integrating the model into clinical decision support systems across various cancers.