Enhanced Multimodal Data Integration Using Early Concatenation Fusion in Mixture of Experts (MoE v3): A Novel Approach for Integrated Analysis of Gene Expression, miRNA, and Methylation Data for Oral Cancer

Abstract

Background: Multi-omics data analysis is a comprehensive approach to understanding biological systems and diseases, particularly in oral cancer. It integrates information from various omics layers, enabling early detection and understanding of tumor heterogeneity, classpath classification, drug response, resistance mechanisms, and epigenetic modifications. Despite challenges like data complexity, it can lead to effective treatment strategies and improved patient outcomes. The study explores a novel approach for integrating gene expression, miRNA, and methylation data for oral cancer using Early Concatenation Fusion in Mixture of Experts (MoE v3).
Methods: The study explores a novel approach for integrating gene expression, miRNA, and methylation data for oral cancer using Early Concatenation Fusion in Mixture of Experts (MoE v3). The MoE v3 model employs an early concatenation fusion strategy for multimodal data integration, combining features from various modalities. This approach improves performance with a 0.72 cross-modal correlation, 45% feature importance distribution, 30% gene expression, 30% miRNA, and 25% methylation.
Results: The original MoE enhanced MoE and enhanced MoE v3 models have different Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (R²) values. The Enhanced MoE v3 model has a significantly lower MSE (0.1463), indicating better performance in minimizing squared errors. The Mean Absolute Error (MAE) values are -0.0139, -0.0156, and -0.8435 respectively. The R-squared (R²) values are -0.0139, -0.0156, and -0.8435, respectively. The Enhanced MoE v3 model shows improved predictive power in MSE, RMSE, and MAE compared to the original and Enhanced MoE models. However, a negative R² value suggests further investigation and validation against additional datasets to confirm its utility and robustness in practical applications.
Conclusion: The study showcases the effectiveness of the Enhanced Mixture of Experts (MoE v3) model in analyzing oral cancer data, highlighting its predictive performance and the importance of biomarkers like BRCA1, MGMT, and TP53 in cancer biology