Trans-Vatabnet: Transcriptomic Signatures For Parkinson's Disease Detection Using Deep Learning

The early detection of Parkinson's disease (PD) is critical for timely and effective clinical intervention and disease management. High dimensional gene expression data redundancy features, class imbalance and limited interpretability are challenges discussed for transcriptomic-based diagnostic approaches. To overcome these drawbacks, TransVaTabNet, new approach that combines transcriptomic analysis, deep learning and explainable artificial intelligence detect PD and discover biomarkers. The RNA-seq data is first pre-processed by low expression gene filtering, Trimmed Mean of M-values normalization, and voom transformation. The differentially expressed genes (DEGs) are determined through the framework of Limma-voom, and are then subjected to stage-specific gene analysis and functional enrichment to map the disease-related biological pathways. The high dimensional gene expression profiles are learned to a compact latent representation using a Variational AutoEncoder, and SMOTE is used to deal with class imbalance. The derived features are then classified with a TabNet model that adopts a sequential attention mechanism to attend to informative transcriptomic patterns. In addition, SHapley Additive exPlanations (SHAP) are used for model interpretability and biomarker identification. Experimental results on the GSE216281 RNA-seq dataset are shown to validate the proposed framework with 90.95% Area Under the ROC Curve (AUC), outperforming the classical machine learning models. The results show that TransVaTabNet can provide an accurate, interpretable and biologically meaningful framework for stage-specific detection PD and biomarker discovery, which can support precision diagnosis and further clinical decisions.