Trans-vatabnet: transcriptomic signatures for Parkinson's disease detection using deep learning.

Abstract

Background: Early detection of Parkinson's disease (PD) is essential for timely clinical intervention. However, transcriptomic-based diagnosis faces challenges associated with high-dimensional gene expression data, redundant features, class imbalance, and limited interpretability.

Objective: This study proposes TransVaTabNet, an integrated framework for PD detection and biomarker identification using transcriptomic analysis, deep learning, and explainable artificial intelligence.

Methods: RNA-seq data were preprocessed using low-expression gene filtering, Trimmed Mean of M-values (TMM) normalization, and voom transformation. Differentially expressed genes (DEGs) were identified using limma-voom, followed by stage-specific gene analysis and functional enrichment to investigate disease-associated pathways. A Variational Autoencoder (VAE) generated compact latent representations, while the Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance. TabNet classified the derived features using sequential attention, and SHapley Additive exPlanations (SHAP) supported model interpretation and potential biomarker identification.

Results: Evaluation on the GSE216281 RNA-seq dataset achieved an Area Under the Receiver Operating Characteristic Curve (AUC) of 90.95%, outperforming the classical machine learning models evaluated in this study.Conclusion: TransVaTabNet combines dimensionality reduction, attention-based classification, and explainability to support PD detection and potential biomarker discovery. The framework offers an interpretable approach to identifying disease-associated transcriptomic patterns and may support future precision diagnostic research.