Objective: This study integrates snRNA-seq and proteomic data to (i) map mtDNA heteroplasmic mutations at single-cell resolution, (ii) identify mitochondrial-driven transcriptional dysregulation, and (iii) classify mitochondrial perturbation subtypes using deep learning.

Background: Mitochondrial dysfunction contributes to Parkinson’s Disease (PD), but its genomic basis remains unclear. Bulk sequencing cannot resolve mtDNA heteroplasmy at the single-cell level, limiting understanding of dopaminergic neuron vulnerability. Advances in snRNA-seq and deep learning enable high-resolution mapping of mitochondrial genomic instability.

Method: We used snRNA-seq data (GEO: GSE157783) with ~19,000 single-nucleus transcriptomes from PD patients and controls and proteomic data from the Human Protein Atlas. Mitochondrial heteroplasmic mutations were inferred by aligning snRNA-seq reads to the human mitochondrial genome, followed by hierarchical clustering to detect mutation-specific expression shifts. A deep temporal convolutional neural network (DTCNN) classified mtDNA-driven transcriptional alterations, while a multi-modal graph attention network (GAT) integrated gene expression and proteomic features to cluster patients into mitochondrial subtypes. Performance was assessed using five-fold cross-validation with AUROC, AUPRC, and sensitivity-specificity metrics.

Results: Our AI model identified 147 novel heteroplasmic mtDNA mutations (q<0.001, FDR-adjusted) linked to PD neurons, revealing mutation hotspots in oxidative phosphorylation genes. Deep learning stratified PD into three mitochondrial subtypes (Mito-PD1, Mito-PD2, Mito-PD3) with 90.8% classification accuracy (95% CI: 88.7–93.1). Mito-PD1 showed oxidative stress hyperactivation (log2FC=+2.14, p=3.7×10⁻⁶), while Mito-PD2 exhibited significant mtDNA depletion (>25.6% loss, p<0.001). The AI-predicted therapeutic targets (NRF2, TFAM, PGC1α) correlated strongly with PD progression and survival (HR=1.69, 95% CI: 1.42–2.01, p<0.0001). Model performance showed AUROC=0.87 (±0.02) and AUPRC=0.81 (±0.03), outperforming bulk sequencing (AUROC=0.66).

Conclusion: This study provides an AI-driven single-cell map of mitochondrial genomic instability in PD, identifying mitochondrial subtypes with therapeutic relevance. Integrating snRNA-seq and proteomic data, we reveal how deep learning enhances biomarker discovery and precision medicine strategies.

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/deep-learning-driven-discovery-of-mitochondrial-genomic-instability-in-parkinsons-disease-a-single-cell-multi-omics-approach/