Deep Learning for Genetic Variant Discovery

We develop deep learning frameworks to analyze whole-genome sequencing (WGS) data from large-scale cohorts such as the Alzheimer's Disease Sequencing Project (ADSP). Our SWAT-CNN approach applies a sliding-window strategy to detect AD-associated genetic variants, achieving an AUC of 0.82. More recently, our Deep-Block framework analyzed 7,416 WGS participants using linkage disequilibrium-informed deep learning, identifying novel SNPs beyond the well-known APOE region. We also developed DuAL-Net, a dual-attention architecture that integrates local genomic patterns with global biological knowledge for improved AD risk prediction from APOE-centered regional WGS data.

Reliable Prediction through Uncertainty Estimation

A critical challenge in applying deep learning to clinical genomics is knowing when a model's prediction can be trusted. Standard classifiers produce point estimates without indicating their own confidence, which limits their utility in high-stakes medical decisions. Our TrUE-Net (Transformer Uncertainty Estimation Network) framework addresses this by combining a transformer encoder and random forest classifier with Monte Carlo dropout to quantify prediction uncertainty.

Applied to WGS data from 1,050 ADNI participants (607 AD, 443 controls), TrUE-Net achieved an overall AUC of 0.664. More importantly, by stratifying predictions based on uncertainty scores, the model identified a high-confidence subset (24.6% of samples) with 72.9% accuracy and F1 score of 0.821. This approach enables clinicians and researchers to distinguish reliable predictions from uncertain ones, providing a practical framework for more informed decision-making in genomic-based disease classification.

Brain Imaging and Early Diagnosis

Tau PET and structural MRI capture pathological changes in the brain years before clinical symptoms of Alzheimer's disease emerge. Our work applies 3D convolutional neural networks to tau PET scans, achieving 90.8% classification accuracy and revealing informative spatial features in the medial temporal and parietal cortices. We have also published a comprehensive review of deep learning applications across multimodal neuroimaging for AD diagnostic classification and prognostic prediction.

Blood-Based Biomarkers for Disease Prediction

Blood-based biomarkers offer a minimally invasive pathway to early Alzheimer's detection. Our longitudinal plasma proteomics study identified a seven-protein signature achieving AUC 0.848 for predicting incident AD dementia, with proteins including ACES, C7, and IL-17C emerging as key predictors. In metabolomics, we developed circular-SWAT (c-SWAT), a deep learning method for serum-based metabolomic classification that reached AUC 0.808, demonstrating the potential of metabolites as accessible, cost-effective biomarkers for population-level screening.

Multimodal Integration for Personalized Prediction

Alzheimer's disease is driven by complex interactions across genetic, molecular, and structural domains. Our precision medicine research integrates genomic, neuroimaging, and metabolomic data into unified AI frameworks to produce individualized risk profiles. By combining the complementary strengths of each data modality, these multimodal models aim to improve prediction accuracy beyond what any single data type can achieve alone, ultimately guiding personalized intervention strategies.