Convolutional Neural Network for Disease Prediction, Biomarker Discovery, and Validation in Alzheimer's Disease - Abstract/Summary Alzheimer’s disease (AD) is a devastating neurodegenerative disorder affecting millions globally, yet it remains without a cure and with limited therapeutic options. Early diagnosis and risk prediction through genetic and clinical biomarkers are critical for timely interventions. Our parent R15 project (R15AG083618-01A1, 06/01/2024-05/31/2027), titled Convolutional Neural Network for Disease Prediction, Biomarker Discovery, and Validation in Alzheimer’s Disease, focuses on developing convolutional neural network (CNN)-based models to predict AD risk and identify genetic biomarkers from large-scale genetic data while training undergraduate students in artificial intelligence (AI) and computational genomics. However, the increasing complexity and volume of multimodal data, including genomics, clinical, and MRI, now exceed our current computational infrastructure, limiting model scalability, reproducibility, and student training opportunities. In response to NOSI NOT-OD-24-078, we propose a revision to transition critical components of our pipeline, data preprocessing, model training, interpretation, and visualization, to a secure, scalable, cloud-based platform using Google Cloud Platform (GCP). This revision builds on our established pipelines and prior success with major data repositories such as NIH dbGaP, NIAGADS, and ADNI. We have demonstrated that machine learning (ML) models using genetic data outperform traditional statistical models and have experience in cloud-based GPU computing environments. Additionally, recent training through the NIH AIM-AHEAD Health Data Science Training Program and participation in the upcoming Multi-Omics NETwork Analysis Workshop (MONET) workshop equips our team with interdisciplinary expertise spanning genetics, clinical research, and cloud-based AI. Aim 1: Evaluate GCP- based ML models for AD risk prediction using expanded PRSs and clinical features. This aim will evaluate more advanced ML models (e.g., Support Vector Machines, Random Forest, XGBoost, CNN, DNN) for AD risk prediction, expanding PRS traits with additional principal components, diagraphic, or clinical features. We will use Terraform for infrastructure-as-code, Docker for containerization, Kubernetes for orchestration, and Vertex AI for training and explainability (e.g., SHAP, Grad-CAM). Aim 2: Improve AD biomarker discovery through the integration of multimodal data using GPU-accelerated cloud tools. This aim will integrate multi- omics, clinical, and MRI data using GCP services like BigQuery and Dataflow, leveraging Vertex AI Workbench with GPU support for scalable multimodal biomarker discovery. Models in both aims will be evaluated on AUC, accuracy, precision, recall, F1-score, runtime, and cost-efficiency, compared to the local server. Impact: This project modernizes research infrastructure through a fully cloud-native ML pipeline. It enhances AD prediction, accelerates biomarker discovery, and provides scalable, reproducible workflows. Equally Importantly, it delivers robust hands-on training for undergraduate students, particularly those from under-resourced institutions, in advanced data science and cloud technologies, positioning them for future careers in biomedical AI.
