Interpreting Cellular Signatures of Alzheimer’s Disease: A Proposed Explainable AI Framework Using Single-Cell Foundation Model
Abstract/Description/Artist Statement
Alzheimer's disease is the most common form of dementia, a neurodegenerative disorder that slowly destroys a person's memory and thinking skills although it's cellular mechanisms remain highly complex and not fully understood. Although single-cell RNA sequencing offers high resolution in these processes, analyzing these vast data often relies on black-box deep learning models that obscure biological understanding. Thus, to understand the complex underlying cellular mechanisms, we propose an explainable AI framework designed to transparently decode the cellular signature of Alzheimer's dementia. We are utilizing the single-cell RNA-seq data from ROSMAP (Religion Orders Study and Memory and Aging Project) dataset. The proposed method involves fine-tuning a pretrained single cell foundational model such as scGPT, Geneformer to classify cellular states across the progression of the disease. To bridge the gap between model complexity and interpretability, we incorporate dual-method interpretability approach. First, we will integrate Shapley Additive exPlanations (SHAP) for rigorous feature attribution, aiming to identify and rank genes responsible for driving model's predictions. Second, to understand the model's internal representations, we will train a Sparse Autoencoder (SAE) on the dense hidden embeddings of fine-tuned single cell foundational model to extract the features learned by the model and understand it's biological relevance. By integrating foundation models, sparse representation learning, and Shapley value analysis, this framework aims to improve both predictive performance and biological interpretability. The goal is to move away from black-box predictions toward explainable AI. This shift ensures our models serve as transparent tools for hypothesis generation. Ultimately, it allows biomedical researchers to truly understand the cellular mechanisms driving cognitive decline.
Faculty Advisor/Mentor
Hong Qin
Faculty Advisor/Mentor Email
hqin@odu.edu
Faculty Advisor/Mentor Department
Computer Science
College/School Affiliation
College of Sciences
Student Level Group
Graduate/Professional
Presentation Type
Poster
Interpreting Cellular Signatures of Alzheimer’s Disease: A Proposed Explainable AI Framework Using Single-Cell Foundation Model
Alzheimer's disease is the most common form of dementia, a neurodegenerative disorder that slowly destroys a person's memory and thinking skills although it's cellular mechanisms remain highly complex and not fully understood. Although single-cell RNA sequencing offers high resolution in these processes, analyzing these vast data often relies on black-box deep learning models that obscure biological understanding. Thus, to understand the complex underlying cellular mechanisms, we propose an explainable AI framework designed to transparently decode the cellular signature of Alzheimer's dementia. We are utilizing the single-cell RNA-seq data from ROSMAP (Religion Orders Study and Memory and Aging Project) dataset. The proposed method involves fine-tuning a pretrained single cell foundational model such as scGPT, Geneformer to classify cellular states across the progression of the disease. To bridge the gap between model complexity and interpretability, we incorporate dual-method interpretability approach. First, we will integrate Shapley Additive exPlanations (SHAP) for rigorous feature attribution, aiming to identify and rank genes responsible for driving model's predictions. Second, to understand the model's internal representations, we will train a Sparse Autoencoder (SAE) on the dense hidden embeddings of fine-tuned single cell foundational model to extract the features learned by the model and understand it's biological relevance. By integrating foundation models, sparse representation learning, and Shapley value analysis, this framework aims to improve both predictive performance and biological interpretability. The goal is to move away from black-box predictions toward explainable AI. This shift ensures our models serve as transparent tools for hypothesis generation. Ultimately, it allows biomedical researchers to truly understand the cellular mechanisms driving cognitive decline.