Sparse Bayesian Multiview Learning for Simultaneous Association Discovery and Diagnosis of Alzheimer's Disease

Shandian Zhe, Zenglin Xu, Yuan Qi, Peng Yu

Introduction

Alzheimer’s Disease (AD) research typically pursues two parallel goals: clinical diagnosis—classifying patients as healthy, having Mild Cognitive Impairment (MCI), or having AD—and association discovery, which identifies genetic markers linked to physical brain changes. Historically, these tasks have been treated as independent. However, they are biologically coupled: the genetic factors driving disease progression are the same ones manifesting as structural changes in the brain.

Traditional approaches often struggle with data heterogeneity—combining discrete Single Nucleotide Polymorphisms (SNPs) with continuous MRI features—and frequently ignore the ordinal nature of AD progression. This paper introduces a unified Sparse Bayesian Multiview Learning framework that bridges these gaps, treating diagnosis and biomarker discovery as a single, integrated task.

Methodology: A Unified Bayesian Framework

The proposed model assumes that genetic data and imaging data are generated from a shared set of "latent features" representing a patient's underlying biological state.

The proposed probabilistic graphical model for sparse Bayesian multiview learning, illustrating the relationships between latent features and observed data.

Key technical innovations include:

  • Common Latent Space: Maps heterogeneous inputs (SNPs and MRI) into a shared representation.
  • Ordinal Regression: Accounts for the natural progression of AD (Normal $\rightarrow$ MCI $\rightarrow$ AD) rather than treating it as a standard classification problem.
  • Spike-and-Slab Priors: Implements Bayesian feature selection to identify only the most relevant biomarkers, effectively filtering out noise.
  • Graph Laplacian Prior: Incorporates biological domain knowledge by encoding Linkage Disequilibrium (LD)—the tendency of specific genetic sequences to be inherited together. This encourages the selection of biologically related gene clusters.
  • Variational Inference: Employs a coordinate-descent algorithm to make the complex Bayesian computation feasible for large-scale datasets.

Experimental Results

The model was evaluated using data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), comprising 618 subjects, 924 SNPs, and 328 MRI features.

Diagnostic Performance

The unified model significantly outperformed existing methods in disease stage prediction. It achieved a 10% improvement over Gaussian Process Ordinal Regression (GPOR) and a 22% improvement over traditional multi-step methods like CCA + Lasso.

Comparison of precision across different models, demonstrating the superior performance of the proposed unified framework compared to baseline methods.

Biomarker Discovery

Beyond accuracy, the model identified clinically relevant biomarkers. It highlighted high-confidence genetic markers such as CAPZB, NCOA2, and BCAR3. Furthermore, it successfully mapped these SNPs to atrophy in brain regions critical for memory, such as the middle temporal lobe and precuneus, aligning with established clinical literature. The model's ability to select correlated features in groups (e.g., bilateral brain structures) further validates its biological consistency.

Discussion and Future Work

By linking association discovery directly to clinical labels, the model ensures that identified biomarkers are not just statistically correlated but clinically significant. The use of biological priors helps navigate the "high-dimensional, low-sample-size" problem common in genomic studies.

While the current model is static, future work aims to incorporate longitudinal data to track how these associations evolve over time. Additionally, the framework is extensible and could be expanded to include other data "views," such as proteomic profiles or cognitive test scores, to further enhance diagnostic precision.

Summary updated

Abstract

In the analysis and diagnosis of many diseases, such as the Alzheimer's disease (AD), two important and related tasks are usually required: i) selecting genetic and phenotypical markers for diagnosis, and ii) identifying associations between genetic and phenotypical features. While previous studies treat these two tasks separately, they are tightly coupled due to the same underlying biological basis. To harness their potential benefits for each other, we propose a new sparse Bayesian approach to jointly carry out the two important and related tasks. In our approach, we extract common latent features from different data sources by sparse projection matrices and then use the latent features to predict disease severity levels; in return, the disease status can guide the learning of sparse projection matrices, which not only reveal interactions between data sources but also select groups of related biomarkers. In order to boost the learning of sparse projection matrices, we further incorporate graph Laplacian priors encoding the valuable linkage disequilibrium (LD) information. To efficiently estimate the model, we develop a variational inference algorithm. Analysis on an imaging genetics dataset for AD study shows that our model discovers biologically meaningful associations between single nucleotide polymorphisms (SNPs) and magnetic resonance imaging (MRI) features, and achieves significantly higher accuracy for predicting ordinal AD stages than competitive methods.
BibTeX
@inproceedings{zhe2015sparse,
  author = {Shandian Zhe and Zenglin Xu and Yuan Qi and Peng Yu},
  title = {{Sparse Bayesian Multiview Learning for Simultaneous Association Discovery and Diagnosis of Alzheimer's Disease}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9473},
}
Figure