TL;DR
This paper introduces a novel Bayesian multiview learning framework to simultaneously discover associations between genetic variations (SNPs) and brain imaging (MRI) data, and to predict Alzheimer's Disease (AD) severity. By jointly modeling these tasks and incorporating prior biological knowledge like Linkage Disequilibrium, the method achieves significantly improved accuracy in diagnosing AD stages compared to existing approaches.
What Problem Does This Paper Address?
Alzheimer's Disease (AD) is a complex neurodegenerative disorder influenced by both genetic factors and observable biological changes. Researchers are keen to understand the intricate relationships between genetic variations (like Single Nucleotide Polymorphisms or SNPs) and phenotypic traits (such as brain atrophy measured by MRI) to better diagnose and potentially predict the progression of AD. Existing methods often tackle association discovery and disease diagnosis separately, or fail to account for the heterogeneous nature of the data (discrete SNPs vs. continuous MRI) and the ordinal nature of disease severity (normal, mild cognitive impairment, AD). Furthermore, valuable biological information, like Linkage Disequilibrium (LD), which describes non-random associations between alleles, is often overlooked. This paper aims to bridge these gaps by proposing a unified framework that can jointly discover associations and diagnose AD, while handling diverse data types and incorporating prior biological knowledge.
Methods
The core of this paper is a Sparse Heterogeneous Multiview Learning Model (SHML), a Bayesian framework that leverages latent variables to connect different data sources. Here's a breakdown of the key methodological contributions:
- Multiview Learning with Latent Features: SHML introduces a common set of latent features ($U$) that are assumed to generate both the genetic (Z) and phenotypic (X) data through sparse projection matrices ($H$ and $G$, respectively). This allows for the discovery of shared underlying factors influencing both data types.
- Sparse Projection Matrices: To ensure that only the most relevant genetic and phenotypic markers are selected, spike-and-slab priors are used for the projection matrices. This encourages sparsity, effectively performing feature selection.
- Heterogeneous Data Handling: The model explicitly accounts for the different data types. For continuous MRI data, a Gaussian distribution is used. For discrete ordinal SNP data, an ordinal regression approach is employed, mapping auxiliary continuous variables to discrete categories.
- Ordinal Regression for Diagnosis: The disease status ($y$) is treated as an ordinal variable, and an ordinal regression model is used to predict it based on the latent features. This captures the natural progression of AD severity (normal, MCI, AD).
- Linkage Disequilibrium (LD) Prior: To incorporate biological knowledge about SNP associations, a graph Laplacian prior based on the LD structure is introduced for the SNP projection matrix ($H$). This helps to regularize the model by leveraging known relationships between SNPs.
- Variational Inference: Due to the complexity of the Bayesian model, an efficient variational inference algorithm is developed to approximate the posterior distributions of the latent variables and model parameters. This involves minimizing the Kullback-Leibler divergence between the approximate and true posteriors.
Datasets
The paper utilizes the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset for its real-world study. This dataset contains genetic (SNP) and neuroimaging (MRI) data from elderly individuals with varying cognitive statuses (normal, MCI, AD). No new benchmark datasets are introduced.
Key Insights and Ideas
- Joint Modeling is Crucial: The paper strongly suggests that treating association discovery and disease diagnosis as separate tasks is suboptimal. By jointly modeling them, the disease labels can effectively guide the discovery of meaningful associations between genetic and phenotypic data.
- Prior Biological Knowledge Matters: Incorporating information like Linkage Disequilibrium (LD) into the model significantly improves its ability to discover biologically relevant associations and enhances prediction accuracy. This highlights the power of integrating domain knowledge into machine learning models.
- Heterogeneity and Ordinality are Key: Explicitly handling the different data types (discrete SNPs, continuous MRI) and the ordinal nature of disease progression leads to more accurate and robust models compared to methods that treat all data as continuous or ignore the severity order.
- Sparsity for Interpretability and Performance: The use of spike-and-slab priors for feature selection not only makes the model more interpretable by identifying key biomarkers but also contributes to improved prediction accuracy.
Key Results
- Association Discovery: The proposed SHML model significantly outperformed existing methods (CCA, Sparse CCA, MRLasso) in discovering associations between the two views (genetic and MRI data), as evidenced by precision-recall curves.
- AD Diagnosis Accuracy: On the ADNI dataset, the SHML model achieved the highest prediction accuracy for ordinal AD stages, outperforming methods like GP ordinal Regression by 10% and CCA+lasso by 22%. The SHML-LD (with Linkage Disequilibrium priors) further improved upon these results.
- Biologically Meaningful Associations: The model identified biologically meaningful associations between SNPs and MRI features. For instance, it highlighted the predictive power of cortical thickness measurements and identified specific genes (e.g., CAPZB, NCOA2, BCAR3) associated with brain atrophy patterns relevant to AD.
Limitations and Future Work
- Training Speed: The authors acknowledge that the proposed framework suffers from slow training speed, which is a common challenge in complex Bayesian models.
- Parameter Tuning: Careful tuning of parameters is required for optimal performance.
- Future Work: The authors suggest that the SHML framework, being a powerful extension of CCA, can be applied to a wide range of applications in biomedical research, such as eQTL analysis, especially when supervised by additional labeling information. The model's ability to handle heterogeneous data and incorporate prior knowledge makes it a versatile tool for complex biological data analysis.
Summary updated