Exploration of biomarkers of Alzheimer's disease based on orthogonal multi-task canonical correlation analysis

Tao Yang, GuangYu Wan, Xiong ZhouBMC Medical Imaging 2025 · DOI

TL;DR

This paper introduces an "orthogonal multi-task sparse canonical correlation analysis" (MTOSCCA) algorithm to identify biomarkers for Alzheimer's disease (AD) by integrating brain imaging, genetic variations (SNPs), and gene expression data. The novel orthogonal constraints in MTOSCCA improve feature selection and lead to more robust AD and mild cognitive impairment (MCI) biomarkers with diagnostic significance.

What Problem Does This Paper Address?

Alzheimer's disease (AD) is a devastating neurodegenerative disorder characterized by memory loss and cognitive decline, with its incidence rising globally. While genetic factors are known to play a significant role, pinpointing specific genetic variations and their interplay with brain changes that lead to AD remains a challenge. Traditional methods often struggle to integrate diverse data types like brain scans and genetic information effectively, leading to potential redundancy and noise. This paper aims to bridge this gap by developing a more sophisticated analytical approach to uncover the complex imaging-genetic signatures associated with AD and its precursor, MCI.

Methods

The core of this research is the development of the Orthogonal Multi-Task Sparse Canonical Correlation Analysis (MTOSCCA) algorithm. This method builds upon existing Multi-Task Sparse Canonical Correlation Analysis (MTSCCA) by introducing orthogonal constraints on the weight vectors for both imaging and genetic data. These constraints are crucial because they encourage the selected features (e.g., specific brain regions, SNPs, or genes) to be linearly independent, thereby preventing redundancy and improving the algorithm's ability to identify unique biological signals.

The MTOSCCA algorithm aims to maximize the correlation between different data modalities while simultaneously enforcing sparsity (selecting a subset of features) and orthogonality. The paper details the iterative optimization process for deriving the weight vectors, ensuring that the algorithm can effectively handle multiple data types (tasks) and extract meaningful associations.

For feature selection prior to MTOSCCA, the researchers employed the Random Forest (RF) algorithm. This was used to identify a subset of SNPs that matched the number of genes, ensuring that the MTOSCCA algorithm received balanced input from both genetic modalities. RF was chosen for its robustness and ability to capture complex relationships.

Datasets

The study utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. Specifically, they focused on data from the ADNI1 phase, including:

  • Structural Magnetic Resonance Imaging (sMRI): Used to extract regional brain information.
  • Single Nucleotide Polymorphism (SNP) data: Representing genetic variations.
  • Gene Expression data: Indicating the activity of genes.

The dataset comprised 138 participants: 102 with MCI and 36 with AD. The data was preprocessed to extract regional information from sMRI (90 brain regions) and to filter and annotate SNPs related to known AD risk genes. Gene expression data was also filtered based on differential expression analysis.

Key Insights and Ideas

  • Orthogonality as a Key to Reducing Redundancy: The introduction of orthogonal constraints in MTOSCCA is a significant innovation. It effectively tackles the problem of highly correlated features that might otherwise obscure true biological signals, leading to more precise biomarker identification.
  • Multimodal Integration for Deeper Understanding: Combining sMRI, SNP, and gene expression data provides a more holistic view of AD pathogenesis than any single modality alone. This integrated approach is essential for understanding the complex interplay of genetic predisposition and brain changes.
  • Biomarkers with Diagnostic Significance: The identified top markers (ROIs, SNPs, and genes) demonstrate not only biological relevance but also significant diagnostic power for differentiating between AD, MCI, and healthy controls.
  • Robustness to Noise: The MTOSCCA algorithm showed improved performance compared to MTSCCA in simulation studies, particularly in its ability to maintain high correlation coefficients even under varying levels of noise.

Key Results

  • The MTOSCCA algorithm successfully identified top biomarkers (ROIs, SNPs, and genes) with diagnostic significance for AD and MCI.
  • The combined predictive performance of the top features from all three modalities achieved a high Area Under the Curve (AUC) of 0.955, indicating excellent diagnostic accuracy.
  • Specific ROC curve analyses showed that many of the top ROIs, genes, and half of the top SNPs had AUC values greater than 0.5 and fell within reasonable confidence intervals.
  • The MTOSCCA algorithm demonstrated a slightly better canonical correlation coefficient (CCC) of 0.347 compared to MTSCCA's 0.316 in their comparative analysis.
  • The algorithm showed better anti-noise performance in simulation studies compared to MTSCCA.

Limitations and Future Work

  • Absence of Healthy Controls: The study did not include a healthy control group, which limits direct comparisons between normal aging and pathological processes.
  • Sample Imbalance: The imbalance in the ratio of MCI to AD participants could potentially bias the algorithm towards the majority class (MCI). While efforts were made to mitigate this, it remains a challenge.
  • APOE Gene Exclusion: Notably, the APOE gene, a well-established AD risk factor, did not appear among the top SNPs or genes. The authors suggest this could be due to algorithmic limitations, sample characteristics, or the fact that APOE's influence might be more pronounced at the functional or metabolic level.

Future work suggested includes:

  • Incorporating a healthy control group for more comprehensive comparisons.
  • Addressing sample imbalance through techniques like SMOTE or adjusting algorithm weights.
  • Adding clinical information as a priori knowledge to the algorithm for more interpretable results.
  • Optimizing the algorithm and data processing, expanding the sample size, and integrating functional imaging data.
  • Investigating the relationship between the APOE gene and AD-related features more thoroughly.

Summary updated

BibTeX
@article{yang2025exploration,
  author = {Tao Yang and GuangYu Wan and Xiong Zhou},
  title = {{Exploration of biomarkers of Alzheimer's disease based on orthogonal multi-task canonical correlation analysis}},
  journal = {BMC Medical Imaging},
  year = {2025},
  doi = {10.1186/s12880-025-01782-2},
}
Figure