Probabilistic Joint and Individual Variation Explained (ProJIVE) for Data Integration

Area of Science:

• Multivariate data analysis
• Bioinformatics
• Neuroimaging analysis

Background:

• Collecting diverse data (genomics, metabolomics, neuroimaging) from common subjects is prevalent in modern science.
• Existing methods for analyzing joint variation across datasets may lack accuracy or comprehensive component isolation.

Purpose of the Study:

• To develop a probabilistic model for the Joint and Individual Variation Explained (JIVE) framework using an expectation-maximization (EM) algorithm.
• To enhance the accuracy of estimating joint and individual variation components simultaneously.
• To apply the developed method, ProJIVE, to neuroimaging and cognitive data in Alzheimer's disease research.

Main Methods:

• Developed an expectation-maximization (EM) algorithm to estimate a probabilistic JIVE model.
• Extended probabilistic principal component analysis (PCA) to accommodate multiple datasets.
• Employed a maximum likelihood approach for simultaneous estimation of joint and individual components.

Main Results:

• The ProJIVE method successfully identified biologically meaningful sources of variation in Alzheimer's disease.
• Joint morphometry and cognition scores derived from ProJIVE showed strong correlations with established, more costly biomarkers.
• The probabilistic JIVE model demonstrated potential for greater accuracy in component estimation compared to existing techniques.

Conclusions:

• ProJIVE offers a robust probabilistic framework for dissecting joint and individual variation in multi-modal datasets.
• The method provides valuable insights into the interplay of brain morphometry and cognition in Alzheimer's disease.
• ProJIVE's findings highlight its utility in biomarker discovery and understanding complex diseases.