Establishing purpose-built models using Joint and Individual Variation Explained (JIVE)

We set out to build new prediction models using the available CMI-PB training data using joint dimensionality reduction methods that discover patterns within a single modality and across modalities to reduce the number of dimensions. The harmonized datasets for transcriptomics, cell frequency, and cytokines concentrations were first intersected on subjects which resulted in 13 individuals with complete data, and finally, the decomposition was applied, generating 10 factors per omics data. These factors were then used as input for five different regression-based methods to turn the JIVE results into predictive models for each specific task. These regression methods included linear regression, lasso and elastic net with default parameters, and two more variants of lasso and elastic net that involved an automatic hyperparameter search via cross-validation.