Cluster-Partial Least Squares (c-PLS) regression analysis: Application to miRNA and metabolomic data

Biomedicine and biological research frequently involve analyzing large datasets generated by high-throughput technologies like genomics, transcriptomics, miRNomics, and metabolomics. Pathway analysis is a common computational approach used to understand the impact of experimental conditions, phenotypes, or interventions on biological pathways and networks. This involves statistical analysis of omic data to identify differentially expressed variables and mapping them onto predefined pathways. Analyzing such datasets often requires multivariate techniques to extract meaningful insights such as Partial Least Squares (PLS). Variable selection strategies like interval-PLS (iPLS) help improve understanding and predictive performance by identifying informative variables or intervals. However, iPLS is suboptimal to treat omic data such as metabolic or miRNA profiles, where features cannot be distributed along a continuous dimension describing their relationships as in e.g., vibrational or nuclear magnetic resonance spectroscopy.

This study introduces a novel variable selection approach called cluster PLS (c-PLS) that aims to assess the joint impact of variable groups selected based on biological characteristics (such as miRNA-regulated metabolic pathway or lipid classes) on the predictive performance of a multivariate model. The usefulness of c-PLS is shown using miRNomic and metabolomic datasets obtained from the analysis of 24 liver tissue biopsies collected in the frame of a clinical study of steatosis.

Results obtained show that c-PLS enables analyzing the effect of biologically relevant variable clusters, facilitating the identification of biological processes associated with the independent variable, and the prioritization of the biological factors influencing model performance, thereby improving the understanding of the biological factors driving model predictions. While the strategy is tested for the evaluation of PLS models, it could be extended to other linear and non-linear multivariate models.