Partial Least Squares with Structured Output for Modelling the Metabolomics Data Obtained from Complex Experimental...

Area of Science:

• Metabolomics
• Bioinformatics
• Chemometrics

Background:

• Partial least squares (PLS) is a standard supervised method for analyzing multivariate metabolomics data.
• Traditional PLS models (regression and classification) struggle with complex experimental designs involving multiple interacting factors.
• Existing coding methods for PLS can lead to ambiguous results when data does not fit pure regression or classification scenarios.

Purpose of the Study:

• To investigate a hybrid target matrix (Y) coding strategy for PLS modeling in metabolomics.
• To develop a coding approach that better represents complex experimental designs compared to standard methods.
• To improve the interpretability of PLS models for multivariate metabolomics data.

Main Methods:

• Developed a novel hybrid target matrix (Y) coding approach for PLS.
• The coding principle is inspired by structural modeling in machine learning.
• Applied and evaluated the new coding strategy using two real-world metabolomics datasets.

Main Results:

• The proposed hybrid Y coding effectively captures complex experimental designs in metabolomics.
• Models built with the new coding demonstrate improved interpretability compared to classic PLS approaches.
• The method provides a more nuanced analysis for multivariate data with multiple interacting factors.

Conclusions:

• The hybrid target matrix (Y) coding offers a superior alternative for PLS analysis in metabolomics with complex designs.
• This approach enhances the clarity and reliability of results derived from multivariate data.
• It provides a valuable tool for researchers dealing with intricate experimental setups in metabolomics.