Implications of confounding from unmodeled interactions between explanatory variables when using latent variable...

Area of Science:

• Chemometrics
• Natural Product Research
• Bioactivity Screening

Background:

• Partial Least Squares (PLS) regression is widely used for analyzing datasets with linear dependencies between explanatory variables.
• Variable importance ranking in PLS models is often used to guide experimental decisions, such as identifying bioactive compounds in natural product extracts.
• A common challenge is that the number of compounds often exceeds the number of samples, leading to correlated concentrations and potential synergistic or antagonistic interactions.

Purpose of the Study:

• To investigate the limitations of standard Partial Least Squares (PLS) regression in interpreting variable importance when dealing with correlated variables and interactions.
• To demonstrate how unmodeled interactions can lead to erroneous conclusions in natural product research.
• To present an improved modeling approach for accurate interpretation and inference.

Main Methods:

• Application of Partial Least Squares (PLS) regression to analyze datasets with linear dependencies.
• Inclusion of interaction terms between explanatory variables in the PLS model.
• Utilizing selectivity ratio plots for enhanced model visualization and interpretation.

Main Results:

• Standard PLS models can produce misleading variable importance rankings when interactions are present and unmodeled.
• A practical example from natural product research illustrates the consequences of confounding due to unmodeled interactions.
• Incorporating interactions into the PLS model and using selectivity ratio plots allows for more reliable interpretation.

Conclusions:

• Standard PLS variable importance can be unreliable in complex natural product mixtures.
• Accounting for interactions is crucial for accurate interpretation of PLS models in such contexts.
• Selectivity ratio plots offer a valuable tool for visualizing and inferring results from interaction-aware PLS models.