Robust latent-variable interpretation of in vivo regression models by nested resampling.

Area of Science:

• Systems biology
• Bioinformatics
• Statistical modeling

Background:

• Multilinear methods like partial least squares regression (PLSR) effectively analyze high-dimensional biological data.
• In vivo data present challenges due to high animal-to-animal variability and nested data structures (cells within tissues within animals).

Purpose of the Study:

• To introduce principled resampling strategies that preserve the hierarchical structure of in vivo biological data.
• To assess the impact of these resampling strategies on the interpretation of multidimensional decompositions and latent variables (LVs).

Main Methods:

• Development and application of nested resampling strategies to account for tissue-animal hierarchy in molecular-phenotypic data.
• Comparison of PLSR model interpretation with and without nested resampling for both in vivo and in vitro datasets.

Main Results:

• Interpretation of decomposed latent variables (LVs) in PLSR models changes significantly when accounting for the in vivo data hierarchy.
• Lagging LVs, which appear significant in global-average models, are unstable under nested resampling, indicating they may not represent robust biological insights.
• Resampling methods are less critical for in vitro data due to lower replicate-to-replicate variance.

Conclusions:

• Nested resampling provides a crucial quality control step for validating regression models derived from in vivo data.
• The findings highlight challenges and opportunities in translating systems biology approaches from in vitro to in vivo studies.
• Careful consideration of data hierarchy and variability is essential for reliable interpretation of biological insights from complex organismal data.