PLSDA-batch

PLSDA-batch removes batch effects from microbiome datasets using a multivariate, non-parametric Partial Least Squares Discriminant Analysis (PLSDA) framework to extract and subtract batch-associated latent components while preserving treatment-related variation.

Key Features:

• PLSDA framework: Implements a multivariate, non-parametric Partial Least Squares Discriminant Analysis tailored for microbiome data.
• Latent component estimation: Estimates treatment and batch variations via latent components.
• Batch-component subtraction: Removes batch-associated components by subtracting them from the dataset.
• Unbalanced-design variant: Provides a variant specifically designed to handle unbalanced batch × treatment designs.
• Variable selection variant: Offers a variant that performs variable selection during component estimation to reduce overfitting.
• Microbiome-specific considerations: Accounts for microbiome data characteristics including zero inflation, overdispersion, and variable correlation.
• Preservation of treatment signal: Aims to preserve critical treatment-related variation while removing batch effects.
• Comparative benchmarking: Reported performance comparisons include removeBatchEffect, ComBat, and Surrogate Variable Analysis (SVA).

Scientific Applications:

• Batch correction in microbiome studies: Removes unwanted batch variation from microbiome datasets for downstream analyses.
• Analysis of zero-inflated and overdispersed data: Applied to datasets exhibiting zero inflation, overdispersion, and correlated variables.
• Unbalanced experimental designs: Corrects batch effects in unbalanced batch × treatment study designs.
• Taxon selection for downstream analysis: Facilitates selection of biologically relevant taxa after batch correction.
• Method validation: Evaluated using simulated data and three case studies to assess removal of batch variation while preserving treatment signals.

Methodology:

Uses PLSDA to estimate latent components representing treatment and batch variation, subtracts batch-associated components from the data, and provides two variants: one for unbalanced batch × treatment designs and one incorporating variable selection during component estimation to prevent overfitting.