Poisson PCA: Poisson measurement error corrected PCA, with application to microbiome data.

Area of Science:

• Statistics
• Bioinformatics
• Computational Biology

Background:

• Principal Component Analysis (PCA) is crucial for dimensionality reduction.
• Poisson noise is common in biological count data, like microbiome data.
• Existing PCA methods struggle with Poisson noise and bias in variance estimation.

Purpose of the Study:

• To develop a semiparametric PCA method for Poisson-distributed data.
• To accurately estimate principal components and scores from latent Poisson means.
• To address bias in variance estimators for transformed and untransformed Poisson data.

Main Methods:

• Developed a semiparametric approach for bias correction in variance estimators.
• Applied methods to both untransformed and log-transformed Poisson means.
• Incorporated corrections for varying sequencing depths and exposure levels.
• Addressed computation of principal scores within the semiparametric framework.

Main Results:

• The semiparametric method effectively corrects bias in variance estimators.
• It accurately identifies principal components of latent log-transformed Poisson means.
• Outperforms the parametric Poisson Lognormal (PLN) model in identifying key components.
• Demonstrates greater robustness to outliers compared to parametric methods.

Conclusions:

• The proposed semiparametric PCA method is a robust and efficient alternative for analyzing Poisson-noisy data.
• It offers significant advantages over parametric approaches, particularly for microbiome and similar datasets.
• The method provides accurate principal component and score estimation with reduced computational cost.