Scalable estimation and regularization for the logistic normal multinomial model.

Area of Science:

• Microbiology
• Statistical Modeling
• Bioinformatics

Background:

• Clustered multinomial data, common in microbiome studies, present challenges with zero and low-count taxa.
• Standard count normalization methods often yield inaccurate abundance estimates for rare taxa.

Purpose of the Study:

• To develop a robust statistical model for analyzing high-dimensional clustered multinomial data, specifically for microbiome composition.
• To improve abundance estimation for low-count taxa by accounting for heterogeneity and overdispersion.

Main Methods:

• Proposed a logistic normal multinomial (LNM) model with an arbitrary correlation structure.
• Developed a scalable parameter estimation approach using a stochastic approximation EM algorithm with Hamiltonian Monte Carlo sampling.
• Implemented a covariance-regularized estimator to address ill-conditioning in high dimensions.

Main Results:

• The LNM model effectively estimates taxa compositions by leveraging information across subjects.
• The proposed computational methods enable scalable parameter estimation in high-dimensional settings.
• Simulations and human gut microbiome data analysis demonstrated the advantages of the new methods.

Conclusions:

• The LNM model offers a significant advancement for analyzing microbiome count data, particularly for rare taxa.
• The developed algorithms provide efficient and robust parameter estimation for complex biological data.
• This approach enhances the accuracy and reliability of microbiome composition analysis.