Exploring the role of normalization and feature selection in microbiome disease classification pipelines

Area of Science:

• Microbiome analysis
• Bioinformatics
• Machine learning

Background:

• 16S rRNA microbiome data presents challenges like high dimensionality, compositionality, and sparsity.
• Small sample sizes often limit the effectiveness of machine learning models.
• Comparative studies on feature selection and normalization in microbiome data are scarce.

Purpose of the Study:

• To evaluate the impact of various feature selection techniques and normalization strategies on microbiome-based disease classification.
• To identify optimal combinations of methods for robust biomarker discovery and improved classification performance.
• To compare the effectiveness of different feature selection algorithms.

Main Methods:

• Evaluation of multiple feature selection techniques (e.g., mRMR, LASSO, Autoencoders, Mutual Information, ReliefF).
• Assessment of different normalization strategies (e.g., centered log-ratio, presence-absence).
• Integration of feature selection with machine learning classifiers (e.g., logistic regression, SVM, random forest).

Main Results:

• Centered log-ratio normalization enhanced logistic regression and SVM performance.
• Random forest models performed well with relative abundances.
• Minimum Redundancy Maximum Relevancy (mRMR) and LASSO provided compact feature sets and comparable performance to other methods, with LASSO offering lower computation times.
• Presence-absence normalization achieved performance similar to abundance-based methods.

Conclusions:

• Feature selection pipelines significantly improve model focus and robustness by reducing the feature space.
• mRMR and LASSO are identified as the most effective feature selection methods for microbiome disease classification across diverse datasets.