A benchmark analysis of feature selection and machine learning methods for environmental metabarcoding datasets.

Area of Science:

• Ecology
• Bioinformatics
• Computational Biology

Background:

• Next-Generation Sequencing (NGS) and DNA metabarcoding generate large ecological community datasets.
• Metabarcoding data present analysis challenges due to sparsity, compositionality, and high dimensionality.
• Feature selection methods theoretically enhance eDNA metabarcoding data analysis by identifying key taxa.

Purpose of the Study:

• To compare the effectiveness of various feature selection methods for environmental DNA (eDNA) metabarcoding datasets.
• To evaluate how feature selection impacts the ability of machine learning models to link microbial community composition with environmental parameters.
• To provide guidelines for selecting appropriate feature selection strategies in metabarcoding studies.

Main Methods:

• A supervised machine learning framework was employed.
• Thirteen diverse environmental metabarcoding datasets were analyzed.
• Workflows included data preprocessing, feature selection, and Random Forest modeling.
• Model performance was assessed based on capturing ecological relationships.

Main Results:

• Feature selection frequently impaired, rather than improved, model performance for tree ensemble models like Random Forests.
• The optimal feature selection strategy was dataset-dependent.
• Using relative sequence counts negatively impacted model performance.
• Current methods for handling data compositionality may be insufficient.

Conclusions:

• Feature selection is not universally beneficial for eDNA metabarcoding analysis with Random Forests and can reduce model accuracy.
• Dataset-specific characteristics influence the outcome of feature selection.
• Novel approaches are essential to effectively manage the compositional nature of metabarcoding data.
• Improving data analysis methods is crucial for robust ecological inference from eDNA metabarcoding.