Classifying high-dimensional phenotypes with ensemble learning.

Area of Science:

• Biological classification
• Machine learning in biology
• Phenotypic data analysis

Background:

• Traditional linear discriminant functions struggle with high-dimensional, complex biological datasets.
• Existing machine learning studies often lack broad applicability across organisms, algorithms, or tasks.
• The potential of ensemble learning for biological classification remains underexplored.

Approach:

• Conducted a meta-analysis of 33 algorithms across 20 high-dimensional shape phenotype datasets (>20,000 samples).
• Utilized an ensemble learning framework for preprocessing, training, and evaluating individual and combined models.
• Assessed performance across binary and multi-class biological classification tasks.

Key Points:

• Discriminant analysis variants and neural networks showed strong performance as base learners, but with dataset-specific variability.
• Ensemble models consistently achieved superior accuracy, increasing performance by up to 3% over top individual models.
• Dataset properties like class R², shape distances, and variance ratios positively impacted performance; covariance distances negatively impacted it.

Conclusions:

• Ensemble models provide a robust, data-agnostic, and highly accurate solution for biological classification challenges.
• Selecting algorithms based on prior studies is unreliable; ensemble approaches offer superior flexibility and performance.
• Understanding dataset and phenotypic properties is crucial for optimizing classification accuracy, with the R package 'pheble' providing accessible tools.