Batch normalization followed by merging is powerful for phenotype prediction integrating multiple heterogeneous...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• Heterogeneity across genomic studies significantly impairs machine learning model performance for cross-study phenotype predictions.
• Developing robust machine learning algorithms requires overcoming data heterogeneity for reproducible prediction performance on independent datasets.

Purpose of the Study:

• To investigate optimal approaches for integrating diverse omics data studies despite various heterogeneities.
• To evaluate the effectiveness of different data integration and batch normalization methods in enhancing cross-study prediction accuracy.

Main Methods:

• Developed a workflow to simulate various types of genomic data heterogeneity.
• Evaluated multiple data integration strategies combined with ComBat batch normalization.
• Applied and validated methods on colorectal cancer (CRC) metagenomic and tuberculosis (TB) gene expression datasets.

Main Results:

• Genomic study heterogeneity negatively impacts machine learning classifier reproducibility.
• ComBat normalization enhanced prediction performance in heterogeneous populations and removed batch effects.
• Prediction accuracy decreased when training and testing populations had different underlying disease models.
• Merging and integration methods showed variable performance depending on the scenario, but improved with ComBat normalization.

Conclusions:

• Batch normalization (e.g., ComBat) is crucial for mitigating population differences and batch effects in cross-study genomic data.
• Both merging strategies and integration methods, when combined with batch normalization, achieve good predictive performance.
• Rank aggregation methods offer a viable alternative for boosting phenotype prediction, comparable to other ensemble learning approaches.