Transcriptome prediction performance across machine learning models and diverse ancestries

Area of Science:

• Genetics and Genomics
• Computational Biology
• Biostatistics

Background:

• Transcriptome prediction methods like PrediXcan and FUSION are vital for complex trait mapping.
• Current models predominantly use linear assumptions (e.g., elastic net) trained on European populations.
• This limits imputation performance across diverse global ancestries.

Purpose of the Study:

• To optimize transcriptome imputation performance across global populations.
• To evaluate non-linear machine learning (ML) algorithms against traditional linear models.
• To assess the impact of ancestry matching on prediction accuracy.

Main Methods:

• Trained transcriptome prediction models using genotype and transcriptome data from the Multi-Ethnic Study of Atherosclerosis (MESA) across African, Hispanic, and European ancestries.
• Employed linear (elastic net - EN) and non-linear ML algorithms (random forest - RF, support vector regression - SVR, K nearest neighbor - KNN).
• Tested model performance using data from the Modeling the Epidemiology Transition Study (METS) in African ancestries and applied to a high-density lipoprotein (HDL) phenotype.

Main Results:

• Prediction performance was highest when training and testing populations shared similar ancestries.
• While EN generally outperformed other ML models, RF showed superior performance for specific genes, especially between disparate ancestries.
• RF imputation demonstrated potential robustness and reduced variability across global populations.
• Integrating RF models into PrediXcan identified potential gene associations for HDL phenotypes missed by EN models.

Conclusions:

• Non-linear ML models, particularly RF, offer complementary imputation strategies for transcriptome prediction.
• Diversifying training populations and incorporating various ML models can enhance the discovery of genes associated with complex traits.
• Improved imputation across diverse ancestries is crucial for advancing complex trait mapping in global health research.