Using pre-training and interaction modeling for ancestry-specific disease prediction using multiomics data from the...

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Genome-wide association studies (GWAS) often lack diversity, leading to underperformance in non-European populations.
• Existing disease prediction models struggle to generalize across ancestries, potentially widening health disparities.

Purpose of the Study:

• To evaluate if interaction modeling and pretraining enhance disease prediction accuracy in diverse ancestries.
• To assess the performance of glinternet and pretrained lasso models using multiomic data.

Main Methods:

• Utilized Group-LASSO INTERaction-NET (glinternet) and pretrained lasso models.
• Trained and validated models on multiomic data from UK Biobank participants (>96,000 individuals) across diverse ancestries.
• Evaluated predictive performance for 8 common diseases using ROC-AUC scores.

Main Results:

• 16 out of 96 models showed statistically significant improvements in predictive performance (ROC-AUC).
• Enhanced accuracy was observed for diseases including diabetes, arthritis, gallstones, cystitis, asthma, and osteoarthritis.
• The benefits of interaction terms and pretraining were modest and inconsistent across all evaluated diseases.

Conclusions:

• Interaction modeling and pretraining can offer incremental improvements in disease prediction accuracy for diverse populations.
• The effectiveness of these methods is disease-specific and requires further investigation.
• The study highlights the need for more inclusive genetic research and improved predictive modeling strategies.