Transfer learning with false negative control improves polygenic risk prediction.

Area of Science:

• Genetics
• Bioinformatics
• Computational Biology

Background:

• Polygenic risk score (PRS) analysis aggregates genetic variants to estimate disease predisposition.
• PRS methods often face challenges with mismatched ancestral backgrounds between training (base) and prediction (target) datasets.
• Leveraging large-scale base data for diverse target populations requires advanced analytical approaches.

Purpose of the Study:

• To develop a transfer learning framework for accurate PRS prediction using knowledge from base data with potentially different ancestral backgrounds.
• To enhance the computational and statistical efficiency of PRS model training.
• To improve the accuracy of trans-data prediction, mitigating issues arising from ancestral heterogeneity.

Main Methods:

• A two-step transfer learning approach is proposed, treating GWAS summary statistics from base data as pre-trained model knowledge.
• Step 1: False Negative Control (FNC) marginal screening is employed to extract relevant knowledge from the base data.
• Step 2: Joint model training integrates knowledge from the base data with target training data for prediction.

Main Results:

• The proposed framework significantly enhances computational and statistical efficiency in joint model training.
• The approach effectively alleviates overfitting issues common in PRS analyses.
• Accurate trans-data prediction is facilitated, even with substantial heterogeneity between base and target datasets.

Conclusions:

• The transfer learning framework offers a robust solution for PRS prediction across diverse ancestral groups.
• This method improves the utility of large-scale genomic datasets for personalized risk prediction.
• The approach demonstrates potential for broader application in genetic epidemiology and precision medicine.