Addressing overfitting bias due to sample overlap in polygenic risk scoring.

Area of Science:

• Genetics
• Neuroscience
• Biostatistics

Background:

• Polygenic risk scores (PRS) for Alzheimer's disease (AD) are widely used but often overlook sample overlap between major cohorts like the International Genomics of Alzheimer's Project (IGAP) and Alzheimer's Disease Neuroimaging Initiative (ADNI).
• This sample overlap can introduce significant overfitting bias, inflating the performance metrics of AD PRS.
• Accurate PRS are crucial for understanding genetic predisposition to AD.

Purpose of the Study:

• To develop and validate a method for adjusting Alzheimer's disease polygenic risk scores (PRS) for sample overlap.
• To mitigate overfitting bias in PRS derived from large, overlapping genetic datasets.
• To improve the accuracy and reliability of PRS for Alzheimer's disease research.

Main Methods:

• Developed an overlap-adjusted PRS (OA-PRS) method to correct for sample overlap in genetic datasets.
• Tested OA-PRS on simulated data with varying proportions of training, testing, and overlapping samples.
• Applied OA-PRS to the IGAP and ADNI datasets and validated results using visual diagnostics.

Main Results:

• OA-PRS effectively adjusted for sample overlap in both simulated and real-world datasets (IGAP and ADNI).
• The original IGAP PRS showed inflated performance (AUROC: 0.915) on overlapping samples, which OA-PRS corrected to 0.726, aligning with non-overlapping sample performance (0.712).
• Visual diagnostics confirmed the successful mitigation of overfitting bias by OA-PRS.

Conclusions:

• OA-PRS successfully adjusted IGAP-based PRS for overlapped ADNI samples, enabling full dataset utilization without overfitting risk.
• The developed OA-PRS method effectively mitigates overfitting bias caused by sample overlap in Alzheimer's disease genetic studies.
• Adjusted PRS demonstrated improved power for association studies with clinical features, enhancing their utility in AD research.