The Construction of Multi-ethnic Polygenic Risk Score using Transfer Learning

Zhangchen Zhao, Lars G. Fritsche, Jennifer A. Smith, Bhramar Mukherjee, Seunggeun LeemedRxiv 2022 · DOI

TL;DR

Researchers have developed TL-PRS, a new method that uses transfer learning to adapt Polygenic Risk Scores (PRS) across different ancestries. By fine-tuning large-scale European-centric genetic data to smaller non-European datasets, the method achieves up to a 42% improvement in predicting disease risk for underrepresented groups, helping to bridge the "ancestry gap" in personalized medicine.

What Problem Does This Paper Address?

Most of what we know about the relationship between genetics and disease comes from Genome-Wide Association Studies (GWAS) conducted on individuals of European ancestry. Because genetic architectures and linkage disequilibrium (how genes are inherited together) vary across populations, a risk model trained on Europeans often performs poorly when applied to people of African, South Asian, or East Asian descent.

This "transferability gap" creates a significant equity issue: as genomic medicine becomes a standard of care, underrepresented populations risk being left behind with less accurate diagnostic tools. The challenge is that we often lack the massive sample sizes for non-European groups needed to train high-performing models from scratch.

Methods

The authors introduce TL-PRS (Transfer Learning Polygenic Risk Score). Instead of starting from zero, this approach treats an existing high-quality PRS (the "baseline" model from a large population) as a starting point and "fine-tunes" it for a target population.

  • The Mechanism: The method uses gradient descent to update the effect sizes (beta coefficients) of genetic variants. It takes the baseline effect sizes as initial values and iterates to minimize a loss function based on the target ancestry's data.
  • Flexibility: TL-PRS is "model-agnostic," meaning it can be applied to various existing PRS frameworks like Lassosum or PRS-CS.
  • Data Efficiency: Crucially, the training step can be performed using GWAS summary statistics rather than requiring sensitive individual-level genetic data, making it more scalable and privacy-friendly.
  • Multi-Source Integration (MTL-PRS): When data from multiple ancestries (e.g., European and East Asian) are available, the method can linearly combine them before fine-tuning, further boosting accuracy.

Datasets

The study validated the method using several major genomic resources:

  • Training Data: Large-scale GWAS summary statistics from the UK Biobank (UKBB) (European ancestry) and Biobank Japan (BBJ) (East Asian ancestry).
  • Target/Evaluation Data: Individual-level data from the UK Biobank for 10,285 South Asian (SAS) and 8,168 African (AFR) ancestry individuals.
  • Traits: The models were tested on eight traits, including six quantitative measures (like BMI, LDL cholesterol, and blood pressure) and two dichotomous diseases (Coronary Artery Disease and Type 2 Diabetes).

Key Insights and Ideas

  • The Power of Fine-Tuning: The "aha moment" is that we don't necessarily need more data; we need to use the data we have more intelligently. By adapting the differences between populations rather than relearning the entire genetic architecture, we can achieve high accuracy with much smaller samples.
  • Summary Statistics are Enough: The authors proved that using summary-level data for training performs almost identically to using individual-level data. This is a huge win for researchers who often face bureaucratic hurdles in accessing raw genetic files.
  • Robustness Across Architectures: TL-PRS proved effective across various "genetic architectures"—whether a trait is influenced by a few high-impact variants or thousands of tiny ones.
  • Optimal Stopping: The researchers found that the model reaches peak performance quickly (around the 5th iteration), after which it begins to overfit the small target dataset.

Key Results

  • Significant Gains: For African ancestry individuals, TL-PRS showed an average relative improvement of 13% to 42% over baseline methods.
  • Consistency: In a head-to-head comparison across 16 trait/ancestry combinations, the multi-source version (MTL-PRS-CS) ranked in the top three most accurate methods in 15 out of 16 cases.
  • Real-World Impact: In predicting Coronary Artery Disease (CAD) for the African cohort, TL-PRS identified a significantly higher cumulant prevalence in the top 10% risk group compared to standard methods (0.16 vs 0.12), demonstrating better identification of high-risk individuals.

Limitations and Future Work

Despite the improvements, the authors note several hurdles:

  • Absolute Accuracy: While relative improvements are high, the absolute prediction accuracy for some traits is still not high enough for standalone clinical use.
  • Admixed Populations: The current model focuses on distinct ancestry groups; future work is needed to adapt the method for individuals with complex, admixed backgrounds.
  • X-Factor: The study did not incorporate data from the X chromosome or rare genetic variants, both of which could harbor additional "missing" heritability.
  • Hyperparameter Tuning: While training doesn't require individual data, the "validation" step (tuning the learning rate) still does, though the authors suggest "pseudo-validation" as a potential future workaround.

Summary updated

Abstract

Abstract As most existing genome-wide association studies (GWAS) were conducted in European ancestry cohorts and as the existing PRS models have limited transferability across ancestry groups, PRS research on non-European ancestry groups is negatively impacted. Here we propose a novel PRS method using transfer learning techniques. Our approach, TL-PRS, uses gradient descent to fine-tune the baseline PRS model from an ancestry group with large sample GWAS to the dataset of target ancestry. In our application of constructing PRS for the six quantitative and two dichotomous traits for 10,285 South Asian and 8,168 African ancestry individuals in UK Biobank, TL-PRS achieved up to 42% average relative improvement compared to the existing methods. Our approach increases the transferability of PRSs across ancestries and thereby helps reduce existing inequities in genetics research.
BibTeX
@misc{zhao2022construction,
  author = {Zhangchen Zhao and Lars G. Fritsche and Jennifer A. Smith and Bhramar Mukherjee and Seunggeun Lee},
  title = {{The Construction of Multi-ethnic Polygenic Risk Score using Transfer Learning}},
  year = {2022},
  doi = {10.1101/2022.03.08.22272114},
  howpublished = {medRxiv},
}
Figure