When Transfer Learning Hurts Genetic Risk Prediction

Learning Outcomes

After studying this material, you should be able to:

1. Explain why polygenic risk scores (PRSs) perform poorly across populations
2. Describe how transfer learning applies to genomic prediction
3. Analyze when cross-population data helps vs. hurts predictive accuracy
4. Distinguish between conserved and population-specific genetic traits
5. Compare three modeling strategies for cross-population PRS development
6. Apply guidelines for choosing the right modeling approach based on sample size and trait type

---

Step 1: Foundation — What Are Polygenic Risk Scores?

Definition

A Polygenic Risk Score (PRS) is a numerical estimate of an individual's genetic predisposition to a disease or trait, calculated by combining the effects of hundreds to millions of genetic variants (typically single nucleotide polymorphisms, or SNPs).

How It Works

Individual's DNA
        ↓
Identify relevant genetic variants (SNPs)
        ↓
Weight each variant by its effect size
        ↓
Sum all weighted variants
        ↓
= Polygenic Risk Score (PRS)

Real-World Example

For BMI prediction, your PRS might combine 500,000 genetic variants, each contributing a tiny effect, to estimate your genetic tendency toward higher or lower body weight.

Key Metric: SNP Heritability

• Measures how much of trait variation is explained by measured genetic variants
• In this study, ranged from 0.07 to 0.28 across eight traits
• Higher heritability = genetics plays a larger role in that trait

---

Step 2: The Core Problem — Why PRSs Fail Across Populations

The European Bias Problem

Historically, genome-wide association studies (GWASs) — the research that identifies which genetic variants matter — have been conducted almost exclusively in European populations.

Historical GWAS Data
├── ~80% European populations
├── ~10% Asian populations
└── ~10% All other populations combined

Three Reasons Cross-Population PRSs Fail

The Practical Barrier

Conducting new, large-scale GWASs from scratch (de novo) across hundreds of thousands of individuals is cost-prohibitive for most healthcare systems.

This creates the central challenge: How do we improve PRS accuracy for underrepresented populations without always starting from scratch?

---

Step 3: The Proposed Solution — Transfer Learning

What Is Transfer Learning?

Transfer learning means taking knowledge learned from one domain and applying it to a related but different domain.

Transfer Learning in Genomics:

Large European GWAS (source)
         ↓
    Transfer knowledge
         ↓
Small Target Population GWAS (target)
         ↓
    Combined, improved PRS

The Study Setup

This research tested transfer learning between two major biobanks:

Eight traits studied:

• Body Mass Index (BMI)
• Systolic blood pressure
• Diastolic blood pressure
• Red blood cell count
• White blood cell count
• HDL cholesterol
• LDL cholesterol
• Blood glucose

Performance measured using: Pearson correlation between predicted and actual trait values

---

Step 4: The Surprising Finding — More Data Isn't Always Better

The Core Discovery

⚠️ Key Insight: Adding European data helps when Japanese sample sizes are small, but actively hurts prediction accuracy once Japanese samples grow large enough.

The Crossover Point

PRS Performance in Japanese Population

High |        ╭──── Japanese-only model
     |       ╱
     |      ╱  ← Crossover (~15,000 samples)
     |     ╱╲
     |    ╱  ╲____ Combined model (European + Japanese)
Low  |___╱
     0    5k   15k   25k   50k   100k
          Japanese Sample Size

What This Means Practically

Why Does This Happen?

When European data is included in training:

• It introduces out-of-distribution noise — genetic patterns that don't apply to Japanese individuals
• As Japanese data grows, this noise outweighs the statistical benefit of larger sample size
• The model gets "confused" by conflicting genetic signals from different populations

---

Step 5: Trait Matters — Conserved vs. Population-Specific Traits

The Concept of Genetic Correlation

Genetic correlation measures how similar the genetic basis of a trait is across two populations.

Genetic Correlation Scale:
0.0 ────────────────────── 1.0
Completely          Identical genetic
different           architecture
(population-        (conserved trait)
specific)

Two Categories of Traits

Category 1: Conserved Traits

(e.g., BMI, blood pressure, blood cell counts)

• High genetic correlation across populations
• Similar genetic variants drive the trait in both Europeans and Japanese
• Benefit from European data up to 25,000–40,000+ Japanese samples

Conserved Trait (e.g., BMI):

Performance
    │                    ╭── Japanese-only
    │                   ╱
    │                  ╱  ← Crossover at ~40k
    │    ╭────────────╱
    │   ╱  Combined model stays competitive longer
    └──────────────────────────
       5k  15k  25k  40k  100k

Category 2: Population-Specific Traits

(e.g., HDL cholesterol, LDL cholesterol, blood glucose)

• Low genetic correlation across populations
• Different genetic variants drive the trait in each population
• European data becomes harmful much sooner (crossover at ~15,000 samples)
• Optimal European sample size is also smaller

Population-Specific Trait (e.g., HDL):

Performance
    │         ╭── Japanese-only
    │        ╱
    │       ╱  ← Early crossover (~15k)
    │  ╭───╱
    │ ╱  Combined model degrades faster
    └──────────────────────────
       5k  15k  25k  40k  100k

Summary Table

---

Step 6: Three Modeling Strategies Compared

Strategy 1: European Discovery + Elastic Net

• Use variants discovered in European GWAS
• Train elastic net model on those variants
• Limitation: Misses variants unique to Japanese population

Strategy 2: Cross-Population Meta-Analysis + Elastic Net

• Run GWAS in both European (UKB) and Japanese (BBJ) populations
• Combine both sets of results to identify candidate variants
• Train elastic net on the expanded variant set

UKB European GWAS ──┐
                    ├── Meta-Analysis → Expanded Variants → Elastic Net
BBJ Japanese GWAS ──┘

Best for: Population-specific traits (HDL, LDL, blood glucose) at moderate sample sizes

Strategy 3: PRS-CSx

• A specialized method that dynamically weights population-specific models
• Theoretically less sensitive to whether a trait is conserved or population-specific
• Requires more data to work well

Performance Comparison

Performance Over Time:

High │                              ╭── PRS-CSx
     │              ╭──────────────╯
     │    ╭─────────╯  Meta-analysis
     │────╯  European-only
Low  └────────────────────────────────
     5k   25k   50k   75k   100k
          BBJ Sample Size

---

Step 7: Practical Decision Framework

When to Use Each Approach

START: What is your target population sample size?
              │
    ┌─────────┴──────────┐
    │                    │
  < 15,000            > 15,000
    │                    │
Use European         Is the trait conserved
data pooling         or population-specific?
                         │
              ┌──────────┴──────────┐
              │                     │
          Conserved           Population-specific
              │                     │
    Continue pooling         Stop pooling European
    European data up         data; use meta-analysis
    to 40k samples           or Japanese-only model

Key Decision Factors

---

Step 8: Broader Implications

What This Research Tells Us

1. Bigger isn't always better — blindly adding out-of-population data can harm predictions

2. Know your trait — genetic correlation between populations should guide your modeling strategy

3. Method selection matters — simple elastic net models can outperform sophisticated methods like PRS-CSx at smaller sample sizes

4. Diverse biobanks are essential — no amount of transfer learning fully compensates for lack of target-population data

The Path Forward

Current State                    Ideal Future State
─────────────────                ──────────────────
European-dominated GWAS    →     Diverse, global biobanks
One-size-fits-all PRS      →     Population-tailored models
Transfer learning as fix   →     Transfer learning as supplement

---

Summary: Key Takeaways