bg-szy Multi-Ancestry PRS Agent
The Multi-Ancestry PRS Agent provides AI-optimized polygenic risk score calculation designed to work across diverse ancestral populations. It addresses the critical limitation of European-biased GWAS by integrating trans-ancestry methods, improving risk prediction for underrepresented populations and enabling equitable precision medicine.
When to Use This Skill
- When calculating PRS for non-European ancestry individuals.
- For developing trans-ancestry risk prediction models.
- To reduce PRS bias across ancestral populations.
- When integrating multi-ancestry GWAS summary statistics.
- For research on PRS portability and equity.
Core Capabilities
Multi-Ancestry PRS: Calculate ancestry-aware polygenic scores.
Trans-Ancestry Optimization: Optimize weights across populations.
Local Ancestry Integration: Account for admixed genomes.
Ensemble Methods: Combine multiple PRS approaches.
Ancestry Calibration: Population-specific score calibration.
Equity Assessment: Evaluate prediction fairness across groups.
Supported Ancestries
| Ancestry | Training Data Availability | PRS Performance |
|---|---|---|
| European (EUR) | Excellent | High |
| East Asian (EAS) | Good | Good |
| South Asian (SAS) | Moderate | Moderate |
| African (AFR) | Limited | Lower |
| Hispanic/Latino (AMR) | Limited | Variable |
| Middle Eastern (MID) | Very Limited | Lower |
Multi-Ancestry Methods
| Method | Approach | Best For |
|---|---|---|
| PRS-CSx | Cross-population shrinkage | Multi-ancestry |
| PRS-Multi | Multi-population training | Large cohorts |
| Ensemble | Combined methods | Best overall |
| Local Ancestry | Ancestry-specific weights | Admixed populations |
| GPSMult | Integrated multi-trait | Correlated traits |
Workflow
Input: Individual genotypes, target ancestry, disease/trait.
Ancestry Inference: Determine genetic ancestry.
Method Selection: Choose optimal PRS approach.
Score Calculation: Compute ancestry-aware PRS.
Calibration: Apply population-specific calibration.
Risk Stratification: Categorize into risk groups.
Output: PRS, percentile, clinical interpretation.
Example Usage
User: "Calculate multi-ancestry coronary artery disease PRS for this admixed individual with African and European ancestry."
Agent Action:
python3 Skills/Precision_Medicine/Multi_Ancestry_PRS_Agent/calc_prs.py \
--genotypes patient_genotypes.vcf.gz \
--ancestry admixed_AFR_EUR \
--local_ancestry lai_segments.bed \
--trait coronary_artery_disease \
--method prs_csx \
--gwas_summary_stats eur_gwas.txt,afr_gwas.txt \
--calibration_cohort 1kg_admixed \
--output prs_results/
Input Requirements
| Input | Format | Purpose |
|---|---|---|
| Genotypes | VCF/PLINK | Individual variants |
| Ancestry | Estimated or self-reported | Method selection |
| GWAS Summary Stats | Multiple ancestries | Score weights |
| Local Ancestry | LAI segments | Admixture handling |
| Reference Panel | Multi-ancestry | LD calculation |
Output Components
| Output | Description | Format |
|---|---|---|
| PRS Score | Raw polygenic score | .csv |
| Percentile | Population-specific ranking | .csv |
| Risk Category | High/Intermediate/Low | .csv |
| Ancestry Breakdown | Component scores | .json |
| Confidence Interval | Score uncertainty | .json |
| Clinical Interpretation | Risk explanation | .md |
Disease-Specific Performance
| Disease | Multi-Ancestry AUC | EUR Only AUC | Improvement |
|---|---|---|---|
| CAD | 0.75-0.80 | 0.70-0.85 | 5-10% in non-EUR |
| Type 2 Diabetes | 0.70-0.75 | 0.65-0.72 | 8-12% in AFR |
| Breast Cancer | 0.65-0.72 | 0.60-0.70 | 5-8% globally |
| Alzheimer's | 0.70-0.78 | 0.65-0.75 | 5-10% in diverse |
AI/ML Components
PRS Optimization:
- Bayesian shrinkage (PRS-CS)
- Cross-population learning
- Neural network weight optimization
Ancestry Inference:
- Supervised classification
- Unsupervised clustering (PCA, ADMIXTURE)
- Local ancestry inference (RFMix)
Ensemble Learning:
- Stacking multiple PRS methods
- Ancestry-stratified weighting
- Uncertainty quantification
Clinical Integration
| Application | PRS Role | Clinical Action |
|---|---|---|
| Primary Prevention | Risk stratification | Screening intensity |
| Risk Communication | Personalized risk | Lifestyle modification |
| Treatment Selection | Predicted response | Drug choice |
| Family Screening | Cascade testing | Genetic counseling |
Prerequisites
- Python 3.10+
- PLINK 2.0
- PRSice-2, LDpred2, PRS-CSx
- Multi-ancestry reference panels
- GWAS summary statistics
Related Skills
- PRS_Net_Deep_Learning_Agent - Deep learning PRS
- Pharmacogenomics_Agent - Drug-gene interactions
- PopEVE_Variant_Predictor_Agent - Variant interpretation
- DiagAI_Agent - Clinical integration
Bias and Fairness
| Bias Type | Cause | Mitigation |
|---|---|---|
| Discovery Bias | EUR-dominated GWAS | Multi-ancestry GWAS |
| LD Variation | Population-specific LD | Local ancestry adjustment |
| Allele Frequency | Differing frequencies | Population-specific weights |
| Effect Size | Heterogeneous effects | Trans-ancestry meta-analysis |
Large-Scale Initiatives
| Initiative | Focus | Contribution |
|---|---|---|
| All of Us | US diversity | 1M diverse participants |
| PAGE | Multi-ethnic GWAS | Discovery in diverse |
| H3Africa | African genomics | Continental diversity |
| Mexican Biobank | Latin American | Admixed populations |
| GBMI | Global Biobank | Multi-ancestry meta-analysis |
Special Considerations
- Self-Reported Ancestry: May not match genetic ancestry
- Admixture: Require local ancestry methods
- Population Stratification: Careful covariate adjustment
- Clinical Validity: Validate in target population
- Health Equity: Consider access disparities
ESC Guidelines Integration (2025)
| Recommendation | PRS Role | Evidence Level |
|---|---|---|
| CV Risk Assessment | Risk modifier | IIa, B |
| Statin Decisions | Borderline risk reclassification | IIa, B |
| Family History Enhancement | Quantify genetic burden | IIa, C |
Limitations
| Limitation | Impact | Research Needed |
|---|---|---|
| AFR Performance | Lower accuracy | More GWAS |
| Rare Variants | Not captured | WGS integration |
| Gene-Environment | Not modeled | Interaction studies |
| Clinical Utility | Limited evidence | Randomized trials |
Author
AI Group - Biomedical AI Platform