Genetic Prediction of Circulating Lipoprotein(a) Levels in Diverse Populations

This study demonstrates that a haplotype-based genetic model using standard genome-wide data can effectively predict elevated circulating lipoprotein(a) levels across diverse populations, offering a highly efficient strategy for opportunistic screening to address the current gap in clinical testing.

The Gist

The Big Picture: The "Hidden Heart Risk" Detective

Imagine your blood carries a specific type of "rust" called Lipoprotein(a), or Lp(a) for short. This isn't regular cholesterol; it's a special kind of "rust" that sticks to your arteries and causes heart attacks and strokes.

Here is the problem:

1. It's mostly genetic: Think of your Lp(a) level like your eye color or height. It's almost entirely determined by your DNA, not by what you eat or how much you exercise.
2. It's a silent killer: Most people have no idea they have high levels of this "rust."
3. The testing gap: Even though doctors know it's dangerous, very few people (less than 1% in the US) actually get their blood tested for it. It's like having a smoke detector that never gets a battery check.

The Solution:
Since we know this "rust" is written in our DNA, why not just read the DNA to find out who is at risk? That's what this study did. They created a new, super-smart "DNA decoder" that can predict your Lp(a) levels using a standard genetic test (the kind you might get from a 23andMe kit or a hospital blood draw) without needing expensive, complex lab work.

---

The Old Way vs. The New Way

The Old Way: The "Broken Map" (Polygenic Scores)

Previously, scientists tried to predict Lp(a) using Polygenic Scores (PGS).

• The Analogy: Imagine trying to navigate a city using a map drawn only for New York, but you are trying to drive in Tokyo.
• The Problem: These old maps worked okay for people of European ancestry, but they were completely useless for people of African, Asian, or Hispanic ancestry. The "traffic rules" (genetic variants) were different in different parts of the world, so the map led people in the wrong direction. In fact, for people of African ancestry, the old maps were practically blank.

The New Way: The "DNA Neighborhood Match" (Haplotype Model)

The authors developed a new method called a Haplotype-based Model.

• The Analogy: Instead of looking at individual street signs (single genes), this new method looks at entire neighborhoods (chunks of DNA).
• How it works:
  1. The scientists looked at a massive database of people (the "All of Us" program) and found specific "DNA neighborhoods" around the Lp(a) gene.
  2. They noticed that if you live in "Neighborhood A," you almost certainly have high Lp(a). If you live in "Neighborhood B," you likely have low levels.
  3. When they tested a new person, they didn't just look at one gene; they checked which "neighborhood" that person's DNA matched.
• The Result: Because they matched the whole neighborhood rather than just single street signs, the map worked perfectly for everyone, regardless of their ancestry. It didn't matter if you were from New York, Tokyo, or Lagos; the neighborhood match was accurate.

---

The Results: Why This Changes Everything

The study tested this new "DNA decoder" on thousands of people in three different hospital systems (Penn, Mass General, and Mount Sinai). Here is what they found:

1. It's incredibly accurate.
The model could predict Lp(a) levels with high precision across all racial and ethnic groups. It fixed the "broken map" problem of the past.

2. It's a "Super-Filter" for doctors.
Imagine a hospital has 1,000 patients.

• Current Reality: Doctors might test 10 people randomly and find 1 person with high Lp(a).
• With the New Model: The computer scans the DNA of all 1,000 patients. It flags the top 128 people who are predicted to have high Lp(a).
• The Magic: When the doctors actually test those 128 flagged people, 104 of them (81%) actually do have high Lp(a).
• The "Need to Test" Number: To find just one person with high Lp(a), doctors only need to test 1.2 people predicted by the model. That is a massive improvement over testing random people.

3. It's a "Time Machine" for existing data.
Millions of people have already had their DNA tested for other reasons (research, ancestry, other diseases). This model allows doctors to go back into those databases, run the new "neighborhood match" algorithm, and instantly identify thousands of people who are at risk but have never been told.

The Bottom Line

Think of this study as building a universal translator for a specific genetic risk.

• Before: We had a translator that only spoke English (European ancestry). If you spoke Spanish or Mandarin (African/Asian ancestry), we couldn't understand your risk.
• Now: We have a translator that speaks every language. It can look at a person's DNA and say, "You have a high risk of heart disease from this specific 'rust'."

Why does this matter?
New drugs are coming soon that can lower this "rust." But you can't take the medicine if you don't know you need it. This new tool allows doctors to find the people who need the medicine before they have a heart attack, simply by looking at genetic data they already have. It turns a "needle in a haystack" search into a targeted, efficient mission to save lives.

Technical Summary

1. Problem Statement

• Clinical Gap: Circulating lipoprotein(a) [Lp(a)] is a highly heritable (70–90%) and causal risk factor for atherosclerotic cardiovascular disease (ASCVD). Despite guideline recommendations for one-time testing, clinical measurement rates in the US remain below 1%.
• Limitations of Current Genetic Approaches:
  • Polygenic Scores (PGS): Traditional PGS for Lp(a) assume additive effects of variants across the genome. These models perform well in European populations but degrade significantly in non-European populations, particularly those of African ancestry (AFR), where Lp(a) levels are often highest.
  • Whole-Genome Sequencing (WGS): While WGS-based models can capture the complex structural variation at the LPA locus (specifically the kringle IV type 2 [KIV-2] repeat polymorphism), they are too expensive for routine clinical implementation or retrospective application to existing biobank cohorts.
• Objective: To develop and validate a cost-effective, haplotype-based prediction model using standard genome-wide genotype data that accurately predicts Lp(a) levels across diverse ancestral populations, enabling "opportunistic screening" in existing genotyped cohorts.

2. Methodology

• Study Design: A multi-cohort study involving model development in the All of Us Research Program (AoU) and external validation in three diverse biobanks: Penn Medicine Biobank (PMBB), Mass General Brigham Biobank (MGBB), and Mount Sinai BioMe.
• Model Development (Haplotype-Based Approach):
  • Reference Panel: Utilized WGS data from the UK Biobank to predict allele-specific Lp(a) concentrations for specific haplotypes.
  • Haplotype Definition: Identified SNP haplotypes flanking the LPA locus (chromosome 6) shared by >20 individuals in the AoU cohort.
  • Prediction Logic: Instead of summing variant effects (as in PGS), the model assigns the mean allele-specific Lp(a) concentration (derived from the WGS-based reference) to each haplotype carried by an individual. The total predicted Lp(a) is the sum of the two allele-specific predictions.
  • Uncertainty: Standard deviations of predicted concentrations were used to generate 95% confidence bounds.
• Data Processing:
  • Genotype data from validation cohorts were phased and imputed to the TOPMed reference panel.
  • Ancestry was inferred using the FRAPOSA method (1000 Genomes + HGDP reference).
  • Clinical Lp(a) measurements were converted to nmol/L (factor 2.15).
• Performance Metrics:
  • Continuous Prediction: Coefficient of determination ($r^2$).
  • Binary Classification (Elevated Lp(a) >125 nmol/L): Positive Predictive Value (PPV), Sensitivity, Specificity, Likelihood Ratios, and Number Needed to Test (NNT).
  • Comparison: Benchmarked against 13 existing Lp(a) PGS from the PGS Catalog.
• Statistical Analysis: Bayesian mixed models with random effects for ancestry and cohort were used to combine performance metrics. Bootstrap resampling (1,000 iterations) was used for confidence interval estimation.

3. Key Contributions

• Novel Methodology: Introduced a haplotype-matching framework that captures the combined effects of common variants, rare variants, and structural polymorphisms (KIV-2 repeats) without requiring WGS or explicit modeling of structural variants.
• Cross-Population Robustness: Demonstrated that this approach overcomes the "transferability gap" inherent in traditional PGS, maintaining high accuracy across African, Admixed American, European, and East Asian ancestries.
• Clinical Utility Framework: Established a strategy for "opportunistic screening" where existing genotype data (from biobanks or direct-to-consumer tests) is used to triage individuals for confirmatory clinical Lp(a) testing, significantly improving detection rates.

4. Results

• Cohort Demographics: Validation included 4,943 participants (PMBB: 1,856; MGBB: 1,401; BioMe: 1,686) with available genotype and clinical Lp(a) data. The mean age was 60, and 51% were female.
• Prediction Accuracy ($r^2$):
  • The haplotype model achieved an overall $r^2$ of 0.46 (95% CrI 0.32–0.60).
  • Performance was consistent across ancestries:
    • European (EUR): $r^2$ = 0.53
    • African (AFR): $r^2$ = 0.45
    • Admixed American (AMR): $r^2$ = 0.47
    • East Asian (EAS): $r^2$ = 0.40 (limited by small sample size, n=8)
  • Comparison to PGS: The haplotype model significantly outperformed all 13 existing PGS. For African ancestry, existing PGS had near-zero performance ($r^2$ < 0.01), whereas the haplotype model maintained $r^2 \approx 0.45$.
• Identification of Elevated Lp(a) (>125 nmol/L):
  • Positive Predictive Value (PPV): 0.81 (95% CrI 0.60–0.89) overall.
  • Number Needed to Test (NNT): 1.2 (95% CrI 1.1–1.7). This means that for every 1.2 individuals predicted to have elevated Lp(a) who undergo confirmatory testing, one true elevation is identified.
  • Likelihood Ratios: Positive Likelihood Ratio (LR+) was 8.9, approaching the diagnostic threshold of 10.
• Detection Rate Improvement:
  • In the full PMBB cohort (n=49,310), the haplotype model identified elevated Lp(a) at a rate of 128 per 1,000.
  • This represents a 14.4-fold improvement over the current rate of clinical assessment (8.86 per 1,000).
  • The improvement was most pronounced in African ancestry populations (184 per 1,000 identified).

5. Significance and Implications

• Bridging the Testing Gap: The model offers a scalable, low-cost solution to identify high-risk individuals in the tens of millions of people who have already undergone genotyping for other purposes (e.g., research biobanks, ancestry testing).
• Equity in Precision Medicine: By correcting the performance disparity seen in traditional PGS, this approach ensures that populations of African ancestry—who bear the highest burden of elevated Lp(a)—are not excluded from genetic risk stratification.
• Therapeutic Readiness: With multiple Lp(a)-lowering therapies (e.g., pelacarsen, olpasiran) in late-stage trials, this tool facilitates the identification of eligible candidates for clinical trials and future treatment.
• Clinical Workflow: The high PPV and low NNT suggest that genetic prediction can serve as an efficient triage tool, prioritizing confirmatory blood tests for those most likely to benefit, rather than relying on universal, low-yield biochemical screening.

Limitations Noted: The study did not assess clinical outcomes (ASCVD events) directly; performance in underrepresented Indigenous populations needs further validation; and the model relies solely on genetics, not accounting for minor environmental influences (e.g., kidney function). Confirmatory laboratory testing remains necessary for clinical decision-making.