How Embryo Selection Technology exposes the Transferability Paradox

Why we must acknowledge the reality of ancestry differences

Mogging the competitors

A new embryo selection tool from Herasight (herasight.com) promises to help parents choose embryos with higher predicted IQ and lower disease risk. This team has achieved an amazing feat, outperforming academic studies and private competitors (Thomson, Mars, Genomic Prediction, Orchid) in genomic prediction, as shown by the authors in their blog post and technical paper. The charts shown below compare the predictive power of polygenic scores computed by Herasight vs their competitors across a range of diseases:

The results of their within-family validation are astounding, achieving the same predictive power as between-families. This is an outstanding result because previous research has found much lower predictive power within families than between families.

As shown in the figure below, the ratio of effect sizes for Herasight’s PGSs observed in population and within-family regression analyses in identical samples, where the dotted line at 1 indicates identical strength of association between and within families.

Their scores had almost identical predictive power when tested with the within-family design for 16 out of 17 diseases.

The European Bias in Genetic Research

The foundation of polygenic prediction rests on genome-wide association studies (GWAS), which identify genetic variants associated with traits like intelligence and disease susceptibility. However, the major ethical and scientific challenge surrounding clinical implementation is that they are much more accurate in European ancestry individuals than others.

Representation of non-European ancestry participants in genome-wide association studies (GWAS) increased, from 4% in 2009 to around 15% in 2025, but this improvement still leaves the vast majority of genetic research centered on European populations. This creates a feedback loop where prediction tools work best for the populations that contributed most to their development.

Why Ancestry Matters in Genetic Prediction

Most GWAS have been conducted in populations of European ancestry, which limits the use of GWAS-derived PGS in non-European ancestry populations. PGS developed using European-ancestry samples tend to perform poorly in non-European ancestry test sets. The Herasight team clearly show this in their white paper, across a range of ancestries. The results are benchmarked against a group of white British descent.

The technical explanation involves linkage disequilibrium patterns, allele frequencies, and population-specific genetic architectures. But the practical result is straightforward: genetic prediction tools are substantially less effective for non-European populations, creating a two-tiered system where some families have access to more powerful genetic optimization than others.

The Race Transferability Paradox

This huge practical and ethical problem highlights a theoretical paradox that unfortunately has been ignored by mainstream academics.

It can be summarized like this: “Human races do not biologically exist, yet the differences between them are too large to use the same genetic variants and polygenic scores to accurately predict phenotypes (physical and behavioral traits) across different racial groups”.

Real-World Impact

Herasight’s platform features an interactive tool that models expected offspring phenotypes based on user-specified embryo counts and parental characteristics. The functionality are:

1. Embryo Quantity Effect
   Increasing the number of embryos expands the selectable range of polygenic scores, enhancing the ability to choose optimal genetic profiles.

2. Selection Strategy
   Users target:

   • Highest PGS for favorable traits (e.g., cognitive ability)

   • Lowest PGS for disease risks (e.g., prostate cancer)

3. Practical Implementation
   For conditions like prostate cancer, the tool identifies embryos with the lowest disease-specific polygenic score to minimize inherited risk.

The risk is reduced from 25.1% to 3%, a 22.1% reduction in chance of getting the disease. However, in relative terms this is reduced by a factor of 8 or 88% (Old Risk / New Risk = 25.1% / 3% ≈ 8.37), or Relative Risk (RR) = New Risk / Old Risk = 3% / 25.1% ≈ 0.1195; 1 - RR = 1 - 0.1195 ≈ 0.881 (or 88.1%).

The absolute reduction is what we should focus on, because the relative risk reduction doesn't tell you how much your actual risk changed. A 50% RRR sounds great, but if your starting risk was only 0.2%, your ARR is only 0.1% – a much smaller real-world benefit. RRR magnifies small benefits when baseline risk is low.

Practical implications of prediction accuracy disparities

Testing HeraSight's embryo selection tool reveals the practical implications of the stark disparities in predictive accuracy across racial groups. When selecting from 20 embryos, parents of European ancestry can expect:

• IQ predictions ranging from 91.5 to 108.5 - a spread of 17 points

Compare this to parents of African ancestry, who see:

• IQ predictions ranging from 95.1 to 104.9 - a spread of only 9.8 points

  

Hence, European parents whose average IQ is 100 can expect to produce a child with an IQ of 108.5 by selecting the best embryo, but 100 IQ African parents can expect a 105 IQ child.

The real-world impact of this technology depends critically on baseline disease rates, which vary between populations. Consider two illustrative examples:

Schizophrenia: Despite a smaller relative risk reduction for African families (47.4% vs 70.5% for Europeans), the absolute benefit is actually larger. The technology reduces schizophrenia risk from 1.9% to 1% in African populations (0.9% absolute reduction) compared to 0.78% to 0.23% in European populations (0.55% absolute reduction). Higher baseline rates in the African population translate to greater absolute benefit.

Melanoma: The pattern reverses when Europeans have higher baseline rates. European families see risk reduction from 2.58% to 0.97% (1.61% absolute reduction), while African families experience a much smaller change from 0.08% to 0.05% (0.03% absolute reduction).

Equal Baseline Scenarios: When baseline risks are similar across populations, European families typically experience larger reductions due to the training bias in the underlying models. The authors demonstrate this pattern using type 2 diabetes as a case study in their analysis.

Why we should acknowledge the reality of race

Addressing this problem requires honest recognition that genetic ancestry - what we commonly call race - produces measurable differences in the effectiveness of modern genetic technologies. The biases and inaccuracies of polygenic risk scores when predicting disease risk in individuals from populations other than those used in their derivation represent a fundamental challenge that cannot be solved by pretending genetic ancestry doesn't exist.

Solutions must include:

Massive investment in diverse genetic research - We need GWAS studies that adequately represent global genetic diversity, not just European populations.

Population-specific prediction models - Rather than one-size-fits-all algorithms, we need tools optimized for different ancestral backgrounds.

Transparent reporting of accuracy differences - Companies offering genetic prediction services should clearly communicate how their accuracy varies across populations.

Equitable access policies - If genetic advantages are distributed unequally by ancestry, society must grapple with how to prevent this from creating permanent caste systems.

Beyond Good Intentions

The developers of embryo selection technology undoubtedly have good intentions. But good intentions don't eliminate the mathematical reality that their tools work better for some racial groups than others.

The genetic revolution promises remarkable benefits for human health and capability. But realizing those benefits equitably requires acknowledging uncomfortable truths about how genetic research has developed and taking concrete steps to correct historical imbalances. The alternative is a future where genetic advantages correlate with racial ancestry, not by biological necessity, but by the accident of which populations happened to fund and participate in early genetic research.

The technology exists. The disparities are measurable. The question now is whether we'll address them before they become permanently entrenched in the next generation's genetic makeup.