Polygenic embryo screening

Polygenic embryo screening, formally preimplantation genetic testing for polygenic conditions (PGT-P), ranks the embryos produced in an IVF cycle by polygenic scores calculated from hundreds of thousands of common genetic variants. Clinics and direct-to-consumer companies use it to report an embryo's relative predicted risk for conditions such as type 2 diabetes, coronary artery disease and schizophrenia, and in some cases for non-disease traits including height and cognitive test performance. The laboratory procedure is a modest extension of routine embryo genotyping. The dispute is about what the resulting rankings are worth.

§How it works

PGT-P attaches to an existing clinical workflow rather than replacing it. Five or six days after fertilisation, an embryologist biopsies five to ten cells from the trophectoderm, the outer layer of the blastocyst that will become placenta rather than fetus. The DNA in that sample is measured in picograms, so it is first amplified across the whole genome, then read on a SNP array or by low-pass sequencing.

The resulting genotype is sparse and noisy. To make a polygenic score usable, the laboratory imputes the missing variants: it genotypes both parents, reconstructs which parental haplotypes each embryo inherited, and fills in the rest by reference to a haplotype panel. An embryo's score is therefore not measured but reconstructed, and imputation error propagates into the ranking. Companies then sum the embryo's variant dosages weighted by effect sizes taken from a published genome-wide association study, and report each embryo's position relative to its siblings or to a population distribution.

Some providers have moved to sequencing the amplified biopsy at higher depth in order to call rare variants alongside common ones, which brings PGT-P closer to a whole-genome report on a sample of a few cells. That does not change the underlying statistical problem, which lies in the effect-size weights rather than the genotyping.

Terminology

PGT-A tests for aneuploidy, PGT-M for a known single-gene disorder, and PGT-SR for structural rearrangements. All three ask a categorical question about one locus or chromosome. PGT-P asks a probabilistic question about the whole genome, which is why it sits in a different evidentiary category.

§Development history

1. 2007–2018
   GWAS reach useful scale
   Consortium studies of hundreds of thousands of participants produce polygenic scores with non-trivial predictive power for height, education and several common diseases.
2. 2017–2019
   First commercial service
   Genomic Prediction, co-founded by physicist Stephen Hsu, begins offering an embryo health score to IVF clinics in the United States.
3. 2019
   Quantitative critique
   Karavani and colleagues model the expected gain from selecting the top-scoring embryo and conclude it is small for both height and cognitive scores.
4. 2020–2021
   First reported birth and formal objections
   A child selected using a polygenic score is publicly reported, and a group of behaviour geneticists and statisticians set out the problems with the practice in the New England Journal of Medicine.
5. 2022
   European societies object
   A joint statement from European genetics and reproductive medicine societies calls the practice unproven and unethical.
6. 2024–2025
   Trait screening goes public
   Journalistic investigations document companies offering embryo ranking on cognitive traits, and at least one firm markets such ranking openly, drawing broad criticism from geneticists.

§What the gain actually is

The size of any benefit is set by three quantities: how well a polygenic score predicts the trait between siblings, how much genetic variation exists among the embryos of one couple, and how many embryos there are to choose from. All three are less favourable than intuition suggests.

Siblings differ only by the recombination and segregation of their parents' chromosomes, so the spread of true genetic values within a sibship is narrower than in the population at large. Selecting the highest-scoring of ten embryos captures only the upper tail of that narrow distribution, and only in proportion to the score's accuracy. Modelling this explicitly, Karavani and colleagues estimated an expected gain of roughly two and a half centimetres of height, or a comparable number of IQ points, when choosing the best of ten embryos.¹ With a more realistic number of viable embryos the expected gain falls further, and the variance around it is wide enough that a substantial fraction of couples would see no gain at all.

For disease, the arithmetic differs because relative risk reduction can be large where absolute risk is small. Lencz and colleagues showed that the apparent utility depends heavily on the selection strategy: excluding embryos in the top few per cent of risk for a highly heritable disorder gives a different picture from ranking all embryos and taking the best.² A large relative reduction in the risk of a condition affecting one person in a hundred still leaves the great majority of the couple's expected outcome untouched.

The comparison that matters clinically is with what selection already does well. Testing for a known single-gene disorder converts a one-in-four risk into something close to certainty of avoidance, because the underlying genetics is categorical. Polygenic testing offers a probabilistic nudge across many conditions at once, and the two are often presented to patients on the same report. Nothing in the laboratory output distinguishes the reliable number from the speculative one, which is the practical complaint clinical geneticists make most often.

Relative versus absolute

A report telling a couple that one embryo carries, say, sixty per cent less risk of schizophrenia than another describes a shift in a probability that was already low, computed from a model that has never been validated against the outcome of a selected birth. Proponents argue that a real if small expected benefit is still a benefit. Critics reply that the reported percentages imply a precision the underlying scores do not have.

§Why prediction weakens where it matters

§Within-family versus population prediction

Polygenic scores are trained on unrelated individuals, where they absorb more than direct genetic effects. Population-scale associations also capture assortative mating, residual population structure, and indirect genetic effects from parents' genotypes acting through the environment. None of these differentiate one embryo from its sibling. Within-sibship analyses consistently find that predictive power drops when the comparison is made between siblings, and the drop is largest for socially patterned traits such as educational attainment.³ A score advertised with its population R² therefore overstates what it can do inside a family, which is the only setting in which embryo selection operates.

§Ancestry portability

Almost all large discovery cohorts are of European ancestry. Because linkage disequilibrium patterns and allele frequencies differ across populations, scores lose accuracy when applied to people of other ancestries, often by a factor of two or more.⁴ Embryos of mixed ancestry are worse still, since no reference panel matches them. A technology whose accuracy tracks the ancestry composition of biobanks distributes its benefit unevenly by construction, a point that connects PGT-P to broader arguments about Access and inequality and Genetic discrimination.

§Pleiotropy and joint optimisation

Variants do not act on one trait. Genetic correlations between traits mean that pushing an embryo's score up for one outcome moves other scores, sometimes in unwanted directions. Companies address this by combining several scores into a single index, which requires the couple, or the company, to assign relative weights to incommensurable outcomes. That weighting is a value judgement dressed as a computation.⁵

§Embryo count

Selection intensity is the binding constraint. A typical stimulated cycle yields a modest number of blastocysts, and fewer still are chromosomally normal; for older patients, whose situation is discussed in Reproductive longevity, the usable number is often one or two. In that case the ranking has nothing to rank. This is why In vitro gametogenesis, which would in principle supply large numbers of embryos, is the development that would most change the calculus.

§The professional response

No major professional body endorses PGT-P for clinical use. A joint statement from European human genetics and reproductive medicine societies described the practice as unproven and unethical, citing the absence of clinical validity data, the misleading precision of the reports, and the burden placed on prospective parents asked to interpret them.⁶ The ethics committee of the American Society for Reproductive Medicine reached a similarly cautious conclusion, holding that the technology is not ready for routine offering and that clinics providing it must disclose its unvalidated status.⁷ American medical genetics bodies have issued comparable points-to-consider documents.

The objections are not only statistical. Reporting a score for a trait such as cognitive ability treats the trait as an optimisation target, which critics connect to the expressivist objection developed in Disability rights and enhancement and to the older arguments catalogued in Bioconservatism. Supporters, several of them associated with Julian Savulescu's work on Procreative beneficence, argue that a parent who would accept a small expected health benefit from prenatal vitamins has no principled reason to refuse a small expected benefit from selection.

There is also a counselling problem that neither side disputes. A PGT-P report presents multiple percentile figures across several conditions, derived from models most clinicians cannot interrogate, to patients already making a decision under time pressure and emotional strain. Genetic counsellors have argued that no realistic consultation can convey the difference between a validated Mendelian result and a polygenic percentile, and that the format of the report itself does persuasive work the evidence does not support. Because PGT-P is ordered as an add-on rather than as a distinct procedure, the discussion is often shorter than the one preceding any proposal to alter an embryo's genome, despite resting on weaker evidence.

§Regulation and access

In the United States, PGT-P reaches patients as a laboratory-developed test, a category historically subject to limited premarket review; no regulator has assessed the clinical validity of any embryo polygenic score. The United Kingdom takes the opposite approach: the Human Fertilisation and Embryology Authority licenses preimplantation testing only for specified serious conditions, and polygenic screening is not among them. That statutory list is the same mechanism by which Britain authorised Mitochondrial replacement therapy under strict conditions, and it illustrates the difference between approving a technique and approving an indication. Several European jurisdictions restrict embryo testing more tightly still, and a few permit it only for conditions on a statutory list. The practical consequence is a market concentrated in a small number of American clinics serving an international clientele, at a price that sits on top of an IVF cycle most health systems do not fund. The resulting picture is the one anticipated in debates over the Governance of human genome editing: a technology governed less by law than by where the clinic is.

§Outlook

Two developments would change the argument. The first is a within-family validated score, trained on sibling comparisons at sufficient scale that its coefficients reflect direct genetic effects rather than the social correlates of ancestry and family background. The second is a large increase in the number of embryos available per couple, which only laboratory-derived gametes plausibly deliver. Together they would move PGT-P from a service selling a fractional expected shift to one capable of a change large enough to argue about on its merits, which is roughly the situation that popular discussion already assumes exists and that the arguments in Bioethics of enhancement were developed to address.

Neither development supplies the missing evidence. Establishing that a selected child actually enjoys the predicted advantage requires following selected and unselected children for decades, against a counterfactual sibling who was never born. No such study is being run, and it is not obvious how one could be. The field is therefore likely to keep arguing about a technology whose central claim is, by its own design, close to untestable.

§See also

• Embryo selection
• Designer babies
• Genetic enhancement of cognition
• In vitro gametogenesis
• Procreative beneficence
• Human germline editing
• Genetic discrimination
• Disability rights and enhancement

§References