Methods
How we estimate risk reduction with family history
The figures use a statistical model to estimate lifetime disease risk before and after selecting the lowest-risk embryo from five sibling embryos. They are projections of disease risk, rather than measured health outcomes in children born after embryo selection.
The calculation
We use the single-condition selection model from Lencz et al. (2021), conditioning on parental disease status. Inputs are Reticular’s liability-scale R², population lifetime prevalence (K), and published liability-scale heritability (h²). Heritability affects the family-history calculation; the population-only calculation does not require total h².
- Starting risk
- Estimated risk for a randomly selected embryo from the same parental disease-status group.
- Risk after selection
- Estimated risk for the embryo with the lowest score for that condition among five sibling embryos.
- Relative risk reduction
- 100 × (1 − risk after selection / starting risk), expressed as a percentage.
- Absolute risk reduction
- Starting risk minus risk after selection, expressed in percentage points.
Assumptions used in the figures
- 5 sibling embryos capable of live birth; no IVF attrition or transfer failure adjustment.
- Optimize one condition at a time; reductions cannot be claimed simultaneously for a multi-condition index.
- Independent parental genotypes/random mating; additive normal genetic liability.
- No family-shared environmental covariance or age-specific parental ascertainment model.
- One affected parent means the other lifetime unaffected for non-sex-limited diseases. Alternative other-unknown estimates are retained in data.
- Sex-limited one-parent scenarios use affected father for prostate/testicular and affected mother for breast cancer; the other genetic parent is unobserved in the corresponding sex-specific liability model.
- Breast/prostate/testicular two-parent scenarios are not estimated; no fabricated numeric points.
- Current external R2 assumed to reflect direct within-family performance; no attenuation or embryo-genotyping correction.
- All h2 point inputs are candidates; numeric integration accuracy does not resolve input or model uncertainty.
- Four illustrative input pairings are marked: POAG transport, gout male h2/pooled inputs, BCC proxy h2, and testicular assumed K=0.004.
- R2 interval fields transform supplied R2 confidence endpoints holding K and h2 fixed; they are not total-uncertainty intervals and are not plotted as such.
Inputs and validation still under review
The figures are preliminary. Type 2 diabetes uses population score performance without adjusting for the measured within-family effect ratio of about 0.86. Alzheimer’s disease and testicular cancer have unstable within-family estimates because of limited case counts.
Asterisks in the family-history figure identify these provisional input pairings:
- Glaucoma: POAG/Taiwan h2 with current target-population inputs.
- Gout: Male Taiwan h2 with pooled R2/K; sex/population pairing assumed.
- Basal cell carcinoma: Nonmelanoma skin-cancer h2 proxy.
- Testicular cancer: K=0.004 assumed; K used to derive reported R2 remains unreconciled.
Changing heritability or prevalence inputs can change starting risk and both risk-reduction measures. Numerical accuracy does not account for uncertainty in those inputs. An unaffected parent today is also not necessarily lifetime unaffected, which is the status assumed in the standard one-parent scenario.
Cross-ancestry liability R²
The cross-ancestry figure starts from the external-permitted score-only AUCs in Table 3. We convert the point estimates and both confidence limits to liability R² by numerically inverting a normal liability-threshold model, following the framework in Wray et al. (2010). We use each condition’s reference prevalence from Table C2 for every ancestry. These are model-derived estimates on a common scale, not directly measured ancestry-specific liability R² or the covariate-adjusted estimates in the comparison figure. Different prevalence assumptions change the converted values.
Results are shown only when the original 95% AUC interval lies entirely above 0.5. This filter highlights results with evidence of discrimination; it does not mean omitted results lack predictive value. The retained results are a selected subset of all evaluations. Testicular cancer is omitted because its reference prevalence has not been reconciled.
Stars flag fewer than 100 cases, including entries reported as ≤20, in the supplied Table B.4. That table reports internal-discovery-only sample sizes, while the AUCs use external-permitted scores. We use it as a caution about limited samples, not as verified sample counts for those scores. A star is not a significance indicator.