A person's genotype predicts their height, cholesterol or heart-muscle mass through two quite different routes. One route uses population-wide effects of many small variants. The other is a shortcut: if the person's sibling or cousin is in the training data, the model partly predicts them by resemblance to that relative. Twin and family studies estimate heritability largely through the shortcut, while genome-wide association studies of unrelated people estimate it through the first route. The difference between the two is a large part of the "missing heritability" puzzle.
Prediction intervals for genetic risk inherit the same split. An uncertainty band calibrated on a cohort full of relatives will be too narrow for a newcomer with no relatives in the data. It will also be too wide for the relatives of probands, the people who are actually tested in family screening clinics. We propose prediction intervals with a distribution-free, finite-sample guarantee that holds separately for each level of relatedness between the test person and the training cohort. The intervals use every labeled individual for both fitting and calibration, and their width as a function of relatedness gives a model-free view of how much predictable variation comes from relatives rather than from population-level variant effects.
The tool is conformal prediction, a wrapper that turns any predictor into intervals guaranteed to contain the true outcome with probability at least $1-\alpha$ under only an exchangeability assumption. The standard version holds out a calibration set. That is harmless at biobank scale but expensive in the settings where relatedness matters most: family cascade-screening cohorts, carriers of rare pathogenic variants, and small non-European or deeply phenotyped cohorts.
Core Claim
For polygenic predictors that are linear in genotype, such as ridge regression on SNPs or GBLUP (a mixed model whose genetic covariance is the genomic relationship matrix between individuals), full conformal prediction can be computed exactly. The model can be written in kernel form over individuals rather than over millions of SNPs. In that form every person's residual is a linear function of the candidate test label, so the interval endpoints follow from a sort, with no retraining and no Hessian inversion over SNPs. This sidesteps the obstacle the file identifies for influence-function methods when $p \approx 10^5$–$10^7$.
On top of this exact computation we add two ingredients. The first is a studentized nonconformity score, which divides each residual by the model's own predicted error scale:
$$s_i(y) = \frac{|y_i - \hat\mu_i(y)|}{\sqrt{v_i + \hat\sigma_e^2}}.$$
Here $y$ is the candidate label for the test person, $\hat\mu_i(y)$ is the fitted value for person $i$ from the model trained on the augmented cohort (training set plus test person carrying label $y$), $v_i$ is GBLUP's prediction error variance for $i$, and $\hat\sigma_e^2$ is the estimated residual variance.
The second ingredient is Mondrian calibration, meaning calibration done separately within groups. Each person is assigned a stratum $k_i$, defined as the binned maximum kinship between $i$ and anyone else in the augmented cohort. The test person's score is ranked only against people in the same stratum. Because $k_i$ is a symmetric function of the augmented data, standard Mondrian full-conformal arguments give coverage of at least $1-\alpha$ conditional on stratum under exchangeability, with nothing held out.
The claim is twofold. Stratified coverage holds where both Gaussian prediction-error intervals and marginal conformal methods miss: they undercover unrelated newcomers and overcover relatives. And the curve of interval width against kinship is an interpretable, assumption-light measurement of the gap between family-based and SNP-based predictability.
Relation to Prior Work
Three papers supply the full-sample calibration machinery we build on. Stable Conformal Prediction Sets bounds how much residuals can move when one point is added, using stability constants of regularized convex models, and computes an outer envelope of the full conformal set from one fit. Approximating Full Conformal Prediction for Neural Network Regression with Gauss-Newton Influence linearizes a network so that scores become piecewise linear in the candidate label, and recovers interval endpoints from changepoints. Conformal prediction interval for dynamic time-series calibrates on out-of-bag residuals from a bootstrap ensemble.
All three were evaluated on low-dimensional tabular, image or time-series data, and all target marginal coverage. We reuse the piecewise-linear endpoint idea but make it exact by working in the kernel form of the genomic model. We also replace marginal calibration with calibration conditioned on genetic relatedness. That conditioning is the variable the file flags as breaking exchangeability in human cohorts and that none of these papers address.
On the genetics side, several papers motivate the work by showing that genomic prediction is accurate on average but untrustworthy for individuals. Fine mapping and accurate prediction of complex traits using Bayesian Variable Selection models applied to biobank-size data shows that multi-locus spike-and-slab models improve prediction correlation by 17–47% over single-marker scores, but their uncertainty statements are prior-dependent posterior intervals. Evaluation of polygenic scores for hypertrophic cardiomyopathy in the general population and across clinical settings shows that polygenic background changes penetrance among pathogenic sarcomeric-variant carriers from 2.3% to 17.2% across quintiles. Those carriers are usually found through family screening, so the people who most need individual-level uncertainty are exactly the related ones.
Another group of papers constrains the design by showing that hidden sample dependence silently inflates apparent accuracy. Overestimated Polygenic Prediction due to Overlapping Subjects in Genetic Datasets reports an AUC gain collapsing from 0.069 to 0.0017 once 432 overlapping subjects are removed. We read relatedness as a graded version of the same leakage: a relative in training is a partial overlap. On Genetic Correlation Estimation With Summary Statistics From Genome-Wide Association Studies shows that the opposite error, attenuation from noise aggregation, can understate genetic signal by an order of magnitude. Our method does not correct either bias in point estimates. Its aim is that the reported uncertainty stays honest despite them.
Nonlinear predictors like A Deep Learning-based Genome-wide Polygenic Risk Score for Common Diseases Identifies Individuals with Risk and the epistasis screen in Epistasis regulates genetic control of cardiac hypertrophy enter only as a secondary extension. In that extension a Hadamard-product (pairwise-epistatic) relationship kernel keeps full conformal exact while allowing non-additive signal.
Research Program
The first component is computational. In kernel ridge or GBLUP form, the augmented-cohort fit is $\hat\mu = H(y_{1:n}, y)$, where $H$ is the hat matrix of the augmented relationship matrix. Every residual is therefore $a_i + b_i y$. We will implement exact full conformal via one eigendecomposition of the training relationship matrix plus a rank-one update for each test person. The cost per test person is $O(n)$ after setup, independent of the number of SNPs. Variance components are estimated once on the training data. We will show, and check empirically, that refitting them in the augmented cohort changes coverage negligibly; otherwise we will handle them with a stability-style envelope following Ndiaye.
The second component is the Mondrian construction. Strata will come from augmented-cohort kinship bins (first-degree, second-degree, third-degree, unrelated). We will prove the stratum-conditional guarantee, and then state plainly the extra assumption needed to transfer it to deployment on a person whose relatedness pattern was not sampled the way the cohort was. That assumption is that, within a stratum, a newcomer's outcome given their genotype behaves like that of cohort members in the same stratum.
The third component is the missing-heritability readout. For each stratum, the calibrated interval width implies a distribution-free bound on predictable variance. Comparing the unrelated stratum with the first-degree stratum gives a predictive analogue of the gap between SNP heritability and pedigree heritability. We will report it with and without a simulated shared family environment, a non-genetic component that also inflates resemblance between relatives.
Baselines
The naive reference point is the Gaussian interval implied by GBLUP itself, $\hat\mu \pm z_{\alpha/2}\sqrt{v + \hat\sigma_e^2}$. Its prediction error variance already widens for people unrelated to the training set, which makes it a serious competitor rather than a straw man. Alongside it we report the interval built from in-sample training residuals, which the file notes undercovers because of overfitting. Both are fit on the identical cohort and relationship matrix.
The strongest Bayesian competitor is BayesC posterior predictive intervals from the BGLR package, as used in Fine mapping and accurate prediction of complex traits using Bayesian Variable Selection models applied to biobank-size data. That paper reports prediction-correlation gains of 17–47% over single-marker regression. No reported coverage result for these intervals was found on Lacuna. We run it with the same SNP set, the paper's block structure, and a chain length matched in wall-clock time to our method's setup cost.
The strongest conformal competitors are two. One is split conformal with whole families assigned to the calibration fold, at 20% and 50% calibration fractions. The other is unstratified exact full conformal with the same studentized score. Comparing against unstratified full conformal is the key ablation: it isolates the Mondrian stratification. A second ablation uses unstudentized absolute residuals with stratification, which isolates the score. For a neural comparator on left ventricular mass we run ACP-GN, which Approximating Full Conformal Prediction for Neural Network Regression with Gauss-Newton Influence reports keeps 95% coverage with tighter intervals than split conformal across 12 UCI benchmarks; no reported result found on Lacuna for it on genomic data.
All methods receive the same individuals, covariates (age, sex, ten ancestry principal components) and SNP set, with hyperparameters chosen by the same inner cross-validation budget. We count our method as beating a baseline only if it meets nominal coverage within Monte Carlo error in every kinship stratum and the competitor does not. Among methods that cover in every stratum, it must also have smaller mean width in the unrelated stratum.
Why not just…?
Why not just hold out whole families for split conformal? This approach is valid and simple. At UK Biobank scale it is close to free, and there we expect it to match us on marginal coverage. It falls short in two places. First, it guarantees only marginal coverage, so a mix of related and unrelated test people can be correctly covered on average while unrelated people are undercovered. Second, in cohorts of a few hundred to a few thousand, such as ADNI or carrier families, the calibration fold within any single kinship stratum becomes tiny. The file's motivating concern about wide, unstable split intervals applies directly. The subsampling experiment below measures where the crossover sample size lies; if it lies below realistic clinical cohort sizes, we will say the method is unnecessary there.
Why not just remove relatives during quality control, as GWAS pipelines do? That discards the people a family-screening clinic actually tests. It also leaves the deployment question unanswered whenever the test person's relative is in the reference data. Our unrelated-stratum result is effectively this baseline without the data loss, so the comparison is direct.
Why not just trust GBLUP's own prediction error variance, which already depends on kinship? This is the real threat. If the additive Gaussian model holds, its intervals are well calibrated in every stratum and our method only adds robustness. The semi-synthetic experiments are designed to settle this. They inject skewed noise, pairwise epistasis, and a shared family environment that the model does not include. Stratum-wise undercoverage of the Gaussian intervals under these departures, together with intact coverage of ours, would justify the method. Matching coverage would mean our contribution reduces to a cheap robustness check, and we would report it that way.
Why not just run ACP-GN on a deep polygenic score? For nonlinear models it may be the right tool. It still calibrates marginally, though, and its Gauss-Newton curvature over genome-scale inputs is the bottleneck the file identifies. Our exact kernel computation needs no curvature. The left ventricular mass comparison shows whether the nonlinear predictor buys enough accuracy to offset approximate, unstratified calibration.
Proposed Experiments
The semi-synthetic study uses real UK Biobank genotypes, including naturally occurring relative pairs. Phenotypes are simulated with known SNP heritability, an optional pairwise epistatic component and an optional shared family environment. Training cohorts of 500, 2,000 and 10,000 people are subsampled with controlled fractions of related individuals. We measure coverage in each kinship stratum, interval width, and the variability of coverage across 200 replicates. The decisive check is coverage for unrelated test people drawn from outside the cohort when the training set is 40% related. We expect the Gaussian and unstratified conformal intervals to undercover there, and stratified full conformal not to.
The real-trait study uses the six blood biomarkers from Fine mapping and accurate prediction of complex traits using Bayesian Variable Selection models applied to biobank-size data. It also uses left ventricular mass in the roughly 29,661 UK Biobank participants with cardiac MRI analyzed by Epistasis regulates genetic control of cardiac hypertrophy, with the epistatic kernel as the nonlinear arm. Coverage there is checked on held-out families never used for anything. We also report the stratum width curve. The missing-heritability claim is supported if the first-degree versus unrelated width gap shrinks once a family-environment covariate (such as shared household, where recorded) is added. The claim is refuted if the gap is unchanged by such adjustment and also matches what additive genetics alone predicts.
The small-cohort study applies the method to ADNI amygdala gray matter density, the endophenotype used in A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study. Here a calibration hold-out is costly, and the comparison with family-held-out split conformal at equal coverage is the test of data efficiency.
As a stretch goal, contingent on access, we will apply the method to hypertrophic cardiomyopathy carrier families. In that setting the method's intended use, an honest interval for a relative of a proband, can be checked against observed penetrance.
References
- Stable Conformal Prediction Sets
- Approximating Full Conformal Prediction for Neural Network Regression with Gauss-Newton Influence
- Conformal prediction interval for dynamic time-series
- Fine mapping and accurate prediction of complex traits using Bayesian Variable Selection models applied to biobank-size data
- Evaluation of polygenic scores for hypertrophic cardiomyopathy in the general population and across clinical settings
- Overestimated Polygenic Prediction due to Overlapping Subjects in Genetic Datasets
- On Genetic Correlation Estimation With Summary Statistics From Genome-Wide Association Studies
- A Deep Learning-based Genome-wide Polygenic Risk Score for Common Diseases Identifies Individuals with Risk
- Epistasis regulates genetic control of cardiac hypertrophy
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study