Genome-wide association studies (GWAS) scan hundreds of thousands of DNA variants one at a time. They explain only part of the trait variation that twin studies attribute to genes, and this shortfall is called missing heritability. The usual yardstick for how much a genome predicts is the variance explained ($R^2$) by a polygenic risk score (PRS), which is a weighted sum of a person's variants. $R^2$ only credits genes that shift a person's average trait value. Genes can also widen or narrow a person's range of plausible values, skew it, or send people into one of two separate states, and $R^2$ gives no credit for any of that. We propose a different yardstick. For each person, build a conformal prediction set: a range of trait values that is guaranteed, with no distributional assumptions, to contain the true value for a chosen fraction of people, say 90%. Then ask how much smaller these sets get once the genotype is known. The sets may be disjoint, meaning a union of separate intervals. This matters because the most informative thing a genome can say is sometimes "this person is near 5 or near 20, but not in between."
The measure has two parts. The first is the fraction of the no-genotype set size that genotype information removes at fixed guaranteed coverage. It gives a finite-sample-valid, individual-level account of genetic predictive information that has no Gaussian assumption built in. The second is the set of individuals whose sets split into several pieces. We argue this group has a specific genetic meaning. If a person's set splits even after conditioning on their genotype, the model has learned that a discrete switch exists but cannot tell which side of it the person is on. That usually points to a large-effect variant the genotyping array does not tag well, or to an interaction the features miss. A split set is therefore a pointer to where the remaining heritability sits, not just a tighter set.
Core Claim
Let $C_0(X)$ be a conformal set built from non-genetic covariates only (age, sex, ancestry principal components). Let $C_g(X)$ be one built with genotype-derived features added. Both are calibrated to the same coverage $1-\alpha$ on the same people. We define the conformal predictive heritability as
$$G = 1 - \frac{\mathbb{E}\,[\operatorname{Leb}(C_g(X))]}{\mathbb{E}\,[\operatorname{Leb}(C_0(X))]},$$
where $\operatorname{Leb}$ is total length, summed over pieces for a disjoint set, and the expectation is over held-out individuals.
The measure has a built-in sanity check. Suppose the trait is Gaussian with constant spread and a purely additive genetic effect. Then interval width scales with the residual standard deviation, so $1-G = \sqrt{1-R^2}$. Any gap above this curve is predictive information that $R^2$ cannot see.
We claim three things:
- On traits with known discrete genetic switches, $G$ computed from flexible, possibly disjoint sets will exceed the Gaussian-implied value by a clear margin.
- The part of that excess due to allowing more than one interval is measurable. It will be near zero for classic polygenic traits such as height and non-trivial for traits dominated by a few large-effect variants.
- People whose sets split will be enriched for carriers of large-effect variants the model was not given. Supplying those variants should collapse the split.
Relation to Prior Work
The conformal methods supply the tools we reuse. Volume Optimality in Conformal Prediction with Structured Prediction Sets proves that no distribution-free method can reach minimum set size in general. It then restricts sets to unions of $k$ intervals and finds the smallest such union with an exact dynamic program over an estimated conditional cumulative distribution. This is our main set constructor because it lets us switch disjointness on and off through $k$. Flexible distribution-free conditional predictive bands using density estimators (CD-split) thresholds estimated conditional densities to produce disconnected intervals. It also groups people by the similarity of their predicted densities, which gives us a second constructor and a route to better coverage within subgroups. JAPAN and CONTRA do the same with normalizing flows, which are invertible neural networks that turn a complex distribution into a Gaussian. We use a one-dimensional conditional flow as the density model and use JAPAN's set-size estimator for any multi-trait extension.
All four of these papers are evaluated on synthetic shapes or engineering data with at most about 20 input features. None addresses the million-variant regime of genomics. We do not try to feed raw variants into a flow. Instead, a separate discovery split compresses the genome into about 10 to 30 summary features, which puts us back in the low-dimensional regime where these methods are known to work.
The genetics papers motivate the question and supply those summary features. Discovering non-additive heritability using additive GWAS summary statistics (i-LDSC) finds non-additive variance from local interactions in 24 of 25 UK Biobank traits. It is a population-level variance split and says nothing about individual prediction. We use its per-trait signal as an external variable that our $G$ excess should track. Epistasis regulates genetic control of cardiac hypertrophy uses iterative random forests to find gene-gene interactions affecting left ventricular mass in 29,661 cardiac MRI participants. It offers no individual uncertainty, and its interaction terms become our non-additive features. Enhancing polygenic risk prediction by modeling quantile-specific genetic effects builds polygenic scores at separate quantiles and shows they help for skewed traits such as triglycerides. It gives no coverage guarantee, its estimated quantiles can cross, and it can only output one connected interval. Its quantile scores become input features to our conditional distribution model, where the dynamic program removes the crossing problem.
One strand constrains the design. The familial hypercholesterolemia study in Analysis of a deeply-phenotyped familial hypercholesterolemia cohort from Mexico is a reminder that genetic effects vary across ancestries. Coverage that holds only on average across a biobank could hide under-coverage in minority groups, so we calibrate separately within ancestry groups.
Research Program
The pipeline uses four disjoint splits of unrelated individuals. We remove relatives because relatedness breaks exchangeability, the condition that calibration and test people are interchangeable, which the coverage guarantee needs.
On the discovery split we compute the genotype summary features:
- an additive PRS;
- quantile-specific scores at several quantiles, following the quantile-PRS approach;
- pathway-level scores in the spirit of A Module-Level Polygenic Risk Score-Based NetWAS Framework;
- a small number of interaction features from iterative random forests;
- ancestry principal components.
On the training split we fit a conditional distribution model for the trait given these features. We use two options. One is a gradient-boosted distributional regression over a grid of thresholds, which gives a conditional cumulative distribution. The other is a small conditional normalizing flow.
On the calibration split we run the $k$-interval dynamic program at $k=1,2,3$ and calibrate the sets. On the test split we measure coverage, set size, $G$, and the fraction of people whose sets split.
The dynamic program's cost grows cubically with calibration size, according to its authors. We therefore cap calibration sets at a few tens of thousands of people, or use coarser grids, and report how sensitive the results are to that choice.
We also run a controlled simulation using real genotype data with simulated traits. In it we plant three kinds of effect:
- an additive effect;
- a variance effect, where a variant changes the spread of the trait;
- an epistatic switch, where two variants jointly move carriers into a separate mode, and one of the two is removed from the feature set.
The simulation tests whether $G$ separates these cases as predicted. It also tests whether split sets concentrate among carriers of the hidden variant.
Baselines
The naive reference is the no-genotype set $C_0$, built by the same dynamic program on covariates only. It defines the denominator of $G$. A method adds nothing if it does not shrink sets relative to $C_0$ at matched coverage.
The standard genetic baseline is an additive PRS with a split-conformal interval of constant width around the prediction. It is the direct analogue of reporting PRS $R^2$, and our sanity curve $1-G=\sqrt{1-R^2}$ says what it should achieve. Beating it means smaller average set size at the same held-out coverage, within the 90% ± 1% band.
The strongest existing competitor is conformalized quantile regression fed the same full feature set as our model. For this baseline the features are the additive PRS, quantile scores, and interaction features. This is the most serious rival because it already captures uneven spread and skew, and it is exactly what a careful statistical geneticist would build from the quantile-PRS paper. Every method gets identical splits, identical features, and the same hyperparameter search budget, tuned on the training split only.
No UK Biobank or KARE conformal set-size result exists to compare against. The file reports no conformal result on Lacuna for any genetic trait. The only reported head-to-head number is on the synthetic benchmark of Romano et al. There, Volume Optimality in Conformal Prediction with Structured Prediction Sets reports an average set size of 0.45 for its interval-union method against 1.29 for the single-interval distributional baseline (DCP-QR*), at 70% coverage. We will reproduce this comparison first, to check our implementation before moving to genetic data.
The key ablation sets $k=1$ in our own pipeline, keeping the model, features, and calibration unchanged. Whatever set-size reduction survives from $k=1$ to $k\ge 2$ is the contribution of disjointness alone. A second ablation removes the non-additive features, to separate the gain from flexible distribution shape from the gain from interactions.
Why not just…?
Why not just report $R^2$ of a better polygenic score? $R^2$ is a statement about means. A variant that changes spread, or splits carriers into two groups with the same average, adds nothing to it. The quantile-PRS paper's gains on triglycerides already show that mean-based accuracy misses real genetic signal. Still, $R^2$ is cheap and standard, so we report it for every model. The height sanity check tests whether $G$ and $R^2$ agree where theory says they should. If $G$ never departs from the Gaussian-implied curve on any trait, the new measure adds nothing, and we will say so.
Why not just use conformalized quantile regression with the same features? This is the real threat. For most polygenic traits the trait distribution given genotype is probably a single broad peak, because genotype explains a modest share of variance. In that case a single adaptive interval will match any disjoint set. We expect this to happen for height and most anthropometric traits. The $k=1$ versus $k\ge 2$ ablation, run on the same pipeline, settles the question trait by trait. The disjointness claim survives only if the gain is concentrated on traits with large-effect or untagged discrete variants and appears in the simulation with a planted switch. If that pattern does not appear, the proposal reduces to its first contribution, conformal predictive heritability with flexible single intervals, and we would report it that way.
Why not just compare log-likelihoods of conditional density models? Held-out log-likelihood gain measures genetic information and is closely related to what we want. It carries no finite-sample guarantee, though, and it gives individuals no range they can interpret. We report it alongside $G$. Where the two disagree, the disagreement is informative, since it usually means the density model is overconfident in its tails.
Why not just add the top variants as covariates in a standard model? When the causal variant is on the array and known, this is correct, and our method should collapse to a single interval. That behaviour is itself one of our tests. Split sets are useful precisely when the switch is unknown or imperfectly tagged. That is the setting that defines missing heritability.
Proposed Experiments
We use unrelated UK Biobank participants on three groups of traits:
- the 25 continuous traits analysed by i-LDSC;
- serum biomarkers with known strong single-gene effects, including total bilirubin, urate, LDL cholesterol, and lipoprotein(a);
- left ventricular mass in the cardiac MRI subset used by the epistasis study.
We replicate on KARE triglycerides and post-challenge glucose, where quantile-specific effects were reported.
Calibration is done both over the whole sample and separately within each ancestry group. We report coverage within each group, within each quartile of the PRS, and among rare-variant carriers.
Four checks decide the outcome:
- Height sanity check. For height, $G$ from the additive model should sit on the curve $1-G=\sqrt{1-R^2}$.
- Disjointness gain on switch-like traits. For bilirubin and lipoprotein(a), the $k\ge 2$ ablation should shrink sets by a margin well beyond the spread across repeated random splits. For height, it should not.
- Agreement with i-LDSC. Across the 25 traits, the excess of $G$ from non-additive features over the additive model should correlate positively with the strength of the i-LDSC non-additive signal. This is an external check that the measure tracks real biology rather than modelling artifacts.
- Collapsing splits. Where exome sequence data are available, people whose sets split should be enriched for carriers of damaging variants in the relevant gene. Adding those variants as features should turn most split sets back into single intervals.
If the fourth check succeeds, split sets become a practical tool for finding which individuals, and which genes, hold the heritability that additive scores miss.
References
- Volume Optimality in Conformal Prediction with Structured Prediction Sets
- Flexible distribution-free conditional predictive bands using density estimators
- JAPAN: Joint Adaptive Prediction Areas with Normalising Flow
- CONTRA: Conformal Prediction Region via Normalizing Flow Transformation
- Discovering non-additive heritability using additive GWAS summary statistics
- Epistasis regulates genetic control of cardiac hypertrophy
- Enhancing polygenic risk prediction by modeling quantile-specific genetic effects
- Analysis of a deeply-phenotyped familial hypercholesterolemia cohort from Mexico shows a role for both rare and common alleles across known dyslipidemia genes and reveals structural variation in a novel locus
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study