Minimax Omnibus Tests for Rare Variant Association

Most of the genetic signal from rare mutations is invisible to single-variant tests, so geneticists pool the rare variants in a gene into one test. The problem is that no single pooling rule works for every gene. Summing variant counts (a "burden" test) is powerful when every variant pushes the trait the same way. Squaring per-variant evidence (a "dispersion" test such as SKAT) is powerful when effects point in mixed directions. Severity-ordered tests are powerful when worse mutations have larger effects. Because the true genetic architecture of a gene is unknown, practitioners run an omnibus test, a fixed rule that merges the p-values of several component tests. Today that rule is chosen by convention: a grid over one mixing parameter, equal-weight Cauchy merging, or a hand-selected ensemble.

We propose to treat omnibus design as an explicit multi-objective optimization. Each candidate genetic architecture defines one objective, the power under that architecture, and these objectives conflict. Any omnibus rule is one point in this space. We will learn the merging weights offline, on simulated cohorts, by maximizing the worst-case power ratio across architectures. That ratio compares the omnibus to the best single test for each architecture, subject to a constraint on false-positive rate. The weights are then frozen and applied to real data, so the learned test stays statistically valid. Concretely, the claim is that a minimax-weighted combination can recover most of the oracle power on architectures that current omnibus tests serve poorly, especially non-monotone ones, without losing power on the common cases.

Core Claim and Why It Is Different

Existing work sits on either side of this idea without occupying it. Learning the kernel for rare variant genetic association test (ecSKAT) learns kernel weights per gene by solving a single-objective quadratic program that minimizes a p-value bound for one trait. It adapts to data but optimizes only one scalar. Excalibur starts from 154 tests, filters them to 59, and picks an ensemble of 36 tests through simulation. It is the closest prior work, but its selection is a filter-then-select heuristic that never states what trade-off across architectures it makes. COAST (An allelic series rare variant association test) merges seven models with a Cauchy combination and assigns fixed severity weights of 1, 2 and 3 to benign missense, damaging missense and protein-truncating variants. SMMAT (Efficient variant set mixed model association tests) merges burden and SKAT p-values with Fisher's method and grid-searches a mixing parameter in SMMAT-O. All of these are, in effect, scalarizations, meaning ways of collapsing the vector of per-architecture powers into one number. None of them chose its scalarization deliberately.

The multi-objective literature tells us which scalarization to choose. Optimal Scalarizations for Sublinear Hypervolume Regret proves that linear weighted sums cannot reach Pareto-optimal points in concave regions of the trade-off front. A Pareto-optimal point is one where no objective can improve without another getting worse. The same paper shows that a min-of-ratios (Chebyshev-type) scalarization with randomized weights recovers the whole front. The Weighted Chebyshev approach in A Tale of Two Problems shows this scalarization working in practice on non-convex fronts. This literature supplies our method.

The motivation comes from Determinants of pleiotropy and monotonic gene dosage responses (FunBurd). Across 43 UK Biobank traits, that study finds that monotonic dosage responses occur in only 4% of cases. Architectures that COAST-style weighting assumes away are therefore common, and an average-power criterion would underweight exactly those architectures.

Formally, let $p_1,\dots,p_m$ be the p-values of $m$ component tests for a gene. Let $w$ be a vector on the probability simplex. The weighted Cauchy combination is

$$T(w) = \sum_{k=1}^{m} w_k \tan\big((0.5 - p_k)\pi\big),$$

which is approximately standard Cauchy under the null for any fixed $w$, even when the components are dependent. For a set of simulated architectures $s \in S$, let $\pi_s(w)$ be the power of $T(w)$ at gene-level significance $\alpha$. Let $\pi_s^\star$ be the power of the best single component under architecture $s$, and let $\epsilon_r(w)$ be the empirical false-positive rate under null regime $r$, for example a given case-control imbalance. We solve

$$\max_{w}\ \min_{s \in S} \frac{\pi_s(w)}{\pi_s^\star} \quad \text{subject to} \quad \epsilon_r(w) \le \alpha\ \text{for every null regime } r.$$

We also trace the full front by sampling Chebyshev weights over $S$, as the hypervolume-scalarization paper prescribes. The constraint is handled with a primal-dual Lagrangian, a choice motivated by the zero-duality-gap result for constrained policy optimization in Constrained Reinforcement Learning Has Zero Duality Gap. We do not claim that the theorem transfers to this setting. We will check the duality gap empirically.

Research Program

The work starts by building a simulation bank. Haplotypes come from COSI 2 coalescent simulation, as in Excalibur, and from real UK Biobank exome genotypes with simulated phenotypes. Over these we simulate a grid of architectures:

  • unidirectional effects with varying causal fractions;
  • mixed-sign effects;
  • monotone allelic series;
  • non-monotone series, where for example damaging missense variants have larger effects than truncating ones, or truncating variants act in the opposite direction, following FunBurd;
  • causal sets concentrated in one protein domain, mimicking the localization setting of DYNATE.

Null regimes include extreme case-control imbalance, which Boutry et al. identify as a major source of inflated false positives.

The component pool consists of the 20 tests Excalibur reports as controlling false-positive rate across all its experiments, plus the COAST components and burden tests under several annotation masks. For every simulated replicate we store the full vector of component p-values once. Power and false-positive rate for any weight vector can then be recomputed from the stored matrix without re-simulating, which makes the optimization cheap. For gradient steps, the significance indicator is replaced by a sigmoid. Monte Carlo noise in power estimates is treated as a noisy objective, which is the setting the hypervolume-regret analysis covers.

Estimating power and false-positive rate at $\alpha = 2.5\times10^{-6}$ needs care. Power is evaluated at effect sizes calibrated so that oracle power lies between 20% and 80%. The false-positive rate in the far tail is checked with roughly $10^{8}$ resampled null p-value vectors per regime, built from permuted phenotypes, and compared against the analytic Cauchy tail.

A secondary objective is cost. Excalibur projects up to 458 hours for a genome-wide scan without parallelization. We will add an L1-style sparsity term and report the front of worst-case power ratio against the number of components with nonzero weight, since only those components need to be run on real data.

Baselines

The naive reference point is SKAT-O, the grid-search omnibus that most rare-variant studies still report, run with its default grid on identical genotypes, covariates and masks. Doing better than it is necessary but not sufficient.

The strongest existing method on real data is COAST. On UK Biobank exomes from 145,735 subjects, COAST reports 29% more Bonferroni-significant associations with circulating lipid traits than SKAT-O on average, and 82% more with blood cell count traits. We will rerun COAST on the same subjects, traits and variant annotations rather than quoting its counts. To beat COAST, the learned omnibus must find more replicated Bonferroni-significant genes on these trait groups at a matched false-positive rate.

SMMAT-E is the strongest simulation baseline for combining burden and dispersion evidence. The SMMAT paper reports that it had the highest power for continuous and binary traits across all tested scenarios with 10,000 samples, though it gives no numeric power figure that we can quote. It will be run inside the same mixed-model null fit as our components.

For Excalibur we found no reported numeric power figure on Lacuna. It is the most direct competitor in spirit, so we will reimplement its 36-test ensemble on our simulation bank and real data.

Two ablations isolate the key ingredient. The first is equal-weight Cauchy combination over exactly our component pool, which tests whether learned weights matter at all. The second is a linear scalarization that maximizes average power across architectures with the same constraint, optimizer and compute, which tests whether the minimax criterion matters beyond simply optimizing.

All learned variants get the same simulation budget and the same component pool. Beating the field means a higher worst-case power ratio on held-out architecture families, no loss of more than a few points of power on the best-served architecture, and controlled false-positive rate in every null regime.

Why not just…?

Why not just use equal-weight Cauchy combination over all components? This is the real threat. Cauchy combination is dominated by the smallest p-value, so with a well-chosen pool, equal weights may already sit near the minimax point. The equal-weight ablation settles the question directly. If its worst-case power ratio on held-out architectures comes within a couple of points of the learned weights, we will report that the value lies in choosing the component pool, not the weights. We will then redirect the work toward the cost objective, where pruning components still saves real compute.

Why not just maximize average power across architectures? It is the obvious single-objective formulation, but Optimal Scalarizations for Sublinear Hypervolume Regret shows that weighted sums miss concave regions of the front. In this problem those regions would correspond to compromise tests that serve both minority and majority architectures. If non-monotone architectures are as common as FunBurd suggests, average-power weights should sacrifice them. The linear-scalarization ablation measures whether this happens on rare-variant power surfaces, which may turn out to be nearly convex.

Why not just use Excalibur's ensemble? Excalibur already controls false-positive rate and was selected for power. Its selection objective, however, is implicit and was never targeted at non-monotone or domain-localized architectures. A head-to-head on held-out architecture families, using its own COSI 2 setup, shows whether an explicit minimax criterion adds anything beyond careful curation.

Why not just learn weights per gene, as ecSKAT does? Per-gene adaptation can outperform any fixed omnibus when its p-value bound is tight. It answers a different question, though, and it requires the bound to correct for adaptation. Our weights are fixed before seeing real phenotypes, so validity follows directly. We will include ecSKAT on the hand grip strength and lupus phenotypes it was evaluated on, and report where per-gene adaptation wins.

Proposed Experiments and Decisive Checks

The first decisive check is held-out architecture generalization. We train weights with one architecture family left out, for example non-monotone series, and evaluate the worst-case power ratio on that family. If the minimax weights do not beat equal-weight and average-power weights there, the simulation-to-reality argument fails.

The second check is calibration. We report QQ behavior and empirical false-positive rate at $\alpha$ values from $10^{-2}$ down to $2.5\times10^{-6}$ under 1:100 case-control imbalance, using UK Biobank genotypes with permuted phenotypes.

The third check is real-data discovery with replication on UK Biobank exomes, for lipid and blood cell count traits as in COAST. Discovery runs on half the cohort and replication on the other half. Success means more replicated genes than COAST and SKAT-O at matched calibration. We will also look for recovery of genes with known common-variant support, such as COAST's ANGPTL4–triglyceride example.

The final check is runtime against Excalibur on a full genome-wide scan. If the sparse front reaches comparable power with a fraction of the components, the method has practical value even when its power gain is modest.