Rare disease genetic diagnosis starts by throwing variants away. A sequenced genome carries millions of differences from the reference, so clinical pipelines discard any variant seen often in healthy people, using a fixed cutoff on its population frequency (for example, below 0.1% for recessive disorders). Nobody knows how often this step discards the variant that actually causes the patient's disease, and the answer almost certainly depends on the patient's ancestry. A variant that is rare in mostly-European reference databases can be common and harmless in a founder population. A real disease allele can drift to a frequency well above what theory predicts and then get filtered out. We propose to treat the frequency cutoff as a statistical decision with a stated error rate. Using a set of solved diagnostic cases, we pick the cutoff so that the probability of discarding the causal variant stays below a user-chosen level, such as 5%. The solved cases are reweighted by how similar their genetic ancestry is to the incoming patient's, so the guarantee holds for patients of underrepresented and admixed ancestries as well as the majority group.
The method rests on conformal prediction, a family of techniques that wrap any scoring rule and use a held-out labeled set to produce outputs with a guaranteed error rate. It also rests on one reframing. Prior work calibrates at the level of individual variants, where pathogenic examples are vanishingly rare and almost absent for minority ancestries. We calibrate at the level of the solved case. Each diagnosed patient contributes exactly one positive: the variant that caused the disease. The class imbalance that breaks variant-level calibration therefore disappears from the calibration problem. The imbalance now shows up only as the efficiency cost, meaning how many benign variants survive the filter.
Core Claim
For a solved case $i$, let $s_i$ be the filter score of its causal variant: its reference allele frequency, or a combination of frequency and a pathogenicity score along a fixed one-parameter path. Let $a_i$ be the patient's ancestry coordinates, taken from genotype principal components. The pipeline keeps every variant whose score is at most $\lambda$. Missing the diagnosis is the event $s_i > \lambda$, and that event can only become less likely as $\lambda$ grows. Under exchangeability, choosing $\lambda$ as the conformal quantile of the calibration scores $\{s_i\}$ guarantees that a new patient's causal variant is kept with probability at least $1-\alpha$. For a binary, monotone loss, conformal risk control as used in Functional protein mining with conformal guarantees reduces to exactly this quantile.
Exchangeability fails across ancestries, so we replace the plain quantile with a weighted one. For a test patient with ancestry $a$, we use
$$\hat\lambda(a) = \text{Quantile}_{1-\alpha}\Big(\sum_{i=1}^n \tilde w_i(a)\,\delta_{s_i} + \tilde w_{n+1}(a)\,\delta_{+\infty}\Big),$$
where $\delta_{s}$ is a point mass at $s$. The normalized weights $\tilde w_i(a)$ are proportional to the estimated ratio of the test-ancestry density to the calibration-ancestry density at $a_i$. In the localized version, this ratio is a kernel in ancestry space centered at $a$. The extra mass at $+\infty$ is the standard finite-sample correction, and it grows as the weights concentrate. When almost no calibration cases resemble the patient, the threshold correctly relaxes toward keeping everything. The pipeline never silently claims a guarantee it cannot back.
The claim has three parts. First, fixed frequency cutoffs and unweighted conformal thresholds miss causal variants at rates above $\alpha$ for ancestries underrepresented in calibration. Second, the ancestry-weighted threshold restores the target miss rate. Third, it does so with a benign-candidate burden close to what a matched population panel achieves, without requiring that panel to exist.
Relation to Prior Work
Two papers motivate the problem. The Ashkenazi reference panel study (Sequencing an Ashkenazi reference panel) sequenced 128 Ashkenazi genomes. It found that filtering against this matched panel removed benign variants eightfold more effectively than filtering against European panels. Its filters are still fixed cutoffs, it gives no bound on the rate of discarding causal variants, and it covers one isolate rather than admixed or continuous ancestry. The population genetics of human disease: recessive, lethal mutations shows that known lethal alleles at non-CpG sites appear at 15-fold (transitions) to 30-fold (transversions) higher frequency than mutation-selection balance predicts. A threshold derived from theory, as in Guidelines for investigating causality or the 0.1% cutoff in Recessive genetic contribution to congenital heart disease, can therefore throw away real disease alleles. We take both findings as reasons to set the threshold from outcomes rather than from theory.
Two lines of work supply the method and constrain it. Act or Defer (STRATCP) shows that marginally valid conformal prediction can still fail badly on a rare class: it reports about 10% coverage of the positive pancreatic-cancer class against a 90% target. It fixes this by calibrating each class separately, and it names the resulting requirement for enough labeled data per class as a limitation for orphan diseases. We adopt its diagnosis but not its remedy. Stratifying by class crossed with ancestry is exactly what scarcity forbids. Case-level calibration with ancestry weighting is our way around that. Classification with Valid and Adaptive Coverage documents that coverage drops below target on hard slices, which motivates measuring miss rate by ancestry rather than on average. The protein-mining paper supplies the risk-control machinery and the habit of reporting an honest "no confident answer" outcome. It also concedes that reference-database bias toward model organisms breaks its guarantees, which is the same shift we target.
The pathogenicity scorers MetaRNN and MVP are inputs to our filter score, not competitors to the guarantee. MetaRNN already restricts itself to rare variants, avoiding the trap of separating common benign from rare pathogenic variants. However, it feeds four gnomAD and ExAC frequency features into the model, so its score inherits the same ancestry skew. Neither tool gives an error budget.
Research Program
We will build the filter in three layers. The base layer computes, for every variant in a genome, a frequency score: the frequency in the reference resource being tested, which may be global, European, or population-matched. The second layer optionally combines that frequency score with MetaRNN along a single path, for example "frequency at most $\lambda$ and MetaRNN score at least $\tau(\lambda)$" for a fixed monotone $\tau$. This keeps calibration one-dimensional and the guarantee intact. The third layer computes ancestry coordinates for every patient and estimates calibration weights. We will compare a classifier-based density ratio (calibration versus target population) with a localized Gaussian kernel. We will report the effective sample size $(\sum_i w_i)^2/\sum_i w_i^2$ next to every threshold, so a clinician can see how much evidence underlies the guarantee for that patient.
We will calibrate separately by inheritance mode (dominant versus recessive), because the guidelines already treat these modes differently and both modes have enough cases to support separate thresholds. For recessive disease, the "causal variant" is the pair. The case is missed if either allele is filtered out, so the loss is the maximum of the two allele scores. This remains monotone.
The main empirical question is whether weighting actually closes the ancestry gap at realistic calibration sizes. The main analytical question is how the guarantee degrades when the patient's ancestry lies outside the support of the calibration ancestries. We will characterize this through effective sample size and through the mass placed at infinity.
Baselines
The naive reference point is current practice. It applies fixed frequency cutoffs, 0.01% for dominant and 0.1% for recessive inheritance as in the congenital heart disease cohort, against gnomAD-style global frequencies. These are run on exactly the same test genomes, so its miss rate and candidate burden are directly comparable. Our method beats it if its miss rate is at or below $\alpha$ in every ancestry group where current practice exceeds $\alpha$, at no more than a modest increase in candidates per genome.
The second baseline is the best population-specific practice: the same fixed cutoffs applied against a matched panel, which here is the Ashkenazi panel for Ashkenazi genomes. This is the strongest existing method for removing benign variants. The only reported result on Lacuna is the eightfold improvement in benign filtering over European panels from the Ashkenazi study. No reported result found on Lacuna for its causal-variant miss rate. Within Ashkenazi genomes, our method "beats" it only if it matches the candidate burden within a pre-stated margin while achieving a controlled miss rate. Outside populations that have panels, this baseline does not exist, and that absence is the point.
The third baseline is unweighted split-conformal calibration on the same solved cases, with identical scores and calibration-set sizes. It isolates the key ingredient, ancestry weighting. If unweighted calibration already holds $\alpha$ in minority groups, the weighting contributes nothing and we will say so.
The fourth baseline is ancestry-stratified conformal calibration, which calibrates within each discrete ancestry group. It is the direct analogue of STRATCP's per-class remedy. STRATCP's strongest reported number is coverage above 99% versus about 10% for standard conformal prediction on EHRSHOT pancreatic cancer. That result comes from a different domain, and no reported result found on Lacuna for stratified conformal calibration of variant filters.
For the pathogenicity score, we use MetaRNN as released (reported AUC 0.9311 on its rare nsSNV test set). We also test a version retrained with its four allele-frequency features removed, to check whether frequency features inside the scorer reintroduce ancestry skew. All methods share the same calibration cases and test genomes. Their only tuned quantity, the kernel bandwidth or density-ratio classifier, is chosen on a separate validation fold with a fixed budget.
Why not just…?
Why not just build a population-specific panel, as was done for Ashkenazi Jews? Where such a panel exists, this is a real threat. It may match our method's benign burden and needs no calibration data at all. It still says nothing about how often the causal variant is discarded, and the 15–30-fold frequency elevations of known lethal alleles suggest that a well-matched panel can also discard real causal variants. Panels also exist for few populations and do not handle admixed individuals. Our Ashkenazi experiment settles the matched case directly. If static filtering against the matched panel already keeps the miss rate below $\alpha$ at equal burden, our contribution narrows to the populations without panels, and we will report it that way.
Why not just stratify calibration by ancestry? A conformal quantile from $m$ calibration cases cannot certify any miss rate below $1/(m+1)$. A group with 19 solved cases therefore cannot support a 5% target at all, and the threshold degenerates to "keep everything." STRATCP's authors flag exactly this scarcity for rare classes. Stratification also forces admixed patients into discrete bins. The stratified baseline will show how often minority-group thresholds become trivial, compared with weighted thresholds at the same data.
Why not just recalibrate MetaRNN's scores per ancestry, for example with isotonic regression? Recalibration fixes how well variant-level probabilities match observed rates. That is not the clinically relevant question, which is whether this patient's causal variant survives. It also gives no finite-sample guarantee and needs pathogenic examples per ancestry, which is the same scarcity problem again. We include it as a secondary baseline, applying a threshold to the recalibrated score chosen to hit $\alpha$ on validation data, and compare miss rates under shift.
Why not just tune the cutoff empirically on the solved cases? This is close to our unweighted baseline, and with large, ancestry-matched calibration sets it should perform nearly identically. The finite-sample correction matters little there. The claim being tested is that empirical tuning on majority-ancestry cases fails for other ancestries, which the shift experiments test directly.
Proposed Experiments
The first experiment is semi-synthetic, which gives controlled ground truth. We spike curated pathogenic alleles into real background genomes. The alleles are ClinVar pathogenic missense variants and indels, plus the 417 lethal or sterile alleles from the ExAC and Counsyl study, each kept with its real population frequencies. The background genomes are the 128 Ashkenazi and 26 Flemish genomes and the 1000 Genomes populations. Each spiked genome is a "case" whose causal variant is known. We calibrate on mostly-European cases and test separately on Ashkenazi, African, East Asian, South Asian, and admixed American genomes. Calibration sizes range from 50 to 2,000 cases, and $\alpha$ is set to 0.05 and 0.10.
The decisive result has three parts. First, fixed cutoffs and unweighted conformal calibration should exceed $\alpha$ in at least the Ashkenazi group, given its founder alleles and its sevenfold-plus excess of shared identical-by-descent segments. Second, the weighted threshold should hold $\alpha$ within binomial error in every group. Third, candidates retained per genome should stay within a stated factor of the matched-panel baseline. If weighting fails to restore coverage at realistic calibration sizes, the core claim is falsified.
The second experiment uses real solved cases with recorded ancestry, drawn from large rare-disease collections of the kind behind the mitochondrial study of over 6,500 families and the recessive congenital heart disease analysis of 5,424 probands. Access to these is a project risk. Here we repeat the shift evaluation with leave-one-ancestry-out calibration. Because real causal variants follow real ascertainment patterns, this tests whether the semi-synthetic gains survive.
The third check concerns the shortcut. We compare miss rates by ancestry when the combined filter uses MetaRNN with versus without its frequency features. If the version with frequency features shows larger ancestry gaps after conformal weighting, then frequency information inside learned scorers is a separate source of bias that calibration alone does not remove. That would be a finding in its own right, and it would point to retraining the scorer rather than wrapping it.
References
- Functional protein mining with conformal guarantees
- Sequencing an Ashkenazi reference panel supports population-targeted personal genomics and illuminates Jewish and European origins
- The population genetics of human disease: The case of recessive, lethal mutations
- Guidelines for investigating causality of sequence variants in human disease
- Recessive genetic contribution to congenital heart disease in 5,424 probands
- Act or Defer: Error-Controlled Decision Policies for EHR Foundation Models
- Classification with Valid and Adaptive Coverage
- MetaRNN: differentiating rare pathogenic and rare benign missense SNVs and InDels using deep learning
- MVP predicts the pathogenicity of missense variants by deep learning