Calibrating Variant Evidence with Age-at-Onset Hazards

Clinical genetics labs sort DNA variants into five tiers, from "pathogenic" to "benign," by adding up points from independent kinds of evidence: how common the variant is in healthy people, what computer predictors say, how it tracks through families, and how it behaves in lab assays. The weak point is the calibration step, which decides how many points a given assay or predictor score is worth. That step is fit against variants that experts have already labeled, and those labels come mostly from European-ancestry databases and from a small set of well-studied genes. We propose calibrating the same scores against something closer to the clinical truth: when carriers actually get sick. In a retrospective cohort, meaning a population whose genotypes and health records already exist, carriers of truly pathogenic variants develop disease earlier than everyone else. Carriers of variants of uncertain significance (VUS) should therefore look like a mixture of two groups: a high-risk group that behaves like known pathogenic carriers and a baseline group that behaves like non-carriers. Estimating the mixing fraction for each band of assay scores yields an outcome-based probability of pathogenicity, which converts directly into evidence points.

The core claim is that time-to-event data, analyzed with proportional hazards models (models in which each risk factor multiplies an unspecified baseline event rate by a constant), can calibrate functional and computational evidence without relying on curated labels. It can also do this separately within each ancestry, which label-based calibration cannot. What is new is the use of survival outcomes as the calibration target for ACMG/AMP evidence strength. Prior work uses outcomes only to check reclassifications after they are made, or treats a pathogenicity score as one more prognostic covariate.

Where the gap is

Three lines of prior work come close to this idea. A scalable approach to resolving variants of uncertain significance is the method we build on. It calibrates multiplexed functional assays (MAVEs) and in silico predictors gene by gene into ACMG/AMP points, using ExCALIBR for the assays, then checks the resulting classifications against static phenotype associations in the All of Us biobank. Outcomes appear there only as validation after calibration, and no time-to-event model is used. Protein Language Model Predicts Mutation Pathogenicity and Clinical Prognosis enters an ESM-1b "Evolutionary Index" as a continuous covariate in a Cox model on TCGA tumors. That tells us whether the score predicts progression. It does not tell us what fraction of variants in a score band are pathogenic, and it does not produce evidence strength. Impact of BRAF, TERT, and novel mutations on the efficacy of lenvatinib for advanced papillary thyroid cancer combines all three directions but runs the logic the other way. Its variant curation is a fixed input to a Cox model of treatment failure, and it reports late-entry and survivor bias, a warning our design has to handle.

Two findings motivate the ancestry angle. The Chinese Han BRCA study (Prevalence and reclassification of BRCA1 and BRCA2 variants) found 13 high-frequency pathogenic variants that are absent as high-frequency variants in Caucasian data. The Chinese Parkinson's cohort (Mutational spectrum and clinical features of GBA1 variants) found carriers had onset about 4 years earlier and that functionally "severe" variants had earlier onset still. In both studies age at onset carries signal about variant class, but both analyzed it cross-sectionally.

Core model

Take one gene. Let $t$ be age, $x_i$ the covariates of person $i$, and $S_B(t\mid x_i)=\exp\{-H_0(t)e^{\beta^T x_i}\}$ the Cox survival function of non-carriers, where $H_0$ is the cumulative baseline hazard. Known pathogenic carriers have survival $S_P = S_B^{\exp(\gamma)}$, with $\gamma$ their log hazard ratio. A carrier of a VUS whose assay score falls in bin $k$ has survival

$$S_i(t) = \pi_k\, S_P(t\mid x_i) + (1-\pi_k)\, S_B(t\mid x_i),$$

where $\pi_k$ is the carrier-weighted fraction of pathogenic variants in bin $k$. Fitting uses expectation-maximization on the likelihood with left truncation at enrollment age, since people in a biobank contribute risk only from the age they joined. This correction is our answer to the late-entry bias flagged in the lenvatinib study. The quantity $\pi_k$ becomes a likelihood ratio relative to the prior odds and then maps onto the points scale from the scalable-VUS paper.

A mixture of two proportional-hazards groups is itself not proportional: the hazard ratio drifts toward 1 as the high-risk members are depleted. This is the formal reason that entering the score as a single Cox covariate is misspecified.

The key identifying assumption is that pathogenic VUS share the penetrance $\gamma$ of known pathogenic variants. Hypomorphic, lower-penetrance variants violate it. A sensitivity analysis will let $\gamma_k$ vary by bin under a prior. If $\pi_k$ and $\gamma_k$ cannot be separated, we will report that openly, because it means survival data alone cannot distinguish "rarely pathogenic" from "weakly penetrant."

Research Program

The primary setting is BRCA1, because saturation genome editing (SGE) scores exist there and the outcome is age at breast or ovarian cancer diagnosis.

The first step is a simulation study that varies cohort size, carrier counts per bin, censoring, and penetrance misspecification. It will measure bias in $\pi_k$ and coverage of its interval, and it will map the sample size at which outcome calibration is informative enough to assign a given point level.

The second step fits the model in All of Us, used as in the scalable-VUS paper. We will use EHR-derived age at diagnosis, enrollment as left truncation, and genetic ancestry groups analyzed separately.

The third step is a case-control adaptation for the Chinese Han BRCA cohort, which has 21,216 patients and 6,434 controls. It assigns weights to cases and controls using external age-specific incidence, and controls are censored at their recruitment age. This design is also available for the GBA1 Parkinson's cohort, with functional severity tiers as bins.

Baselines

The naive reference is to do nothing: VUS stay VUS and receive zero points. Any calibrated method has to show that its reclassifications separate carriers' hazards better than this.

The strongest existing method is gene-specific ExCALIBR calibration against ClinVar labels, applied to the same SGE and MAVE scores and the same bins. That paper reports reclassifying 75% (12,135 of 16,115) of ClinVar VUS across 40 genes with under 1% discordance against definitive ClinVar calls. On time-to-event outcomes we found no reported result on Lacuna for ExCALIBR. The comparison is therefore run on held-out cohort outcomes with identical inputs and bin boundaries. Our method beats ExCALIBR if its reclassified groups show a larger held-out hazard ratio between the pathogenic-called and benign-called groups, and better calibration of predicted cumulative incidence, especially in non-European strata.

The second strong baseline is the covariate approach from the protein-language-model paper: a Cox model with the score entered linearly and also through restricted cubic splines, as in the Wuhan BMI study (Association of body mass index with severity and mortality of COVID-19 pneumonia). Both models get the same covariates and the same cross-validation.

Two ablations isolate the time component. One replaces the survival likelihood with a binary ever-diagnosed logistic mixture, removing timing information. The other removes left truncation, removing the retrospective-bias correction.

Why not just…?

Why not just use case-control odds ratios, as the All of Us validation did? This is a real threat. For high-penetrance genes a binary outcome may capture most of the signal. Time adds information only when onset age differs substantially between groups, which the GBA1 data suggest it does. The binary-mixture ablation settles the question: if its interval for $\pi_k$ is no wider than the survival version's, the time component is not earning its place, and we will say so.

Why not just put the score in a Cox model? A hazard ratio per score unit is not a probability of pathogenicity and cannot be turned into ACMG points. The mixture structure also predicts a specific departure from proportional hazards, one that splines on the score cannot represent. Goodness-of-fit comparisons with Schoenfeld-type residual checks will test this directly.

Why not just recalibrate ExCALIBR on ancestry-specific labels? Too few curated non-European labels exist, which is the problem documented by the Chinese BRCA and GBA1 studies. Outcomes are available in any genotyped cohort, whereas labels are available only where experts have already worked.

Proposed Experiments

The first decisive check is masked recovery. We hide the labels of known pathogenic and benign BRCA1 variants in All of Us, treat them as VUS, and require the model to return $\pi$ near 1 and near 0, with calibrated intervals. The second is agreement: outcome-derived points should agree with ExCALIBR points in European strata, and any disagreement in other strata should be examined as possible label bias rather than assumed to be error. The third is the efficiency comparison against the binary ablation, alongside the head-to-head against the baselines above.

A negative result would still be informative. If the simulation shows that biobank carrier counts are too small for outcome calibration to exceed "supporting" strength, that bounds what retrospective cohorts can contribute to variant interpretation.