Rare missense variants hid more heritability than burden tests showed: a flexible kernel lifts the lower bound from 0.54% to at least 1.27%
Synopsis
Rare missense contributions to complex traits have been underestimated because collapsing variants within a gene into one burden score averages away effect differences. The Kernel Allelic Series Model (KASM) builds within-gene effect covariance kernels from protein-structure and variant-effect annotations; across 786,871 exomes and 26 continuous traits it raises the lower bound on rare missense heritability from 0.54±0.04% to at least 1.27±0.09%, and to 1.83±0.11% under the most flexible model, above the 0.87±0.05% for loss-of-function variants.
Interpretation
Rare missense variants explain more than twice the complex-trait heritability estimated by burden approaches, and exceed the contribution of loss-of-function variants in the same data. Earlier burden-based estimates put missense heritability near 0.54±0.04%, with loss-of-function burden heritability roughly twice that of missense. Across 26 continuous traits in 786,871 UK Biobank and All of Us exomes, a BHR-like model fitted to the same inputs reproduced 0.54±0.04%, an exon-varying model gave 1.27±0.09%, and adding independent effects gave 1.83±0.11%, against 0.87±0.05% for loss-of-function variants.
Missense variants in the same gene have effects that are neither identical nor independent, and their similarity can be described by structural annotations and variant-effect scores. The burden model assumes perfectly correlated effects within a gene and therefore yields a single collapsed score. The model combines 15 deleteriousness predictors with 20 structural and functional annotations into kernels, recovers simulated kernel heritability and enrichment, and returns near-zero estimates in null simulations with common-variant effects only.
Trait-relevant missense effects concentrate at structurally important protein sites, with the strongest enrichment for charge-reversing substitutions at protein-protein interfaces. Prior structural enrichment work focused on Mendelian disease mutations; which features matter across complex-trait allelic series was unclear. In significant genes, charge-reversing substitutions at interchain hydrogen-bond residues were enriched 27±2-fold versus 3.9±0.2-fold for other substitutions there; 14 PIEZO1 inter-subunit hydrogen-bond variants, 0.8% of heterozygosity, explained 78% of that gene's HbA1c heritability.
Perspective
The analysis covers 26 well-powered continuous traits in UK Biobank and All of Us, about 17,000 genes and 15 million rare coding variants below 0.1% frequency; four UK Biobank binary disease traits were excluded from later analyses because of low observed-scale heritability. Conclusions are limited to protein-coding variants, and regulatory variants would need a separate variant-to-gene grouping scheme. Per-gene enrichment was fitted separately on 299 significant gene-trait associations, and the PIEZO1 and FGFR3 examples use UK Biobank exome association statistics.
The higher estimate under the most flexible model rests on additional assumptions: confounding effects are equally strong for missense and synonymous variants at the same frequency and do not vary with gene-level constraint, which are plausible but not fully testable; that model's independent-effect kernel is not frequency stratified, so the estimate remains a lower bound on total missense heritability. Missense and loss-of-function effects are positively correlated but more weakly than a uniform partial loss-of-function model predicts, and existing annotations do not systematically identify gain-of-function effects. The method depends on existing annotations, and more accurate deleteriousness or mechanism annotations could yield higher estimates.