Which Part of Silent Kidney Decline Is Inherited?

Chronic kidney disease progresses for years without symptoms. The record of that progression usually sits in routine lab tests and never becomes a diagnosis. A computable phenotype built from lab values found that over 80% of CKD cases lacked an ICD billing code (Shang et al.). Genetic studies that define disease by billing codes therefore place most affected people among the "healthy controls." They also reduce a decades-long trajectory to a yes/no label. Both choices push heritability estimates down. They are a plausible, measurable contributor to "missing heritability": the gap between how heritable a trait looks in twin studies and how much of it genotype arrays can explain.

This proposal asks a narrow, testable question. Suppose we describe each person's silent kidney decline by a few trajectory features: kidney function level, rate of decline, and curvature, computed only after removing acute injuries. Suppose we then search for the weighted combination of those features that is most heritable. Does that combination capture more genetic variance than static diagnoses or a simple slope? The test uses held-out hospital sites, so the search for heritability cannot grade its own homework. And what does the combination tell us about which aspect of progression the genome actually controls?

Core Claim

Chip heritability ($h^2_{\text{SNP}}$) is the share of trait variance explained jointly by all genotyped SNPs. Our claim is that the chip heritability of asymptomatic CKD is substantially underestimated because of two separable errors. The first is labeling: undetected cases are counted as controls. The second is temporal collapse: a trajectory is reduced to a stage. Each error can be repaired and measured on its own.

The method itself is deliberately simple. We take the inverse phenotype-optimization idea of Refining multivariate disease phenotypes for high chip heritability and An Effective Method to Identify Heritable Components from Multivariate Phenotypes. Those papers find weights over clinical indicators that maximize chip heritability, and in cocaine dependence they raised $h^2_{\text{SNP}}$ from 0.41 for symptom counts to 0.70, with replicated loci. We move it from cross-sectional symptom surveys to longitudinal lab trajectories.

If the genetic covariance matrix $\Sigma_g$ and the phenotypic covariance matrix $\Sigma_p$ of the trajectory features are estimated, the heritability of a composite $y = Xw$ is the ratio below:

$$h^2(w) = \frac{w^\top \Sigma_g w}{w^\top \Sigma_p w},$$

Here $X$ is the person-by-feature matrix and $w$ is the weight vector. Maximizing this ratio is a generalized eigenvalue problem. We add an $L_1$ penalty on $w$, as Sun et al. do, to keep the composite interpretable.

The novelty is not a new architecture. It is three things together. First, applying heritability-targeted phenotype construction to sparse, irregular, chronicity-filtered EHR trajectories. Second, a split-by-site design that prevents the optimizer from overfitting heritability. Third, explicit ablations that attribute recovered heritability to relabeling, to trajectory features, or to the optimization.

Research Program

The data come from the eMERGE-III network: 105,108 genotyped individuals with linked EHRs at 8 centers. Shang et al. already deployed their CKD phenotype there and replicated known loci at UMOD and APOL1 (Shang et al.).

We first apply their chronicity rule. Abnormal values must be separated by at least 90 days, and episodes coinciding with acute kidney injury, sepsis or shock are excluded. We also apply their ordinal model, which places dipstick, urine albumin-to-creatinine ratio (UACR) and 24-hour urine protein measurements on a common albuminuria scale.

From the cleaned series we fit a mixed-effects growth model per person, with age as the time axis. This model handles irregular sampling by pooling information across patients. It yields about ten interpretable features: estimated kidney filtration rate (eGFR) at a reference age, eGFR slope, eGFR curvature, the same three for albuminuria, residual variability, and age at first sustained abnormality.

We deliberately also compute healthcare-process features: test frequency, number of excluded acute episodes, and length of follow-up. These are not meant as phenotype. They are a probe for confounding. In EHRs, testing is driven by clinical suspicion (Forecasting Potential Diabetes Complications documents under 2% test coverage). Healthcare use also correlates with ancestry and socioeconomic status, so an optimizer could "find" heritability in who gets tested rather than in kidney biology.

The genetic covariance matrices are estimated with the blocked moment-matching estimator of Phenome-wide heritability analysis of the UK Biobank. That estimator handles 100,000 people in under 30 seconds on a CPU. This speed matters because we evaluate many candidate phenotypes and bootstrap replicates. Analyses are run within ancestry groups, adjusting for genetic principal components.

The optimization is fit on a subset of sites. All reported heritabilities and GWAS hits come from sites the optimizer never saw.

As a secondary arm, we replace the hand-built trajectory features with the latent embedding of the Longitudinal Variational Autoencoder. That model uses Gaussian-process priors to separate shared aging trends from individual deviation. This arm tests whether nonlinear latent representations add genetic signal beyond interpretable features. That idea is motivated by the nonlinear TWAS gains in Inference of nonlinear causal effects with application to TWAS with GWAS summary data, which found 18 Alzheimer's genes versus 11 for linear two-stage least squares.

Relation to Prior Work

Shang et al. motivate the proposal and supply the cleaning pipeline. They demonstrated the case-detection gap and used their phenotype for pedigree heritability and standard GWAS. However, they kept the fixed clinical staging grid and never asked which aspect of progression is most heritable.

Sun et al. supply the optimization. Their work is cross-sectional and survey-based, and an unvalidated jump to EHR data risks optimizing toward confounders. Our site-held-out design and process-feature probe are designed to address exactly this risk.

The Function-on-Scalar LASSO with Applications to Longitudinal GWAS showed that SNP effects on blood pressure vary with age, using densely scheduled Framingham exams. It does not address irregular multi-analyte EHR data and does not estimate total heritability. It motivates our hypothesis that genetic variance lives partly in rates of change.

The age-related decline in heritability reported for UK Biobank, averaging 3.5% per decade across 33 of 50 physical traits (Ge et al.), is the observation we try to reinterpret. If slope carries heritability that static values lose, the late-life "decay" is partly an artifact of measuring a single time point.

Baselines

The naive reference is ICD-coded CKD as a binary trait, analyzed on the same held-out sites with the same estimator. This is what most biobank analyses use.

The strongest existing phenotype is Shang et al.'s A-by-G staging, treated as an ordinal trait. Its clinical validity is established: 95% positive predictive value for cases on chart review, and 87% sensitivity and 97% specificity in independent validation (Shang et al.). No reported result was found on Lacuna for the chip heritability of this phenotype or of any eGFR trajectory in eMERGE. We will therefore compute it ourselves, and the comparison stands on our own held-out estimates.

The strongest simple trajectory baseline is the per-person eGFR slope after chronicity filtering, with no optimization. It is the single most obvious "progression" trait.

The strongest optimization baseline is Sun et al.'s method applied only to cross-sectional features: the latest values and the person-level means. The reference number from their work is cross-validated chip heritability of 0.70 versus 0.41 for symptom counts (An Effective Method to Identify Heritable Components from Multivariate Phenotypes). That number is for a different trait and is cited only as evidence of the scale of gain the method can produce.

The key ablations use identical sites, estimator, covariates and penalty-tuning budget. The full optimizer is run without the chronicity filter, isolating the cleaning step. It is run without slope and curvature features, isolating trajectory information. It is run with process features included and then excluded, exposing confounding.

Beating a baseline means a held-out heritability difference whose bootstrap 95% interval excludes zero. It must be accompanied by at least as many held-out loci, with UMOD and APOL1 recovered within the relevant ancestry groups.

Why not just…?

Why not just use the chronicity-filtered eGFR slope? This is the real threat, and it may well win. If the filtered slope matches the optimized composite within error on held-out sites, the honest conclusion is the following: the recoverable heritability comes from relabeling and from measuring rate of change, and heritability optimization adds nothing for CKD. That result is useful in its own right, because it tells biobanks to compute filtered slopes rather than run optimizers. The ablation table is built to give exactly this answer.

Why not just fix the labels and keep static stages? Correcting the more than 80% of missed cases may by itself recover most of the gap. We measure this directly as the step from ICD-coded CKD to the A-by-G stage. Only the further step to trajectory phenotypes is attributed to temporal information.

Why not just use a deep latent model like L-VAE end to end? Its reported advantages are in imputation and prediction, not genetics (Longitudinal Variational Autoencoder). No quantitative figure was found on Lacuna. Its latent dimensions are also hard to interpret biologically. We run it as the secondary arm and report it as useful only if it beats the interpretable-feature composite on held-out sites.

Why not just run a multivariate GWAS on all features? Joint GWAS gains power for locus discovery. However, it does not produce a single interpretable trait or a heritability estimate that can be compared with clinical definitions. We report joint-GWAS loci as a secondary check on whether the optimized composite misses loci.

Decisive Checks

Four results settle the question. The first is the heritability ladder on held-out sites: ICD code, then A-by-G stage, then filtered slope, then optimized cross-sectional composite, then optimized trajectory composite. Each step is reported with bootstrap intervals.

The second is the confounding probe. If process features receive substantial weight, or if the composite shows genetic correlation with test frequency, the gains are not credited to kidney biology.

The third is the age analysis. We stratify by age and test whether the heritability of slope declines more slowly with age than the heritability of the latest eGFR value. This directly tests whether the reported age-related decline is partly an artifact of measuring a single time point.

The fourth is prospective utility. A polygenic score trained on the optimized composite at discovery sites should predict later coded CKD at held-out sites better than a score trained on ICD labels.