Polygenic risk scores add up thousands of small genetic effects into one number per person. Each person receives the same weights. Yet the same score moves biomarkers by different amounts in different people: genetic effects on glucose and triglycerides are larger in the upper tail of the distribution (Kim et al.), and genetic associations with brain chemistry or heart attack weaken with age (Bustillo et al., Isgut et al.). We propose that this heterogeneity has a mechanical cause that can be measured. When a feedback-controlled system, such as the loop that regulates blood glucose, approaches a tipping point, its restoring force weakens. A weaker restoring force does two things. A steady push, such as a genetic load, shifts the resting level further. Random day-to-day disturbances also take longer to fade, which shows up as larger and more slowly decaying fluctuations over time. Linear stability analysis, the study of how small perturbations around an equilibrium grow or decay, says both effects are governed by the same number.
The note turns this into a falsifiable test. Across people, the local slope of a biomarker on polygenic score should rise in proportion to that person's biomarker autocorrelation time, estimated from repeated lab measurements. If it does, a person's "stability margin" becomes a quantity that can be estimated, and it should predict when compensation fails, for example the onset of type 2 diabetes, better than either the genetic score or the current biomarker level.
Core Claim
Consider one homeostatic variable $x$, such as HbA1c or fasting glucose, near its equilibrium $x^*$. To first order it obeys
$$ dx = \big(-k\,(x - x^*) + b\,g\big)\,dt + \sigma\, dW, $$
where $k>0$ is the restoring rate (minus the relevant Jacobian eigenvalue), $g$ is the polygenic score, $b$ converts genetic load into a steady push, and $\sigma\,dW$ is environmental noise. The equilibrium shift per unit of score is $b/k$. The stationary variance is $\sigma^2/(2k)$, and the autocorrelation time is $1/k$; both follow from the standard solution of this equation (an Ornstein–Uhlenbeck process). Eliminating $k$ gives the prediction we test:
$$ \frac{\partial x^*}{\partial g} = b\,\tau, \qquad \tau = 1/k . $$
The genetic slope should scale linearly with the individual's autocorrelation time $\tau$, with an approximately constant proportionality factor $b$.
Near a saddle-node bifurcation, the point where a stable equilibrium merges with an unstable one and disappears, $k$ shrinks toward zero. Slopes then steepen in the tail, which matches the quantile pattern. Past the bifurcation the person sits in a different basin of attraction. There the linear relation around the healthy equilibrium no longer holds, and the prediction is that the slope collapses in people who have already decompensated. This matches the age pattern.
Other explanations predict something different. A change of measurement scale or a variance-controlling locus predicts slope heterogeneity with no link to how slowly fluctuations decay. That difference is what makes the claim testable rather than a redescription.
Relation to Prior Work
The intersection is empty in a specific sense. Barcenas, Bocci, and Nie estimate Jacobians along single-cell pseudotime and locate a saddle-node tipping point in epithelial–mesenchymal transition. They use somatic transcriptomes with no germline variation, and we reuse their logic that the leading eigenvalue marks the transition. Biroli, Bunin, and Cammarota show that large interacting systems settle at marginal stability, where susceptibility is large. This motivates expecting small restoring rates in real physiology.
On the genetics side, three lines relax the one-weight-for-everyone assumption statistically:
- quantile-specific scores (Kim et al.),
- dominance-aware boosting (GenoBoost),
- multi-dimensional risk vectors (PRV).
All three describe heterogeneity at the genotype or phenotype-quantile level. None ties it to a measurable dynamical quantity of the person. Module-level scores (Luo et al.) and pathway scores (Barbu et al.) localize risk to biological subsystems but treat networks as static graphs.
The obvious bridge would be to fit a gene-regulatory ODE per patient from their score. We deliberately avoid it: its parameters are not identifiable from genotypes and cross-sectional data, a limitation Barcenas et al. already hit with far richer data. Our proposal needs only one eigenvalue per person, and that eigenvalue is estimable from repeated clinical labs.
Research Program
Stage 1: validate the estimators in simulation. We simulate the bistable toggle and tristable circuits used by Barcenas et al. across a synthetic population. Genetic load enters as a shift in a bifurcation parameter, and noisy, irregularly spaced "lab draws" are sampled. We then check two things. First, whether per-person autocorrelation time, estimated from 4–10 irregular draws with a continuous-time autoregressive model, recovers $1/k$. Second, whether the slope-versus-$\tau$ relation is recoverable at biobank sample sizes. This stage also gives the power calculation, and it shows how the relation degrades when measurement error, trends, or medication breaks the stationarity assumption.
Stage 2: estimate the quantities in real data. In UK Biobank we compute a standard T2D/glycemic polygenic score from public weights and extract repeated HbA1c and glucose values from linked primary-care records. That repeated-measure coverage must be verified before committing; the file documents linked health records but not their density. For each person we estimate $\tau$ and within-person variance from medication-free, pre-diagnosis windows, using a model that separates slow trends from fluctuations. KARE (Kim et al.) serves as a second cohort with skewed glycemic traits, if repeat visits are available.
Stage 3: test the identity. We regress the biomarker on score × $\tau$ with covariates, and bin people by $\tau$ to estimate a genetic slope per bin. The core prediction is a monotone, approximately linear rise of slope with $\tau$ and a flat implied $b$. The secondary prediction is that slopes collapse among people past diagnosis.
Stage 4: prediction. We combine score, current HbA1c, and estimated margin $k$ into a time-to-event model for incident T2D among normoglycemic people.
Baselines
The naive reference is a standard additive polygenic score with a single slope, entered linearly alongside HbA1c and clinical covariates. It represents "genetics adds a constant shift." We will also report the score after log-transforming the biomarker, since a scale change is the cheapest explanation of tail-dependent effects.
The strongest existing competitor for heterogeneous effects is the quantile-specific approach of Kim et al.. It directly models tail-varying genetic effects on the same glycemic traits. The file reports gains in $R^2$ for triglycerides and post-load glucose, but I found no reported numerical value on Lacuna for glycemic traits. For incident T2D discrimination with a polygenic score I found no reported result on Lacuna either. The closest reported benchmark of how much a score adds to clinical risk is for coronary disease: C-statistic 0.718 for clinical equations, rising to 0.753 with the integrated score, with NRI 0.093 (iPRS study). We cite that only as a scale reference, not a target.
GenoBoost and PRV will be run as non-additive genotype-level competitors. For fairness, all models see the same individuals, the same genotype weights source, and the same repeated-lab window, with equal hyperparameter search budgets on a held-out tuning split. Beating the baselines has two parts. For the mechanistic test, the score × $\tau$ interaction must explain slope heterogeneity that quantile-specific scores fit but cannot attribute. For prediction, we need a C-statistic gain with a confidence interval excluding zero over the best baseline among people with normal HbA1c.
The key ablation replaces the estimated $\tau$ with within-person variance alone, and then with a shuffled $\tau$. This isolates whether the temporal-correlation ingredient, which is specific to the stability account, carries the signal.
Why not just…?
Why not just log-transform the trait? If tail-steepening were a scale artifact, a suitable transform would flatten the slopes. It would not, however, create a correlation between slope and autocorrelation time, because a monotone transform of a stationary series does not change how quickly fluctuations decorrelate. We will report slopes on raw and transformed scales. If the slope–$\tau$ link disappears after transformation, the central claim fails.
Why not just fit a variance-controlling genetic score or quantile model? These describe heteroscedasticity well, and Kim et al. show that they improve fit. They say nothing about time, though. A variance locus predicts larger spread without slower recovery; the stability account predicts both, linked through the same $k$. This is a genuine competitor for prediction. If quantile-specific scores match our predictive gains, the contribution reduces to the mechanistic identity, and we will say so.
Why not just use glycemic variability as a biomarker, without genetics? This is the most serious threat to the prediction stage. Within-person variability may already predict diabetes, and our margin estimate is partly a variability measure. The ablation settles it. If variance alone matches the full model, the predictive novelty is small. The genetic test remains distinct, because variability alone says nothing about why genetic effects vary.
Why not a deep model on raw genotypes? DeepPRS reaches a genotype-only AUC of 0.7245 for Alzheimer's disease. A genotype-only network cannot see a person's current distance from a tipping point, since that depends on age and environment. It is complementary, not competing, and we can feed its score in as $g$.
Decisive Checks
The idea stands or falls on three results:
- Simulations show that $\tau$ is recoverable from sparse labs at the needed precision.
- In UK Biobank, the genetic slope rises with $\tau$ on both raw and transformed scales, with shuffled-$\tau$ controls flat.
- Slopes collapse after diagnosis.
A clean negative on the second check would itself be informative. It would imply that tail-varying genetic effects in glycemic traits come from genotype-level or measurement heterogeneity, not from proximity to instability.
References
- Enhancing polygenic risk prediction by modeling quantile-specific genetic effects
- Risk-Conferring Glutamatergic Genes and Brain Glutamate Plus Glutamine in Schizophrenia
- Highly elevated polygenic risk scores are better predictors of myocardial infarction risk early in life than later
- Tipping points in epithelial-mesenchymal lineages from single-cell transcriptomics data
- Marginally stable equilibria in critical ecosystems
- A polygenic score method boosted by non-additive models
- Polygenic risk vectors (PRV) improve genetic risk stratification for cardio-metabolic diseases
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- Association of Whole-Genome and NETRIN1 Signaling Pathway–Derived Polygenic Risk Scores for Major Depressive Disorder and White Matter Microstructure in the UK Biobank
- Polygenic risk score improves the accuracy of a clinical risk score for coronary artery disease
- A Deep Learning-based Genome-wide Polygenic Risk Score for Common Diseases Identifies Individuals with Risk