Railway engineers know that a train's late arrival does not simply carry over into a late departure. The scheduled turnaround slack absorbs the delay first, and only the leftover passes downstream. Two trains that never interact can suddenly delay each other once they compete for the same platform or rolling-stock unit. The delay that comes out is a kinked, "max with zero" function of the delays that go in. We propose that a large class of gene–gene interactions, called epistasis, has the same structure. Genetic variation reduces the capacity of an enzyme or transporter. A cell's spare capacity absorbs small losses, so they stay silent. Only when the combined load from two genes that feed the same metabolite exceeds that slack does the metabolite level, and then the trait, shift. If this is right, the shape and sign of an interaction should be predictable before looking at phenotype data. Two genes arranged in series along a pathway should show one kind of interaction. Two genes acting as parallel, redundant routes should show the opposite kind. Genes that share no metabolite should show none.
This matters for the "missing heritability" problem. That is the gap between how heritable a trait looks in twin and family studies and how much variance single genetic variants explain in genome-wide association studies, which test one variant at a time. Non-additive variance is now detectable. Interaction LD score regression finds it in all 25 UK Biobank traits it tests (Discovering non-additive heritability using additive GWAS summary statistics), and forest-based scans have found validated interactions in cardiac hypertrophy (Epistasis regulates genetic control of cardiac hypertrophy). Searching all gene pairs, however, is combinatorially hopeless, and the interactions found carry no mechanism. A network-derived prior on which pairs interact, and in which direction, would shrink the search space and make each detected interaction a testable biochemical statement.
Core Claim
Delay propagation contributes a specific functional form, not a metaphor. In timetable models, the delay downstream of a buffer of size $s$ is $d_{out} = \max(0, d_{in} - s)$. When two upstream activities feed one resource, the effective input is the larger of their delays if they act in series, or a pooled load if they act as parallel routes. We carry this over directly. Let $\ell_A$ and $\ell_B$ be an individual's genetically predicted capacity losses at genes $A$ and $B$, and let $y$ be the level of a metabolite both genes touch. The slack model is
$$ y = \beta \, \max\big(0,\; f_{\tau}(\ell_A, \ell_B) - s\big) + \text{covariates} + \varepsilon , $$
where $s$ is the cell's spare capacity, $\beta$ is the effect once slack is exhausted, and $f_{\tau}$ is fixed in advance by the topology $\tau$ of the pair in a metabolic network. In series, $f = \max(\ell_A, \ell_B)$, because the slowest step limits flux. With parallel redundancy, $f = \ell_A + \ell_B$, which is silent until both routes are degraded. The model has one free nonlinear parameter, $s$, where a generic interaction test has a free coefficient per pair.
The claim is that $\tau$, read from a whole-body metabolic model, predicts three things in human cohorts. It predicts which pairs carry interaction signal on a shared metabolite, the sign of that interaction, and that the signal is specific to the pair and to that metabolite. The last prediction is what separates the proposal from a trait-wide change of measurement scale.
We should be candid that flux-based explanations of dominance and epistasis have a long history in biochemistry. The proposal does not claim that mechanism as new. The new parts are three. First, a slack-parameterized, topology-conditioned test run in human genotype–metabolite cohorts. Second, an estimator validated on a domain (rail) where the slack mechanism is known to hold. Third, a direct confrontation with the alternative that "epistasis" is a scale artifact.
Research Program
Mechanism source. The whole-body metabolic models used by Coupling metabolomics and exome sequencing reveals graded effects of rare damaging heterozygous variants on gene function and human traits simulate enzyme knockouts and reproduce clinical profiles of inborn errors of metabolism. We reuse them in two ways. We extract, for every gene pair sharing a metabolite node, whether the pair sits in series, in parallel, or shares only a cofactor. We also simulate graded double knockdowns to check that the simulator itself produces kinked, slack-like responses. The same paper motivates the work. It shows that rare damaging heterozygous variants have graded "allelic series" effects on metabolite flux. One example is the sulfate transporters SLC13A1 and SLC26A1, which affect plasma sulfate and height. A pair of transporters handling one solute is exactly a candidate parallel-redundancy pair.
Capacity proxies with power. Rare damaging variants make double carriers vanishingly rare in a cohort of a few thousand people. We therefore build a continuous capacity-loss score per gene. It combines a rare-variant burden term with genetically regulated expression predicted from cis-eQTL weights, the transcriptome-wide association idea summarized in the file, using GTEx v8 tissues as in The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution. This gives every individual a nonzero load at both genes, so the region where both are loaded is populated.
Estimator and positive control. We fit the slack model by profile likelihood over $s$, and we calibrate the test on synthetic data. We then apply the identical code to the Italian railway records (A Railway Network Dataset Incorporating Multi-Type Train Operation Records and Train Scheduling), which cover 2.67 million records over 3,324 trains. There the slack is the scheduled dwell and turnaround margin, which we know from the timetable. The control asks whether the estimator recovers known buffers from arrival–departure pairs. That paper reports a "strict linear correlation" between arrival and departure delays. The slack model predicts that this linearity holds above the buffer and flattens below it, so recovering the kink near the scheduled margin is a meaningful check. Shared-resource coupling between consecutive trips will be used only where the timetable data make the continuation unambiguous. The dataset may not label rolling-stock assignment, so we treat that part as opportunistic.
Baselines
The naive reference is the additive model. It regresses each metabolite on $\ell_A$ and $\ell_B$ with no interaction term. Every interaction claim must improve on it out of sample.
The strongest competitor for detecting pairwise interaction in individual-level data is a generic product-term scan. It adds $\gamma \ell_A \ell_B$ for every candidate pair sharing a metabolite and is run on the same capacity scores. It is the honest competitor because it uses the same pathway restriction and differs only in lacking our functional form and sign prior. It gets the same pairs, covariates, and cross-validation folds. We count our model as beating it if, at matched false discovery rate on held-out folds, it finds more replicated interactions, or the same number with correct predicted sign at a rate above chance. There is no reported result on Lacuna for pairwise epistasis detection in GCKD, so this comparison is new. The relevant reported anchor is that rare-variant aggregation found 235 significant gene–metabolite associations in that cohort (Coupling metabolomics and exome sequencing reveals graded effects of rare damaging heterozygous variants on gene function and human traits). Those genes define our candidate set.
We also compare against two ML and summary-statistic methods. The first is low-signal signed iterative random forests (Epistasis regulates genetic control of cardiac hypertrophy), run on the same features with an equal tuning budget. Its reported evidence is experimental: dual silencing of CCDC141 and IGF1R gave an 8.5% decrease in median cell size with non-additivity at $p < 10^{-4}$. It has no metabolite benchmark number. The second, for trait-level questions in UK Biobank, is interaction LD score regression (Discovering non-additive heritability using additive GWAS summary statistics). It reports significant non-additive cis-interaction variance in 25 of 25 UK Biobank traits and 23 of 25 BioBank Japan traits, with $r^2 > 0.85$ against additive LDSC. Its cis-only window cannot see trans pairs such as two transporters on different chromosomes. Our model can, so trans pairs are where we must add signal.
The ablation that isolates the key ingredient keeps the slack form and refit procedure but permutes topology labels across pairs, holding pair counts fixed. If performance survives shuffled topology, the network contributes nothing. A second ablation uses a single global nonlinear link on the summed additive score, described below.
Why not just…?
Why not just transform the phenotype, since kinked interactions are often scale artifacts? This is the most serious threat. A threshold on an additive load looks like epistasis on the raw scale and disappears under the right transformation. We address it in two ways. Every analysis is repeated under log and rank-inverse-normal transforms. We also fit the "global" model, in which one nonlinear link applies to the summed additive score across all genes affecting a metabolite. A scale artifact predicts interactions of the same sign among all pairs affecting that metabolite. Slack epistasis predicts opposite signs for series and parallel pairs, and none for pairs that do not share the metabolite. If the global model fits as well as the topology-conditioned one, we will report that the effect is a scale effect and drop the mechanistic claim.
Why not just run an unbiased forest or deep model and let it find interactions? These methods detect co-occurrence, but they pay the full combinatorial cost and return no sign prediction. The lo-siRF authors themselves describe epistasis discovery as limited by search space and signal-to-noise (Epistasis regulates genetic control of cardiac hypertrophy). The comparison at matched budget settles whether a one-parameter structured model wins at cohort sizes near 4,700, where we expect flexible models to be underpowered.
Why not just use static network modules and module-level scores? Module-level polygenic scores (A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study) and stratified heritability over eQTL communities (The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution) locate where additive signal concentrates. They are additive and cannot predict interaction sign. They motivate our candidate selection: heritability concentrates in local hubs, and those are natural shared nodes. They do not compete on the interaction question. A related warning comes from the delay literature, where static degree centrality underperformed dynamic contagion models at finding critical airports (Identification of Critical Nodes for Delay Propagation in Susceptible-Exposed-Infected-Recovered (SEIR) and Genetic Algorithm (GA) Route Networks). That is suggestive, but it comes from another domain, and we do not rely on it.
Proposed Experiments
The first decisive check is in silico. In the whole-body metabolic model, we sweep graded double knockdowns for all candidate pairs. We confirm that series pairs produce max-type responses and parallel pairs produce silent-then-sharp responses, and that the slack form fits better than a product term. If the simulator itself does not show the kink, the human analysis loses its prior and we stop.
The second is the rail positive control described above. It succeeds if the recovered buffers track scheduled margins across stations.
The third and central experiment is in the German Chronic Kidney Disease cohort: 4,737 participants with exome sequencing, 1,294 plasma and 1,396 urine metabolites. For pairs drawn from the 235 reported gene–metabolite associations and their network neighbours, we test three predictions. Interaction evidence should be enriched among topology-predicted pairs on their shared metabolite, compared with matched pairs that do not share a node. The observed interaction sign should agree with the predicted sign above 50%, using a binomial test. And the interaction should be absent on metabolites the pair does not share. The SLC13A1/SLC26A1 sulfate pair is the pre-registered flagship case.
Finally, for pairs that pass, we ask in UK Biobank whether the slack interaction on capacity proxies propagates to the downstream trait, such as height for the sulfate pair. We also ask whether this trans signal is outside what cis-only interaction LD score regression can capture.
A negative result at the third stage would also be informative. If sign agreement is at chance, it would suggest that most detectable human epistasis on metabolites is not organized by local network topology. The scale-artifact and omnigenic explanations would then gain ground.
References
- Discovering non-additive heritability using additive GWAS summary statistics
- Epistasis regulates genetic control of cardiac hypertrophy
- Coupling metabolomics and exome sequencing reveals graded effects of rare damaging heterozygous variants on gene function and human traits
- The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution
- A Railway Network Dataset Incorporating Multi-Type Train Operation Records and Train Scheduling
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- Identification of Critical Nodes for Delay Propagation in Susceptible-Exposed-Infected-Recovered (SEIR) and Genetic Algorithm (GA) Route Networks