Fitting Pathway Wiring to Genotype–Metabolite Data with Disjunctive Programming

Genetic studies of blood and urine metabolites keep finding genes whose rare damaging variants shift a metabolite's level. They then explain these shifts gene by gene, as if each gene acted alone. Real metabolism is a plumbing network. Transporters and enzymes sit in series or in parallel, and backup routes absorb a broken part. That wiring decides how variants in different genes combine: two genes in series compound each other's damage, while two genes in parallel buffer it. This gene-by-gene combining, called epistasis, is one place where "missing heritability" (the gap between how heritable a trait is and how much of that variation known genetic variants explain) can hide. Additive models cannot see it. Generic interaction scans see it only as an unstructured statistical residue.

We propose to borrow superstructure optimization from chemical process design and use it in reverse. In process design, engineers write one model containing every plausible unit and connection. A solver then picks which units to switch on (discrete choices) while tuning flows and capacities (continuous choices), with proofs of how far the answer can be from the best possible. Here, the candidate "units" are genes that may act at particular steps of a metabolic route. The "operating data" are the measured metabolite levels of thousands of sequenced people. The solver picks the wiring that best explains how genotypes map to metabolite levels. The output is an explicit, testable pathway structure, together with an estimate of how much genetic variance that structure explains beyond an additive model.

Core Claim

Treating each person, or more precisely each group of people who share a genotype at the modeled genes, as an operating scenario of a single shared plant turns pathway inference into a multi-scenario generalized disjunctive program. A generalized disjunctive program is an optimization model in which "either this step exists with these equations, or it does not" is written directly as a logical choice. Solving it with logic-based decomposition recovers pathway wiring that additive and generic interaction models miss. The recovered wiring predicts metabolite levels in held-out double-variant carriers better than those models do.

The closest existing work comes at this from different sides. Exploiting the pathway structure of metabolism to reveal high-order epistasis (NetKO) enumerates minimal knockout sets on a fixed, known E. coli stoichiometric network. It found that over 95% of the 11,706 cut sets disabling biomass have size 5 or more. That result motivates our view that network structure generates hidden interactions. But NetKO never fits a structure to human population data. Coupling metabolomics and exome sequencing reveals graded effects of rare damaging heterozygous variants on gene function and human traits supplies our data regime and our first target: rare variants in sulfate transporters SLC13A1 and SLC26A1 with graded effects on plasma sulfate and height. Its whole-body models are run forward, one knockout at a time, and are never inverted. Stem cell biomanufacturing under uncertainty: A case study in optimizing red blood cell production shows that superstructures can carry biological kinetics and parameter uncertainty, but there the goal is designing a bioreactor rather than explaining genetic data. Our contribution is the inversion: the structure is unknown, the data are a population, and the objective is a likelihood.

Research Program

For one metabolite (sulfate first, then other gene–metabolite clusters among the 235 associations reported by Scherer et al.), we build a superstructure. It contains every candidate gene that genome databases and the whole-body model place near that metabolite's production, transport, or excretion. Each gene is offered at each step where it could plausibly act, alongside alternative routes, such as parallel reabsorption paths in the kidney. Each person's genotype at a candidate gene sets a continuous capacity multiplier, using rare-variant burden and, where available, expression-predicting common variants. The model is

$$\min_{Y,\theta}\; \sum_{c} n_c\,\ell\big(\bar m_c,\, m(x_c)\big) + \lambda \sum_i [Y_i]$$

$$\text{s.t. } S v_c = b_c,\quad \begin{bmatrix} Y_i \\ v_{c,i} = k_i\, r(g_{c,i};\theta_i)\, \phi_i(x_c)\end{bmatrix} \vee \begin{bmatrix}\neg Y_i \\ v_{c,i}=0\end{bmatrix}\;\; \forall i, c.$$

The symbols are as follows. $c$ indexes genotype classes, $n_c$ is the number of people in class $c$, $\bar m_c$ is their observed mean plasma and urine metabolite levels, and $x_c$ is the model-predicted internal state with predicted readouts $m(x_c)$. $\ell$ is a Gaussian negative log-likelihood. $Y_i$ is the Boolean choice of whether candidate step $i$ exists, and $\lambda$ penalizes structural complexity in the way capital cost does in plant design. $S$ is the stoichiometry and $b_c$ the intake and excretion terms. $v_{c,i}$ is the flux through step $i$. $k_i$ is a baseline capacity, $r(g;\theta_i)$ maps genotype burden $g$ to a capacity fraction, and $\phi_i$ is a saturating rate law.

Grouping people by genotype is what makes this tractable. Rare damaging variants produce few distinct classes, so the scenario count stays in the hundreds rather than the thousands. Common-variant dosages are binned.

We solve the model with GDPOpt's logic outer-approximation and logic branch-and-bound from the Pyomo ecosystem described in Advances in Generalized Disjunctive and Mixed-Integer Nonlinear Programming Algorithms and Software for Superstructure Optimization.. In that work, the 64-alternative HDA flowsheet defeated commercial global solvers on its MINLP reformulation but was solved globally by GDPOpt in under a minute. The structure suits our problem: fixing the wiring leaves a nonlinear fit that separates by scenario, and each new wiring is proposed from cuts. CORAMIN relaxations give lower bounds, so we can report an optimality gap for the chosen structure. No random-forest or graph-learning method gives such a gap. We will also report the runner-up structures within a likelihood margin, because biological identifiability, not solver optimality, is the real limit.

The heritability readout comes from the fitted model's predictions on held-out people. We project those predictions onto an additive per-gene basis. The variance explained by the projection is the additive part. What the projection misses is structure-induced non-additive variance: the part of genetic variance that exists only because of how the genes are wired together.

Baselines

The naive reference point is the additive per-gene burden model, which is the rare-variant aggregation test used by Scherer et al. extended to a joint linear regression of the metabolite on all candidate-gene burdens. It is fit on the same genotype classes and the same training split. Every gain must be measured against it.

The strongest statistical competitor is the same linear model augmented with all pairwise burden-by-burden terms under ridge or lasso shrinkage, plus a marginal-epistasis variance component in the spirit of Detecting epistasis with the marginal epistasis test in genetic mapping studies of quantitative traits. That method tests each variant's aggregate interaction with the genome background. In GEUVADIS expression data it attributed roughly 10% of variance to epistasis versus 3.5% to additive effects. That figure is on a different benchmark, and no reported result found on Lacuna for held-out metabolite variance explained in GCKD by any interaction model. The strongest nonparametric competitor is low-signal signed iterative random forests from Epistasis regulates genetic control of cardiac hypertrophy, which recovered CCDC141–IGF1R non-additivity for left ventricular mass. All competitors get the same features, the same splits, and nested cross-validation with a matched hyperparameter budget.

Two ablations isolate the key ingredient. The first keeps the mechanistic model but fixes the wiring to the textbook pathway, which tests whether structure search matters at all. The second uses greedy forward or backward stepwise search over the same superstructure, which tests whether the global solver matters. Beating the competitors means higher held-out variance explained overall and, decisively, lower error on held-out double carriers, whose joint effect no model saw in training.

Why not just…?

Why not just fit pairwise interaction terms? This is the main threat for small pathways, and the experiments are designed to answer it. Pairwise terms need data on each pair, and double carriers of rare variants are scarce. A mechanistic structure predicts a double carrier's metabolite level from single-carrier data plus wiring, whether the genes act in series or in parallel. If the regularized interaction model matches us on held-out double carriers, our structure adds interpretation but no predictive value, and we will report that.

Why not just enumerate candidate wirings and compare them by an information criterion? For a pathway of five or six genes, this works and the solver is unnecessary. We will run full enumeration there as a correctness check on the solver. The case for disjunctive programming rests on larger superstructures: with 30 or more placement choices, enumeration means billions of nonlinear fits. NetKO's finding that relevant redundancies involve five to ten simultaneous elements says the interesting structures live at that scale. The greedy-search ablation shows whether a cheap heuristic already finds the same answer. If it does, we would recommend the heuristic.

Why not just use a flexible learner such as random forests or a graph neural network like BayReL? BayReL infers undirected cross-omics association graphs regularized by known graphs. Those graphs respect no mass balance and give no directional wiring. Flexible learners may match us on overall variance explained. Where they should fall short is extrapolation to unseen genotype combinations and conservation-consistent joint prediction of plasma and urine, which is where we will compare them.

Proposed Experiments

The first test is structure recovery on synthetic data with known ground truth. We generate genotypes with GCKD-like allele frequencies, simulate metabolite levels through whole-body-model subnetworks whose wiring is known, add measurement noise, and score how often the solver recovers the true wiring as cohort size and rare-variant frequency vary. This tells us where the approach is identifiable before we use any real data.

The second test is the sulfate system in GCKD, with 4,737 participants and plasma and urine metabolomics. We check whether the recovered wiring places SLC13A1 and SLC26A1 consistently with their known kidney and gut roles. We then report held-out variance explained against all baselines and the double-carrier extrapolation test. We also replicate the predicted plasma-sulfate effects in UK Biobank carriers. Access to individual-level GCKD data is a dependency; if it fails, the synthetic study and UK Biobank carry the work.

The third test scales to other gene–metabolite clusters from the 235 associations. Following The Importance of Regulatory Network Structure for Complex Trait Heritability and Evolution, which found trait heritability concentrated in a minority of network modules, we restrict each superstructure to one module, and we record solve time, optimality gap, and runner-up structures. The decisive outcome is whether structure-induced non-additive variance is consistently non-zero and replicates. If it is, network wiring is a measurable source of missing heritability. If the additive projection captures nearly everything, metabolite-level heritability is not hiding in the wiring.