Many genetic effects depend on the environment. A variant can raise body weight in sedentary people and do nothing in active ones. Geneticists call this gene–environment interaction (G×E). A recurring trap is that much apparent G×E is not biology at all. It is a scale artifact: an interaction that appears only because the trait was measured on a curved scale, or because its variance grows with its mean. For body mass index in UK Biobank, a genetic risk score's apparent interaction with environment explained 1.9% of variance, and matched "fake" controls showed that this signal was non-specific Quantification of the overall contribution of gene-environment interaction for obesity-related traits. Meanwhile, the standard way to interpret genomic hits is functional enrichment analysis, which asks whether the flagged genes fall into known biological pathways more often than chance. Enrichment inherits every artifact in the gene list it is handed. If the list is inflated by scale effects, the enriched pathways will look like mechanism when they are measurement.
We propose a pathway-level G×E test whose null hypothesis is "this pathway interacts with the environment no more than a random gene set with the same main-effect strength would, given the trait's scale." The test compares each real pathway against many matched fake pathways and fake phenotypes. That calibration is expensive: every null draw means refitting a penalized interaction model, and refit times are heavy-tailed. We make it affordable with a compute allocator borrowed from structured procrastination, an algorithm-configuration scheme that runs candidates under short time limits and pushes unfinished ones to the back of the queue with doubled limits. The claim that matters is biological. Enrichment should serve as a specificity control on G×E rather than as a descriptive afterthought. The scheduler is what makes that control computationally feasible across thousands of pathways.
Core Claim
Current practice runs in one direction. Studies detect interacting loci or differentially expressed genes, then run Gene Ontology or KEGG enrichment on the result. The oyster growth study Identification of SNPs and Candidate Genes Associated with Growth Using GWAS and Transcriptome Analysis in Portuguese Oyster (Magallana angulata) and the poplar salt-stress transcriptomes Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings both work this way. They motivate our proposal, because they show that environmental responses organize into coherent pathway programs such as reactive-oxygen scavenging in roots and ion transport in leaves. They also expose the gap: nothing in that pipeline asks whether a pathway's interaction signal exceeds what scale alone would produce.
We claim that pathway-level G×E should be declared only when it survives two matched negative controls, both lifted from the gene level to the pathway level. The first, "fake pathways," are random gene sets matched to the real pathway on size and main-effect strength. The second, "fake phenotypes," are phenotypes simulated from the fitted main-effects model and passed through the trait's estimated nonlinear transformation, with no interaction built in. These controls follow the Fake GRS and Fake Y designs of Quantification of the overall contribution of gene-environment interaction for obesity-related traits, which applied them only to one aggregate risk score per trait. We apply them to each of thousands of pathways.
Concretely, let $T_P$ be the interaction statistic for pathway $P$. This is the penalized-likelihood gain from adding that pathway's G×E terms to a model that already contains all main effects. Let $\tilde{T}_P^{(b)}$ be the same statistic for the $b$-th matched null, drawn from either control. The specificity p-value is
$$ p_P = \frac{1 + \#\{b \le B_P : \tilde{T}_P^{(b)} \ge T_P\}}{1 + B_P}, $$
where $B_P$ is the number of null draws spent on pathway $P$. False discovery control is then applied across pathways. The central object is $B_P$. Uniform allocation wastes almost all compute on pathways that are obviously null after a few draws. Naive early stopping, meanwhile, biases the null.
Research Program
The interaction model for each fit is the hierarchy-respecting penalized regression of Structured gene-environment interaction analysis. That method writes each interaction coefficient as a main effect times an interaction weight, so an interaction cannot enter without its main effect. It penalizes both with the minimax concave penalty and fits by coordinate descent. For expression data it adds a Laplacian penalty from a gene network. We reuse this model unchanged as the per-fit engine. Our contribution sits one level above it: pathway-level hypothesis testing with scale-matched calibration, which that paper does not attempt. It reports selected genes and interactions, not calibrated pathway-level specificity.
Matched null generation works as follows. Fake pathways are sampled by stratifying genes on marginal main-effect strength and local correlation (linkage structure for SNP data), then drawing sets with the same size and strata profile as $P$. Fake phenotypes come from the fitted no-interaction model. We estimate a monotone transformation linking the linear predictor to the observed trait, together with a variance function, and simulate traits carrying the same curvature and heteroscedasticity but zero true interaction. A pathway is called specific only if it clears both controls.
Compute allocation is where procrastination scheduling enters. Efficiency Through Procrastination: Approximately Optimal Algorithm Configuration with Runtime Guarantees gives each candidate a short time limit and requeues unfinished candidates with doubled limits. The authors prove worst-case dominance over incumbent-driven schemes, which can sink unbounded time into slow candidates. They report no empirical benchmarks. Our setting has two budgets that need this treatment. The first is the number of null draws per pathway. A pathway whose running p-value is clearly above or below the decision band stops, while ambiguous pathways are requeued with doubled $B_P$. The second is wall-clock time per penalized refit. Coordinate descent on large pathways under strong interaction can converge slowly.
The key design choice is that fits exceeding their time limit are deferred, not discarded. Discarding them would bias calibration, because slow convergence plausibly correlates with the very structure being tested. This is exactly the failure that separates our allocator from off-the-shelf sequential p-value stopping, which assumes every draw completes. The allocator's guarantee is inherited in spirit from structured procrastination: no pathway's decision can be starved by a few pathological fits. We will state and check a version of that bound in which the "configurations" are pathways.
The other two directions play smaller roles. Functional enrichment supplies the hypothesis units, which are the curated sets across GO, KEGG, Reactome and WikiPathways in the 12,612-set corpus described by Contrastive Learning for Gene Set Enrichment Analysis Post-Processing. That paper's embedding-based grouping (gtCLIP) is used only after testing, to collapse redundant specific pathways into interpretable themes. It matters because over 80% of related pathway pairs share no genes, so overlap-based grouping would fragment the output. The energy-oriented side of procrastination scheduling, meaning idle-time aggregation as in WAGES: Workload-Aware GPU Sharing System for Energy-Efficient Serverless LLM Serving, is not part of the claim. We report CPU-hours, not joules.
Baselines
The naive reference is the practice the file documents repeatedly. We run a per-gene G×E test, take genes below a p-value threshold, and apply hypergeometric over-representation analysis against GO and KEGG. We will run it on identical data with the same gene universe and the same false discovery threshold. On fake phenotypes it should produce many pathway calls, which is the failure we target. Our method beats it if its pathway false-positive rate on fake phenotypes sits at the nominal level while the naive method's does not, and if its power on planted interactions is not much lower.
The strongest existing method is the structured penalized G×E model of Structured gene-environment interaction analysis, fit once on real data. Its selected genes are aggregated to pathways by over-representation analysis. It is the strongest competitor because it already encodes both the main-effect hierarchy and network structure. On GENEVA diabetes data it reports an Observed Occurrence Index (a selection-stability measure) of 0.69 against 0.21 to 0.47 for competitors. On TCGA melanoma it reports a survival C-statistic of 0.65, against 0.54 for marginal analysis and 0.59 to 0.64 for other penalized models. We use its published tuning procedure and give it the same total compute as our full pipeline, spent on extra stability resampling. To beat it, we must match its C-statistic when our specific pathways are used as predictors, and substantially reduce pathway calls under fake phenotypes. No reported result was found on Lacuna for any pathway-level specificity or false-positive metric for this method, so that comparison is new.
The ablation that isolates the calibration ingredient is our pipeline with fake-pathway controls only, dropping fake phenotypes. A second ablation isolates the scheduler. We replace procrastination allocation with, in turn, uniform $B_P$; sequential count-based stopping in which timed-out fits are dropped; and sequential stopping in which timed-out fits are retried at a fixed limit. Each variant gets the same CPU-hour budget. The scheduler earns its place only if it yields more correct decisions per CPU-hour than the best sequential variant, or better calibration under heavy-tailed fit times.
Why not just…?
Why not just rank-inverse-normal transform the phenotype and skip the fake controls? This is a serious threat. If transformation alone removes scale artifacts, the fake-phenotype control is redundant. The objection has two weak points. A monotone transform cannot remove heteroscedasticity that depends on genotype. It can also shrink genuine interactions. The UK Biobank analysis Quantification of the overall contribution of gene-environment interaction for obesity-related traits relied on simulated controls rather than transformation alone to separate BMI's non-specific signal from the nine traits with specific G×E. We test the alternative directly: the transformed-phenotype pipeline is a baseline in every simulation and real-data comparison. If it matches our false-positive rate and power, we will say so and recommend it.
Why not just run standard enrichment on the output of a good G×E model? Enrichment on a gene list cannot tell whether that list's interaction signal is specific. It measures annotation overlap, not interaction beyond a matched null. The strongest-baseline comparison above measures exactly this gap on fake phenotypes.
Why not just use sequential Monte Carlo stopping for p-values instead of procrastination? For pathways whose fits all finish quickly, count-based sequential stopping probably does as well as our allocator, and we expect it to. The difference appears only when refit times are heavy-tailed and dropping slow fits biases the null. We measure refit-time distributions on real data first. If the tails are light, we adopt sequential stopping and present the scheduler as unnecessary. The scheduler ablation decides this.
Why not just learn interactions with a neural network? As High-Dimensional Gene–Environment Interaction Analysis notes, prediction-based methods do not enforce the main-effect hierarchy and do not give calibrated tests. Our question is which pathways specifically interact, not how well the phenotype can be predicted.
Proposed Experiments
We will start with simulations built on real genotype matrices from GENEVA. Each simulation plants interaction in chosen pathways under three trait-generating regimes: linear with no interaction; nonlinear or heteroscedastic with no interaction; and nonlinear with true pathway interaction. The measures are pathway false-positive rate, power, and CPU-hours to finish all decisions. The decisive check is that under the second regime the naive and structured baselines produce inflated pathway calls while ours stays near nominal.
The real-data work uses three sources. On UK Biobank Quantification of the overall contribution of gene-environment interaction for obesity-related traits gives an external reference with known answers. BMI should yield few specific pathways despite its apparent 1.9% interaction variance. The nine traits reported to carry specific G×E, such as leg impedance, should yield more. A reversal of this pattern would count against the method. On GENEVA we additionally report selection stability. On TCGA melanoma, with age and stage as environments, we report the survival C-statistic of models restricted to called pathways, against the published 0.65 Structured gene-environment interaction analysis. Finally, specific pathways are grouped with gtCLIP embeddings Contrastive Learning for Gene Set Enrichment Analysis Post-Processing and checked for sign coherence of interaction effects within groups.
The project fails, and would be reported as failing, if the transformed-phenotype baseline matches our calibration, or if our specific pathways lose predictive value relative to the structured baseline.
References
- Quantification of the overall contribution of gene-environment interaction for obesity-related traits
- Identification of SNPs and Candidate Genes Associated with Growth Using GWAS and Transcriptome Analysis in Portuguese Oyster (Magallana angulata)
- Tissue-Specific Transcriptome Analysis Reveals Multiple Responses to Salt Stress in Populus euphratica Seedlings
- Structured gene‐environment interaction analysis
- Efficiency Through Procrastination: Approximately Optimal Algorithm Configuration with Runtime Guarantees
- Contrastive Learning for Gene Set Enrichment Analysis Post-Processing
- WAGES: Workload-Aware GPU Sharing System for Energy-Efficient Serverless LLM Serving
- High-Dimensional Gene–Environment Interaction Analysis