Most genetic association studies test one DNA variant against one trait at a time and then apply a harsh multiple-testing correction. This approach throws away a useful fact: many traits are correlated, and a variant that nudges several related traits at once (pleiotropy) is easier to detect if those traits are analyzed together. Methods that analyze traits jointly need a trait map, a graph saying which traits should share genetic effects. Today that map is usually built from raw phenotype correlations, which are noisy in small cohorts and mix genetic with environmental sharing. Language-model agents that read the phenotype table and the biomedical literature could plausibly draw a better map. They also create a hazard. An agent that can peek at interim association results will steer toward whatever looks significant, and the reported p-values stop meaning anything.
This note proposes a way to get the benefit while keeping valid statistics. We split the work between agents, and the split works as an information barrier. The agents that propose trait structure never see genotypes. A separate, non-language executor fits the joint model and calibrates it by permutation. Because the agents' output is a function of the phenotypes and the literature only, shuffling genotypes relative to phenotypes leaves every agent decision unchanged. A permutation test of the whole pipeline is therefore exact without rerunning the agents. The question the work answers is whether knowledge-informed trait maps buy real, replicable power over correlation-derived maps once validity is enforced this way.
Core Claim
Consider a joint penalized regression over $K$ traits, in the form used by graph-guided fused lasso. Here $y_k$ is trait $k$, $X$ is the genotype matrix, $b_k$ holds the variant effects on trait $k$, $B$ stacks all $b_k$, and $E$ is a set of trait edges with weights $w_{km}$ and signs $s_{km}$:
$$\min_{B}\ \sum_{k=1}^{K}\lVert y_k - X b_k\rVert_2^2 + \lambda\lVert B\rVert_1 + \gamma\sum_{(k,m)\in E} w_{km}\sum_j \lvert b_{jk} - s_{km}\, b_{jm}\rvert .$$
The last term pulls a variant's effects on linked traits toward each other. Our claim has two parts. The first part is statistical. If $(E, w, s)$ is computed from the phenotypes $Y$ and external text alone, then under the null that genotypes carry no signal, permuting individuals' genotype rows (after regressing out covariates), or circularly shifting variant positions, leaves $E$ fixed. The permutation null of any statistic built on the fit then accounts for every agent choice. This follows directly from the invariance and needs no modeling of the agents' behavior. The second part is empirical. A small team of role-separated, genotype-blind agents can produce trait maps that give more replicable discoveries than maps built from empirical correlations. The advantage should appear specifically at small sample sizes and where environmental correlation distorts the raw correlation matrix. If the second part fails, the first still yields a validity result: a safe design for agentic multi-trait analysis.
Where This Sits
The statistical side supplies the method and motivates the gap. Statistical Estimation of Correlated Genome Associations to a Quantitative Trait Network introduced the graph-guided fused lasso and applied it to 543 severe-asthma patients, 53 traits and 34 IL-4R variants. It recovered the known Q551R variant and two markers univariate testing missed. Its trait network is a thresholded empirical correlation graph, and the file notes the authors' warning that errors in this initial network could bias results. We keep their estimator and change only who draws $E$ and what they are allowed to see. Two related lines share the dependence on a fixed prior structure. Leveraging the genetic correlation between traits improves the detection of epistasis pools interaction signal across correlated traits. networkGWAS aggregates variants over protein-interaction neighborhoods and calibrates with a dual permutation. That permutation, circular for variants and degree-preserving for the graph, is the calibration device we reuse. networkGWAS also reports that incomplete networks erode its advantage, which is the failure mode a knowledge-informed map should address. Contaminated Chi-square Modeling constrains us: it shows that analytic p-value models break under correlation among features. This is why we calibrate empirically rather than trust nominal p-values from a joint fit.
On the agent side, MESHAgents is the closest work. It has specialist agents for anatomical regions and confounders, plus a coordinator that collects independent opinions before discussion. It selects cardiac imaging phenotypes whose disease-prediction AUC comes within −0.004 ± 0.010 of human experts across 9 diseases. It uses no genotypes and says nothing about calibrated genetic tests; the authors list genetic integration as future work. We borrow its "opinions before discussion" consensus protocol and move its output from disease classifiers to genetic priors. AutoBA and BioAgents automate bioinformatics pipelines end to end: AutoBA reports 87.5% end-to-end execution success on 40 omics tasks. Neither separates what an agent may see from what it decides, and an end-to-end agent looking at results is exactly the design our validity argument forbids. BioAgents' finding that critique loops degrade after three rounds constrains us. We cap deliberation at three rounds and give all numeric iteration, including optimizer convergence, to deterministic code. Role decomposition itself follows MACM, where proposer, verifier and executor roles lifted GPT-4 Turbo on hard MATH problems from 54.68% to 76.73%. We adopt the proposer–judge–executor split but give the judge a statistical mandate.
Research Program
We will build three agent roles and one deterministic component. A physiology agent receives trait names, units, measurement protocols and literature retrieval, and proposes candidate edges with a sign and a written rationale. A confounder agent receives the phenotype table and covariates but no genotypes. It flags edges likely driven by shared environment, medication or measurement artifacts, in the spirit of MESHAgents' non-imaging agents. It can downweight such edges using partial correlations after covariate adjustment. A judge agent reconciles the two after each has submitted independently and emits $(E, w, s)$ and a frozen hash of that output. The executor is ordinary code. It fits the joint model, checks convergence numerically, and runs the permutation null. The firewall is enforced by the file system: agent sandboxes have no read access to genotype files or to any association output.
The analysis has three strands. The first establishes calibration: whole-pipeline type I error under the permutation null, and a deliberate breach condition. The second measures power and replicability against correlation-based and genetics-based maps. The third characterizes when knowledge helps, by varying sample size and by constructing settings where environmental and genetic correlation disagree. Scalability is a known limit of fused-lasso estimators. We therefore run the joint model within gene-centred or linkage-block windows, as the original method did on a single gene. Any genome-wide screen is a separate stage whose threshold is also permutation-calibrated.
Baselines
The naive reference is univariate testing of each trait with a linear mixed model and a Bonferroni threshold of $5\times10^{-8}/K$. This is what the field does by default, and any joint method must beat it on replicated hits to justify itself.
The strongest existing method on this exact question is graph-guided fused lasso with the empirical-correlation network, in both its uniform-weight and correlation-weighted forms Statistical Estimation of Correlated Genome Associations to a Quantitative Trait Network. It is the strongest competitor because it uses the same estimator and differs from us only in the trait map. Any gain we report can then be attributed to the map. We will give it the same windows and covariates, the same permutation calibration, and the same grid budget over $\lambda$, $\gamma$ and the correlation threshold. Tuning for all methods happens by cross-validated prediction on phenotypes within the discovery split. For its power on SARP, no reported result found on Lacuna for a numeric power or false-discovery figure. The file reports only the recovery of Q551R, rs3024660 and rs3024622, so we will reproduce the baseline ourselves rather than quote a number. A second strong competitor is a sparse partial-correlation trait graph estimated by graphical lasso on covariate-adjusted phenotypes. It is also genotype-blind and so enjoys the same exact-permutation property, which makes it the cleanest statistical rival. A third is a genetic-correlation graph. Estimated on the same genotypes, it forfeits the exact-permutation guarantee, so we estimate it on a held-out genotype split and pay the sample-size cost honestly. For context on agent-selected phenotypes, the MESHAgents result of −0.004 ± 0.010 mean AUC difference versus experts shows agents can match human phenotype selection. It is not a genetic benchmark.
We run four ablations. The first is a single-prompt condition: one model instance gets all the same inputs and returns a graph. This isolates role decomposition from mere access to an LLM. The second is a no-literature condition, where agents see only the phenotype table. It isolates external knowledge from simple re-derivation of correlations. The third is the breach condition, where the physiology agent may see interim per-trait association statistics and revise the graph. It isolates the firewall itself. The fourth is a random-graph control with matched edge density. Beating the correlation-graph baseline means a statistically significant increase in replicated loci, at equal whole-pipeline type I error, in at least the small-sample regime.
Why not just…?
Why not just use the empirical correlation graph, as GFlasso already does? This is the real threat. At biobank scale, correlations among 15 to 50 traits are estimated almost perfectly, and an LLM's prior may add nothing or even add error. We therefore expect any advantage to shrink with sample size, and we will measure that directly with subsampling curves. If the agent map fails to beat the correlation map at the SARP-like scale of a few hundred people, or where environmental confounding is strong, we will report the knowledge-prior claim as negative. The contribution then rests on the validity result and the breach experiment.
Why not just run a single strong model on the whole analysis, as AutoBA does for omics pipelines? An end-to-end agent sees results and can iterate toward significance. Its choices then depend on genotypes, and no cheap permutation argument covers them. The breach ablation measures how large the resulting inflation is. The single-prompt ablation separately tests whether decomposition matters when the firewall is held fixed. The MACM and LEGO gains over single-prompt baselines, 76.73% versus 54.68% and F1 0.666 versus 0.615 LEGO, suggest decomposition may help, but those are reasoning benchmarks, not graph quality. We treat this as an open question the ablation settles.
Why not just estimate genetic correlations and build the graph from those? Genetic correlation is arguably the right quantity for a pleiotropy prior. On the same data, though, it reuses genotypes, and valid use requires sample splitting, which costs discovery power. We include this baseline with an honest split. If it beats the agent map even after paying that cost, the conclusion is that agents are not worth it for this task.
Why not just use graphical lasso and skip language models entirely? It is genotype-blind, cheap and deterministic, so it shares our calibration property. The no-literature ablation tells us whether the agents are doing anything beyond reinventing it. Only a gain of the full team over graphical lasso supports the claim that external knowledge contributes.
Proposed Experiments
The first experiment is calibration. We use real genotypes from WTCCC1 chromosome 22 and the AraGWAS SNP matrix (2,029 accessions, about 10.7 million markers AraPheno and the AraGWAS Catalog 2020). These are paired with real, permuted phenotype panels to create a global null. We run the full pipeline hundreds of times and check that the rejection rate matches nominal $\alpha$ for firewalled agents. We expect measurable inflation in the breach condition. Agent stochasticity is handled by freezing one graph per phenotype panel, which the invariance argument permits.
The second experiment tests power with known truth. We take real phenotype panels with meaningful names: the 15 hematological traits in the mouse Heterogeneous Stock data used by mvMAPIT, and a curated subset of AraPheno traits. We then plant pleiotropic variants on real genotypes whose effect-sharing pattern follows a physiologically defined grouping. We add environmental noise that correlates traits across group boundaries. This creates the regime where raw correlation misleads. Metrics are area under the precision–recall curve for recovering planted variants, following networkGWAS's evaluation style (it reports about 76% AUPRC on simulated Arabidopsis data), and power at calibrated $\alpha$. One caveat matters. The planted grouping is chosen by us from physiology, so this experiment tests whether agents recover physiological structure. It does not test whether nature follows it.
The third experiment tests real replication. We use UK Biobank blood-cell or cardiac-imaging-derived traits, following the cohort splits MESHAgents used (26,893 and 38,309 participants). Discovery runs on subsamples of 500, 2,000 and 10,000 people, with replication in a disjoint set. The outcome is the number of discovery hits that replicate at a pre-specified threshold under each trait map, all at equal permutation-calibrated error. Replication in held-out individuals avoids the contamination problem that published loci may already sit in the language model's training data. We will also rerun the SARP IL-4R analysis as a qualitative check, knowing it is contaminated.
The decisive checks are two. The firewalled team should match nominal type I error while the breach condition inflates it. The team should also beat both the correlation graph and graphical lasso on replicated hits at small sample sizes, with the gap closing as sample size grows. If the calibration holds but the power gain does not, the paper reports a negative result for knowledge priors and a positive one for the barrier design. If the breach condition shows no inflation, the firewall is less necessary than argued, and we will say so.
References
- Statistical Estimation of Correlated Genome Associations to a Quantitative Trait Network
- Leveraging the genetic correlation between traits improves the detection of epistasis in genome-wide association studies
- <tt>networkGWAS</tt> : a network-based approach to discover genetic associations
- Contaminated Chi-square Modeling and Its Application in Microarray Data Analysis
- Multi-agent Reasoning for Cardiovascular Imaging Phenotype Analysis
- An AI Agent for Fully Automated Multi‐Omic Analyses
- BioAgents: Democratizing Bioinformatics Analysis with Multi-Agent Systems
- MACM: Utilizing a Multi-Agent System for Condition Mining in Solving Complex Mathematical Problems
- LEGO: A Multi-agent Collaborative Framework with Role-playing and Iterative Feedback for Causality Explanation Generation
- AraPheno and the AraGWAS Catalog 2020: a major database update including RNA-Seq and knockout mutation data for Arabidopsis thaliana