A polygenic risk score (PRS) estimates a person's genetic liability for a trait by adding up thousands of small effects from common DNA variants. The newest personalized versions do this by keeping several separate scores, called "experts," for example one trained on men and one on women, or one per ancestry group. A small model then decides how much to trust each expert for a given person. This is the same design that multi-agent systems research calls static role assignment. Each expert's role is fixed in advance by an identity label, and those systems spent a decade learning that such fixed roles break when the situation that actually matters is not the label. We propose to apply that field's main fix to polygenic prediction. Experts become shared, learnable modules whose roles are discovered from data. Each person is routed among them according to their current physiological and treatment state, not only by who they are demographically.
The clearest evidence that this matters comes from the closest existing system. MoEPRS mixes sex- and ancestry-stratified PRS experts using a linear gate over age, sex and ten ancestry principal components. For LDL cholesterol, its gate's behaviour with age was traced to statin prescriptions masking genetic risk in older participants. The relevant partition of people was therefore a treatment state. The model had no expert for that state and recovered it only indirectly, through age. In short, we ask whether polygenic models should discover their own subgroup structure from state, and when explicit interaction terms already suffice.
Core Claim
The claim is narrow. Suppose the population splits along a state that is unmeasured or only partly measured, such as medication, metabolic status, or disease stage, and that state changes how genetic effects show up in the trait. Then a mixture whose experts are learned jointly and routed on state should beat two alternatives: mixtures whose experts are pre-assigned by demographic label, and simple PRS-by-covariate interaction models. When the modifying state is fully measured and enters roughly linearly, we expect no advantage. In that case explicit interactions should tie, and the method is not worth its complexity. The proposal is designed to tell these two regimes apart rather than to assume the first.
How the Two Literatures Fit Together
The multi-agent literature motivates the framing and supplies the method. Early work showed that roles reduce conflict but fail when local constraints shift (Cammarata et al.). Bayesian role discovery learned roles from data but then froze them for the whole mission (Wilson et al.). Freezing a learned role is exactly what label-stratified PRS experts do. ADMN resolves the tension between full sharing and full specialization by giving all agents a shared bank of modules and each agent a router that picks modules from its current observation. It adds a mutual-information penalty so that different routes produce genuinely different behaviour rather than collapsing to the same modules. ACORM learns roles as behaviour patterns with a contrastive loss. Its known weakness is a fixed number of K-means clusters. We borrow two things from these papers: shared modules with individual routing, and an anti-collapse regularizer. We do not use any reinforcement learning. There are no trajectories or rewards here, and treating each person as an "agent" would add nothing.
The genetics literature supplies the inputs and the constraints. MPRS-NetWAS builds module-level scores from protein-interaction networks that track Alzheimer's stages. QPRS builds quantile-specific scores for skewed traits. The NETRIN1 pathway score of Barbu et al. is another example of a pathway-restricted score. All three still combine their sub-scores with fixed weights that are the same for everyone, which is the static-role assumption at the level of pathways. The causal transfer-learning work on treatment response (Polygenic prediction of treatment efficacy) supports the idea that treatment changes how genetic risk is expressed.
The main constraint comes from MoEPRS's own authors. They note that adding parameters risks overfitting on small minority validation sets. DeepPRS shows that nonlinear genome-wide models can work, but only with large individual-level data and heavy compute. Our design therefore keeps the learned part small.
Research Program
Each person $i$ has a fixed bank of $S$ precomputed scores $s_i \in \mathbb{R}^S$ and a state vector $z_i$. The bank includes a genome-wide LDpred-type score, network module scores in the MPRS-NetWAS style, pathway scores, and QPRS quantile scores. The state vector contains age, sex, ancestry principal components, baseline biomarkers, medication records and, where available, disease stage. The model is
$$\hat y_i = c(x_i) + \sum_{k=1}^{K} g_k(z_i)\, w_k^\top s_i,$$
where $c(x_i)$ is the usual covariate adjustment, $w_k \in \mathbb{R}^S$ is the weight vector defining expert $k$, and $g_k(z_i)$ is a softmax router giving person $i$'s weight on expert $k$. The router is a small multilayer network. The experts $w_k$ are learned jointly with the router rather than pre-trained on demographic strata.
Because experts are combinations over a few dozen scores, not over $10^6$ variants, the total parameter count is about $KS$ plus the router. This is what makes training feasible on biobank subsets and on small minority cohorts. The variant-level estimation stays where it already works well, in summary-statistics methods.
Training adds a role-diversity term in the spirit of ADMN. It rewards high mutual information between the router's assignment and each person's residual phenotype under the base score, which keeps the experts from collapsing into copies of one another. It also adds a sparsity prior on routes so that most people get a dominant expert, which keeps assignments interpretable. To avoid ACORM's fixed-K problem, we choose $K$ by held-out likelihood with experts allowed to merge.
The analysis has two parts. First, for each discovered role, we report which observed states predict assignment to it. Second, we test whether roles align with states held out from the router, such as statin use. A role that lines up with a withheld state is evidence that the model has found the modifying state rather than merely fitting noise.
Baselines
The do-nothing reference is a single genome-wide PRS plus standard covariates in a linear model. We report incremental $R^2$, or incremental AUC for binary traits, over a covariates-only model.
The strongest existing competitor is MoEPRS itself. It is the only published PRS method with individual-level routing over experts, and its LDL finding is the phenomenon we target. We run it on the same UK Biobank and CARTaGENE splits, with the same score bank available as inputs to its experts and the same hyperparameter search budget. The file reports no numeric accuracy for MoEPRS. The only reported results are qualitative: 100% gate weight on sex-matched experts for testosterone, and the age-related LDL accuracy drop. We will therefore reproduce it ourselves. Beating it means a statistically significant gain in held-out incremental $R^2$ on LDL and triglycerides, together with equal or better calibration within ancestry and age strata.
For binary outcomes we include DeepPRS as the high-capacity nonlinear reference. The file reports a genotype-only AUC of 0.7245 for Alzheimer's disease in UK Biobank and 0.8624 with clinical covariates (DeepPRS). For Crohn's disease, PleioPred-anno's reported prediction correlation of 0.307 against LDpred's 0.247 (PleioPred) gives the bar for the score bank's genome-wide component. We expect our method to inherit that component, not to improve it.
Three ablations isolate the key ingredient. The first keeps our router but fixes the experts to label strata, which tests whether learning the roles matters. The second keeps learned experts but restricts the router to MoEPRS's demographic inputs, which tests whether routing on state matters. The third removes the diversity term, which tests whether the anti-collapse regularizer matters.
Why not just…?
Why not just add PRS-by-statin and PRS-by-age interaction terms to a linear model? This is the most serious threat. When the modifying state is measured and acts roughly linearly, it should match us, and it is far simpler. We will run it with every state variable available to our router, and we will report a tie as a negative result for that regime. Our claim survives only if we beat it in two settings. The first is the withheld-state experiments, where the modifier is not handed to either model. The second is traits where the modifiers interact with each other, such as medication combined with metabolic status. Pre-specified interactions cannot cover combinations nobody wrote down.
Why not just give MoEPRS's linear gate more covariates? This addresses the routing half of the idea but keeps experts tied to labels. Ablation one tests exactly this. If label experts with a rich router match learned experts, then role discovery adds nothing, and the honest recommendation is to extend MoEPRS.
Why not just fit gradient-boosted trees on the score bank plus state? Trees can capture state-dependent effects. However, they do not produce a small set of reusable risk profiles that can be inspected, and they tend to overfit small minority cohorts, the regime MoEPRS's authors flag as risky. We include tree models as a baseline matched on tuning budget. If they win on accuracy, we will say so and argue only for interpretability.
Proposed Experiments
The first check is semi-synthetic. On real UK Biobank genotypes we simulate phenotypes in which a hidden binary state, present in about 30% of people, scales down the effects of one module score, mimicking treatment masking. We also include a QPRS-style variance-QTL (variants that change a trait's spread rather than its average) case. Success means the learned roles recover the hidden state with high adjusted Rand index. They should also beat label-expert MoEPRS and linear interactions on held-out mean squared error, while tying the oracle model that is given the hidden state.
The second check uses LDL cholesterol in UK Biobank with statin status deliberately withheld from the router. We then ask two questions. Do the discovered roles separate statin users from non-users? Does held-out incremental $R^2$ approach that of an interaction model that is given statin status?
The third check covers testosterone as a sanity case, where roles should reduce to sex, and triglycerides as a skewed trait. Both are evaluated in UK Biobank, with CARTaGENE as an external cohort for transfer and minority-ancestry calibration.
As a secondary test, we apply the model to the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, using MPRS-NetWAS module scores and clinical stage as state. This tests whether module weights should shift across the five diagnostic stages that cohort defines.
The overall result counts as positive only if the semi-synthetic recovery and the withheld-statin LDL test both succeed. If explicit interactions tie us everywhere the state is observed, we will conclude that learned roles help only when the modifier is unmeasured, and report that scope.
References
- Personalized polygenic risk prediction and assessment with a Mixture-of-Experts framework
- Strategies of Cooperation in Distributed Problem Solving
- Bayesian Policy Search for Multi-Agent Role Discovery
- ADMN: Agent-Driven Modular Network for Dynamic Parameter Sharing in Cooperative Multi-Agent Reinforcement Learning
- Attention-Guided Contrastive Role Representations for Multi-agent Reinforcement Learning
- A Module-Level Polygenic Risk Score-Based NetWAS Framework for Identifying AD Genetic Modules Mediated by Amygdala: An ADNI Study
- Enhancing polygenic risk prediction by modeling quantile-specific genetic effects
- 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 prediction of treatment efficacy with causal transfer learning
- A Deep Learning-based Genome-wide Polygenic Risk Score for Common Diseases Identifies Individuals with Risk
- Joint modeling of genetically correlated diseases and functional annotations increases accuracy of polygenic risk prediction