A polygenic risk score adds up thousands of small genetic effects into one number that predicts disease risk. In clinics and in lower-resource settings, the score is often not computed from a genome-wide chip. Instead it runs on a small targeted panel: a manufactured assay that can measure only a limited number of genetic variants, and only variants that are physically compatible with one another on the same assay. The PRS-15 venous thromboembolism study is an example, with a 15-variant score deployed in Chinese orthopedic patients. Choosing which variants go on such a panel forces a trade-off. The variants that predict best in one ancestry group are often not the ones that predict best in another, because allele frequencies and correlation patterns between nearby variants differ. This portability decay is documented in Polygenic risk score portability for common diseases across genetically diverse populations and On cross-ancestry cancer polygenic risk scores.
This proposal treats panel design as a two-objective problem: accuracy in ancestry A against accuracy in ancestry B, under a fixed panel capacity and assay compatibility rules. It also asks for something no current tool provides. The designer should get a continuous family of optimal panels, in which neighboring trade-off points differ by only a handful of variants. A laboratory serving a shifting patient mix, or re-manufacturing a panel as new association data arrive, could then move along the trade-off without redesigning the assay from scratch. In short, we want a front of panels that is smoothly navigable in the space of physical assays, not just a scatter of good panels.
Core Claim
Without the panel constraint, cross-ancestry score weighting is close to a convex problem. A weighted sum of the two objectives traces the whole trade-off, and the optimal weights change continuously as the weighting changes. Path-connectedness of optima is then nearly automatic. This is the regime covered by Xu's quasiconvexity results, which prove connected and path-connected efficient sets under convex feasible sets. It is also why "trace a Pareto path of polygenic weights" is not by itself a contribution.
The panel constraint is what breaks this. Choosing a subset of variants is combinatorial. Compatibility rules forbid certain pairs, for example variants too close in the genome to be assayed in the same reaction. Together they make the feasible set non-convex. Makarov et al. show this is exactly the condition under which efficient sets fracture into disconnected pieces.
Our claim is empirical and methodological. First, under realistic capacity and compatibility constraints, independently optimized panels at adjacent trade-off weights differ in a large fraction of their variants. We call this panel churn. Second, coupling the panels along a parameterized path can cut that churn sharply while losing little of the two-ancestry accuracy front. The coupling works in a relaxed selection space with the assay rules built into the optimization rather than applied afterward. The new object is a navigable front of manufacturable panels. We are not proposing a new score estimator.
Research Program
The building block is a summary-statistic accuracy model. For a score with weight vector $w$ over candidate variants, the expected predictive $R^2$ in ancestry $a$ is approximately
$$R^2_a(w) = \frac{(w^\top c_a)^2}{w^\top \Sigma_a w},$$
where $c_a$ is the vector of marginal variant–trait covariances estimated from that ancestry's association study, and $\Sigma_a$ is that ancestry's correlation matrix between variants, taken from a reference panel. Candidate variants and starting weights come from existing estimators: LDpred-style posterior means, the sparse spike-and-slab LASSO, and functionally informed shrinkage from LDpred-funct. We reuse these estimators. The contribution is the panel layer that sits on top of them.
The panel is represented by relaxed selection variables $s \in [0,1]^M$, so the score is $w \circ s$. Two pieces are borrowed from manufacturing-aware design. To drive selections toward 0 or 1 without hand-tuned schedules, we adapt the neighborhood product operator of the normalized field product method. In topology optimization that operator yields near-binary material layouts with a built-in minimum feature size. Here the "neighborhood" is the set of variants within the assay's incompatibility window, so the product term both binarizes and discourages selecting forbidden neighbors together. For hard rules, namely panel capacity and per-reaction plex limits, we use the DiffOPC approach: gradient steps are scaled toward zero as the design nears a forbidden configuration, and a straight-through estimator handles rounding. The point of both choices is that every panel the optimizer returns already satisfies the assay rules.
The continuous-deformation component follows MM-BezEA, which represents an approximation set as Bézier curves rather than a point cloud, so a decision-maker can slide along it without jumps. We parameterize selection logits as a Bézier curve $s(t)$ over trade-off weight $t \in [0,1]$ and optimize
$$\min_{s(\cdot)} \int_0^1 \Big[-\big(t\,R^2_A + (1-t)\,R^2_B\big)(s(t)) \Big]\,dt \;+\; \lambda \int_0^1 \Big\|\tfrac{d}{dt}\,\mathrm{bin}(s(t))\Big\|_1 dt,$$
where $\mathrm{bin}$ is the rounded panel and $\lambda$ prices panel churn, the number of variants swapped per unit of trade-off. We do not rely on linear scalarization to reach non-convex parts of the front: the curve is evaluated by hypervolume over its rounded panels, and a Chebyshev-style variant of the scalarization is included. The relaxed objective itself is approximately path-connected, but the binarized front is not, and we measure that gap directly. Multiple curves can be seeded where valley-detection clustering finds genuinely separate panel families, as MM-BezEA does with Hill-Valley clustering.
The phase-field path-connectedness work enforces connectedness inside a single physical design. Our use of connectedness is the other kind, between solutions. That paper therefore motivates the distinction rather than supplying the method. The sufficient conditions of Borum et al. rely on scale invariance, which subset selection does not have. We use them only as contrast.
The data are UK Biobank. Its large European-ancestry majority is used as the source population, matching the setup of LDpred-funct. Its smaller African- and South-Asian-ancestry participants serve as the second objective and as held-out validation. Traits are lipid levels, motivated by Natarajan et al. showing polygenic burden dominates severe hypercholesterolemia, plus height and type 2 diabetes. Panel capacities run from 15 to 1,000 variants. Compatibility windows are a swept parameter rather than a claim about any single assay vendor.
Baselines
The naive reference is the deployment default: build the score in the European source data with pruning and thresholding, keep the top $K$ compatible variants, and use that one panel for everyone. It sets the floor for the second-ancestry accuracy objective.
The strongest existing competitor has two parts. The weighting comes from LDpred-funct, the best genome-wide estimator in the file. Panel design is then done independently at each trade-off weight with greedy compatible forward selection, followed by post-hoc removal of any rule violations. This is the most likely thing a capable lab would do today. LDpred-funct reports an average prediction $R^2$ of 0.144 across 21 UK Biobank traits, a 4.6% relative gain over SBayesR, in genome-wide European validation (LDpred-funct). That figure is an unconstrained ceiling, not a panel result. We found no reported result on Lacuna for panel-constrained, cross-ancestry score accuracy or for panel churn along a trade-off, so both quantities will be measured here for the first time.
The multi-objective search baseline is NSGA-II-style evolution directly on binary panels with a repair operator. It gets the same number of accuracy-model evaluations as our method.
All methods share the candidate variants, reference correlation matrices, held-out individuals, and a matched tuning budget measured in accuracy-model evaluations and wall-clock time.
Two ablations isolate the ingredients. The first removes path coupling ($\lambda = 0$) and optimizes each trade-off weight separately with our relaxed method. Any churn reduction must come from coupling, not from the relaxation. The second replaces built-in constraints with post-hoc filtering, which tests whether proactive integration matters in genomics the way it does in photonics.
Beating the strongest baseline means three things together. Churn between adjacent panels must fall by at least half. The two-ancestry hypervolume, computed on held-out individuals, must stay within a few percent. And every panel must be rule-compliant without repair.
Why not just…?
Why not just use greedy forward selection, which already gives nested panels? Greedy selection does give nested panels across budgets $K$ at a fixed trade-off weight. That is a real strength, and nesting along the budget axis is reported as a secondary result. But greedy selection says nothing about stability across the ancestry trade-off. As the trade-off weight moves, the greedy order can reshuffle entirely. This is a genuine threat to the idea: if greedy churn along $t$ turns out to be small, the problem we target is mild. The decisive check below measures exactly that before any claims are made.
Why not just run a weighted-sum lasso path and drop incompatible variants afterward? On the relaxed problem a weighted sum is adequate, as argued above. The failure comes from the post-hoc step. Dropping variants after optimization removes correlated tagging partners and shifts what the remaining weights mean. This is the genomic analogue of the corruption that post-processing causes in nanophotonic resonances (Neuroshaper) and in mask cost (DiffOPC). Whether the analogy holds for scores is an open empirical question. The second ablation answers it directly by comparing held-out $R^2$ loss under post-hoc filtering against built-in constraints.
Why not just genotype on a cheap genome-wide array and impute? Where genome-wide arrays are affordable and imputation panels match the population, this is the better answer, and we say so. Our setting is small targeted assays like the PRS-15 panel. In that setting imputation is unavailable, and reference panels for the target ancestry are thin, which is also where portability already fails. The genome-wide LDpred-funct score is reported alongside every panel as the ceiling, so readers can see how much accuracy the panel constraint costs.
Why not just run multi-trait methods like wMT-SBLUP or PleioPred? wMT-SBLUP and PleioPred borrow strength across correlated traits to produce one static predictor. They are complementary, not competing. Their weights can be the input $w$ to our panel layer. Neither addresses panel capacity, assay rules, or movement along a trade-off.
Proposed Experiments
The first experiment is the existence check. Using the strongest baseline, we compute rounded panels at 21 evenly spaced trade-off weights for each trait and capacity. We report the Jaccard distance between adjacent panels and count how many disconnected panel families valley-detection clustering finds. If adjacent churn is below about 10% across settings, the fragmentation premise fails. In that case we redirect the work to the budget-nesting question.
The main experiment compares all baselines and ablations on four measures, all computed on held-out individuals: two-ancestry hypervolume, mean adjacent churn, the MM-BezEA smoothness score, and the fraction of rule-compliant panels. It sweeps capacity and compatibility-window width, so we can see where coupling helps most. We expect the benefit to be largest at small capacities and wide windows, where the feasible set is most fragmented.
A robustness experiment re-estimates association statistics on bootstrap halves of the source data. It asks whether navigable fronts also change less under data refresh, which is the practical re-manufacturing scenario.
The result that would most clearly settle the claim is a matched-hypervolume comparison: at equal held-out two-ancestry accuracy, does the coupled front require substantially fewer variant swaps to move from a European-optimized panel to an African-ancestry-optimized one than the independently optimized front does?
References
- Predictive Performance and Clinical Application of PRS-15 for Venous Thromboembolism Risk in Chinese Orthopedic Patients: A Multicenter Prospective Case-Control Study
- Polygenic risk score portability for common diseases across genetically diverse populations
- On cross-ancestry cancer polygenic risk scores
- Connectedness and Path Connectedness of Weak Efficient Solution Sets of Vector Optimization Problems via Nonlinear Scalarization Methods
- Arcwise Connectedness of Closed Efficient Point Sets
- Sparse polygenic risk score inference with the spike-and-slab LASSO
- Incorporating functional priors improves polygenic prediction accuracy in UK Biobank and 23andMe data sets
- Normalized Field Product Approach: A Parameter‐Free Density Evaluation Method for Close‐To‐Binary Solutions in Topology Optimization With Embedded Length Scale
- Differentiable Edge-based OPC
- Obtaining Smoothly Navigable Approximation Sets in Bi-objective Multi-modal Optimization
- Keeping it together: A phase-field version of path-connectedness and its implementation
- Sufficient Conditions for a Path-Connected Set of Local Solutions to an Optimal Control Problem
- Deep-coverage whole genome sequences and blood lipids among 16,324 individuals
- Shaping freeform nanophotonic devices with geometric neural parameterization
- Improving genetic prediction by leveraging genetic correlations among human diseases and traits
- Joint modeling of genetically correlated diseases and functional annotations increases accuracy of polygenic risk prediction