Biobanks increasingly run polygenic risk scores (PRS: a weighted sum of a person's genetic variants that estimates their risk for a trait) as a shared computing service. Many traits, several ancestry groups, and a dozen competing methods all draw on one cluster budget. The methods differ in cost by more than two orders of magnitude. The PRS-hub benchmark measured 0.84 CPU hours per trait for simple pruning-and-thresholding and about 359 CPU hours per trait for SBayesR (PRS-hub). The expensive, ancestry-aware methods are often what rescues accuracy in under-represented groups. GAUDI, for example, improved white-blood-cell-count prediction in African Americans by 63.8% over PRS-CSx (GAUDI). So when compute is finite, the order in which jobs run and the method chosen for each (trait, ancestry) pair decide who gets a good score first, and who ever gets one.
We propose to treat this as a fairness–throughput scheduling problem in the sense of the queueing literature. In that literature, "stretch" means a job's waiting time divided by its own processing time. It was introduced so that small and large jobs are judged on a common scale. Our analogue for genomics is the accuracy shortfall of each (trait, ancestry) group, measured relative to the best accuracy that group's own data can reach. A scheduler then chooses which method to run for which group, and when. It uses cheap predictions of each method's accuracy gain and cost, and it keeps a round-robin safety reserve so that bad predictions cannot starve any group. The claim to test is specific. Under realistic budgets, part of the minority-ancestry accuracy gap in a PRS service is caused by scheduling. A shortfall-aware scheduler can shrink the worst group's shortfall substantially while losing only a few percent of population-weighted accuracy delivered per CPU hour.
Core Claim and Objective
Let $g$ index a (trait, ancestry) group. Let $R^*_g$ be the best held-out $R^2$ (the fraction of trait variance the score explains) that any method in the portfolio achieves for $g$, determined in hindsight. Let $R_g(\tau)$ be the best $R^2$ delivered for $g$ by compute time $\tau$. The shortfall $s_g(\tau) = 1 - R_g(\tau)/R^*_g$ lies between 0 and 1. It plays the role of stretch: it normalizes by what is attainable for that group, not by the raw $R^2$ scale, so a European group with $R^* = 0.15$ and an African group with $R^* = 0.02$ count equally. The scheduler minimizes
$$J_k = \int_0^{B} \Big( \sum_g s_g(\tau)^k \Big)^{1/k} d\tau,$$
where $B$ is the total compute budget and $k \ge 1$ is the trade-off knob. With $k=1$ the objective favors total improvement. As $k \to \infty$ it minimizes the worst group's shortfall over time. This follows the way higher-order flow-time norms penalize extreme delays (Moseley et al.; batra et al.). The knob $k$ plays the same role as the tunable $\Delta$ in Fair Transaction Processing For Multi-Tenant Databases: it dials between strict fairness and maximum throughput. Throughput is measured separately as population-weighted $R^2$ delivered per CPU hour. The result of interest is the frontier traced as $k$ varies.
Relation to Prior Work
Two systems papers supply the method. Competitive Fair Scheduling with Predictions minimizes max-stretch using predicted job sizes. It sorts jobs into short and long by a geometric-mean threshold, and its $RG^+$ variant pairs this with round robin so that worst-case behaviour stays bounded whatever the prediction error. We borrow the two-class prediction idea and the round-robin reserve. We differ in that our jobs carry a predicted value (accuracy gain for a specific group) as well as a size, and the scheduler also chooses among alternative jobs for the same group. The paper's guarantees are for a single machine. We make no theoretical claim for our multi-core cluster setting and evaluate empirically. DRFT (Fair Transaction Processing) supplies a second ingredient: accounting for cost that depends on the schedule. A DRFT transaction's cost depends on locks held by others. In our setting, a method's cost depends on whether the linkage-disequilibrium matrices it needs have already been computed by an earlier job. (Linkage disequilibrium is the correlation between nearby variants; many PRS methods need these correlation matrices as input.) Work on slowdown in single-server queues (Aalto and Scully) motivates normalizing by intrinsic size rather than using absolute quantities.
On the genomics side, PRS-hub is the closest existing work, and it motivates the gap. It runs 13 methods over 36 UK Biobank traits for 377,490 European and 8,214 African individuals through a static Cromwell/WDL workflow engine. It reports that LDpred2-meta beat PRS-CSx and X-Wing across ancestries, that LD reference panels plateau beyond about 1,000 samples, and that a curated set of 1.1 million SNPs often beats 6.6 million. PRS-hub runs everything and reports the table. It never asks what to run first, or what to skip, under a budget. Its plateau and SNP-filtering findings matter here because they turn panel size and SNP set into cheap actions the scheduler can take.
The methods X-Wing, LOG-TRAM, and GAUDI are the expensive jobs whose value to minority groups we try to predict. We do not modify any of them. The pooled-versus-meta-analysis comparison of Dias et al. is out of scope. That is a question about statistical power under data-sharing rules, not about compute. Our proposal is deliberately about the compute layer, where a real shared-resource trade-off exists.
Research Program
The first deliverable is a ground-truth table. For every (trait, ancestry, method, SNP set, panel size) cell we record held-out $R^2$ (or odds ratio per standard deviation for diseases), CPU hours, peak memory, and which LD artifacts the cell produces or consumes. We reproduce the PRS-hub portfolio on UK Biobank European and African samples and add an East Asian arm from Biobank Japan summary statistics, following the X-Wing and LOG-TRAM setups. Where access permits, we add an admixed African American arm from the Women's Health Initiative so that GAUDI enters the portfolio.
The second deliverable is a replay simulator. Trait requests arrive over time on a cluster of fixed size. Each group starts with a P+T score, which gives a nonzero floor for $s_g$. A policy then repeatedly picks the next (group, method, configuration) job.
The third deliverable is the predictor. It forecasts each job's $R^2$ gain and cost from features that are cheap to compute before running the job: GWAS sample size per ancestry, SNP-heritability and cross-ancestry genetic-correlation estimates, and the P+T $R^2$ already observed. Following $RG^+$, the scheduler discretizes predicted gain-per-cost into two classes and runs the highest predicted shortfall-reduction-per-cost job from the favoured class. It keeps a fraction $\epsilon$ of cores cycling round-robin over groups whose shortfall has not moved.
The analysis traces the $k$-indexed frontier of worst-group shortfall against throughput and reports where on it each baseline sits.
Baselines
The naive reference point is the PRS-hub-style static workflow (PRS-hub). It runs every method for every group in submission order, first come first served, and deliveries are recorded as jobs finish. It represents current practice, and any scheduler that cannot beat it at intermediate budgets is pointless.
The strongest competitor on efficiency is a throughput-greedy knapsack. It uses the same predictor and the same feature set, and it runs jobs in order of predicted population-weighted $R^2$ gain per CPU hour. This isolates exactly what fairness costs. Beating it means matching it on throughput within 5% while lowering time-integrated worst-group shortfall. We borrow the 5% threshold as a target from DRFT's reported overhead of under 5% (Fair Transaction Processing), which comes from a different domain.
The strongest competitor on fairness is static per-ancestry quotas (fair-share partitions), plus plain round robin over groups. DRFT reports that static quotas waste capacity when a tenant goes idle. $RG^+$ reportedly beat round robin and Greedy-with-Rounding on cluster traces (Competitive Fair Scheduling), but the file gives no numeric margin.
For the headline question, no reported result was found on Lacuna for fairness-aware or budget-constrained scheduling of multi-ancestry PRS pipelines. The only relevant reported ordering is PRS-hub's finding that LDpred2-meta outperforms PRS-CSx and X-Wing, stated without a numeric margin in the file. We therefore also include a "single best method" baseline that runs LDpred2-meta for every group.
All policies share the same simulator, cluster size, arrival sequence, and ground-truth table. Every learned policy uses the same predictor, trained on a held-out split of traits with an identical hyperparameter search budget.
Three ablations isolate the proposal's ingredients. An oracle scheduler with true gains and costs gives the achievable upper bound. A non-clairvoyant version, with no gain predictions and only the round-robin reserve, measures what prediction adds. A version that uses raw $R^2$ instead of normalized shortfall tests whether the stretch-style normalization is the key ingredient. We expect that version to neglect low-heritability, small-sample groups. We also remove the round-robin reserve to measure how much the safety fallback protects against predictor error. For this we deliberately corrupt the predictions at increasing noise levels, mirroring the error parameter $\eta$ in the $RG$ analysis.
Why not just…?
Why not just run LDpred2-meta everywhere? This is the most serious threat. PRS-hub found it beat PRS-CSx and X-Wing in both European and African cohorts. If it is also cheap relative to the budget and within a few percent of $R^*_g$ for every group, the scheduling problem mostly disappears. The ground-truth table settles this directly. We will report the fraction of groups where LDpred2-meta is within 5% of the best, and the budget at which "run it everywhere" becomes feasible. If that fraction is near one, our claim shrinks to two cases: budgets below that point, and portfolios that include individual-level methods such as GAUDI, whose admixed-population gains PRS-CSx does not match (GAUDI). We will say so plainly.
Why not just use cluster fair-share quotas per ancestry? Quotas equalize CPU, not accuracy. A small African cohort and a large European cohort need very different compute to approach their own $R^*_g$. DRFT's evidence that static quotas leave resources idle (Fair Transaction Processing) suggests throughput will also suffer. The quota baseline above measures both effects on the same workload.
Why not just fit a per-ancestry ensemble and be done? PRS-hub's ensembles gained 5–10% in odds ratio per standard deviation. But an ensemble needs its component methods to be run first, so it does not avoid the allocation problem. In our design it is simply one more job in the portfolio, and the scheduler decides when it is worth its cost.
Why not just greedy by value per CPU hour? Greedy is the throughput end of our frontier, and it is included as a baseline. If its worst-group shortfall is already close to the $k \to \infty$ policy, the fairness mechanism adds nothing and the frontier will show that.
Decisive Checks
The proposal succeeds if three things hold together. First, at budgets between 10% and 50% of the full-portfolio cost, the shortfall-aware scheduler at large $k$ reduces time-integrated worst-group shortfall relative to throughput-greedy. Second, it stays within 5% of greedy's population-weighted $R^2$ per CPU hour. Third, the raw-$R^2$ ablation loses most of that gain. It fails if LDpred2-meta-everywhere is near-optimal for all groups at affordable budgets, or if predictions from cheap features are so poor that the non-clairvoyant variant matches the predictive one. A secondary check exercises the LD-sharing cost model on held-out traits. It asks whether ordering jobs to reuse LD matrices, together with the PRS-hub findings on the 1,000-sample panel plateau and the 1.1M-SNP set, measurably raises throughput without raising shortfall.
References
- Benchmarking PRS methods for risk prediction of 36 complex traits in UK Biobank by establishment of a large-scale PRS computation platform PRS-hub
- Improving polygenic risk prediction in admixed populations by explicitly modeling ancestral-differential effects via GAUDI
- The Complexity of Scheduling for p-norms of Flow and Stretch
- Constant Factor Approximation Algorithm for Weighted Flow Time on a Single Machine in Pseudo-Polynomial Time
- Fair Transaction Processing For Multi-Tenant Databases
- Competitive Fair Scheduling with Predictions
- Minimizing the mean slowdown in the M/G/1 queue
- Quantifying portable genetic effects and improving cross-ancestry genetic prediction with GWAS summary statistics
- Leveraging the local genetic structure for trans-ancestry association mapping
- Evaluating multi-ancestry genome-wide association methods: Statistical power, population structure, and practical implications