Most genetic risk models are fit on people of European ancestry. When applied to other populations they lose accuracy, largely because the markers they rely on are proxies for causal variants, and the proxy relationship (linkage disequilibrium, or LD: the correlation between nearby variants inherited together) differs between ancestries. We propose building a sparse risk model for a small, underrepresented target cohort in a specific way. The fitting algorithm may only use variants already chosen in large source cohorts, and it may add a new variant only when the target data show, beyond what noise could produce, that the source variants are insufficient. The algorithm we adapt, lazy Frank-Wolfe, already works like this for computational reasons. It keeps a "cache" of past solutions, checks the cache first, and falls back to an expensive global search only when the cache fails.
Our proposal changes what that fallback threshold means. It becomes a significance level. A cache hit then says a locus's European signal transfers to the target. A certified miss says the target needs its own variant there, and it names that variant. The same procedure yields three things: a portable sparse polygenic score, a genome-wide map of which loci transfer and which do not, and the speed that caching normally buys.
Core Claim
In Lazifying Conditional Gradient Algorithms, a weak separation oracle scans a cache of previously used vertices for one that makes enough progress. When it finds none (a "negative call"), it returns a dual certificate that the current point is within an accuracy $\Phi_t$ of optimal. Our claim is that in cross-ancestry sparse regression, $\Phi_t$ should not follow the optimizer's halving schedule. It should be set to the sampling-noise level of the target cohort.
Two consequences follow. First, a negative call becomes a hypothesis test. It asserts that no variant outside the source-derived cache improves the target fit by more than chance at level $\alpha$. Second, restricting the small target cohort to this cache is a structured form of transfer. It should cut false selection of ancestry-specific tag variants, while still admitting target-specific causal proxies when the evidence is strong. What is new is not caching in genomics for speed. It is the cache-miss event, reinterpreted as a calibrated, locus-resolved statement about portability.
Concretely, for ancestry $e$ with in-sample LD matrix $R_e$ (estimated from an ancestry-matched panel such as TOP-LD) and standardized marginal effects $b_e$ from its summary statistics, we fit
$$\min_{\|\beta\|_1 \le \tau} \; f_e(\beta) = \tfrac12 \beta^\top R_e \beta - \beta^\top b_e.$$
Here $\beta$ is the vector of joint variant effects and $\tau$ is the sparsity budget. The vertices of this feasible region are single variants, $\pm \tau e_j$. The gradient entry $g_j = (R_e\beta)_j - b_{e,j}$ is the residual association of variant $j$ once the current model is accounted for.
The cache $C$ is seeded with the supports of fits in large source cohorts. Variants outside the cache are admitted only when their residual association beats a noise threshold:
$$\max_{j \notin C} |g_j| > \Phi_\alpha, \qquad \Phi_\alpha = \text{the } (1-\alpha) \text{ quantile of } \max_{j\notin C}|\xi_j|,\;\; \xi \sim \mathcal{N}(0, R_e/n_e),$$
where $n_e$ is the target sample size. The Gaussian null holds because summary-statistic noise is approximately $\mathcal{N}(0, R_e/n_e)$. The threshold is computed per LD block by simulation, which is cheap because blocks are small.
Research Program
The first component is the estimator. We will implement source-seeded lazy Frank-Wolfe on top of the lazy and active-set machinery in FrankWolfe.jl, with LD matrices handled block-wise. The computational point is real but secondary. Checking a cached variant costs work proportional to the current support size. A full scan instead touches every variant's LD row, which matters with sequencing-based panels: TOP-LD indexes 146.5 million European autosomal variants. We expect hit rates below the "often exceeds 90%" that Braun et al. report for deterministic benchmarks. How far below is itself a measurement.
The second component is the portability map. For each LD block and target ancestry, we record three outcomes: all hits (the source signal transfers), certified admissions (a target-specific variant is needed), or no signal. We will then check whether the map agrees with independent evidence. That evidence includes LD divergence between the EUR and AFR panels in TOP-LD, and the local genetic correlation estimates of LOG-TRAM. LOG-TRAM partitions the genome into 1 Mb segments and borrows strength between ancestries only where local architecture is shared. Our map is a sparse, variant-level counterpart: it names the replacement variant rather than a segment-level correlation.
The third component is a downstream use in causal inference. Mendelian randomization uses genetic variants as natural experiments to estimate the effect of an exposure on a disease. For it, we will take the cache-hit variants as candidate cross-ancestry instruments and the certified admissions as ancestry-specific ones. We will then test whether this split reduces bias in the semi-synthetic All of Us GLP1R setting of Miao et al..
Prior work plays three roles here. Contrastive Losses and Solution Caching for Predict-and-Optimize and LAVA (Solver-Free Decision-Focused Learning) supply the evidence that inner approximations built from past solutions lose little quality. Mandi et al. call the exact solver only 5% of the time. Both assume a fixed feasible region and a fixed objective. In our setting the objective itself changes with ancestry through $R_e$, and this is exactly what we measure, not an assumption we make.
PRISM motivates the work. It shows that sparser, biologically aligned models transfer better to African-ancestry individuals, and it uses functional priors to get there. We impose sparsity through cross-ancestry evidence instead, and the two can be combined by seeding the cache with PRISM's prioritized variants.
Ancestry calibration constrains how we evaluate. Post-hoc rescaling against continuous West African ancestry fixes score distributions but cannot recover unmapped variants. It is also needed to avoid the "covariate trap" (unstable effect sizes when raw ancestry is added as a covariate to uncalibrated scores). We therefore apply it to every method before comparing them.
MR-SENSEMAKR is a warning about large biobanks. There, tiny residual stratification yields significant results. So our admission test must be checked for inflation on real target data, not only on simulations.
Baselines
The naive reference is a European-only sparse score applied unchanged to the target. We fit it by the same $\ell_1$-constrained regression on UK Biobank European summary statistics, then apply ancestry calibration. It shows how much any transfer method gains.
The strongest existing competitor for African-ancestry prediction in the file is PRISM. Its reported result is an average 13.10% improvement in $R^2$ (p = 1.6×10⁻⁵) in UK Biobank African-ancestry individuals over annotation-agnostic models, with 36.5% for FEV1/FVC. No reported absolute $R^2$ values were found on Lacuna for PRISM. We will therefore rerun it and our method on the same traits, the same African-ancestry test split and the same LD panels. Both will get an equal hyperparameter search budget over $\tau$ and PRISM's penalties. Beating it means one of two things. Either our method matches or exceeds its African-ancestry $R^2$ without functional annotations, or it adds a significant gain when seeded with PRISM's variants.
For East Asian transfer, the strongest competitor is LOG-TRAM. It reports 17% to 24% $R^2$ improvement in the independent WeGene Chinese cohort. WeGene may be unavailable to us, so we will compare on a held-out Biobank Japan split, using identical UKBB and BBJ summary statistics for both methods.
A multi-ancestry group lasso, which ties each variant's effects across ancestries under one penalty, is the soft-sharing counterpart to our hard cache. It gets the same tuning budget.
Three ablations isolate the key ingredient. The first uses the seeded cache with the optimizer's usual $\Phi_t$ schedule instead of the noise-calibrated threshold, which isolates calibration. The second uses a cold cache seeded only from the target, which isolates source seeding. The third refits the target on the source support with no admissions at all, which isolates the fallback.
Why not just…?
Why not just warm-start coordinate descent with an active set, as standard lasso solvers do? For speed alone this is roughly equivalent, and we concede it. Active-set methods with periodic optimality checks do much of what lazy Frank-Wolfe does. Our claim is not a runtime win over such solvers. It rests on the calibrated admission rule and the certificate it produces, which a standard warm start does not provide. We will report runtime against a warm-started coordinate descent run under the same schedule, so that speed claims stay honest.
Why not just use a multi-ancestry group lasso or PRISM-style priors to encourage shared supports? This is the most serious threat to the prediction claim. A tuned group lasso may match our $R^2$. If it does, the predictive contribution collapses and what remains is the portability test. Group lasso trades sharing against fit through a single penalty and offers no statement with a known error rate about any locus. The experiments settle this directly: head-to-head $R^2$ under equal tuning, plus the calibration checks below, which group lasso cannot be subjected to.
Why not just calibrate a European score post hoc? Vargas et al. show calibration helps substantially. In REGARDS it raised the Type 2 Diabetes top-versus-bottom decile odds ratio from 7.97 to 10.77. But the same authors note it cannot add variants absent from the European model. We apply calibration to all methods, so any remaining gain comes from certified admissions.
Why not just use the invariant autoencoder of Miao et al. for instruments? It removed bias of up to 0.3 in its semi-synthetic setting, but it produces dense latent instruments rather than named variants. Its authors also note it fails when ancestral environments are not distinct enough. We compare against it only in the Mendelian randomization extension, and only on bias.
Decisive Experiments
The first check is calibration, and the project depends on it. In semi-synthetic data, we plant shared and ancestry-specific causal variants on real TOP-LD EUR and AFR LD, then simulate summary statistics at realistic target sizes. We require the false-admission rate per null block to stay at or below $\alpha$, while power to admit planted ancestry-specific variants increases with $n_e$. We repeat this with the threshold computed from mismatched European LD. The purpose is to quantify the reference-panel sensitivity the file flags for MR-LDP-style methods.
The second check is agreement. On 29 traits in the UKBB-to-BBJ setting, blocks we flag as non-portable should be enriched for low local genetic correlation under LOG-TRAM and for high EUR–AFR LD divergence. If the cache-miss map is uncorrelated with both, the interpretation fails.
The third check is prediction. We measure African-ancestry and East Asian $R^2$ against the baselines above, after calibration.
The fourth check is null behavior on real data. Following MR-SENSEMAKR, we run the admission test on permuted-phenotype and negative-control traits in UK Biobank. The aim is to detect inflation from residual stratification, which simulations would miss.
A positive outcome would pair three results: calibrated admission, agreement with LOG-TRAM, and prediction at least matching group lasso. If calibration holds but prediction only ties, the result is still useful: a cheap, certified locus-level portability test.
References
- Lazifying Conditional Gradient Algorithms
- TOP-LD: A tool to explore linkage disequilibrium with TOPMed whole-genome sequence data
- FrankWolfe.jl: a high-performance and flexible toolbox for Frank-Wolfe algorithms and Conditional Gradients
- Leveraging the local genetic structure for trans-ancestry association mapping
- Addressing Instrument-Outcome Confounding in Mendelian Randomization through Representation Learning
- Contrastive Losses and Solution Caching for Predict-and-Optimize
- Solver-Free Decision-Focused Learning for Linear Optimization Problems
- PRISM: ancestry-aware integration of tissue-specific genomic annotations enhances the transferability of polygenic scores
- Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults
- Robust Mendelian randomization in the presence of residual population stratification, batch effects and horizontal pleiotropy