A polygenic risk score predicts a trait or disease from millions of small genetic effects. These scores transfer poorly across ancestries. One cause is that roughly 86% of genome-wide association study participants are of European descent (Optimal strategies for learning multi-ancestry polygenic scores vary across traits). Another is that some genetic effects truly differ between populations. The strongest recent fix, DendroPRS (Multi-ancestry polygenic risk scores using phylogenetic regularization), gives each population its own effect sizes. Those effect sizes are built as a shared "root" vector plus sparse corrections along the branches of a human ancestry tree. On UK Biobank BMI for non-White groups, this lifts prediction correlation from 0.10 (separate per-population models) and 0.03 (naive pooling) to 0.14. But DendroPRS needs every genotype in one computer's memory, and real biobanks cannot legally pool genotypes. The federated genomic work in hand (Efficacy of federated learning on genomic data) runs only plain federated averaging, which produces one global model and discards ancestry structure. Federated learning here means training a shared model by exchanging parameter updates, never raw data.
The naive bridge, "federated DendroPRS with one site per ancestry," is close to trivial. If each site holds exactly one population, the objective splits by site. Each site can then send its summary statistics or its gradient, and the centralized solution is recovered exactly. The hard and realistic case is different, and it is the one we propose to solve. Real sites, such as the 19 UK Biobank assessment centres, hospitals, or national biobanks, each hold a mixture of ancestries. Many individuals are admixed, meaning their genomes descend from several branches of the tree. Minority-ancestry samples are scattered thinly across sites, where their per-site statistics are small and easy to re-identify. We treat ancestry as a soft, per-individual position on the tree, estimated without pooling. We train tree-structured effect sizes from clipped, noise-protected gradients. We then ask whether the tree structure protects minority-ancestry accuracy when a formal privacy budget is imposed.
Core Claim
Each individual $i$ receives an individual-specific effect vector
$$\beta_i = \beta_{\text{root}} + \sum_{b} a_{ib}\,\delta_b,$$
The terms are defined as follows:
- $\beta_{\text{root}}$ is the effect vector shared by everyone.
- $\delta_b$ is a sparse correction attached to tree branch $b$.
- $a_{ib} \in [0,1]$ is the fraction of individual $i$'s ancestry that descends through branch $b$.
Training minimizes, summed over all sites $s$ and their individuals, the squared error $(y_i - x_i^\top \beta_i)^2$. Here $x_i$ is the genotype and $y_i$ the trait. The penalties are an L1 penalty on $\beta_{\text{root}}$, an L1 penalty on each $\delta_b$, and the group penalty of DendroPRS. With hard 0/1 memberships, this is exactly the DendroPRS objective. With soft memberships, admixed people contribute to every branch they descend from, in proportion to their ancestry.
The gradient of individual $i$'s loss with respect to $\delta_b$ is $a_{ib}$ times the gradient with respect to $\beta_i$. So any site, whatever its ancestry mix, can compute its contribution to every tree parameter locally.
We make two testable claims:
- Soft tree membership beats hard ancestry labels for admixed and mislabelled individuals.
- Under a fixed differential-privacy budget, the tree parameterization loses less minority-ancestry accuracy than flat federated averaging, with or without local fine-tuning. Differential privacy is a formal guarantee that limits how much any one person's data can change what is released.
The reasoning behind the second claim is as follows. The root is estimated from everyone, so injected noise is small relative to its signal. The branch corrections need to recover only a few large-effect loci, such as the African-specific DARC hotspot for mean corpuscular volume or the direction-flipping MSRA variant for BMI. Sparse penalties can still find such loci under noise. Flat averaging, by contrast, cancels opposite-signed effects outright.
Research Program
The first component estimates ancestry without pooling. We run federated principal component analysis, using the P-STACK approach shown in Efficacy of federated learning on genomic data to match centralized PCA accuracy when compressing ~65,000 SNPs to 20 components. This gives every individual coordinates in a shared genetic space. Reference leaf centroids come from the public 1000 Genomes superpopulations. Each site converts its individuals' coordinates into branch memberships $a_{ib}$ by non-negative projection onto the tree's leaves, and then accumulates along root-to-leaf paths. Memberships never leave the site.
The second component is the training protocol, a federated proximal gradient method:
- The server broadcasts $(\beta_{\text{root}}, \{\delta_b\})$.
- Each site computes per-individual gradients for root and branches, clips each individual's contribution, and adds Gaussian noise calibrated to a target privacy level. Under secure aggregation, as discussed for genomics in Privacy-preserving federated GWAS via dynamic sampling, the server would see only the sum across sites.
- The server applies the soft-thresholding steps implied by the L1 and group penalties.
Because the model is linear, local drift is limited. We still adopt one local step per round, following the finding that 32 local epochs caused loss spikes across superpopulation sites (Efficacy of federated learning on genomic data).
The third component handles variant screening and tuning. Variants are pre-screened with relaxed, summed marginal statistics computed jointly across sites, not by intersecting per-site non-significant lists. This avoids the premature discarding of jointly important variants that the dynamic-sampling GWAS design risks. Penalty strengths are tuned per trait using site-local held-out folds whose validation losses are summed by ancestry. This is where the trait-dependence documented in Optimal strategies for learning multi-ancestry polygenic scores vary across traits should emerge on its own: large branch penalties for height, small ones for mean corpuscular volume.
Baselines
The naive reference points are two. The first is isolated per-population local models. The second is centralized naive pooling. On UK Biobank non-White BMI these reach correlations of 0.10 and 0.03, as reported in Multi-ancestry polygenic risk scores using phylogenetic regularization. We rerun both on our splits so the numbers are comparable.
The strongest existing method is centralized DendroPRS, which reports 0.14 on that task. It is the ceiling: our non-private federated version should match it within seed-to-seed variation. It shares the same variant set, the same tree, the same tuning grid, and the same number of optimizer passes.
The strongest federated competitor is flat federated averaging of a LASSO-type model, as in Efficacy of federated learning on genomic data. That paper reports only that its R² is "nearly indistinguishable" from centralized training. It gives no per-trait numbers and no multi-ancestry breakdown, so no reported result was found on Lacuna for flat federated averaging on per-ancestry polygenic prediction, and we will produce it. We also run flat averaging followed by per-site fine-tuning, the standard personalization recipe.
The fairness rules apply to every method:
- identical splits;
- identical round counts and per-round communication;
- identical hyperparameter search size;
- under privacy, identical privacy accounting and clipping.
Beating flat federated averaging means higher correlation on the non-European ancestry groups at matched privacy, without losing more than the noise floor on Europeans.
Two ablations isolate the key ingredients:
- The first replaces soft memberships with hard self-reported labels, which isolates the admixture handling.
- The second replaces the tree with a "star": every leaf is attached directly to the root, giving the generic "shared plus per-group offset" model. This isolates whether tree structure, rather than merely shared-plus-specific parameters, carries the gain.
Why not just…?
Why not just have each site send its summary statistics and fit DendroPRS centrally? For a linear model, per-ancestry Gram matrices $X^\top X$ and cross-products $X^\top y$ are sufficient statistics. When sites map one-to-one to ancestries and privacy is not formally required, this is exact, takes a single round, and we expect it to tie with our method. We include it as a baseline and do not argue otherwise. It weakens in our target regime for two reasons. With soft memberships, every site must release a membership-weighted Gram matrix for each pair of branches. Under differential privacy, noising a $p \times p$ matrix costs far more accuracy than noising $p$-dimensional gradients when minority counts per site are in the tens. The privacy-sweep experiment below compares noisy sufficient statistics against noisy gradients directly. If the noisy sufficient statistics win, the contribution narrows to the membership model, and we will report that.
Why not just run flat federated averaging and fine-tune locally at each site? Fine-tuning personalizes by site, not by ancestry. A mixed-ancestry site fine-tunes toward its majority. A minority group spread over 19 centres never has enough local data to fine-tune its DARC-like effects. The file already shows that a small African-only model beats a 300,000-person European model for mean corpuscular volume. A site-level fine-tune that starts from a European-dominated average must then undo that average with little data. The per-ancestry breakdown on mean corpuscular volume and BMI will show whether it can.
Why not just a shared vector plus one free offset per ancestry, without a tree? This is the most serious threat, and it is exactly the star ablation. The tree should help most for small groups whose nearest relative is data-rich. For example, a South Asian or Middle Eastern leaf can borrow from a shared internal branch rather than from the global root. Tree-based regularization has been shown to recover sparse mechanism changes far better than unstructured baselines: a mean correlation coefficient of 0.98 versus ~0.45 on semi-real GTEx data (Leveraging Structure Between Environments). That evidence is centralized, though, and not on human ancestry. If the star matches the tree on UK Biobank and Geuvadis, we will conclude that soft membership and privacy-robust sparsity are the contributions and the tree topology is not.
Why not just reweight minority samples in the federated loss? Inverse-frequency reweighting inflates variance and shrinks effective sample size, and the file reports it falls short of structural regularization (Optimal strategies for learning multi-ancestry polygenic scores vary across traits). Under privacy noise this gets worse, because upweighted individuals need larger clipping bounds. We include it in the privacy sweep as a cheap check.
Proposed Experiments
The primary benchmark is UK Biobank split by its 19 real assessment centres, which naturally mixes ancestries across sites. Evaluation uses the six DendroPRS traits plus mean corpuscular volume and height, which test the two architectural extremes. Results are reported per Pan-UKBB ancestry group, with admixed individuals scored separately.
The first decisive check is that non-private federated training on centre splits recovers centralized DendroPRS. The target is the 0.14 non-White BMI correlation, and the claim fails if we land materially below it.
The second check is a privacy sweep over several privacy levels. It plots per-ancestry correlation for the following methods:
- the tree model;
- the star ablation;
- flat averaging with and without fine-tuning;
- noisy sufficient statistics;
- reweighting.
The hypothesis predicts that the gap between the tree and flat averaging widens for non-European groups as privacy tightens.
The third check is locus recovery. Branch corrections should recover the DARC signal on the African branch for mean corpuscular volume and the MSRA direction flip for BMI. Height should remain root-dominated.
The admixture test uses 1000 Genomes with Geuvadis expression for the 67 genes on which DendroPRS improved prediction. Sites are built as random ancestry mixtures, and Admixed American individuals are held out. Soft membership should beat hard labels on these individuals, and the hard-label model should reproduce DendroPRS when sites equal populations.
Throughout, we report communication per round, rounds to convergence, and wall-clock time. The goal is for the comparison with one-round summary statistics to be judged on cost as well as accuracy.
References
- Optimal strategies for learning multi-ancestry polygenic scores vary across traits
- Multi-ancestry polygenic risk scores using phylogenetic regularization
- Efficacy of federated learning on genomic data: a study on the UK Biobank and the 1000 Genomes Project
- Privacy-preserving federated genome-wide association studies via dynamic sampling
- Leveraging Structure Between Environments: Phylogenetic Regularization Incentivizes Disentangled Representations