A polygenic risk score predicts a person's disease risk by adding up thousands of small genetic effects. The effects must be estimated from genome-wide association study (GWAS) summary statistics, which are per-variant association estimates published without the underlying genotypes. Nearby variants are inherited together; this correlation is called linkage disequilibrium (LD). Because of it, the summary statistics pin down some combinations of effects very precisely and others almost not at all. Current Bayesian methods such as LDpred2 and PRS-CS treat all of these combinations with one hand-chosen prior and one sampler. When the inputs are slightly wrong, for example a misstated sample size or a reference LD panel from the wrong population, their samplers go unstable precisely in the poorly determined combinations.
This note proposes to estimate polygenic effects with a diffusion model, a generator that learns to turn random noise into samples by gradually removing it. The noise schedule is asymmetric: each direction of the LD correlation matrix gets its own noise rate, and that rate is set by how well the GWAS determines that direction. With this schedule, the observed summary statistics are, up to rotation, exactly a sample from the diffusion's forward process at a known intermediate time. Posterior inference then becomes "start the reverse process from the data and denoise." The reverse process fixes well-determined directions first and fills in poorly determined ones last, drawing on a learned prior over genetic architectures. In molecule diffusion the order of denoising is chosen by hand. Here it is not a heuristic: the statistics of GWAS dictate it.
Core Claim
Write the within-block LD matrix as $R = V \Lambda V^\top$, with eigenvectors $V$ and eigenvalues $\lambda_k$. The standard summary-statistic likelihood is $\hat\beta \mid \beta \sim \mathcal{N}(R\beta, R/N)$, where $\hat\beta$ are the marginal GWAS effects, $\beta$ the joint effects and $N$ the GWAS sample size. The research file writes the noise covariance as $I/N$; the $R/N$ form is the one LDpred2 and PRS-CS actually use. Rotating into the eigenbasis gives a set of independent noisy observations:
$$y_k \equiv \frac{v_k^\top \hat\beta}{\lambda_k} = u_k + \epsilon_k,\qquad u = V^\top\beta,\quad \epsilon_k \sim \mathcal{N}\!\left(0, \tfrac{1}{N\lambda_k}\right).$$
Now define a forward diffusion on $\beta$ in which eigen-direction $k$ receives noise of variance
$$\sigma_k^2(t) = \frac{t}{N\lambda_k}, \qquad t \in [0, T],\ T > 1.$$
At $t=1$, the diffusion state has exactly the distribution of the observation $y$. Running the learned reverse process from $t=1$, initialized at $Vy$, therefore samples from the posterior $p(\beta \mid \hat\beta)$ under whatever prior the diffusion model has learned. This needs no approximate guidance term and no Gibbs sampler.
Several properties follow from the schedule. Directions with large $N\lambda_k$ start nearly clean and barely move. Directions with small $N\lambda_k$ start very noisy, and the prior resolves them last, conditioned on the already-fixed structure. These are the directions where LDpred2 needs ad hoc repairs like shrink_corr and allow_jump_sign (Identifying and correcting for misspecifications in GWAS summary statistics and polygenic scores). A misstated global sample size only moves the correct starting time away from 1, and that starting time can be estimated from the data.
The claim to be tested is this: a diffusion prior trained on simulated architectures and sampled with this schedule matches or beats LDpred2-auto on real traits and degrades substantially less under summary-statistic misspecification. The expected gain is in robustness first and accuracy second.
Relation to Prior Work
The mechanism comes from molecular diffusion. There, asymmetric schedules noise one component faster than another so the reverse process builds the slow component first. MolDiff noises bonds faster than coordinates to fix atom–bond inconsistency (MolDiff). MiDi noises atom identities faster than the 3D skeleton (MiDi). DeCoDe delays noise on a ligand's center of mass relative to its internal conformation (DeCoDe). We reuse their core idea that the order of destruction sets the order of creation. We differ in that our split, well-determined versus poorly determined eigen-directions, follows from the likelihood rather than from domain intuition.
Equivariant Asynchronous Diffusion learns per-atom noise rates adaptively and reports that such schedules are sensitive to the asynchrony ratio (Equivariant Asynchronous Diffusion). That finding constrains our design. We avoid any free ratio because $N\lambda_k$ fixes every rate. DiffPack shows that building domain structure into the diffusion space lets a model with 3 million parameters beat one with 208 million (DiffPack). This motivates our bet that structure, not scale, is what a genetic generative model needs.
On the genetics side, LDpred2 (LDpred2: better, faster, stronger) and PRS-CS (Polygenic prediction via Bayesian regression and continuous shrinkage priors) supply the likelihood, the block-diagonal LD approximation and the baselines. They differ from us only in the prior: point-normal or a fixed global-local shrinkage family, where ours is learned and non-parametric. They also differ in the inference engine, Gibbs sampling versus a single reverse pass.
PRSformer is a constraint, not a competitor. It shows that deep discriminative models on raw genotypes beat LDpred2 only beyond roughly one million individuals (PRSformer). We sidestep this barrier by keeping the model additive in its output and learning only the prior over effects from simulation. We make no claim about epistasis.
The closest diffusion-for-genotype work is G2PDiffusion. It generates insect images from DNA barcodes with a standard symmetric schedule and does not touch effect-size inference or LD (G2PDiffusion). The other "diffusion" papers in this area use diffusion MRI or graph heat diffusion, which overlap only in name.
Research Program
The prior is a conditional denoiser over SNP effects within an LD block. It is a small transformer with local attention over SNP positions, with the window sized to roughly 100 kb following PRSformer's neighborhood design. Its inputs are the noisy state $x_t = \beta + V\,\mathrm{diag}(\sigma_k(t))\,\xi$, where $\xi$ is standard Gaussian noise, together with the time $t$, per-SNP features (allele frequency, LD score) and the block's LD matrix through its eigenvectors. It is trained by standard denoising score matching.
Training data are simulated effect vectors placed on real LD blocks from the 1000 Genomes and UK Biobank reference panels. The simulated architectures deliberately span wide ranges: the fraction of causal variants runs from very sparse to infinitesimal, effect distributions range from Gaussian to heavy-tailed, the mixtures include a few large-effect loci over a dense background, and heritability varies. The model is conditioned on two architecture summaries, total block heritability and a sparsity index. At inference these are estimated per trait, in the spirit of LDpred2-auto, or marginalized over a grid.
Inference proceeds in steps. First, rotate each block's summary statistics. Second, floor very small eigenvalues at a fixed fraction of the largest, the same regularization lassosum and LDpred2 apply to $R$, so that the $1/\lambda_k$ division stays bounded. Third, estimate the starting time $t^*$ by maximizing the model's marginal likelihood over a one-dimensional grid; $t^*=1$ when the sample size is stated correctly. Fourth, run the reverse process from $t^*$. Repeating the reverse pass gives posterior samples of the effects, and scoring those samples gives each individual a score with a credible interval.
The analysis side asks three questions. The first is whether the learned prior outperforms parametric priors on real traits. The second is whether estimating $t^*$ absorbs sample-size misspecification. The third is whether the eigen-ordering itself matters, which the ablations below isolate.
Baselines
The naive reference points are the unadjusted marginal score, which sums GWAS effects with no LD correction, and pruning and thresholding. Pruning and thresholding costs 0.84 CPU hours per trait in the PRS-hub benchmark (Benchmarking PRS methods for risk prediction of 36 complex traits in UK Biobank). Any method that fails to clear these is not worth discussing.
The strongest existing competitor is LDpred2, run as both -auto and -grid. It was the best single method in 14 of 19 diseases and 9 of 17 quantitative traits in PRS-hub. On eight real UK Biobank diseases it reports a mean AUC of 65.1%, against 63.8% for lassosum, 62.9% for PRS-CS and 61.5% for SBayesR (LDpred2). It is the right bar because it uses exactly our likelihood with a well-tuned prior, so any gain we show must come from the prior or the inference, not from extra information.
PRS-CS-auto and lassosum2 are secondary baselines. PRS-CS matters most in the misspecification tests, where it is reported to be relatively robust to sample-size errors.
For fairness, every method receives the same summary statistics, the same LD reference, the same SNP set and the same validation individuals for any tuning. The curated 1.1-million-SNP set that PRS-hub found better than 6.6 million SNPs is a natural shared choice. Our own hyperparameters, meaning the floor fraction and the architecture grid, get a tuning budget no larger than LDpred2-grid's grid. We report wall-clock and CPU hours alongside accuracy.
Beating LDpred2 means two things. On real traits, we need a mean AUC or $R^2$ gain that is significant under paired bootstrap across traits. Under misspecification, we need a clearly smaller loss.
Two ablations isolate the key ingredient. The first keeps the identical learned prior but uses a synchronous, isotropic schedule, with conditioning on the summary statistics done by a standard gradient-guidance approximation. The second keeps the asymmetric schedule but replaces the learned prior with an independent Gaussian prior per eigen-direction, which is analytically close to LDpred2-inf. The first ablation tests whether the schedule matters. The second tests whether the learned prior matters.
Why not just…?
Why not just use LDpred2-auto with its existing robustness fixes? It is a real threat. If real genetic architectures are close to point-normal, a learned prior buys little, and LDpred2 with shrink_corr already tames small-eigenvalue directions. The real-trait comparison settles this directly. We commit in advance to reporting a null result, which would indicate that the prior's shape is not the accuracy bottleneck in current PRS. Our specific edge should appear in the misspecification settings, where LDpred2 needs manual correction and our starting-time estimate is automatic.
Why not just shrink each eigen-direction independently with an empirical-Bayes prior? It is cheaper and uses the same rotation. However, sparsity in SNP space becomes dense, dependent structure in the eigenbasis: a single causal SNP loads on many directions. An independent per-direction prior therefore cannot represent "few SNPs matter." The second ablation is exactly this model. If it matches the full method, the diffusion network is unnecessary and we would say so.
Why not just train a standard diffusion prior and condition it with the usual guidance? Guided posterior sampling for linear inverse problems is the default. In this setting the guidance gradient involves $R$, and its approximation error is largest in the ill-determined directions that cause trouble in the first place. The first ablation measures this directly through accuracy, sampling steps and failure rates under a mismatched LD panel.
Why not just stack existing scores in an ensemble? Stacking gives a 5–10% median increase in odds ratio per standard deviation in PRS-hub. It is complementary rather than competing, since our score can enter the ensemble. We will report whether it adds to a stack of LDpred2, lassosum2 and PRS-CS.
Proposed Experiments
The first experiment is a set of simulations on UK Biobank genotypes that follow the LDpred2 designs. We include its low-causal-fraction setting, where LDpred2 reports an AUC of 81.7% against 73.6% for LDpred1. We also include architectures held out from our training distribution, such as a heavier tail or unseen sparsity, to measure how badly the prior breaks when the simulations are wrong. This is the main risk of the whole approach.
The second experiment uses the eight real UK Biobank diseases from the LDpred2 paper and the 36 PRS-hub traits, comparing against the 65.1% mean AUC reference.
The third experiment reproduces the misspecification settings of Privé and colleagues. In their sample-size case, LDpred2-low-h2 improved from 0.134 to 0.164 once the maximum sample size was used. In their LD-panel case, lassosum2's squared correlation dropped from 0.173 to 0.166 with a South European reference. The decisive check is whether our method, given the uncorrected sample size, lands near the corrected value without manual intervention. A second decisive check is whether the starting time $t^*$ it recovers tracks the true misspecification factor. Heterogeneous per-variant sample-size errors are not a single global scale, so the method is not expected to absorb them fully. We will report that case separately.
The fourth experiment, secondary and exploratory, applies European-trained summary statistics to the 8,214 African-ancestry PRS-hub individuals. It asks whether resolving the well-determined directions first, which rest on the most strongly estimated signal, transfers better than LDpred2. Nothing in the file gives a reported result for diffusion-based effect estimation on any of these benchmarks.
We will also check calibration by testing whether 90% credible intervals for individual scores cover the true genetic values 90% of the time in simulation, against LDpred2-auto's samples. Finally, we will compare compute against SBayesR's roughly 359 CPU hours per trait.
References
- Identifying and correcting for misspecifications in GWAS summary statistics and polygenic scores
- MolDiff: Addressing the Atom-Bond Inconsistency Problem in 3D Molecule Diffusion Generation
- MiDi: Mixed Graph and 3D Denoising Diffusion for Molecule Generation
- DeCoDe: Decoupling Binding Position and Molecular Conformation in 3D Ligand Diffusion for Structure-Based Drug Design
- Equivariant Asynchronous Diffusion: An Adaptive Denoising Schedule for Accelerated Molecular Conformation Generation
- DiffPack: A Torsional Diffusion Model for Autoregressive Protein Side-Chain Packing
- LDpred2: better, faster, stronger
- Polygenic prediction via Bayesian regression and continuous shrinkage priors
- PRSformer: Disease Prediction from Million-Scale Individual Genotypes
- G2PDiffusion: Genotype-to-Phenotype Prediction with Diffusion Models
- Benchmarking PRS methods for risk prediction of 36 complex traits in UK Biobank by establishment of a large-scale PRS computation platform PRS-hub