Learning the Shape of Pleiotropy from GWAS Summary Statistics with Flow Matching

Many diseases share genetic risk. A variant that raises schizophrenia risk often shifts depression or bipolar risk too, and statistical geneticists summarize this sharing mostly with linear tools: one genome-wide correlation number per trait pair, or a linear factor model that splits each disorder's risk into a shared part and a unique part. These tools hide a basic object. Each genetic variant has a vector of true effects, one per trait. Across the genome, those vectors form a distribution in $K$ dimensions, where $K$ is the number of traits. That distribution is the shared genetic architecture. Most variants sit at zero. Some lie along a single trait's axis. Some point in a shared direction. Some point in opposite directions for different traits. We propose to estimate this $K$-dimensional distribution directly, as a flexible generative model, from the only data most consortia release: noisy per-variant association z-scores that are correlated across neighboring variants.

The tool is flow matching, a way to train continuous generative models by plain regression on straight-line paths between noise and data, with no differential-equation solver inside the training loop. We adapt it to a setting it has not been used in. The training samples here are never observed cleanly. Each is seen only through a known, correlated noise process set by linkage disequilibrium (LD), the correlation between nearby variants, and by overlapping samples across studies. Once the effect distribution is learned, it serves as a prior for per-variant posteriors. Those posteriors give multi-trait risk scores and disorder-specific scores that assign each variant to shared or specific structure individually, instead of subtracting a genome-wide linear factor. The concrete test is whether this removes the loss of predictive power that current disorder-specific scores pay.

Core Claim

The linear factor approach, Genomic Structural Equation Modeling (GenomicSEM), was used in Enhancing the discriminatory power of polygenic scores for ADHD and autism to build disorder-specific polygenic scores across 10 psychiatric traits. Those scores removed false cross-disorder associations: six unrelated outcomes dropped to zero in the Philadelphia Neurodevelopmental Cohort (PNC). The cost was a loss of explained variance. Our hypothesis is that this cost comes from a modeling error, not from biology. A linear factor model treats every variant as a scaled copy of the same shared-plus-unique split. The true effect distribution is a mixture of a point mass at zero, trait-specific axes, and shared directions, and its weights depend on functional annotation. A per-variant posterior under the correct prior can keep a variant's full effect when that variant is trait-specific and discount it when it is shared. That should give specificity without the drop in explained variance.

The methodological claim is that this prior can be learned without assuming a parametric shape. We write it as

$$g(\beta \mid a) = \pi_0(a)\,\delta_0 + (1-\pi_0(a))\,q_\theta(\beta \mid a),$$

where $\beta \in \mathbb{R}^K$ is one variant's vector of true standardized effects, $a$ is its annotation vector (functional category, tissue-specific expression, local LD score), $\pi_0(a)$ is the probability that the variant is null, $\delta_0$ is a point mass at zero, and $q_\theta$ is a conditional flow. The explicit point mass answers a known weakness: push-forward flows with Gaussian base distributions handle the exact sparsity of polygenic architectures poorly. The research file flags this sparse, heavy-tailed regime as a gap, citing Sparse polygenic risk score inference with the spike-and-slab LASSO for spike-and-slab structure and Conditional Flow Matching for Bayesian Posterior Inference for the heavy-tail difficulty.

Within one LD block $b$, the observed z-scores follow

$$Z_b = R_b B_b D + E_b,$$

with the following terms:

  • $Z_b$ is the variants-by-traits matrix of z-scores.
  • $R_b$ is the LD matrix from a reference panel.
  • $B_b$ holds rows drawn independently from $g(\cdot \mid a_j)$ for each variant $j$.
  • $D$ is a diagonal matrix of $\sqrt{N_k}$, the square root of each trait's GWAS sample size.
  • $E_b$ is Gaussian noise with covariance $R_b$ across variants and $C$ across traits. $C$ is the sample-overlap correlation, estimated from null variants.

Everything except $g$ is known. Learning $g$ is therefore a deconvolution problem with a known, correlated corruption.

Research Program

We train by stochastic expectation–maximization, an alternating scheme with two regression steps and no backpropagation through a solver.

The first step estimates effects from observed scores. We draw synthetic blocks from the current prior, corrupt them through the known forward model, and fit a conditional flow that maps observed block z-scores to posterior draws of each variant's effect vector. The flow is trained by the conditional flow matching regression of Flow Matching for Generative Modeling. Its conditioning structure follows Conditional Flow Matching for Bayesian Posterior Inference, which learns posterior transport from simulated (parameter, observation) pairs.

The second step updates the prior. We refit $q_\theta$ and $\pi_0$ by flow matching on posterior draws for the real variants. The general idea of training generators from corrupted observations under a known noise process comes from Inverse Flow and Consistency Models. Our contribution is to make that corruption the GWAS summary-statistic model, including cross-variant LD and cross-trait sample overlap.

A known constraint shapes the posterior network. The conditional flow matching posterior work reports that it degrades when observations are very high-dimensional. A full block posterior over hundreds of variants times $K$ traits is exactly that case. We therefore never generate the full block jointly. Each variant's $K$-dimensional posterior is conditioned on its own z-score vector plus an encoding of the block that pools neighbors' z-scores weighted by their LD. The generated object stays $K$-dimensional, and LD still enters. The approximation is testable. In small simulated blocks we can compute the exact posterior by enumeration and measure the gap directly.

The learned $g$ supports three analyses.

  1. Architecture shape. We report how prior mass divides among null, trait-specific, shared-concordant, and shared-opposing regions. Linear correlations cannot show this last region, because opposing effects cancel. We report the division overall and as a function of annotation.
  2. Multi-trait scores. Per-variant posterior means give multi-trait polygenic scores.
  3. Disorder-specific scores. For each variant, the posterior probability of lying in trait $k$'s specific region reweights its effect into a disorder-specific score, variant by variant.

Local correlation hot spots such as the hub regions of Detecting local genetic correlations with scan statistics then become a check on the model, not a primary output. Our block-level posterior shared-mass should concentrate in those regions.

Baselines

The naive reference point is the standard polygenic score built separately for each trait, using independent per-trait priors with no cross-trait information. Sparse polygenic risk score inference with the spike-and-slab LASSO (SSLPRS) is the strongest single-trait instance in the file. It reports matching or exceeding Lassosum and VIPRS on 9 UK Biobank quantitative traits (337,205 individuals, 1.1 million variants), selecting 5% fewer variants, and running genome-wide in about 3 minutes under 1 GB of RAM. The file gives no absolute prediction-accuracy value. We will run SSLPRS on the same summary statistics, LD reference, and held-out individuals as our method, with its standard tuning. Beating it means higher held-out $r^2$ on traits that have genetically correlated partners. Losing on traits without correlated partners would be acceptable and informative.

The strongest existing method for the specificity question is GenomicSEM as applied in Enhancing the discriminatory power of polygenic scores for ADHD and autism. It reduced the autism score's significant non-focal associations in PNC from six to zero. It also made the ADHD score predictive of comorbid ADHD in the clinical SPARK autism cohort, where the traditional score failed. The file reports no numeric $r^2$ for the power loss. We will refit their HiTOP-structured model on the same 10 summary statistics. Our disorder-specific scores win only if two conditions both hold in PNC. They must also reach zero non-focal associations. They must also recover explained variance for the focal trait to at least the unrefined score's level, where GenomicSEM's refined score falls short.

The ablation that isolates the flow replaces $q_\theta$ with a Gaussian mixture. The mixture gets many components, learned covariances, and annotation-dependent weights from the same neural network. It is embedded in the identical expectation–maximization loop with the same LD-aware posterior network, point mass, data, and GPU budget. No reported result was found on Lacuna for mixture-prior deconvolution of multi-trait GWAS effects, so we build it ourselves. This is the comparison that decides whether flow matching earns its place.

Why not just…?

Why not just use GenomicSEM with more factors? More linear factors can capture more covariance. They cannot represent a distribution in which some variants lie exactly on one trait's axis while others share a direction. In a linear model every variant is loaded identically, so specificity is bought by subtracting shared variance from every variant, including trait-specific ones. That is the mechanism we believe causes the power loss Li et al. report. The PNC and SPARK comparison tests this directly. If a higher-factor GenomicSEM closes the $r^2$ gap while keeping zero non-focal hits, our hypothesis about the source of the trade-off is wrong.

Why not just fit a Gaussian mixture prior? This is the real threat. A rich mixture approximates heavy tails and axis-plus-shared structure well. For two or three traits without annotations we expect it to match the flow, and we will say so. Our prediction is narrower: the flow wins only at $K$ of 7 to 10 with annotation conditioning. There, the number of mixture components needed to cover curved or annotation-dependent structure grows quickly, and fitting becomes unstable. The ablation above measures three things across $K$: held-out z-score likelihood on a withheld chromosome, recovery of simulated architectures, and downstream score accuracy. If the mixture matches the flow at $K=10$, we will adopt it. The lasting contribution would then be the LD-aware deconvolution loop and the per-variant specific-score construction, not the flow.

Why not just use local scan statistics like LOGODetect? LOGODetect finds regions of pairwise local correlation. On the 7 neuropsychiatric traits it found 227 regions, and 17 of 24 replicated in UK Biobank. It produces no effect distribution, no per-variant posterior, and no risk score. Because it is built on inner products of z-scores, it summarizes the concordance direction within a window rather than resolving which individual variants drive shared versus opposing effects. We use its regions as external validation, not as a competitor for prediction.

Why not just take density ratios between trait-specific liability distributions, as scRatio does for cell states? Flow-Based Density Ratio Estimation for Intractable Distributions with Applications in Genomics reports that its ratios become unstable when distributions barely overlap. Effect distributions with point masses at zero and disjoint axis components are exactly that case. Modeling the joint distribution once and reading off regions avoids forming unstable ratios.

Proposed Experiments

Controlled simulation. Simulated genotypes follow the HAPGEN2 / 1000 Genomes setup used by LOGODetect, with 100,000 individuals. We generate effects from known non-Gaussian $K$-trait architectures: axis-plus-shared mixtures, opposing-effect components, annotation-dependent null fractions, and a nonlinear shared curve. We then compute summary statistics with controlled sample overlap. The metrics are:

  • sliced Wasserstein distance between the learned and true $g$;
  • error in the mass assigned to each region;
  • coverage of per-variant credible intervals;
  • the gap between our approximate LD-aware posterior and the exact posterior in small blocks.

The decisive check is whether, at $K=7$ and $K=10$ with annotations, the flow recovers region masses that the mixture ablation misestimates. The check must survive equal tuning budgets.

Real data. On the 7 neuropsychiatric summary statistics we report the learned architecture shape. We test whether block-level shared mass is enriched in LOGODetect's regions, including those near DRD2 and NCAM1. On the 10 HiTOP disorders, with PNC and SPARK, we run the specificity test against GenomicSEM. On the 9 UK Biobank quantitative traits we compare multi-trait posterior-mean scores against SSLPRS by held-out $r^2$. We also report wall-clock time and memory against SSLPRS's roughly 3 minutes and under 1 GB, since a large computational penalty needs to be justified by accuracy.

Stopping rule. Two results would end the project as proposed. One is a mixture matching the flow at high $K$. The other is our specific scores failing to beat GenomicSEM's power loss in PNC while keeping zero non-focal associations.