SNPgen: Phenotype-Supervised Genotype Representation and Synthetic Data Generation via Latent Diffusion

Andrea Lampis; Michela Carlotta Massi; Nicola Pirastu; Francesca Ieva; Matteo Matteucci; Emanuele Di Angelantonio

cs.LGMar 11, 2026

SNPgen: Phenotype-Supervised Genotype Representation and Synthetic Data Generation via Latent Diffusion

Authors: Andrea Lampis, Michela Carlotta Massi, Nicola Pirastu, Francesca Ieva, Matteo Matteucci, Emanuele Di Angelantonio

Organizations: DEIB, Politecnico di Milano, Milan 20133, Italy · Health Data Science Centre, Human Technopole, Milan 20157, Italy · Genomics Research Centre, Human Technopole, Milan 20157, Italy · MOX - Department of Mathematics, Politecnico di Milano, Milan 20133, Italy ·

Abstract

Polygenic risk scores and other genomic analyses require large individual-level genotype datasets, yet strict data access restrictions impede sharing. Synthetic genotype generation offers a privacy-preserving alternative, but most existing methods operate unconditionally, producing samples without phenotype alignment, or rely on unsupervised compression, creating a gap between statistical fidelity and downstream task utility. We present SNPgen, a two-stage conditional latent diffusion framework for generating phenotype-supervised synthetic genotypes. SNPgen combines GWAS-guided variant selection (1,024-2,048 trait-associated SNPs) with a variational autoencoder for genotype compression and a latent diffusion model conditioned on binary disease labels via classifier-free guidance. Evaluated on 458,724 UK Biobank individuals across four complex diseases (coronary artery disease, breast cancer, type 1 and type 2 diabetes), models trained on synthetic data matched real-data predictive performance in a train-on-synthetic, test-on-real protocol, approaching genome-wide PRS methods that use $2$ - $6\times$ more variants. Privacy analysis confirmed zero identical matches, near-random membership inference (AUC $\approx 0.50$ ), preserved linkage disequilibrium structure, and high allele frequency correlation ( $r \geq 0.95$ ) with source data. A controlled simulation with known causal effects verified faithful recovery of the imposed genetic association structure.

Figures & tables

Figure 1: SNPgen pipeline. UK Biobank genotypes and external GWAS summary statistics are combined for GWAS-guided SNP selection, producing a compact set of trait-associated variants. In Stage 1, a 1D VAE encoder compresses the selected SNPs into a continuous latent space

\mathbf{z}

. In Stage 2, a latent diffusion model (1D UNet DDPM) generates novel latent representations conditioned on binary disease labels via cross-attention, which are then decoded into synthetic trait-conditioned genotypes.

Figure 2: Downstream risk prediction (ROC-AUC) across four UK Biobank traits. Bars show three models (XGBoost, XGBoost Balanced, PRS Univariate) under four data conditions: Real, Reconstructed, Synthetic, and Synthetic Augmented. Error bars: 95% CI from 5-fold CV. Dashed lines: External PRS baseline. Cohort sizes: CAD

n{=}458{,}724

(35,306 cases); BC

n{=}248{,}987

(19,634 cases, female only); T1D

n{=}458{,}724

(4,593 cases); T2D

n{=}458{,}724

(38,541 cases).

Trait	PRSice-2	LDpred2	SNPgen
	(best config)	(auto)	(best XGB)
CAD	0.606	0.539	0.594
BC	0.636	0.661	0.598
T1D	0.628	0.559	0.671
T2D	0.599	0.628	0.593

Table 1: Genome-wide PRS baselines (PRS-only AUC).

Figure 3: Pairwise LD ( $r^{2}$ ) for T2D (first 200 SNPs). Left: original. Centre: reconstructed. Right: synthetic. Block-diagonal structure is preserved.

Trait	Condition	NNDR	MI AUC	NNAA	DCR <5% (%)	MAF $r$
CAD	Reconstructed	0.263	1.000	0.468	100.0	0.989
CAD	Synthetic	0.983	0.500	0.537	3.5	0.988
BC	Reconstructed	0.984	0.856	0.507	73.7	0.950
BC	Synthetic	0.990	0.501	0.507	46.3	0.951
T1D	Reconstructed	0.247	0.979	0.460	86.3	1.000
T1D	Synthetic	0.934	0.503	0.572	2.3	0.999

Table 2: Privacy evaluation summary. IMR: Identical Match Rate (

0\%

= no copies). NNDR: Nearest Neighbour Distance Ratio (higher = more private). MI: Membership Inference AUC (

{\approx}0.5

= no leakage). NNAA: Nearest Neighbour Adversarial Accuracy (

{\approx}0.5

= no leakage). DCR <5% : fraction of synthetic samples closer to training than the 5th percentile of holdout–training distances (lower = more private). MAF

r

: allele frequency correlation (higher = better population fidelity).

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Trait	$n$	Cases	Controls	Sample prev. (%)	Pop. prev. (%)
BC	248,987	19,634	229,353	7.9	12.0
CAD	458,724	35,306	423,418	7.7	6.0
T1D	458,724	4,593	454,131	1.0	0.4
T2D	458,724	38,541	420,183	8.4	8.0

Appendix

Table S1: Cohort demographics by trait. Sample prevalence is the case fraction in the UK Biobank cohort; population prevalence is used for liability-scale

R^{2}

conversion.

Trait	GWAS source	Clumping window	$r^{2}$ threshold	Final $L$
CAD	Nikpay et al. (2015)	10 kb	0.5	2,048
BC	Zhang et al. (2020)	500 kb	0.1	1,024
T1D	Michalek et al. (2024)	10 kb	0.5	2,048
T2D	Scott et al. (2017)	10 kb	0.5	2,048

Appendix

Table S2: Per-trait SNP selection parameters. BC required a looser clumping configuration due to sparser GWAS signal in the available summary statistics.

Trait	Condition	XGBoost	XGBoost (Balanced)	PRS Univariate
Simulated	Real	0.967 $\pm$ 0.002	0.975 $\pm$ 0.002	0.903 $\pm$ 0.006
	Reconstructed	0.960 $\pm$ 0.003	0.967 $\pm$ 0.003	0.898 $\pm$ 0.007
	Synthetic	0.902 $\pm$ 0.012	0.897 $\pm$ 0.006	0.849 $\pm$ 0.008
	Syn Augmented	0.905 $\pm$ 0.011	0.898 $\pm$ 0.006	0.849 $\pm$ 0.008
CAD	Real	0.592 $\pm$ 0.031	0.601 $\pm$ 0.009	0.575 $\pm$ 0.014
	Reconstructed	0.559 $\pm$ 0.036	0.578 $\pm$ 0.014	0.566 $\pm$ 0.019

Appendix

Table S3: Downstream risk prediction performance (ROC-AUC

\pm

95% CI). Bold : best result per trait and model (excluding Real). Underline : second best.

Config	N SNPs	AUC (PRS Only)	AUC (Age+Sex)	AUC (Age+Sex+PRS)	$\Delta$ AUC	$R^{2}_{\text{liab}}$
10 kb / 0.1	19,867	0.587	0.713	0.709	$-$ 0.004	0.052
10 kb / 0.5	23,517	0.587	0.713	0.708	$-$ 0.004	0.052
50 kb / 0.1	3,878	0.593	0.713	0.710	$-$ 0.003	0.060
50 kb / 0.5	5,706	0.596	0.713	0.711	$-$ 0.002	0.064
100 kb / 0.1	1,724	0.596	0.713	0.711	$-$ 0.002	0.065
100 kb / 0.5	4,954	0.602	0.713	0.713	$+$ 0.000	0.072

Appendix

Table S4: PRSice-2 results for Coronary Artery Disease (CAD).

n=458{,}724

(

35{,}306

cases). Bold : best AUC.

Config	N SNPs	AUC (PRS Only)	AUC (Age+Sex)	AUC (Age+Sex+PRS)	$\Delta$ AUC	$R^{2}_{\text{liab}}$
10 kb / 0.1	110,465	0.595	0.581	0.626	+0.045	0.077
10 kb / 0.5	109,954	0.600	0.581	0.629	+0.048	0.084
50 kb / 0.1	617	0.604	0.581	0.632	+0.051	0.089
50 kb / 0.5	15,193	0.613	0.581	0.639	+0.058	0.107
100 kb / 0.1	458	0.613	0.581	0.639	+0.058	0.108
100 kb / 0.5	9,493	0.623	0.581	0.647	+0.066	0.127

Appendix

Table S5: PRSice-2 results for Breast Cancer (BC).

n=248{,}987

(

19{,}634

cases). Bold : best AUC.

Config	N SNPs	AUC (PRS Only)	AUC (Age+Sex)	AUC (Age+Sex+PRS)	$\Delta$ AUC	$R^{2}_{\text{liab}}$
10 kb / 0.1	1,040	0.616	0.586	0.643	+0.057	0.201
10 kb / 0.5	1,632	0.623	0.586	0.648	+0.062	0.230
50 kb / 0.1	356	0.619	0.586	0.645	+0.060	0.223
50 kb / 0.5	799	0.625	0.586	0.650	+0.064	0.248
100 kb / 0.1	234	0.623	0.586	0.649	+0.063	0.246
100 kb / 0.5	646	0.626	0.586	0.652	+0.066	0.261

Appendix

Table S6: PRSice-2 results for Type 1 Diabetes (T1D).

n=458{,}724

(

4{,}593

cases). Bold : best AUC. Note: the kb_500_r2_0.5 configuration was not available for this trait.

Config	N SNPs	AUC (PRS Only)	AUC (Age+Sex)	AUC (Age+Sex+PRS)	$\Delta$ AUC	$R^{2}_{\text{liab}}$
10 kb / 0.1	133,010	0.590	0.619	0.639	+0.020	0.056
10 kb / 0.5	156,002	0.593	0.619	0.641	+0.022	0.060
50 kb / 0.1	1,383	0.589	0.619	0.638	+0.019	0.056
50 kb / 0.5	20,591	0.597	0.619	0.643	+0.024	0.064
100 kb / 0.1	1,137	0.589	0.619	0.638	+0.019	0.056
100 kb / 0.5	5,007	0.597	0.619	0.643	+0.024	0.066

Appendix

Table S7: PRSice-2 results for Type 2 Diabetes (T2D).

n=458{,}724

(

38{,}541

cases). Bold : best AUC.

Trait	$n$	Cases	OR/SD	AUC (PRS Only)	AUC (A+S)	AUC (A+S+PRS)	$\Delta$ AUC	AUC (Full)	AUC (Full+PRS)	$R^{2}_{\text{liab}}$	PR-AUC
BC	248,987	19,634	1.822	0.661	0.581	0.679	$+$ 0.098	0.582	0.679	0.221	0.147
CAD	458,724	35,306	1.165	0.539	0.713	0.705	$-$ 0.007	0.708	0.708	0.012	0.089
T1D	458,724	4,593	1.235	0.559	0.586	0.599	$+$ 0.013	0.584	0.602	0.046	0.013
T2D	458,724	38,541	1.602	0.628	0.619	0.665	$+$ 0.046	0.621	0.671	0.115	0.134

Appendix

Table S8: LDpred2-auto results.

\Delta

AUC: improvement over Age+Sex baseline.

R^{2}_{\text{liab}}

: liability-scale

R^{2}

increment of PRS.

Trait	Condition	Class	NNDR	MI AUC	NNAA	DCR <5% (%)	MAF $r$
Simulated	Reconstructed	Controls	0.260	1.000	0.438	100.0	0.989
	Reconstructed	Cases	0.260	1.000	0.334	100.0	0.989
	Synthetic	Controls	0.982	0.502	0.583	5.3	0.979
	Synthetic	Cases	0.983	0.502	0.546	34.5	0.984
CAD	Reconstructed	Controls	0.262	1.000	0.464	100.0	0.989
	Reconstructed	Cases	0.249	1.000	0.000	100.0	0.989

Appendix

Table S9: Class-specific privacy metrics. All metrics are computed within each class independently: controls (label 0) and cases (label 1). NNAA computed on up to 50,000 samples per set within each class. Abbreviations as in Table 2 of the main text.

Figure S1: Pairwise LD ( $r^{2}$ ) for CAD (2,048 SNPs). Left: original. Centre: VAE-reconstructed. Right: LDM-generated synthetic.

Figure S2: Pairwise LD ( $r^{2}$ ) for BC (1,024 SNPs). Left: original. Centre: VAE-reconstructed. Right: LDM-generated synthetic.

Figure S3: Pairwise LD ( $r^{2}$ ) for T1D (2,048 SNPs). Left: original. Centre: VAE-reconstructed. Right: LDM-generated synthetic.

Figure S4: Pairwise LD ( $r^{2}$ ) for T2D (2,048 SNPs). This is the full matrix corresponding to the 200-SNP subset shown in the main paper (Figure 3 ). Left: original. Centre: VAE-reconstructed. Right: LDM-generated synthetic.

Figure S5: Pairwise LD ( $r^{2}$ ) for the simulated phenotype (2,048 SNPs, same variant set as CAD). Left: original. Centre: VAE-reconstructed. Right: LDM-generated synthetic.

Figure S6: LD decay with physical distance for all five traits. Mean

r^{2}

(solid line)

\pm

SEM (shaded band) is shown for original (blue), VAE-reconstructed (green), and LDM-generated synthetic (orange) data, computed over same-chromosome SNP pairs binned into 50 log-spaced distance intervals.

Trait	Dataset	Same-chrom. pairs (%)	Missing $r^{2}$ (%)	Spearman $\rho$
CAD	Original	6.7	0.0	$-$ 0.602
	Reconstructed	6.7	0.9	$-$ 0.638
	Synthetic	6.7	0.5	$-$ 0.642
BC	Original	6.1	0.0	$-$ 0.395
	Reconstructed	6.1	17.2	$-$ 0.140
	Synthetic	6.1	0.8	$-$ 0.096

Appendix

Table S10: LD decay diagnostics. Same-chrom. pairs: fraction of all SNP pairs on the same chromosome (used for decay estimation). Missing

r^{2}

: fraction of same-chromosome pairs with undefined

r^{2}

(monomorphic loci). Spearman

\rho

: rank correlation between

\log_{10}(\text{distance})

and

r^{2}

over all valid same-chromosome pairs (

p<10^{-10}

for all).

Figure S7: Effect size preservation (simulated trait). PRS univariate betas from reconstructed (left, blue) and synthetic (right, orange) data vs. real-data betas. Synthetic data show higher agreement with real betas (Pearson $r=0.835$ ) than reconstructed data ( $r=0.726$ ), indicating that phenotype-conditioned generation preserves marginal associations more faithfully than unconditional VAE reconstruction.

Figure S8: XGBoost ROC-AUC on the held-out real test set as a function of augmentation proportion (fraction of the class-balanced synthetic dataset used for training) across five traits. Three subsampling strategies are compared: random (blue, mean

\pm

std over 3 seeds), diversity-aware (orange), and PCA-stratified (green). Dotted horizontal lines indicate non-augmented baselines.

Synthetic Data Generation

Synthetic data generation creates artificial examples, simulations, and datasets to train, test, or protect machine-learning systems.

See all 625 papers

HELICS: Biobank-scale Conditional Synthetic Genome Generation via Latent Flow Matching
Oct 26, 2026Ahmad Abdel-Azim, Xihong LinSynthetic Data Generation Genomics
Department of Biostatistics, Harvard University, Boston, MA 02135 · Department of Statistics, Biostatistics, Harvard University, Boston, MA 02135
PSyGenTAB: A Privacy-Preserving Framework for Synthetic Clinical Tabular Data Generation via Constrained Optimization
Jun 16, 2026Arshia Ilaty, Hossein Shirazi, Manasi Chitale +5Synthetic Tabular Data Generation Tabular Data Synthesis
San Diego State University, San Diego, CA 92182 USA. · University of California, Irvine, CA 92697 USA.
Sequential versus fully conditional specification for tree-based synthesis of health data
Oct 26, 2026R{é}my Chapelle, Mohammed SedkiSynthetic Clinical Data Generation Tabular Data Synthesis
CB, CESP, EVDG · Université Paris-Saclay, UVSQ, Inserm, CESP, 94807, Villejuif, France · Université Paris-Saclay, Université Paris Cité, ENS Paris-Saclay, CNRS, SSA, Inserm, Centre Borelli, 91190, Gif-sur-Yvette, France +2

Oct 26, 2026J/K move · Enter open · S save

HELICS: Biobank-scale Conditional Synthetic Genome Generation via Latent Flow Matching

Ahmad Abdel-Azim, Xihong Lin

Department of Biostatistics, Harvard University, Boston, MA 02135 · Department of Statistics, Biostatistics, Harvard University, Boston, MA 02135

Synthetic Data Generation Genomics Conditional Generative Modeling Flow Matching

Open paper AskPDF arXiv

Individual-level genotype data underpin genetic discovery, disease-risk modeling, and therapeutic development, yet many critical settings remain data-limited, including underrepresented ancestries, rare or imbalanced disease cohorts, and restricted-access biobanks. Conditional synthetic genome generation would support privacy-aware benchmarking, simulation, and data augmentation, but requires learning an ultra-high-dimensional discrete distribution over hundreds of thousands to tens of millions of correlated single-nucleotide polymorphisms (SNPs), or sites of DNA variation. We introduce HELICS: High-fidelity gEneration via Latent Interpolant Conditional Synthesis, a phasing-free latent flow matching framework for genome-wide SNP-level generation from biobank-scale genotype data. HELICS tokenizes chromosomes with convolutional autoencoders, reducing >600K SNPs to 4,096 continuous latent tokens, and trains a transformer parameterized conditional flow matching model over the concatenated multi-chromosome latent representation. Trained on over 417K genomes in the UK Biobank, HELICS generates genome-wide common-variant synthetic cohorts without haplotype phasing and achieves stronger fidelity to real genotypes than existing haplotype-based simulators, including high allele-frequency agreement in external evaluation and superior preservation of local and long-range SNP correlation structure. Conditioning on ancestry and genetic liability summaries enables targeted generation across genetic risk profiles; in perturbation analyses, by increasing the conditioned disease trait liability, HELICS induces SNP-level dosage changes in generated genomes that are highly correlated with published association weights. HELICS provides a scalable generative modeling framework for SNP-level genome data and a path toward conditional synthetic cohorts for data-limited genetic studies.

Abstract

Synthetic Data Generation

Explore similar work

HELICS: Biobank-scale Conditional Synthetic Genome Generation via Latent Flow Matching

PSyGenTAB: A Privacy-Preserving Framework for Synthetic Clinical Tabular Data Generation via Constrained Optimization

Sequential versus fully conditional specification for tree-based synthesis of health data