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

Ahmad Abdel-Azim; Xihong Lin

q-bio.GNOct 26, 2026

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

Authors: Ahmad Abdel-Azim, Xihong Lin

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

Abstract

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.

Figures & tables

Figure 1: Chromosome-specific conditional autoencoder architecture. Unphased SNP dosages are encoded into chromosome-level latent tokens and decoded back to genotype logits (left) via residual downsampling and upsampling blocks (right).

Figure 2: Latent flow matching for conditional genome generation. (a) A conditional USiT model architecture is used to learn a velocity field over concatenated chromosome latent tokens. (b) Synthetic genomes are generated by integrating over the learned velocity field, splitting generated tokens by chromosome and decoding with trained decoders.

Approach	GW	No. SNPs	MAF $R^{2}$ $\uparrow$	Low MAF $R^{2}$ $\uparrow$	MAF MAE $\downarrow$	LD diff. $\downarrow$	Membership $\downarrow$	Gen. NN IBS $\downarrow$
HAPNEST [ 56 ]	✓	10,509	0.942	0.207	0.025	0.101	0.030	0.914
Sim1000G [ 19 ]	✗	9,899	0.988	0.637	0.011	0.122	0.011	0.839
HAPGEN2 [ 50 ]	✗	17,117	0.976	0.444	0.016	0.069	0.005	0.879
GAN-based [ 60 ]	✗	6,513	0.947	0.129	0.023	0.129	0.060	0.923
RBM-based [ 60 ]	✗	7,531	0.829	0.191	0.032	0.110	0.001	0.927
HCLT-based [ 18 ]	✗	6,827	0.846	0.140	0.036	0.152	0.029	0.906

Table 1: Chromosome 15 benchmarks against external CCDG cohort, with UKB and 1KG as reference baselines. GW: whether a method is designed for genome-wide generation. Each method uses its available overlap with the harmonized common chr15 reference panel. Low-MAF

R^{2}

R^{2}

for

0.01<\mathrm{MAF}<0.05

; LD diff.: Frobenius norm of the LD difference; Membership:

|\mathrm{AUC}-0.5|

for train-validation distinguishability by distance from synthetic set; Gen. NN IBS: generated nearest-neighbor IBS similarity.

Figure 3: Synthetic genome fidelity on chromosome 15. Panels show (a) LD decay, (b) genotype-PC projections, (c) local LD heatmaps (chr15:24,321,414–24,848,545), and (d) MAF agreement. Restricted regional baselines are omitted from these chromosome-scale evaluations.

Approach	Scope	No. SNPs	MAF $R^{2}$ $\uparrow$	Low MAF $R^{2}$ $\uparrow$	MAF MAE $\downarrow$	LD diff. $\downarrow$	Membership $\downarrow$	Gen. NN IBS $\downarrow$
HAPNEST [ 56 ]	Genome-wide	349,564	$0.926\pm 0.003$	$0.177\pm 0.012$	$0.027\pm 0.001$	$0.118\pm 0.003$	$0.028\pm 0.020$	$0.820\pm 0.004$
Sim1000G [ 19 ]	Regional	$4,973\times 22$	$0.985\pm 0.001$	$0.579\pm 0.034$	$0.012\pm 0.0004$	$0.127\pm 0.026$	$0.020\pm 0.017$	$0.843\pm 0.006$
HELICS	Genome-wide	600,336	$0.980\pm 0.009$	$0.489\pm 0.199$	$0.023\pm 0.006$	$0.109\pm 0.013$	$0.013\pm 0.006$	$0.893\pm 0.027$
HELICS-EUR	Genome-wide	600,336	$0.976\pm 0.009$	$0.502\pm 0.069$	$0.023\pm 0.003$	$0.107\pm 0.025$	$0.030\pm 0.010$	$0.866\pm 0.013$

Table 2: Metrics and abbreviations follow Table 1 . Values are means

\pm

SD across 22 chromosomes or, for Sim1000G, independently simulated regions for each autosome.

Figure 4: Genome-wide latents preserve human genetic structure. (a) UKB genotype PCs for $100{,}000$ individuals colored by self-reported ancestry. (b) UMAP of held-out validation latents ( $N=5{,}000$ ) using token-standardized Frobenius distances. (c) Real and generated latents overlap in the same embedding. (d) The embedding colored by genotype PCs shows preserved axes of genetic variation.

Figure 5: Conditional PRS control schematic. HELICS generates

\widetilde{\boldsymbol{G}}_{i}

conditional on a validation profile

\boldsymbol{c}_{i}

. High correlation between generated and observed PRS indicates effective conditional control.

Figure 6: HDL conditioning produces association-consistent SNP-level responses. (a) Trait-guided perturbed generation controls HDL polygenic profiles with minimal off-target effects. (b) Among the largest HDL PRS-contributing shifts, most generated dosage changes $\Delta G_{j}$ move in the direction predicted by the published HDL weight $\beta_{j}$ , yielding positive contributions $\beta_{j}\Delta G_{j}$ . (c) Across HDL score variants, bin-10 versus bin-1 dosage shifts are positively associated with published HDL weights, with large effects at known lipid loci.

Trait	Variants	Gen-Obs. $r$	Bin 10 $>$ 1	Top-20
HDL	168	0.66	0.98	0.98
LDL	606	0.75	1.00	0.94
T1D	16,570	0.63	1.00	1.00
RA	14,728	0.71	0.99	1.00
ApoB	43,001	0.71	0.95	0.80

Table 3: Conditional-control summary across HDL, LDL, T1D, rheumatoid arthritis (RA), and apolipoprotein B (ApoB). For each trait, we report the number of variants included in the PRS, validation-profile PRS recovery (Gen–Obs.

r

; see Figure 5 ), low-to-high perturbation success (Bin 10

>

1), and contribution-weighted directional alignment among the top 20 SNP contributors to PRS shift.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Chromosome	$L_{c}$	$L_{c}^{\mathrm{work}}$	$m_{c}$	$T_{c}$	$N_{c}^{\mathrm{down}}$
1	47,555	49,152	3	256	6
2	48,530	49,152	3	256	6
3	40,846	40,960	5	256	5
4	38,621	40,960	5	256	5
5	36,643	40,960	5	256	5
6	42,638	49,152	3	256	6

Appendix

Table B1: Chromosome-specific autoencoder configurations.

L_{c}

is the number of retained SNPs,

L_{c}^{\mathrm{work}}

is the padded working length,

T_{c}

is the allocated latent token count, and

N_{c}^{\mathrm{down}}

is the number of residual factor-two downsampling blocks.

Hyperparameter	Value
Optimizer	AdamW, $\beta=$ $(0.9,0.99)$
Learning rate	$2\times 10^{-4}$
Learning-rate schedule	None
Batch size	$128$
Number of epochs	50
Embedding dimension $d_{e}$	$8$

Appendix

Table B2: Autoencoder training hyperparameters.

Figure B1: Regularization ablation for chromosome-specific autoencoders. For each chromosome, we compare validation metrics at the last epoch before $\lambda_{\rm reg}>0$ , against the final epoch after regularization. Regularization consistently reduces latent total variation, robustness loss, and stability loss, while continuing to improve validation reconstruction MSE.

Figure B2: USiT block schematic demonstrating long-range skip connections.

Parameter	Value
Hardware	$4\times$ NVIDIA H200 GPUs
Training budget	72 Hours
Input latent shape	$(4096,256)$
Backbone	USiT
Transformer hidden width	$1152$
Transformer layers	$29$

Appendix

Table B3: Latent flow matching training and sampling configuration.

Figure C1: Effect of genotype-space PC alignment on chromosome-wise MAF fidelity. Each panel compares UKB reference MAF to synthetic MAF for one chromosome. Raw generated genotypes show distortions in marginal allele frequencies, consistent with a global population-structure misalignment. The PC-alignment correction substantially improves marginal agreement across chromosomes, and the final discretized calls preserve most of this improvement.

Figure C2: Effect of retained PC dimension $k$ and genotype discretization strategy after PC alignment. (a) MAF $R^{2}$ correspondence across several values of $k$ for all variants, variants with $0.01<MAF<0.2$ , and $0.01<MAF<0.05$ . (b) Top row: low-frequency MAF ( $<0.1$ ) agreement under stochastic genotype calling versus deterministic rounding. Representative LD matrices are also shown after stochastic calling (middle row) and deterministic rounding (bottom row).

Figure C3: Effect of PC alignment on local LD structure. A representative LD block (chr15:24,321,414–24,848,545) is shown for the UKB reference, raw generated genotypes, PC-aligned generated genotypes, and PC-aligned genotypes after final rounding. A symmetrized LD-decay curve, computed from the center of the LD block, is also shown.

Figure C4: Effect of token-wise spectral truncation on multi-ancestry population structure. (a) Overlay of generated samples on UKB genotype-PC space. (b) UKB ancestry structure based on chromosome 15 variants. (c) PC-projected generated samples without spectral thresholding; generated samples collapse in PC space and ancestry groups are poorly separated. (d) Generated samples after spectral thresholding.

Trait	PGS Catalog accession
Standing height	PGS002332
Coronary artery disease	PGS003725
Alzheimer’s disease (dementia)	PGS003958
Major depression	PGS003333
Body mass index	PGS005198
High-density lipoprotein cholesterol	PGS004914

Appendix

Table D1: Polygenic scores used to construct trait-conditioning variables.

PGS002960 (external) variant set	Variants scored	Generated-observed $r$	Real cross-score $r$
All matched variants	2,627	0.189	0.300
After LD exclusions	2,608	0.135	0.194

Appendix

Table D2: External HDL-score validation. Generated-observed

r

correlates continuous PGS002960 scores between generated genomes and their paired real genomes under validation-profile conditioning. Real cross-score

r

summarizes the association between PGS002960 and PGS004914 within the real cohort. The second row excludes external-score variants with

r^{2}\geq 0.8

to any matched conditioning-score variant on the same chromosome.

Figure D1: Classifier-free guidance strengthens PRS control in HELICS. Each panel compares standardized generated HDL PRS to standardized observed HDL PRS for held-out validation conditioning profiles under a different guidance scale

s

; see Figure 5 for an illustration of this procedure. As

s

increases from

0

2

, generated HDL PRS becomes more strongly correlated with the conditioning profile, indicating tighter conditional control.

Training $N$	Training chr.	MAF $R^{2}\uparrow$	Low-MAF $R^{2}\uparrow$	MAF MAE $\downarrow$	LD diff. $\downarrow$
2,500	22	0.978	0.187	0.0199	0.205
5,000	22	0.975	0.169	0.0220	0.221
10,000	22	0.968	0.179	0.0284	0.226
100,000	22	0.983	0.368	0.0173	0.210
251,465	1-22	0.991	0.678	0.018	0.120

Appendix

Table F1: Training-size sensitivity under external CCDG-EUR evaluation.

Flow model (parameter count)	MAF $R^{2}\uparrow$	Low-MAF $R^{2}\uparrow$	MAF MAE $\downarrow$	LD diff. $\downarrow$
Tiny (1.18M)	0.221	0.002	0.187	0.251
Small (7.98M)	0.668	0.013	0.112	0.257
Medium (25.5M)	0.978	0.219	0.027	0.249
Large (58.9M)	0.983	0.260	0.024	0.248

Appendix

Table F2: Flow-model capacity sensitivity on chromosome 22 under external CCDG-EUR evaluation. The autoencoder, training data (

N=100{,}000

), and training schedule are fixed. Parameter counts refer to the flow model.

Reference	MAF $R^{2}$	MAF MAE	Low-MAF $R^{2}$	LD diff.
1000G-EUR	$0.985\pm 0.001$	$0.012\pm 0.0004$	$0.579\pm 0.034$	$0.127\pm 0.026$
UKB-EUR	$0.990\pm 0.001$	$0.010\pm 0.0004$	$0.691\pm 0.050$	$0.122\pm 0.026$

Appendix

Table F3: Sim1000G reference-panel sensitivity against CCDG-EUR. Values are means

\pm

standard deviation across 22 independently simulated regions, with one region sampled from each autosome and

4{,}682

evaluated variants per region on average.

Synthetic samples vs. training reference
Relatedness category	Sim1000G	HELICS	HAPGEN2	HAPNEST
MZ twins $\downarrow$	0	0	0	0
1st degree $\downarrow$	0	0	0	0
2nd degree $\downarrow$	0	0	0	0
3rd degree $\downarrow$	15,656	21,645	1,779	29,114
Unrelated $\uparrow$	4,984,344	24,978,355	24,998,221	24,970,886

Appendix

Table G1: KING kinship analysis on the chromosome 15 benchmark panel. Genetic relatedness is estimated from genotype data, and pairs are categorized by the inferred degree of relatedness. The validation-to-training panel provides a matched baseline for each method’s variant overlap set. Validation samples are unrelated held-out reference individuals compared against the same training reference used for each method. Fewer close relatives and more unrelated pairs indicate lower train proximity.

Chr	MAF $R^{2}$ $\uparrow$	Low-MAF $R^{2}$ $\uparrow$	MAF $<0.2$ $R^{2}$ $\uparrow$	MAF MAE $\downarrow$	LD diff. $\downarrow$	LD blocks
1	0.9805	0.4967	0.9205	0.0248	0.1414	753
2	0.9735	0.5144	0.9037	0.0229	0.1452	808
3	0.9765	0.5307	0.9117	0.0269	0.1393	639
4	0.9774	0.5203	0.9106	0.0258	0.1327	631
5	0.9739	0.4972	0.8955	0.0262	0.1327	597
6	0.9793	0.6279	0.9197	0.0261	0.1362	636

Appendix

Table G2: Chromosome-specific fidelity metrics for HELICS-EUR based on 5,000 generated samples. Metrics are computed separately for each autosome using harmonized UKB training and generated variants. MAF

R^{2}

and MAF MAE compare generated and UKB training minor allele frequencies across all variants on the chromosome. Low-MAF

R^{2}

is the

R^{2}

restricted to low-frequency variants defined by UKB MAF

<0.05

; MAF

<0.2

R^{2}

is the corresponding

R^{2}

for variants with UKB MAF

<0.2

. LD diff. is the mean normalized Frobenius norm of the discrepancy between UKB and generated SNP–SNP LD correlation matrices within predefined LD blocks; lower values indicate better LD fidelity. LD blocks denotes the number of evaluated blocks on each chromosome.

Chr	MAF $R^{2}$ $\uparrow$	Low-MAF $R^{2}$ $\uparrow$	MAF $<0.2$ $R^{2}$ $\uparrow$	MAF MAE $\downarrow$	LD diff. $\downarrow$	LD blocks
1	0.9790	0.2032	0.8503	0.0330	0.1616	753
2	0.9876	0.5716	0.9566	0.0179	0.1082	808
3	0.9871	0.8617	0.9830	0.0170	0.0908	639
4	0.9750	0.7342	0.9585	0.0232	0.0909	631
5	0.9886	0.3162	0.9159	0.0240	0.1428	597
6	0.9868	0.7083	0.9546	0.0168	0.1113	636

Appendix

Table G3: Chromosome-specific fidelity metrics for HELICS based on 5,000 generated samples. Metrics are computed separately for each autosome using harmonized UKB training and generated variants. MAF

R^{2}

and MAF MAE compare generated and UKB training minor allele frequencies across all variants on the chromosome. Low-MAF

R^{2}

is the

R^{2}

restricted to low-frequency variants defined by UKB MAF

<0.05

; MAF

<0.2

R^{2}

is the corresponding

R^{2}

for variants with UKB MAF

<0.2

Ancestry	$N_{\mathrm{train}}$	MAF $R^{2}\uparrow$	MAF MAE $\downarrow$	Low-MAF $R^{2}\uparrow$	LD diff. $\downarrow$
EUR	391,125	$0.980\pm 0.010$	$0.023\pm 0.006$	$0.488\pm 0.199$	$0.109\pm 0.013$
SAS	9,033	$0.807\pm 0.067$	$0.046\pm 0.007$	$0.080\pm 0.026$	$0.112\pm 0.011$
EAS	5,576	$0.718\pm 0.068$	$0.049\pm 0.006$	$0.008\pm 0.006$	$0.117\pm 0.008$
AFR	7,291	$0.710\pm 0.084$	$0.051\pm 0.007$	$0.069\pm 0.054$	$0.116\pm 0.017$

Appendix

Table G4: Ancestry-specific genotype fidelity of multi-ancestry HELICS across all 22 autosomes.

N_{\mathrm{train}}

denotes the number of training individuals in each ancestry group: European (EUR), South Asian (SAS), East Asian (EAS), and African (AFR). Values are means

\pm

standard deviations across chromosomes. Allele-frequency fidelity is weaker in underrepresented groups, particularly for low-frequency variants.

Figure G1: Chromosome-resolved EUR genotype fidelity against the external CCDG-EUR cohort across all 22 autosomes. Rows show (top to bottom) overall MAF $R^{2}$ , low-frequency MAF $R^{2}$ ( $0.01<\mathrm{MAF}<0.05$ ), MAF MAE, and mean normalized Frobenius LD discrepancy. HELICS, HELICS-EUR, and HAPNEST generate genome-wide samples; Sim1000G is evaluated on one independently simulated region per chromosome. Each method uses its available overlap with the harmonized reference panel. Higher $R^{2}$ and lower MAE and LD discrepancy indicate closer agreement.

Ancestry	$N_{\mathrm{train}}$	Sliced $W_{2}$	RBF MMD 2
White	391,125	0.002 $\pm$ 0.0001	0.002 $\pm$ 0.0003
South Asian	9,033	0.042 $\pm$ 0.001	0.315 $\pm$ 0.016
SE Asian	5,576	0.078 $\pm$ 0.003	0.331 $\pm$ 0.016
Black	4,230	0.075 $\pm$ 0.002	0.846 $\pm$ 0.037
African	3,061	0.085 $\pm$ 0.003	0.564 $\pm$ 0.016

Appendix

Table G5: Within-ancestry alignment between real and HELICS-generated genome-wide latent representations in the multi-ancestry model.

N_{\mathrm{train}}

denotes the number of training individuals from each ancestry group used to fit HELICS. For each ancestry, we compare real and generated latent distributions using sliced Wasserstein distance and RBF MMD 2 , where smaller values indicate closer agreement. Means and standard errors are estimated from 250 bootstrap replicates, each formed by sampling 542 real and 542 generated latents within ancestry. The results show very close agreement for the majority White ancestry group, which is also the largest training group, and systematically weaker alignment for underrepresented ancestry groups. This pattern is consistent with the reduced representation of these groups during training and suggests that latent-space fidelity tracks training sample size in the multi-ancestry setting.

Figure G2: UMAP visualization of HELICS latent representations for real validation individuals and generated individuals, shown separately for each chromosome and for the concatenated genome-wide latent space. Each chromosome-specific panel shows a two-dimensional UMAP embedding of the corresponding chromosome latent tokens for $1{,}000$ held-out real individuals and $1{,}000$ generated individuals. The large panel at right shows the same comparison using the concatenated latent representation across all chromosomes. Across nearly all chromosomes, generated latents overlap closely with real latents and do not form a clearly separated cluster, suggesting that the latent flow model covers the support of the chromosome-specific latent spaces rather than generating isolated or low-diversity samples. The bar plots summarize chromosome-wise sliced Wasserstein distance and RBF MMD 2 between real and generated latent distributions; lower values indicate better alignment. These quantitative summaries mirror the visual UMAP overlap and show that latent fidelity is broadly consistent across chromosomes.

Figure G3: Trait-guided perturbation analyses for low-density lipoprotein (LDL; a), apolipoprotein B (ApoB; b), rheumatoid arthritis (RA; c), and type 1 diabetes (T1D; d). Each row repeats the same analysis used for HDL in Figure 6 . Left: PRS computed for generated genomes conditioned on validation sample conditioning profiles is compared to the observed PRS of the corresponding real validation individual; the reported $r$ measures conditional control success. Second: the target PRS conditioning bin is swept from 1 to 10 while ancestry and all other trait conditions are fixed; violins show the distribution of generated PRS and points show the mean with 95% confidence intervals. Third: the top 20 variants with the largest absolute PRS-dosage shifts contributions ( $\beta_{j}\Delta G_{j}$ ) are decomposed into the published PRS weight $\beta_{j}$ , the generated effect-allele dosage shift $\Delta G_{j}=\mathbb{E}[\widetilde{G}_{j}(10)]-\mathbb{E}[\widetilde{G}_{j}(1)]$ . Right: across matched PRS variants, generated dosage shifts are compared to published PRS weights; points outside the shaded low-effect region are used for the displayed linear fit, and color denotes $\beta_{j}\Delta G_{j}$ . Across all traits, HELICS recovers individual-level PRS variation, produces monotone low-to-high PRS shifts under conditioning-bin perturbations, and induces SNP-level dosage changes that are broadly concordant with large published association directions, despite being trained only on discretized PRS summaries rather than variant-level PRS weights.

Trait	Ancestry	$N_{\mathrm{corr}}$	$N_{\mathrm{sweep}}$	Matched PGS variants	Gen.-obs. $r$	Bin 10 $>$ bin 1 mean	Top-20 alignment
HDL	EUR	9,366	950	168	0.684	0.977	0.950
HDL	SAS	213	28	168	0.230	1.000	0.950
HDL	AFR	172	23	168	0.244	0.913	0.900
LDL	EUR	9,366	950	606	0.762	0.997	0.800
LDL	SAS	213	28	606	0.221	1.000	0.750
LDL	AFR	172	23	606	0.308	0.957	0.850

Appendix

Table G6: Ancestry-stratified conditional-control results from multi-ancestry HELICS for European (EUR), South Asian (SAS), and African (AFR) ancestry groups. The perturbation analysis partitions the same

1{,}001

validation profiles used in Table 3 .

N_{\mathrm{corr}}

denotes the number of paired real and generated genomes used to compute the Pearson correlation between continuous PRSs under validation-profile conditioning (Gen.-obs.

r

);

N_{\mathrm{sweep}}

denotes the number of profiles used for each perturbation condition. Perturbations change only the target trait’s PRS decile, holding ancestry and other trait conditions fixed. Bin 10

>

bin 1 is the proportion of profiles whose generated PRS under bin-10 conditioning exceeds that under bin-1 conditioning for the same profile. Top-20 sign fraction is the unweighted fraction of the 20 variants with the largest absolute PRS contributions

|\beta_{j}\Delta G_{j}|

satisfying

\beta_{j}\Delta G_{j}>0

, with dosage shifts and rankings computed separately within each ancestry. Low-to-high PRS responses and SNP-level directional alignment remain broadly preserved across groups, while generated-observed PRS correlations are lower in SAS and AFR.

Synthetic Data Generation

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

See all 625 papers

Mar 11, 2026

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

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.

Andrea Lampis, Michela Carlotta Massi, Nicola Pirastu +3

DEIB, Politecnico di Milano, Milan 20133, Italy · Health Data Science Centre, Human Technopole, Milan 20157, Italy · Genomics Research Centre, Human Technopole, Milan 20157, Italy +2

Synthetic Data Generation Genomics Conditional Generative Modeling Privacy-Preserving ML Latent Diffusion Models

Ask PDF

Oct 26, 2026

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

Disseminating synthetic data in place of original ones is a promising way to enforce privacy while complying with open science principles, especially in the healthcare domain. To keep pace with the rate at which researchers collect such data, there is a need for fast and flexible synthesis procedures. A popular class of such procedures is based on models of conditional distributions induced by decision trees. Such models can be combined into a synthetic distribution following two main approaches: sequential specification and fully conditional specification. However, there is currently no recommendation on which specification to use to optimize the risk-utility tradeoffs. In this work, we use both specifications to synthesize medical datasets by combining tree-based models of conditional distributions. We compare the quality of the resulting synthetic data using several metrics of risk and utility. Our results indicate no general superiority of either specification over the other. Sequential synthesis seems sensitive to the ordering of the variables, while the fully conditional specification involves more hyperparameters and incurs higher computational costs, which may limit its suitability for off-the-shelf synthesis procedures.

R{é}my Chapelle, Mohammed Sedki

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

Synthetic Clinical Data Generation Tabular Data Synthesis Synthetic Data Evaluation Conditional Generative Modeling

Ask PDF

Jun 5, 2026

Accelerating Reproducible Research in Synthetic EHR Generation

The generation of high-fidelity synthetic Electronic Health Records (EHR) is crucial for advancing medical research while preserving patient privacy. However, head-to-head comparison of existing generative models is hindered by disjointed codebases, incompatible data loaders, conflicting library dependencies, and inconsistent evaluation protocols. To address these gaps, we introduce a lightweight, end-to-end benchmarking framework for reproducible synthetic EHR evaluation, organized as a unified pipeline spanning data ingestion, standardized model training, and architecture-agnostic evaluation. Our current implementation targets the generation of longitudinal ICD diagnosis codes -- the most commonly studied modality in this literature -- and is built on the community-maintained PyHealth library. We reimplement and unify strong baselines (MedGAN, CorGAN, PromptEHR, HALO) under full ICD-9 vocabulary granularity, and add a lightweight GPT-2 baseline from the general-purpose sequence-modeling literature. We contribute a rigorous, architecture-agnostic privacy-utility evaluation suite that applies identically to GAN- and transformer-based generators, and report bootstrapped confidence intervals across all metrics. We further analyze the poor long-tailed performance of existing models and discuss the extensibility of our framework beyond diagnosis codes. By lowering the engineering barrier to running, extending, and evaluating under a single pipeline, we introduce a starting point for community-driven reproducibility and benchmarking synthetic EHR models.

Jalen Jiang, Chufan Gao, Ethan Rasmussen +2

University of Illinois Urbana-Champaign

Synthetic Data Generation ML Reproducibility Longitudinal EHR Modeling Synthetic Data Evaluation

Ask PDF

Abstract

Synthetic Data Generation

Explore similar work

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

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

Accelerating Reproducible Research in Synthetic EHR Generation