Research map / Image generation / Fast diffusion sampling
HELICS: Biobank-scale Conditional Synthetic Genome Generation via Latent Flow Matching
Ahmad Abdel-Azim, Xihong Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HELICS, a latent flow matching model for conditional synthetic genome generation at biobank scale, aiming to produce individual-level genotypes for data-limited settings.
Problem
Individual-level genotype data is scarce for underrepresented ancestries, rare or imbalanced disease cohorts, and restricted-access biobanks, and conditional synthetic genome generation requires learning an ultra-high-dimensional genotype distribution.
Approach
Uses latent flow matching to model ultra-high-dimensional genotypes in a compressed latent space, enabling conditional generation of synthetic genomes (details beyond the abstract not available).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train HELICS only on EUR samples instead of multi-ancestry data (ablation) | HELICS · Chr15 EUR vs CCDG | LD difference: 0.115 -> 0.067 | Improved |
| train HELICS only on EUR samples instead of multi-ancestry data (ablation) | HELICS · Chr15 EUR vs CCDG | low-MAF R^2: 0.362 -> 0.493 | Improved |
| train HELICS only on EUR samples instead of multi-ancestry data (ablation) | HELICS · Chr15 EUR vs CCDG | MAF R^2: 0.992 -> 0.962 | Hurt |
| add MAF and LD penalties and condition-prediction loss to the autoencoder | HELICS autoencoder · UKB | cohort-level fidelity | No clear effect |
| add late-stage latent regularization (TV, robustness, stability) to the autoencoder (ablation) | HELICS autoencoder · UKB validation (per chromosome) | validation reconstruction MSE | Improved |
| add late-stage latent regularization to the autoencoder (ablation) | HELICS autoencoder · UKB validation (per chromosome) | latent TV, robustness and stability penalties | Improved |
| HELICS (genome-wide, unphased) versus Sim1000G on regional MAF matching | HELICS · 22 autosomes | MAF R^2: 0.985 vs 0.980 | Hurt |
| HELICS versus HAPNEST on LD discrepancy | HELICS · 22 autosomes | LD difference: 0.118 -> 0.109 | Improved |
| condition on HDL PRS decile 1 to 10 | HELICS · UKB validation | generated HDL PRS: Bin 10 > 1 fraction 0.98 | Improved |
| post-training spectral and PC-alignment refinement of multi-ancestry generator (ablation) | HELICS · multi-ancestry | low-frequency MAF agreement | Mixed |
| fewer training samples for underrepresented ancestries | HELICS · ancestry-stratified UKB | latent alignment and fidelity | Hurt |
Closest papers by meaning
- LineageFlow: Flow Matching for High-Fidelity Family-Aware Protein Sequence Generation (ICML 2026 · similarity 0.84)
- Training Flexible Models of Genetic Variant Effects from Functional Annotations using Accelerated Linear Algebra (ICML 2025 · similarity 0.83)
- Sequence-Augmented SE(3)-Flow Matching For Conditional Protein Generation (NeurIPS 2024 · similarity 0.83)
- DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genomes (ICLR 2024 · similarity 0.83)
- Hyperbolic Genome Embeddings (ICLR 2025 · similarity 0.82)
- Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
