39: Scaling whole-genome polygenic scores with VIPRS

Zabad S et al., The American Journal of Human Genetics - This episode covers Zabad et al.'s methods to scale summary-statistics-based polygenic risk score (PRS) inference to millions of variants. The authors introduce compressed LD storage, memory-efficient coordinate-ascent variational algorithms, and multi-level parallelism to cut storage, runtime, and RAM by orders of magnitude while retaining competitive prediction accuracy. Key terms: polygenic risk scores, linkage disequilibrium, variational inference, LD compression, VIPRS.

Study Highlights:
The authors design a compact LD-matrix format (CSR stored in Zarr with quantization) and algorithmic optimizations that reduce LD storage by over 50-fold. They reimplement coordinate-ascent variational updates in C/C++ using single-precision floats, triangular-LD updates, dequantize-on-the-fly, and two layers of parallelism to cut runtime and memory use by orders of magnitude. VIPRS v0.1 can run variational Bayesian regression on 1.1M HapMap3 variants in under a minute and converges genome wide on up to 18M variants in tens of minutes using <15 GB RAM. The paper also analyzes spectral causes of numerical instability in LD matrices and gives practical recommendations to improve stability and prediction accuracy.

Conclusion:
The updated VIPRS toolkit enables fast, memory-efficient whole-genome PRS inference at biobank scale with competitive accuracy and provides storage formats and numerical safeguards to improve reproducibility and portability.

Music:
Enjoy the music based on this article at the end of the episode.

Article title:
Toward whole-genome inference of polygenic scores with fast and memory-efficient algorithms

First author:
Zabad S

Journal:
The American Journal of Human Genetics

DOI:
10.1016/j.ajhg.2025.05.002

Reference:
Zabad S., Haryan C.A., Gravel S., Misra S., Li Y. (2025). Toward whole-genome inference of polygenic scores with fast and memory-efficient algorithms. The American Journal of Human Genetics 112, 1–19. https://doi.org/10.1016/j.ajhg.2025.05.002

License:
This episode is based on an open-access article published under the Creative Commons Attribution 4.0 International License (CC BY 4.0) – https://creativecommons.org/licenses/by/4.0/

Support:
Base by Base – Stripe donations: https://donate.stripe.com/7sY4gz71B2sN3RWac5gEg00

Official website https://basebybase.com

On PaperCast Base by Base you'll discover the latest in genomics, functional genomics, structural genomics, and proteomics.

Episode link: https://basebybase.com/episodes/viprs-whole-genome-prs

QC:
This episode was checked against the original article PDF and publication metadata for the episode release published on 2025-06-07.

QC Scope:
- article metadata and core scientific claims from the narration
- excludes analogies, intro/outro, and music
- transcript coverage: Audited the transcript's presentation of VIPRS architecture (LD storage, quantization, DQF, triangular LD), memory/performance benchmarks, parallelism, numerical stability guards, and cross-ancestry/cross-biobank findings against the original article.
- transcript topics: Polygenic risk scores and LD challenges; LD matrix compression via upper-triangular storage; CSR storage and Zarr cloud-native format; Quantization to int8/int16 and scale quantization; Dequantize-on-the-Fly (DQF) memory management; Coordinate ascent updates and OpenMP parallelism

QC Summary:
- factual score: 10/10
- metadata score: 10/10
- supported core claims: 8
- claims flagged for review: 0
- metadata checks passed: 4
- metadata issues found: 0

Metadata Audited:
- article_doi
- article_title
- article_journal
- license

Factual Items Audited:
- LD matrix compression reduces storage by >50-fold; 1.4M HapMap3 variants stored in ~300 MB
- LD matrices stored in CSR format with quantization to int8/int16 (scale quantization)
- Dequantize-on-the-Fly (DQF) streams data and avoids full in-memory decompression, reducing memory usage
- Triangular LD mode reduces memory usage by about 40% compared with symmetric LD mode
- Two layers of parallelism: across chromosomes and within coordinate-ascent; ~30% total runtime reduction with 4 threads
- VIPRS v0.1 can infer 1.1M HapMap3 variants in under a minute; converges on up to 18M variants in tens of minutes using <15 GB RAM

QC result: Pass.