bpred LD reference (bpred_ldref)

A pruned, windowed, block-diagonal linkage-disequilibrium (LD) reference panel for 1,141,514 HapMap3+ variants, packaged as the runtime LD input consumed by bpred (Bayesian polygenic-score training from GWAS summary statistics). It is a reformatted, pruned derivative of Florian Privé's published LDpred2 HapMap3+ LD reference.

Contents

The archive bpred_ldref.zip (ZIP64, uncompressed / ZIP_STORED, ~4.77 GB) unpacks to:

bpred_ldref/
  snps.parquet         variant index, genome order (1,141,514 rows)
  blocks/chr{1..22}.npz windowed block-diagonal LD, one file per chromosome

snps.parquet columns:

column	dtype	meaning
ldref_id	Int64	0…n−1, contiguous in genome order (the panel's row key)
rsid	string	dbSNP rsID
chr	string	chromosome (1–22)
pos_hg19	Int64	position, GRCh37/hg19
pos_hg38	Int64	position, GRCh38/hg38 (nullable; 827 variants have no map)
ref	string	reference allele (a0 in the LDpred2 map)
alt	string	alternate allele (a1)
af	Float32	allele frequency (af_UKBB, UK Biobank)

Each blocks/chr{c}.npz stores the chromosome's LD as per-row bands, partitioned into the LD blocks the Gibbs sampler parallelizes over. Within a block, correlation is windowed to a fixed genetic-distance neighborhood, so each row is a contiguous band around the diagonal.

Note: the correlation matrix is symmetric, so only the lower triangle of each per-row band is stored (columns [start_i, i]). This halves the on-disk and download size; the full symmetric band is reconstructed on load (bpred.core.block_ld.load_block_full). Consumers other than bpred must mirror the lower triangle to recover the upper half.

How it was built

Derived from the LDpred2 HapMap3+ reference (see Source & attribution) by:

1. Greedy LD pruning within each native LD block at r² > 0.99 (the higher-MAF variant kept), removing near-redundant variants.
2. Windowing / banding each kept block to its in-window neighborhood and storing only the per-row bands.
3. Re-serialization to snps.parquet + per-chromosome .npz.

The build is deterministic: re-running it on a fixed toolchain reproduces a byte-identical archive.

• Archive sha256: dc068176b88302f070f0bb44c689c649bf4a38da2fbf846bb852c049a94c8237

Usage

bpred fetches and verifies this archive automatically:

python -m bpred.utils.download_ldref   # download + sha256 check + unpack into {ld_dir}/bpred_ldref

Or load snps.parquet directly with any Parquet reader; the .npz block files are read by bpred.core.block_ld (see the repository for the on-disk band layout).

Source & attribution

This dataset is a derivative of:

Florian Privé. LD reference for HapMap3+. figshare dataset. https://figshare.com/articles/dataset/LD_reference_for_HapMap3_/21305061 (DOI: 10.6084/m9.figshare.21305061), licensed CC BY 4.0.

Changes made: greedy within-block r² > 0.99 pruning (higher-MAF kept), distance-windowed banding, and re-serialization from .rds sparse matrices to snps.parquet + per-chromosome .npz. The underlying LD values are Privé's.

License

CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), matching the source reference. You may share and adapt with attribution and an indication of changes. (The bpred software itself is separately licensed MIT; this license covers the data only.)

Citation

If you use this LD reference, please cite the originating LDpred2 work and the source dataset:

Privé F, Albiñana C, Arbel J, Pasaniuc B, Vilhjálmsson BJ. "Inferring disease architecture and predictive ability with LDpred2-auto." American Journal of Human Genetics 110(12):2042–2055, 2023.

Florian Privé. LD reference for HapMap3+. figshare. DOI: 10.6084/m9.figshare.21305061.

And, if you use the bpred software: https://github.com/bkorpan/bpred (see CITATION.cff).