LD scores

LD scores

SUMMIT estimates LD scores from a reference genotype panel. Use a panel suited to the ancestry and variant set of the GWAS being analyzed.

Genome-wide LD scores

summit --geno reference.bed --out results/reference \
  --nvecs 1000 --block-size 1000 --num-threads 8

--nvecs controls the number of random vectors. More vectors reduce Monte Carlo error and increase computation. Set --seed to reproduce the randomization.

Add annotations and covariates when needed:

summit --geno reference.bed --annot annotations.tsv --covar covariates.tsv \
  --out results/partitioned --nvecs 1000 --block-size 1000 --num-threads 8

Outputs are <out>.gw.ldscore.gz, <out>.gw.M, and <out>.gw.log. <out>.gw.mc.tsv summarizes random-vector error by annotation.

Windowed LD scores

summit --geno reference.bed --ld-wind-kb 1000 \
  --out results/reference.1mb --num-threads 8

--ld-wind-kb sets the genomic distance limit in kilobases. Windowed estimation is deterministic and does not use --nvecs. Annotation and covariate options work as above. Outputs use .win.ldscore.gz, .win.M, .win.M_5_50, and .win.log.

Dosage input

Use an explicit .pgen path in either command. SUMMIT reads stored REF dosages, mean-imputes missing values, and estimates LD in that dosage matrix. Hard calls and imputed dosages can therefore yield different LD scores.

Genome-wide PGEN requires CPU dense computation, mean imputation, and --ddof 1. It does not support Mailman, HWE imputation, skew correction, or kernel-moment output. Windowed PGEN uses --win-panel-cols and --win-cache-mb to limit its decoded panels and cache.

Choosing resources

--num-threads sets the CPU thread count. --block-size controls variants per block; smaller blocks reduce temporary memory. --memory-gib sets the sketch-panel or windowed-LD memory budget in GiB. Use an explicit budget on a cluster when the process cannot detect its scheduler allocation.

CUDA is available for supported BED genome-wide runs. The default Monte Carlo diagnostic makes phase 2 run on the CPU; --skip-ld-mc keeps the CUDA path.

Monte Carlo error

The default .gw.mc.tsv reports the RMS per-SNP Monte Carlo standard error and its size relative to the RMS LD score. This measures random-vector error conditional on the reference panel. It excludes uncertainty from sampling that panel.

For per-SNP variances and approximate 95% intervals, add --write-ld-mc-var. This needs at least two vectors and additional storage. The intervals are pointwise. They may extend below zero.

The Methods page gives the variance formula. For a runnable synthetic example, use bash example/estimate_partitioned_gwldscore.sh.