Input files

Input files

Use the same genome build, SNP naming, and annotation definitions throughout an analysis. Input examples below use invented IDs.

Genotypes

--geno accepts a .bed or .pgen filename, or a prefix with one complete trio:

Format	Files
PLINK 1	.bed, .bim, .fam
PLINK 2	.pgen, .pvar, .psam

PGEN input must contain biallelic diploid variants and a plain-text PVAR. Stored fractional dosages are retained. Missing values are mean-imputed. Convert BGEN or compressed PVAR input before running SUMMIT. If both genotype formats share a prefix, specify the desired filename.

PGEN is supported for genome-wide, windowed, and one-environment G×E LD scores, and PGS. The direct generalized reference executor currently reads BED.

GWAS summary statistics

Supply a whitespace-delimited table with a header:

Meaning	Accepted columns
SNP ID	SNP, ID, snp, id
Effect allele	A1, ALT
Other allele	A2, REF
Sample size	N, OBS_CT
Effect estimate	BETA, beta
Standard error	SE, STDERR, se, stderr

SNP A1 A2 N BETA SE
sim_variant_000001 A C 50000 0.01 0.005

For h² and rg, provide BETA and SE; Z alone is insufficient. These methods reconstruct marginal linear-regression score statistics, so verify suitability before using statistics from a different association model.

Use each SNP's observed sample size in N; do not replace varying values with the file maximum. SUMMIT accounts for these differences in its null moments and regression weights. For rg, see the sample-overlap assumptions in Methods.

Optional COV_RANK or P_EFF gives the number of non-intercept GWAS covariates; --cov-rank overrides it. The rg calculation uses this metadata. The current h² calculation uses a covariate-rank value of zero, including when h² supplies the denominator for rg. Methods explains the calculation.

Binary PCGC uses a separate joint .binary.npz artifact containing compatible trait and reference moments. Prepare it from a FID IID Y table, population prevalence, population genotype scales, and optional risk covariates; see Binary traits and PCGC. Its annotation TSV has SNP followed only by named nonnegative weight columns.

LD scores and annotations

LD-score files begin with CHR SNP BP, optionally CM, followed by LD-score columns. An annotation file contains either:

• CHR BP SNP, optional CM, and named annotation columns; or
• a numeric matrix with one row per input SNP, optionally headed by column names.

Row order must match the genotype or LD-score SNP order. Omit --annot for a single-component model. Disjoint bins and overlapping annotations have different interpretations; see Heritability and genetic correlation.

For chromosome-split h²/rg files, @ expands to chromosomes 1–22, for example reference/chr@.gw.ldscore.gz.

Covariates, environments, and phenotypes

These tables use FID IID followed by named data columns:

FID IID cov1 cov2
SIM000001 SIM000001 0.2 -0.4

Keep IDs as strings. Covariates must be numeric; encode categorical fixed effects before using the main summit command. The contextual Python and PGS APIs can apply saved categorical coding recipes.

For one-environment G×E, an environment table has exactly one selected environment column. Trait scoring requires finite selected phenotypes on the reference sample set. For PGS, each trait can declare its own sample subset; keep files also have a FID IID header.