Heritability and genetic correlation

Heritability and genetic correlation

These analyses require GWAS summary statistics, reference LD scores, and optional SNP annotations. See Input files for column names.

For binary traits, PCGC prepares ascertainment-aware moments from individual data. Ordinary logistic beta/SE files cannot substitute for those moments. To estimate covariance between two traits' environmental genetic responses, use the cross-trait research API.

Heritability

summit --h2 trait.sumstats.gz --ldscores reference.gw.ldscore.gz \
  --out results/trait --njack chr --num-threads 4

For partitioned heritability, add --annot annotations.tsv and use the corresponding multi-column LD scores. --h2 also accepts a directory of traits or chromosome-split paths.

Results are written to <out>.results.tsv and <out>.log. The table includes component estimates, standard errors, and enrichment where applicable.

Genetic correlation

For studies with no overlapping participants:

summit --rg trait1.sumstats.gz,trait2.sumstats.gz \
  --ldscores reference.gw.ldscore.gz --overlap-covariance-rg 0 \
  --out results/trait1.trait2 --njack chr

For overlapping studies, either supply --overlap-covariance-rg or let SUMMIT estimate the overlap covariance from summary statistics by omitting that option. The supplied value is the standardized phenotype cross-product on the shared samples divided by sqrt(N1*N2). It is not the number or proportion of overlapping participants.

When N varies across SNPs, the overlap correction assumes a constant overlap covariance on the local score scale. This is appropriate for shared genotype missingness with similar missing fractions in both traits and their overlap. Different cohort contributions or trait-specific missingness require additional overlap information; per-SNP N alone is insufficient.

If you have participant-level inputs locally, --pheno-rg a.tsv,b.tsv and optional --pheno-rg-cov a.cov,b.cov compute this value. Phenotype tables have FID IID and the phenotype as their last column.

Partitioned analyses with summary-only overlap estimation need a scalar regression LD score:

summit --rg trait1.sumstats.gz,trait2.sumstats.gz \
  --ldscores partitioned.gw.ldscore.gz --annot annotations.tsv \
  --ldscores-reg total.gw.ldscore.gz --out results/pair --njack chr

For disjoint annotations, partition LD scores can be summed to obtain total LD. For overlapping annotations, supply a separately computed total LD score.

Allele alignment

SUMMIT aligns alleles by default. It handles allele swaps and strand complements, changing the second trait's effect sign when required. Invalid or incompatible alleles are excluded. A/T and C/G variants are excluded by default because their strand is ambiguous.

--keep-ambiguous uses the supplied literal allele orientation. Use --no-align-alleles only after independently harmonizing the inputs.

Standard errors and filtering

--njack chr deletes one chromosome at a time. An integer requests SNP blocks. The default jackknife supports varying per-SNP N. The optional --rg-se-method kmoments and full SCORE normal-equation export require constant per-SNP sample sizes because they assume one shared sample matrix.

--max-chisq auto uses max(80, 0.001*Nmax); --chisq-action chooses drop, clip, warn, or none. Review the retained SNP counts in the log.

The default estimation mode is --weight-mode he. --weight-mode ldsc uses an iterative weighted LD-score regression with a fixed univariate intercept and a separately resolved overlap covariance. It is a different estimator; its formulas and limits are in Methods.

Interpretation

With disjoint annotation bins, components describe each bin's contribution. With overlapping annotations, a coefficient is conditional on the other columns. Component rg values are ratios of those coefficient contributions; they are not rg restricted to all SNPs carrying that annotation.

Raw estimates can be negative, and rg may be undefined when its h² denominator is nonpositive. Do not turn those cases into zero or a bounded correlation. Strongly collinear annotations may make components unidentifiable.

See Batch analyses for shared loading across many traits.