Skip to content

v8.0.0

Choose a tag to compare

@michael-denyer michael-denyer released this 08 Sep 01:28
5b40bac

JAMMA 8.0.0 closes the September 2026 mathematical audit against GEMMA v0.98.5 and the code-quality review that followed it. Four changes are visible to a user, which is why this is a major release. The full record is the [8.0.0] section of CHANGELOG.md.

What changes for a user

gwas() returns PipelineResult and takes phenotype_columns. GWASResult and GWASTiming are gone; gwas(phenotype_column=2) is gwas(phenotype_columns=[2]). The CLI flags are unchanged.

jamma -gk selects SNPs the way gemma -gk does. The matrix spans every sample in the .fam, and the MAF and missingness filters default to 0.01 and 0.05, measured over the samples whose selected phenotype columns and covariates are present. jamma -gk 1 matches GEMMA's reference kinship on mouse_hs1940 at kinship_rtol. Multi-column -n under -gk intersects the columns instead of raising.

A covariate file without a constant column gets an intercept. As GEMMA's CheckCvt does, a column of 1s is appended when no column is constant over the analysed samples. A file whose first column is not 1s no longer silently fits a model without an intercept.

A kinship passed to the runner API is centred and consumed. run_lmm_association_numpy(kinship=K) centres the analysed kinship before eigendecomposition, as the pipeline already did, so MLE, LRT, Score and PVE match GEMMA through the direct API. The array is overwritten in place and must be writeable; pass K.copy() to keep it.

Numerical fixes

  • Only an exactly zero P_yy is replaced with 1e-8, as GEMMA v0.98.5 does. The former absolute floor made p-values depend on the phenotype scale.
  • Weighted eigenvector files written under -widv carry the row scaling, so a reload through -d/-u reproduces the weighted run.
  • A multi-chunk run with the progress bar on no longer deletes its finished output.
  • The C accelerator computes logdet(H) as a mantissa product with an exact exponent sum, within 1.1e-12 of a 40-digit reference.

Internals

The C accelerator has one compute entry point per kernel family and one workspace creator per covariate shape (ABI_VERSION 15). The memory estimators are pure and the gate reads the machine once. One analysed-sample basis serves -lmm and -gk, one intercept rule serves every path, one P_yy guard serves each language, and the overlapped chunk driver is a generator. Docs prose is gated by a banned-word list.

Upgrading

Linux and Windows users on ILP64 numpy-mkl install in the documented order (dependencies, then numpy-mkl, then jamma --no-deps) so LP64 numpy does not overwrite the ILP64 build. See docs/USER_GUIDE.md.