Skip to contents

Changelog

Source: NEWS.md

gauss 1.1.1

Bug fixes

  • The package could not be built on Windows. src/bgzf.c reads the reference panel through zlib, but nothing passed -lz to the linker, so ld reported deflate, deflateEnd, deflateInit2_, crc32, inflateInit2_, inflate and inflateEnd as undefined and the install failed outright.

    ELF and Mach-O leave undefined symbols in a shared object to be resolved at load time, and R has already linked zlib by then, so Linux and macOS built either way and the omission was invisible there. A Windows DLL has to resolve every symbol at link time.

    The dependency used to arrive through PKG_LIBS = $(GSL_LIBS) and was lost in 1.1.0, when hand-written src/Makevars and src/Makevars.win replaced the configure-generated src/Makevars.in. Both now set PKG_LIBS = -lz.

Build and packaging

  • The vignettes build without the reference panel. R CMD build executes them, and every analysis they ran needed the 33KG panel – 43 GB that are not distributed – so a clean checkout could not build the package at all.

    Each analysis call is now gated on GAUSS_REF_33KG. Where it points at a panel the real call runs; where it does not, the result that call produced is read back from inst/extdata. Everything downstream renders from whichever result arrives, so the tables and figures are the same either way. data-raw/make_vignette_results.R regenerates the saved results.

  • Suggests now declares what the vignettes use: ggplot2, dplyr and data.table. tidyverse, kableExtra and reshape2 were dropped rather than declared – kableExtra::kable() is knitr::kable() re-exported, tidyverse was pulled in for ggplot2 and dplyr alone, and reshape2 covered one melt() of a matrix with no dimnames.

  • The vignettes are built and checked on CI again, on Linux, macOS and Windows.

Documentation

  • The pkgdown site is rebuilt. It had not been since January, so it still published the ancestry-weighted numbers that the 1.1.0 mixture covariance fix replaced, and was missing the data set help pages, the changelog and the corrected package author.

  • The changelog is reachable from the navigation bar.

gauss 1.1.0

Bug fixes

  • CalWgtCov() now computes the mixture covariance the DISTMIX paper specifies, which changes the results of every mixture analysis. Affected: distmix(), computeLD(), qcatmix(), jepegmix() and prep_recessive_impute(). Single-population and superpopulation analyses (dist(), qcat(), jepeg()), ancestry estimation (afmix(), cpw2(), zmix()) and fiqt() are unaffected.

    The ancestry-informed covariance is Equation (2) of Lee et al., Bioinformatics 31(19):3099-3104 (doi:10.1093/bioinformatics/btv348),

    C = sum_k w_k C_k + sum_k w_k (mu_k - mu)(mu_k - mu)^T

    where C_k is the variance-covariance matrix of reference population k and w_k its mixing proportion. The within-population term sum_k w_k C_k was being scaled by m_k^2, the square of that population’s panel sample size, because (m*sumxy - sumx*sumy) was never divided by m*(m-1).

    Two consequences, both now gone. Populations were effectively re-weighted by w_k * m_k^2 instead of w_k; since panel sizes span 52 to 6360 samples, user-supplied mixing proportions were largely overridden by panel sample sizes. And the between-population term, which is the reason a mixture LD matrix differs from a within-population one, was m_k^2 times smaller than the inflated first term and so numerically discarded – mixture LD behaved like a size-weighted within-population average.

    On the package’s test fixture (3 populations of 20 samples, weights 0.5/0.3/0.2), the previous release’s numbers differ from the corrected ones by up to 0.41 in distmix() Z-scores (3 of 101 change sign), 0.20 in computeLD() correlations (14 of 100 change sign), 0.48 in qcatmix() chi-squares and 0.37 in jepegmix() gene chi-squares. The distortion grows with panel sample size, so differences on the 33KG panel are larger.

    Users who need to reproduce results from gauss 1.0.1 or earlier should pin that version.

  • Variants that the reference panel does not carry are handled correctly. MakeSnpVec() and MakeSnpVecMix() looped over every SNP in the map, including those present in the user’s input but absent from the panel. Such a SNP has no genotype record, so its stored file offset is still the default -1; bgzf_seek() failed, no call site checked the return value, and the next read returned whichever record the stream happened to be sitting on. The variant therefore acquired an unrelated variant’s allele frequency, and that number then decided — via the minor-allele-frequency cutoff, on evidence belonging to a different variant — whether it appeared in the output at all.

    Such variants are now passed through to the output with type = 2 and an NA allele frequency, which is what every analysis already assumed: each one carries an explicit “if (type == 2) do nothing” branch excluding them from the linear algebra. Their own Z-score reaches the output untouched, and the count is reported on the console.

    This adds rows to dist(), distmix(), qcat(), qcatmix(), prep_qcat() and prep_recessive_impute() output — one per input variant in the reporting window that the panel lacks — and those rows carry NA in the af1ref / af1mix column. Results for panel variants are unchanged: on a 1 Mb window of chromosome 22 against the 33KG panel, one variant of 5869 is recovered this way and the other 5868 rows are identical.

  • Every seek into the reference panel is checked. A file offset that is negative, or a seek that fails, now raises an error naming the offset instead of silently leaving the stream somewhere else and parsing the wrong record. This covers all seven live call sites.

  • pop_wgt_df is now validated, and every function that takes ancestry proportions rescales them to sum to one before doing any work. Equation (2) assumes sum_k w_k = 1; its between-population term is a centred outer product, and above 1 it can go negative far enough to make a variant’s mixture variance negative, at which point the square root in the LD normalisation produces NaN that spreads through the whole LD matrix. The m_k^2 inflation described above had been masking this, so the two fixes belong together.

    This was not hypothetical: the mixing proportions shipped as data(PGC2_SCZ_ANC_Prop), used throughout the README and the vignettes, summed to 1.061. On the real 33KG panel that produced 402 NaN entries across every row of a 202-variant LD matrix once the covariance was corrected.

    Rescaling is reported on the console rather than raised as a warning: afmix() and cpw2() return unrescaled weights by construction, so this is ordinary preprocessing, not an exceptional condition. Only the scale changes — each population’s share relative to the others is preserved, and passing the same proportions at any scale gives identical results.

    Populations named in pop_wgt_df that the reference panel does not carry are reported by name instead of having their weight silently dropped; a mixture in which no population matches the panel, a negative or non-finite weight, and weights summing to zero are all now clear errors.

Data

  • data(PGC2_SCZ_ANC_Prop) is now rescaled to sum to 1 (it summed to 1.061). Only the scale changed: every population’s share relative to the others, and the set of 21 populations, are exactly as published. The derivation is now recorded in data-raw/make_PGC2_SCZ_ANC_Prop.R, and the help page documents the provenance and the column meanings.

Build and packaging

  • src/Makevars is now shipped directly rather than generated by a configure script whose only job was to copy src/Makevars.in into place, and it sets PKG_CXXFLAGS with = instead of +=. += is a GNU make extension that the package never declared in SystemRequirements, so a build under a non-GNU make could silently drop the flag.

  • stats and utils are declared in Imports, and the symbols fiqt() uses (pnorm, qnorm, p.adjust) are imported explicitly. They previously resolved only because those packages happen to be attached in a default R session. RcppEigen has moved out of Imports: it is a headers-only dependency and LinkingTo already covers it.

  • fiqt() spells out lower.tail = FALSE where it previously relied on partial matching of low =. The computation is unchanged.

  • src/zmixrcpp.cpp_ and src/etc/, both superseded before the Rcpp rewrite, are kept out of the source tarball.

Documentation

  • The pop_wgt_df argument is documented properly for distmix(), computeLD(), qcatmix(), jepegmix(), simulateLD() and prep_recessive_impute(): that it is read positionally, that the weights are rescaled, and that afmix() returns three columns whose first must be dropped before use.

  • ?gauss is a real package help page — what the package does, the four groups of entry points, the positional pop_wgt_df convention and the DIST reference — in place of the package.skeleton() placeholder text that had been shipping since the first release.

  • The five data sets the vignettes use have help pages for the first time: JEPEG_SNP_Annotation.v1.0, PGC2_3Mb, PGC2_Chr22_ilmn1M_AF1, PGC2_Chr22_ilmn1M_Z and PGC3_SCZ_ilmn1M_Z. Each records its columns, its extent, which function it is the example input for, and the study it comes from — PGC SCZ2 (Nature 511:421-427) for the PGC2_* sets, PGC SCZ3 (Nature 604:502-508) for PGC3_SCZ_ilmn1M_Z, and the JEPEG paper (Bioinformatics 31:1176-1182) for the annotation.

gauss 1.0.1

Bug fixes

  • jepeg() and jepegmix() no longer terminate the R session when the SNP annotation file cannot be opened. A mistyped annotation_file path now raises an ordinary R error, leaving the session and any unsaved work intact.

  • Annotation rows whose categ value is not one of PROTEIN, TFBS, WTH_HAIR, WTH_TARGET, CIS_EQTL or TRANS_EQTL are now skipped, and the number skipped is reported as a warning. Previously such a row was assigned to whichever category the preceding row used, which silently corrupted the JEPEG weight matrix and the resulting gene-level p-values. Results are unchanged for annotation files that use only the six documented categories.