NEWS.md

In StoreyLab/lfa: Logistic Factor Analysis for Categorical Data

lfa 2.0.0.9000 (2020-09-18)

Major overhaul from last version (1.9.0, last updated 2018-02-11). Overall, added unit testing to all functions, which resulted in the identification and fixing of several edge-case bugs, and also minor improvements.

• User-facing changes
• Removed redundant/obsolete exported functions:
  • model.gof: entirely redundant with sHWE in this same package
  • read.tped.recode: obsolete file format; instead, use plink for file conversions!
  • read.bed: instead, use read_plink from the genio package!
  • center: only worked for matrices without missingness (useless for real data), no dependencies in code, plus centering is trivial in R
• Renamed remaining functions, replacing periods with underscores.
  • Only specific change: trunc.svd -> trunc_svd
  • NOTE af_snp and pca_af already had underscores instead of periods (unchanged). All other functions unchanged.
• Function trunc_svd
  • Debugged d=1 case (output matrices were dropped to vectors, which was a fatal error in lfa after it called trunc_svd).
  • Added option force that, when TRUE, forces the Lanczos algorithm to be used in all cases (most useful for testing purposes).
• Function lfa
  • Improved documentation.
• Functions af_snp and af
  • Fixed a bug in which large logistic factor coefficients resulted in NaN allele frequencies instead of 1 as they should be in the limit.
  • Improved code to "impute" allele frequencies for NA genotypes. Original version preserved NA values (a genotype that was NA for a particular individual and locus combination resulted in an NA in the corresponding allele frequency only, and conversely non-NA genotypes never resulted in NA allele frequencies). The new version takes advantage of knowing the LFs of all individuals (regardless of genotype missingness), and LFs and their coefficients are never NA, permitting allele frequencies to always be imputed into non-NA values!
• Function pca_af
  • Similarly imputes allele frequencies for NA genotypes (internal change was trivial)
  • Debugged d=1 case, which incorrectly returned an intercept column matrix instead of an allele frequency matrix.
• Function check_geno
  • Debugged testing for matrix class (original code when run in R 4.0 generated warning "the condition has length > 1 and only the first element will be used")

• Function sHWE (through internal inverse_2x2)

  • When a test was "singular" at a single SNP, function used to die; now that SNP gets an NA p-value.
  • Other previous NA cases here are avoided now that af never returns NA values.

• Internal changes

• Separated R functions into one source file each.
• Added .gitignore files from another project.
  • Removed src/lfa.so from version control tracking.
• Added unit tests for all functions using testthat.
• Updates to C code solely to pass latest R CMD check requirements.

lfa 2.0.1.9000 (2020-11-11)

• Function lfa added support for BEDMatrix objects for the genotype matrix X.
• This consumes lower memory when the number of loci m is very large, so it enables analysis of larger datasets.
• Algorithm for BEDMatrix case is different: instead of Lanczos truncated SVD, covariance matrices are computed explicitly and truncated eigendecomposition performed. This means runtime and memory complexity are very different here as the number of individuals n gets larger.
• Added RSpectra package dependency (for fast truncated eigendecomposition).

lfa 2.0.2.9000 (2020-11-12)

• More functions updated to support BEDMatrix inputs for the genotype matrix X. Although BEDMatrix is supported, in these cases there are minimal memory reduction advantages as outputs or intermediate matrices are necessarily as large as the input genotype data.
• Function af. Although there is memory saving by not loading X entirely into memory, the output individual-specific allele frequency matrix P has the same dimensions so memory usage may still be excessive for in large datasets, negating the BEDMatrix advantage.
• Function pca_af. Note same memory usage issue as af.
• Function sHWE. A worse memory problem is present, since internal code calculates the entire individual-specific allele frequency matrix P, then simulates B random genotype matrices of the same dimensions as input (each in memory) from which LF and ultimately HWE statistics are obtained.

lfa 2.0.3.9000 (2020-12-16)

• Fixed an integer overflow error that occurred in sHWE (in internal function compute_nulls), which only occurred if the number of individuals times the number of loci exceeded the maximum integer size in R (the value of .Machine$integer.max, which is 2,147,483,647 in my 64-bit machine).
• Function lfa added rspectra option (FALSE by default), as an alternative way of calculating SVD internally (for experimental use only).
• Function trunc_svd is now exported.
• Minor, user-imperceptible changes in other functions.

lfa 2.0.4.9000 (2020-12-22)

• Function sHWE fixed bug: an error could occur when internal statistics vector included NA values.
• Original error gave this obscure message, which occurred because an index went out of bounds due to a discrepancy in vector lengths due to the presence of NA values:

Error in while ((i2 <= B0) & (obs_stat[i1] >= stat0[i2])) { : 
  missing value where TRUE/FALSE needed

• Now empirical p-value code is separated into new internal function pvals_empir, and its tested against a new naive version pvals_empir_brute (slower brute-force algorithm, used to validate outputs only) in unit tests including simulated data with NA values.
• Also refactored other internal sHWE code into a new internal function gof_stat, which by itself better handles BEDMatrix files (though overall memory savings are not yet there on the broader sHWE).
• Spell-checked this news file (edited earlier entries).

lfa 2.0.5.9000 (2021-02-16)

• Documentation updates:
• Fixed links to functions, in many cases these were broken because of incompatible mixed Rd and markdown syntax (now markdown is used more fully).

lfa 2.0.6.9000 (2021-03-01)

• Functions af_snp, af, and sHWE added parameters max_iter (default 100) and tol (default 1e-10).
• Previous version of code had these parameters hardcoded.
• NOTE: max_iter = 20 used to be the default value, which in downstream tests was not routinely sufficient to converge with comparable numerical accuracy to glm fits (not in this package lfa, but in downstream packages gcatest and jackstraw, which require calculating deviances).

lfa 2.0.7 (2021-06-16)

• Lots of minor changes for Bioconductor update.
• Function trunc_svd:
  • Removed seed, ltrace, and V options.
  • Added maxit option.
  • Reduced default tol from 1e-10 to .Machine$double.eps (about 2e-16)
• Function lfa:
  • Reduced default tol from 1e-13 to .Machine$double.eps (about 2e-16)
• Added more examples in function docs.
• DESCRIPTION:
  • Updated to Authors@R.
  • Lengthened "Description" paragraph.
  • Increased R dependency from 3.2 to 4.0.
• Updated README.md.
• Reformatted this NEWS.md slightly to improve its automatic parsing.
• Added published paper citation to vignette, README.md, inst/CITATION.
  • First two used to point to arXiv preprint, last one didn't exist.
• Updated vignette to reflect that read.bed has been removed.
• Corrected spelling.
• Resurrected and deprecated functions that were exported in last Bioconductor release but deleted on GitHub:
  • center
  • model.gof
  • read.bed
  • read.tped.recode
• Internal changes:
  • All unexported functions are now prefixed with a period.
  • Replaced 1:x with seq_len(x) several functions.
  • Reformatted all code with package reformatR and otherwise match Bioconductor guidelines.
  • Split some functions up so individual functions have less than 50 lines.
  • Removed unexported function inverse_2x2, probably speeding up sHWE slightly.
  • Removed unexported function mv (all instances called C code directly instead of this R wrapper).
  • Cleaned up trunc_svd source considerably.

lfa 2.0.8 (2021-06-18)

• Minor updates:
• Added LICENSE.md.
• Edits to README.md.
• Vignette now also suggests BEDMatrix for loading data.

lfa 2.0.9 (2022-11-11)

• Fixed critical bug that prevented compilation of C code in latest R-devel. Documenting here path that led to debugging as it may be informative to maintainers of other packages that have written similar code.
• Here's error message, abbreviated:

fastmat.c: In function ‘mv’:
fastmat.c:22:14: error: too few arguments to function ‘dgemv_’
   22 |     F77_CALL(dgemv)(&tr,dimA,dimA+1,&alpha,A,dimA,v,&one,&zero,ret,&one);
      |              ^~~~~
/home/biocbuild/bbs-3.17-bioc/R/include/R_ext/RS.h:77:25: note: in definition of macro ‘F77_CALL’
   77 | # define F77_CALL(x)    x ## _
      |                         ^
/home/biocbuild/bbs-3.17-bioc/R/include/R_ext/BLAS.h:107:10: note: declared here
  107 | F77_NAME(dgemv)(const char *trans, const int *m, const int *n,
      |          ^~~~~
...
make: *** [/home/biocbuild/bbs-3.17-bioc/R/etc/Makeconf:176: fastmat.o] Error 1
ERROR: compilation failed for package ‘lfa’

• Bug manifested after R-devel commit r82062 (2022-04-02): R CMD check --as-cran now uses FC_LEN_T (I was testing locally using --as-cran, perhaps it manifests later otherwise.)
• Googling for FC_LEN_T led me to R news, which pointed me to Writing R Extensions: 6.6.1 Fortran character strings, which shows that an argument of type FC_LEN_T now has to be added to specify the length of a string passed to Fortran code.
• Eventually text-searched for dgemv in the R source code and came across array.c examples where it sufficed to append the C macro FCONE to my existing dgemv call, and that solves it! (FCONE, defined in R_ext/BLAS.h, equal to ,(FC_LEN_T)1 if FC_LEN_T has been defined, otherwise it is blank.)

lfa 2.0.10 (2022-11-11)

• Minor non-code updates to fix check --as-cran notes:
• Package description cannot start with package name.
• README.md updated an http link to https to which it redirects.
• Function sHWE documentation used \doi instead of direct link.

lfa 2.1.10 (2023-05-25)

• Version bump for bioconductor devel.

lfa 2.1.11 (2023-06-20)

• Commented out excessive test for internal function .lreg against glm, which differ more often than expected due to poor or lack of convergence.
• Removed unused LaTeX package dependencies from vignette to prevent errors restricted to specific testing platforms.

StoreyLab/lfa documentation built on March 7, 2024, 9:59 p.m.