NEWS.md
run_admixture() renamed its outputs to a doubly-nested, non-existent path when path was relative (it setwd()s into path, then built the rename target with file.path(path, ...) again). The rename failed silently, so successive same-K runs overwrote each other’s .Q/.P. path is now resolved to an absolute path and the rename is verified (warns on failure).as_snpmatrix() used a shifted byte encoding, so genotype 0 was stored as missing, missing as BB, and heterozygotes/alt-homozygotes were off by one. Any snpStats statistic (call rate, MAF, HWE, GRM/PCA) computed on its output was therefore wrong. It now uses the canonical SnpMatrix encoding (0x00 = missing, 0/1/2 -> 0x01/0x02/0x03), matching getGeno().check.snp.same.position() was defined twice; the shadowing copy lacked guards. It is now a single vectorized definition that groups SNPs by chromosome + position, safe for single-SNP chromosomes and missing positions.check.identical.samples(), pairs2sets() and check.mendelian.inconsistencies() no longer error on empty or single-element inputs (1:n / 2:n off-by-one guards). check.identical.samples() also extracts pairs vectorially instead of growing a data.frame in a nested loop.check.snp.no.position() now treats a SNP as unmapped when its position is missing (NA), blank, non-numeric, or zero (previously only == 0, which also returned spurious NA names when positions were missing). qcSNPs() uses the same check consistently before the same-position filter.check.identical.samples.by.block() now returns the pairs identical across every block (data.frame), instead of only the last block’s pairs, and is guarded against fewer than two samples.get.correl.fc() (result unchanged).qcSNPs() drops the missing_ind and missing_snp arguments, which were never implemented. Use min_snp_cr for per-SNP call rate and qcSamples(smp_cr = ...) for per-individual call rate.combineSNPData() now assembles the combined genotype matrix through rbindSnpFlexible() instead of a duplicated inline implementation. The returned object is byte-identical to before (covered by a new equivalence test suite); only the intermediate “Adding N missing SNPs” message is no longer emitted.rbindSnpFlexible(), rbind_SnpMatrix() and cbind_SnpMatrix() are now internal helpers (no longer exported). Use combineSNPData() for the supported, high-level combining workflow.testthat suite, including a broad characterization battery that checks combineSNPData() against a reference copy of the pre-refactor implementation across overlapping/partial/disjoint/multi-object/randomized inputs and edge cases.runPCA() runs the genotype PCA on a SNPDataLong object without any clustering, returning a prcomp-like object and the selected top principal components. It uses the same PCA engine as runAnticlusteringPCA() (standardised SNPs, Gram-matrix eigendecomposition, optional matrix-free RSpectra fast path), so scores are directly comparable. runAnticlusteringPCA() now calls runPCA() internally.doPCA() is deprecated in favour of runPCA(). It still works but emits a deprecation message. doPCA() used an unscaled GRM (snpStats::xxt) on a raw SnpMatrix; runPCA() is the recommended, standardised PCA on a SNPDataLong.runAnticlusteringPCA() is now dramatically faster and lighter on memory for wide genotype data. When there are more SNPs than individuals (the usual case), PCA is computed from the small n x n Gram matrix instead of a full SVD on the n x p matrix, avoiding the construction of the huge p x n rotation matrix (which could be several GB and was never used). The genotype matrix is also no longer converted to a data.frame before PCA. Scores and standard deviations are unchanged. runAnticlusteringPCA() no longer uses anticlust’s removed features argument.
runAnticlusteringPCA(): when a fixed number of PCs is requested and the optional RSpectra package is installed, only the top n_pcs components are computed with a matrix-free solver, so the n x n Gram matrix is never formed at all. This is the dominant remaining cost for very wide data. Falls back to the Gram-matrix eigen decomposition when RSpectra is absent or a proportion of variance is requested (n_pcs < 1).
runAnticlusteringPCA() gains an anticlust_method argument. The default "exchange" preserves current behaviour; "fast" uses anticlust::fast_anticlustering, which scales to large numbers of individuals.
getGeno(): the chromosome column of the returned map is now always named Chromosome (previously it kept the original header name, which could be Chr or Chromosome). This fixes a match.names ... names do not match error in combineSNPData() / import_geno_list() when combining datasets whose SNP_Map.txt files used different chromosome column names.
getGeno() now builds the SnpMatrix directly from the data.table::fread output instead of calling snpStats::read.snps.long. The latter’s internal search does not scale to very large long-format FinalReport.txt files (millions of lines / hundreds of samples): it silently read only the first sample and reported all remaining rows as “not found”, producing a matrix in which every sample but one was empty. Building the matrix from fread is reliable regardless of file size and is also robust to malformed lines – empty or unreadable confidence (GC Score) fields simply parse to NA and are treated as no calls, so no line repair or temporary file is required and the original file on disk is never modified.
combineSNPData(): fixed spurious "object has no names" warning from snpStats when filling missing SNPs with NA. The SnpMatrix block is now constructed with dimnames set at creation time.
SNPDataLong: relaxed validation of the xref_path slot to accept character vectors of any length (one entry per individual), resolving an error in import_geno_list() when datasets contained more than one individual.
qcSNPs(): fixed a warning (max returning -Inf) and potential incorrect removal of all SNPs in a same-position group when all MAF values were NA. The first SNP in the group is now kept in that case.