Skip to contents

Changelog

Source: NEWS.md

genoaligner 2.0.0.9000 (development)

  • msa_codon() gains backbone = "protein": translate each sequence in frame 0, build the alignment backbone at amino-acid level with the same progressive engine (msa_align(mode = "protein")), then back-project every residue to its codon — output is a codon grid with whole-codon indels by construction. Simulation benchmarks show substantially higher sum-of-pairs accuracy than the codon-token backbone at high divergence (SPS ≈ 0.95 vs ≈ 0.5–0.7 on n=50, d=0.3 datasets) while preserving codon phase purity 1.0. backbone = "codon" (default) is unchanged. Nucleotides outside complete codons (trailing partial codons) are kept as end insertion columns and counted in qc$partial; refine = TRUE runs the stage-2 codon refinement on the back-projected grid.
  • msa_codon(refine = "auto") — selective stage-2 refinement: only sequences whose length is not a multiple of 3 may emit frameshift event columns (codon_refine_sel_cpp runs the per-row phase DP on flagged rows against the full profile; every other row contributes its backbone placement verbatim). In-frame input is returned exactly as the un-refined backbone, so “auto” cannot degrade clean CDS — on simulated frameshift data it improves SPS over the un-refined path and recovers frameshift events (mean fs_recall ≈ 0.3–0.5 on the protein backbone, ≈ 0.1 on the codon backbone).

genoaligner 2.0.0

  • Breaking: align() renamed to align_edit() — the generic name collides with Biostrings/IRanges exports in mixed sessions. align() remains as a deprecated alias returning identical results.
  • New multiple-sequence-alignment API over the same portable C++17 core (vendored host engine, verified bit-exact against the GPU driver in the sibling C++ repository):
    • msa_align(seqs, mode = "dna" | "protein") — progressive profile-profile MSA with a deterministic NJ guide tree.
    • msa_codon(seqs, gc = 1 | 2, refine = FALSE) — MACSE-class codon-aware MSA: whole-codon indels preserve the reading frame; refine = TRUE runs a frameshift-refinement pass producing a nucleotide MSA.
  • Both return S3 objects of class genoaligner_msa ($aligned character matrix, $qc per-sequence QC — frame / stop codons / partial codons in codon mode —, $params), with print/as.matrix/as.character methods and optional ape::DNAbin coercion via as_dnabin().
  • msa_align()/msa_codon() gain iterative refinement and FFT anchors (MAFFT FFT-NS-i class): iter_refine = k realigns the two induced profiles of every guide-tree edge and keeps a candidate only when the sum-of-pairs score strictly improves; fft_band > 0 restricts each realignment DP to a band around Fourier-detected homology anchors (faster, exact-rescored, falls back to the full DP on weak signal). On 37 real Lepus cytb sequences two rounds improve the SP objective 620,764 → 628,384 at ~30% less refinement time with fft_band = 32.
  • Re-vendored engine: genoaligner-devel devin/msa-iter-refine @ 2b15a7a (PR #13) — see inst/VENDORED.txt.

genoaligner 1.0.0.9000

  • New dataset lepus_cytb: 37 real NCBI cytochrome b records for three hare species plus an Oryctolagus outgroup (mixed lengths, includes two mislabelled records used as a QC teaching case).
  • New vignettes: lepus-cytb (real-data case study: locus QC, common-window cropping, barcode gap, smax semantics) and simulated-use-cases (bounded off-target scoring, adapter scan, read dereplication).

genoaligner 1.0.0

CRAN release: 2026-09-24

Initial release.

  • Edit-distance (Levenshtein / WFA-equivalent) and Smith-Waterman local alignment from a self-contained C++17 CPU core; both with score and CIGAR.
  • Pipeline-friendly, vectorised align() and align_sw() over data.frame/tibble pair columns; smax-bounded resolution that reports unresolved pairs rather than returning silently wrong scores.
  • CRAN-buildable and checkable without a GPU toolchain; R CMD check clean.
  • Apptainer recipe (containers/) to reproduce the build in a pinned cluster environment; verified on a KU HPC Q6000 node with --gres + --nv.