Run devtools::document()
 

 update github workflow to Bioc3.24
 

 version 2.3.2. replace scran::findMarkers with internal findMarkersSimple
 

 disabled sample-level junction list
 

 added some comments.
 

 IRanges 2.47.2: Use new transform() signature in method definitions
 

 added vbuf to many file pointers
 

 Bump version
 

 Add check for degenerate model_matrix in predict.glmGamPoi for se.fit=TRUE
 

 fix showClass typo
 

 NEWS: frame 1.32.0 entries relative to the last release

The release notes should describe net changes since the last
release (1.30.x), not changes between intra-development states.
Two entries were written against mid-cycle states and don't make
sense to a user upgrading from the last release:

- Dropped the standalone "merge_similar2(return.clusters) now
  returns a universalmotif_df instead of a data.frame" entry:
  merge_similar2() is itself new this cycle, so there is no prior
  released behavior to migrate from. The return.clusters output is
  now described as part of the merge_motifs2()/merge_similar2()
  feature entry.

- Dropped the rev_comp_mat()/view_motifs2() reverse-complement
  bug-fix entry: view_motifs2() is new this cycle, so that bug
  never shipped in a release.

Also narrowed the negative-letter-stack rendering fix to
view_motifs() (the released function that actually carried the
bug; view_motifs2() is new).
 

 vignettes: cross-reference v2 functions throughout the prose

Mention the v2 siblings (compare_motifs2, merge_motifs2,
merge_similar2, scan_sequences2, enrich_motifs2, view_motifs2,
motif_tree2) alongside their v1 counterparts where the docs
introduce or describe each function, so readers discover the
leaner alternatives in context. Also tidies assorted phrasing
and trims yamtk-subcommand-specific references to the bare
project link.
 

 view_motifs: fix overlap in negative-letter stacks of 3+

flip_neg_letters() positions each negative letter by shifting it
down from the baseline by the cumulative height of every letter
above it in the stack. The shift table is computed at the bottom
of the function:

  letMins <- c(0, letMins)
  for (i in seq_along(lets)) d$y[j_i] <- d$y[j_i] + letMins[i]

`letMins` here is the per-letter min-y (i.e. each letter's height
expressed as a negative number after the flip). With this version,
letter i was shifted by *only* letter (i-1)'s height instead of
the sum of heights for letters 1..i-1, so any 3-or-more letter
negative stack had letters 2..N piling on top of each other. With
a CWM display the bug was glaring (3 equal-magnitude negative
contributions per core column rendering as 2 visible rows);
PWM logos with 3+ significant negative letters had the same
defect but more rarely.

The fix is the one-line cumsum the previous developer left
commented out directly above:

  letMins <- c(0, cumsum(letMins))

For 1- and 2-letter negative stacks the two formulas are
arithmetically identical, so PWM logos that used to render
correctly continue to do so. Full test suite still passes with
no new failures.
 

 add CWM (Contribution Weight Matrix) motif support

Adds CWM as a first-class fifth motif type alongside PCM / PPM /
PWM / ICM, with TF-MoDISco / DeepLIFT-style attribution scores
in mind: signed per-position per-letter contribution values, no
column-sum constraint, no probabilistic interpretation.

C++ layer (src/universalmotif-class.cpp,
src/utils-exported.cpp): accept "CWM" as a 5th type string in
both TYPES2_e and TYPES_e enums; matrix-constraint check
(check_motif_and_type) falls through (no positivity, no
column-sum check) just as it does for PWM; auto-detection of
type from a matrix is unchanged (CWM is explicit-only);
position_icscoreC and get_consensusC / get_consensusAAC each get
a CWM branch that column-normalises absolute values into a
PPM-shaped column before the existing IC / consensus logic runs.

R conversion (R/convert_type.R, R/utils-motif.R): new
MATRIX_cwm_to_ppm() helper computing
ppm[i,j] = |cwm[i,j]| / sum_i |cwm[i,j]| (TF-MoDISco convention,
with a uniform fallback for all-zero columns). convert_type()
gains "CWM" as an accepted source; convert_from_cwm() routes
CWM -> PPM directly and CWM -> PWM / ICM / PCM via PPM (so
CWM -> PWM produces a real log-odds matrix, not a relabel).
The reverse direction (PPM/PWM/ICM/PCM -> CWM) errors with a
clear message.

I/O (R/read_meme.R, R/write_meme.R): both gain a CWM = FALSE
argument. With CWM = TRUE, read_meme() tags the parsed matrices
as type = "CWM" and skips the PPM normalisation check;
write_meme() emits CWM values verbatim into the
letter-probability matrix block (matching what TF-MoDISco-lite
writes). With CWM = FALSE (default), CWM motifs are first
converted to PPM via |cwm|/sum before serialisation, keeping
the output strictly MEME-spec-compliant for downstream FIMO /
AME / CentriMo consumers. No "# CWM" comment marker is added,
matching TF-MoDISco-lite's output.

Display (R/view_motifs.R, R/view_motifs2.R): both accept
use.type = "CWM"; the renderer captures the original CWM
matrices before convert_type_internal() coerces them to PPM,
then splices them back at display time so signed contributions
show with their actual heights. Y-axis labelled "contribution".
prep_single_motif_plot_data() learns a CWM branch that scales
letter heights by the per-motif max(|cwm|). PWM-specific blocks
(zero-baseline geom_hline, signed-axis breaks/ylim) widen to
also cover CWM.

Trimming (R/trim_motifs.R, R/trim_cwm.R NEW): trim_motifs()
automatically works on CWM input via the new C++ IC branch
(no R-side change needed; type tag is preserved). The new
trim_cwm() function is a dedicated CWM-aware trimmer that drops
edge columns by absolute column sum. Two modes:
fraction-of-peak (trim.threshold = 0.3 default, matching
TF-MoDISco-lite's --trim_threshold) or absolute-value cutoff
(abs.threshold non-NULL takes precedence). Walks inward from
the chosen edges only, mirroring trim_motifs()'s edge semantics.

Hardcoded type whitelists in create_motif(), get_consensus(),
get_consensusAA(), view_motifs(), view_motifs2(), and
convert_type() all extended to include "CWM".

Every v1 and v2 analysis function that routes its input through
convert_type_internal() now accepts CWM motifs automatically:
compare_motifs(), compare_motifs2(), scan_sequences(),
scan_sequences2(), enrich_motifs(), enrich_motifs2(),
merge_motifs(), merge_motifs2(), merge_similar(),
merge_similar2(), motif_pvalue(), motif_tree(), motif_tree2(),
motif_rc(). Out of scope (continue to require probabilistic
types): motif_finder(), add_multifreq(), create_sequences(),
shuffle_sequences(), get_bkg(), match_bkg(),
sequence_complexity().

Tests (tests/testthat/test_cwm.R, tests/testthat/test_trim_cwm.R):
32 new assertions covering create_motif type="CWM",
convert_type CWM<->PPM contract (including the error on
PPM->CWM), read_meme/write_meme round-trip with CWM=TRUE,
display via use.type="CWM" in v1 and v2, motif_rc/trim_motifs
type preservation, trim_cwm default vs abs.threshold,
trim.from variants, and a no-op-on-uniformly-strong-CWM
invariant. Full test suite still passes with no regressions.

DESCRIPTION version 1.31.22 -> 1.31.23.
 

 BiocGenerics 0.59.5: Add S4 generics sequence() and transform()
 

 Run devtools::document()
 

 Avoid duplicated .normalize_array_path() in write_zarr_array()
 

 Add dimension_names input checking
 

 Bump version
 

 Make 0 a valid and default zstd compression lvl
 

 Make it clearer that both v2 and v3 are supported
 

 Move more functions to store_utils
 

 Mark 4.2 as min required R version
 

 get rid of .onUnload hook
 

 bump version to 1.3.1 and document plotVenn fix in NEWS
 

 Merge remote-tracking branch 'upstream/devel'
 

 update README
 

 add general link for zenodo doi
 

 mention hexSticker R package for the logogo
 

 Use texttt instead of code
 

 fix rev_comp_mat() row-name handling for name-indexed consumers

rev_comp_mat() reverses columns and then permutes rows via
m[c(4,3,2,1), ], which physically rearranges the data AND carries
the input row names over in the new T,G,C,A order. That worked for
merge_motifs2()'s per-position accumulator (which adds the columns
positionally into an out_mat with alphabetical A,C,G,T row names,
ignoring the source row names) but produced visually wrong logos in
view_motifs2(): the logo renderer reads each column by row name,
so a position whose physical row-1 data was "the original T row"
got drawn as a "T" letter instead of an "A" (its proper RC).

The MotifDatabaseCuration vignette's AP-1 cluster view exposed the
bug: motifs flagged [RC] rendered as the reverse of TGACTCA
(ACTCAGT) instead of its proper RC (TGAGTCA). Restoring the
alphabetical row names after the permutation fixes it. Physical
data placement is unchanged, so merge_motifs2() behaviour and
tests are unaffected.
 

 add view_motifs2() and motif_tree2()

Two v2-backend siblings of the package's most user-facing
visualisation functions, following the established v2 pattern
(compare_motifs2, scan_sequences2, enrich_motifs2,
merge_motifs2, merge_similar2): leaner Pearson-correlation
alignment via compare_motifs2_align_cpp, fewer arguments, and a
small, focused public surface.

view_motifs2: picks the highest-IC motif as the anchor (same as
merge_motifs2 / view_motifs), aligns every other motif against
it in one compare_motifs2_align_cpp call, applies reverse
complement where the best alignment lands on the - strand,
pads each oriented motif into a shared column frame, and runs
the result through the same logo renderer as view_motifs.
Drops the v1-only arguments (method, score.strat, min.mean.ic,
min.position.ic, normalise.scores). DNA / RNA only.

motif_tree2: builds the distance matrix from
compare_motifs2(matrix.out = "score") and converts it to a
symmetric distance via (1 - score) / 2, then feeds it into the
same hclust / ape::as.phylo / ggtree::ggtree pipeline that
motif_tree uses. The is(motifs, "dist") short-circuit path is
preserved, so users can pre-compute their own distance matrix
and pass it in directly. Drops the v1-only arguments
(method, use.type, score.strat, relative_entropy,
min.mean.ic, min.position.ic).

Both default to nthreads = 1 to match every other user-facing
function in the package.

The v1 functions view_motifs() and motif_tree() are untouched
beyond a one-sentence cross-reference pointing readers to their
v2 counterparts. Adds tests covering the single/multi-motif
paths, RC detection on non-palindromic motifs, sort.by /
names.pos branches, alphabet rejection, dist short-circuit, and
distance-range invariants. DESCRIPTION version 1.31.20 to
1.31.21; new files added to the Collate field.
 

 version name change
 

 version updated after adding ISF annotation
 

 addtions to allow ISF annotated results to be saved
 

 ExperimentHubData 1.39.1: Drop dependency on utils

Package utils doesn't seem to be used anywhere in ExperimentHubData, so
can go away. This will also make the following 'R CMD check' warnings go
away:

* checking whether package ‘ExperimentHubData’ can be installed ... WARNING
Found the following significant warnings:
  Warning: replacing previous import ‘utils::data’ by ‘BiocGenerics::data’ when loading ‘ExperimentHubData’
  Warning: replacing previous import ‘utils::findMatches’ by ‘S4Vectors::findMatches’ when loading ‘ExperimentHubData’
 

 Merge pull request #52 from federicomarini/patch-2

Update panels.Rmd 

 Update iSEEtree.Rmd

to have the name rendered correctly on Bioc 

 fix the intergenic release used in the vignette
 

 updated ggplot2 deprecated fns
 

 avoiding deprec warnings
 

 Remove importing utils::data() b/c (1) is not used and (2) it has been promoted to BiocGenerics; see issue #22
 

 vignette and readme rendering
 

 Bump version
 

 Document S3 support in H5Fopen()
 

 enable remote H5open
 

 Set R CMD check level back to WARN
 

 Reduce duplication in h5checktype
 

 Add de.NBI ROR
 

 version
 

 Increment version number to 0.99.10
 

 Fix README
 

 since cellCounts now works well, the env check is removed.
 

 removed debug output
 

 Fixed many bugs in memory-realign buffer
 

 version bumps and news
 

 Merge branch 'devel' of https://github.com/npcooley/SynExtend into devel

update version number
 

 Merge pull request #82 from ahl27/master

Documentation fixes 

 update documentation
 

 add function, fix indexing bug
 

 Fix issue with Standardize_Subset where something inexplicable was happening.

Not sure what I was even intending to do. Previous state of this system was doing
something crazy when passed input of `Subset` that was an integer vector.

I've simplified this codepath to make it more understandable and left some
documentation. Should be working correctly now.
 

 bumphunter 1.55.1: No need to import data() from utils

Package internal code doesn't use utils::data() so there's no need
to import it.
 

 Bump version
 

 Fixed in the documentation
 

 Fix parameter whdim in call to .check_ondisk()

Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> 

 Fixed saveHDF5GSVAranks() and loadHDF5GSVAranks()
 

 Added dropExistingAssays parameter to gsvaRowNorm() and gsvaColRanks(). Added corresponding unit tests
 

 Fixed gsvaEnrichment(), updated show() method, updated scRNA-seq vignette
 

 Added gsvaColRanks() and gsvaColScores() functions
 

 load index once only
 

 bump version
 

 update description
 

 Use rlang::inject() instead of .create_replace_call()
 

 Use global variables for supported codecs
 

 Refactor json S3/local reading
 

 add ChIP-seq workflow vignette + view_motifs sort.by + match_bkg genome path

New end-to-end ChIPseqWorkflow.Rmd vignette walks through a complete
motif analysis on real ChIP-seq peaks, using a newly bundled BED
(VAL1-GFP peaks in Arabidopsis from Yuan et al. 2021, NAR 49:98-113)
that lives in inst/extdata. The workflow: import peaks via
rtracklayer, extract sequences from BSgenome.Athaliana.TAIR.TAIR9,
build a composition-matched genomic background with match_bkg(),
discover motifs with motif_finder(bkg.sequences=...), deduplicate
with merge_similar2(), label clusters by their biological identity
(RY, G_box, ...), compare to JASPAR2022 plants via MotifDb,
validate enrichment with enrich_motifs2(), test central positional
bias with motif_peaks(), and examine RY-self and RY-other spatial
clustering with motif_coocc() plus per-pair distance histograms.
The biological story: six distinct real TF families (B3-domain
ABI3, bZIP ABF3, BPC1, DOF CDF5, RVE8, SPL7) come out of one
peak set.

match_bkg() gains a `genome` argument (BSgenome object) plus
companion `exclude` (GRanges to skip), `n.candidates`, and
`chromosomes` knobs. When supplied, the function samples random
non-overlapping genomic windows internally, replacing the
sample-and-getSeq boilerplate that was otherwise duplicated in
every ChIP-seq workflow. Mutually exclusive with `universe`.

view_motifs() gains a `sort.by` argument with options
"none" (default; input order preserved while alignment still
uses the highest-IC anchor), "ic" (descending information
content), and "similarity" (hierarchical clustering on Pearson
correlation between aligned columns, so visually similar motifs
are adjacent).

DESCRIPTION: adds rtracklayer and BSgenome.Athaliana.TAIR.TAIR9 to
Suggests; version 1.31.13 -> 1.31.16. New BibTeX entry yuan2021val.
Added 11 new tests (7 for match_bkg, 4 for the genome path) and
all 643 tests pass.
 

 drop plot_motif_coocc and the motif_clusters stub

- plot_motif_coocc() removed; users with the data.frame can build any
  ggplot2 heatmap they want without a thin wrapper.
- R/motif_clusters.R removed -- unfinished v1 stub (empty function
  body, never exported, never tested). Dropped from Collate.
- NEWS, vignette, tests, NAMESPACE, man/ all cleaned up.
 

 motif_coocc(): make spatial mode descriptive, not part of the Fisher test

The previous implementation filtered the `both` cell by spatial
distance but kept `a_only` and `b_only` based on the unfiltered
presence sets, leaving the 2x2 internally inconsistent
(a_only + b_only + both could exceed n.sequences, neither clamped to
zero). The resulting Fisher p-value lost its meaning.

The conceptual issue is bigger than the math: a proper spatial test
(SpaMo-style) is conditional on co-presence and tests the distance
distribution against a null -- not a 2x2 Fisher.

New semantics: the Fisher test always answers the SET question (do
these motifs co-occur in the same sequences?). When `max.distance`
is supplied, two descriptive columns are added:
  - both.clustered: subset of `both` with a within-distance pair.
  - median.distance: median nearest-pair distance.

Read together: pvalue/qvalue say "do they co-occur?", median.distance
hints "when they co-occur, do they cluster?". Heterodimer-like
arrangements show up as small q AND small median.distance.
 

 vignette: tone down implant_motifs() blurb
 

 drop rng.seed from match_bkg() and implant_motifs()

Both are pure-R and used set.seed(rng.seed) inside the function body,
which silently mutated the caller's global .Random.seed -- a real
side-effect bug.

Caller now controls reproducibility via set.seed() before the call,
which is the standard R idiom. Tests and vignette examples updated to
match. rng.seed remains on functions that pass the seed into C++ RNG
state (create_sequences, motif_pvalue, motif_finder, shuffle_sequences,
enrich_motifs, enrich_motifs2, shuffle_string) where it doesn't touch
R's global state.
 

 mark candidate sites for RcppThread::isInterrupted() in long C++ hot paths
 

 match_bkg: fix sample() length-1 gotcha + ring-cap fallback

Two related bugs caused match_bkg() to return wildly-mismatched
backgrounds when the target distribution was narrow and the home
bin's available pool dropped to a single index:

  1. R's `sample(x, n)` treats a length-1 numeric `x` as `1:x` and
     returns sample.int(x, n). So `sample(c(201), 1)` returns a
     random integer in [1, 201], not the value 201. The find_candidates
     loop was hitting this when a cell had only one universe
     sequence available after dedup -- which then produced an
     arbitrary universe index unrelated to the target's bin.

  2. The original unbounded ring-expansion (up to max(n.gc.bins,
     n.len.bins) rings) was too aggressive: once nearby cells got
     drained via unique=TRUE, ring expansion would walk far out to
     bins compositionally unlike the target before finally hitting
     the global random fallback.

Fixes:
  - All sample() calls switched to `pool[sample.int(length(pool), n)]`
    which is immune to the length-1 gotcha.
  - Ring expansion is now capped at MAX_LOCAL_RING = 2. Beyond that,
    the function falls back to with-replacement sampling from the
    accumulated *local* pool (rings 0-2). This preserves the
    composition match at the cost of occasional duplicates, instead
    of degrading into a global random sample.

On the canonical demo fixture (target GC ~0.68/length ~400, universe
GC ~Beta(2,2)/length ~U(200,800)):
  - Before: matched bkg GC = 0.52 +/- 0.22, length = 474 +/- 154
  - After:  matched bkg GC = 0.67 +/- 0.05, length = 389 +/-  61

Includes the linter's incidental polish to match_bkg.R / man /
vignette (drop the chromVAR reference, tighten markdown emphasis).
 

 add match_bkg() and plot_match_bkg(): composition-matched backgrounds

`match_bkg()` samples background sequences from a universe XStringSet
that match the input set on GC fraction and sequence length, the
HOMER-style binned approach:

  1. Bin the universe by (GC equal-width, length quantile-based).
  2. For each target, find its (GC, length) cell and sample
     n.per.target universe sequences from it.
  3. If the home cell is empty / undersized, expand outward in
     Manhattan-distance rings (with a warning). Fall back to global
     with-replacement sampling only if the universe is truly too
     small to cover the target set.

Returns either the matched XStringSet (default) or a data.frame of
per-pair match indices (return.indices = TRUE) for traceability.

Complements (does NOT replace) shuffle_sequences(): the shuffle path
randomises each input in place; match_bkg() instead samples real
sequences from a larger genomic universe. Use the latter when the
input has composition bias relative to the genome that you want to
control for in enrich_motifs2() / motif_finder().

plot_match_bkg() is a companion ggplot helper that overlays GC and
length density curves for target vs matched-background, useful for
visually verifying the match landed. Same theme conventions as
plot_motif_peaks() (no interior gridlines, L-frame axes).

20 unit tests including: shape/class, GC-match quality, length-match
quality, determinism, unique-without-replacement, indices output,
n.per.target = N, AA rejection, alphabet mismatch, small-universe
fallback, empty-bin ring expansion, plot ggplot class. Integration
test through enrich_motifs2 round-trips cleanly.

No new package dependencies. No C++ code.
 

 motif_peaks: add v1 GRanges test + polish plot_motif_peaks() output

- Add a unit test covering scan_sequences() (v1) GRanges round-trip:
  motif_peaks() auto-resolves seq.length from seqlengths(hits) for v1
  GRanges the same way it already did for v2 GRanges.

- plot_motif_peaks() polish:
    * black bars (was grey25)
    * no interior grid lines
    * L-frame axes only (kill panel border, draw bottom + left axis
      lines explicitly)
    * bottom x-axis line, ticks, and tick labels on every facet (via
      axes = "all_x" / axis.labels = "all_x")
    * strip-title outline removed
    * per-motif p-value / enrichment annotation moved into the strip
      label via a labeller -- avoids the overlap that happened when
      the tallest bar landed at the top-left corner of a panel
 

 rewrite motif_peaks(): CentriMo-style positional enrichment

Wholesale replacement of the old kernel-density-with-permutation-null
implementation, which never matched the function's intended CentriMo
semantics. The new version takes its input directly from
scan_sequences() / scan_sequences2() (either data.frame or GRanges)
and answers the actual question: "are this motif's hits non-uniformly
distributed along the sequences?"

Algorithm (Bailey & Machanick 2012):
  1. Normalise hits (best score per (motif, sequence) pair).
  2. For each motif and each candidate window width, compute the
     one-sided binomial p-value for hit-count-in-window under the
     null "hits uniform over [1, seq.length]".
  3. Pick the best (smallest-p) window per motif; Bonferroni-correct
     over windows tested; BH-correct across motifs.

Two modes:
  - mode = "central" (default): only centre-of-sequence windows.
  - mode = "local": window centre is also varied across the sequence.

Companion plot_motif_peaks() returns a ggplot faceted by motif with a
hit-centre histogram and the best window shaded. It consumes the new
`centers` list-column on motif_peaks()'s return, so users don't have
to thread their scan results back in.

The C++ helpers used by the old stub (peakfinder_cpp, linbin_cpp in
src/motif_peaks.cpp) are gone -- they were unique to that stub and
unreferenced elsewhere.

This is a hard incompatible rewrite under the same name: the v1
signature (hits, seq.length, seq.count, bandwidth, max.p, peak.width,
nrand, plot, BP) has no v2 counterpart. NEWS notes the break.
 

 docs: polish merge_motifs/similar wording + cross-link v1 docs to v2
 

 add merge_motifs2() and merge_similar2(): significance-based motif merging

Companion v2 functions to merge_motifs() and merge_similar(), built on
compare_motifs2(). Both DNA/RNA only.

merge_motifs2(motifs):
  - Anchor-based progressive alignment via compare_motifs2()'s C++
    alignment finder. Picks the highest-IC input motif as the anchor;
    every other motif's best alignment offset / strand / overlap is
    computed against the anchor in a single C++ call.
  - Motifs whose best alignment is on the - strand are reverse-
    complemented before averaging, so the merged motif is always in
    the anchor's orientation. This removes the input-order dependence
    of merge_motifs()'s iterative pairwise scheme.
  - Column-union averaging in R: each output column at frame position p
    is the (optionally @nsites-weighted) mean of the per-motif PPM
    columns covering p. No automatic IC-trim -- users who want that
    can pipe through trim_motifs().

merge_similar2(motifs, qvalue):
  - Calls compare_motifs2(matrix.out = "qvalue") to get an N x N
    pairwise q-value matrix.
  - Symmetrises Q via pmin(Q, t(Q)) (pair linked if either direction
    is significant; per-query BH adjustment in compare_motifs2 makes
    Q asymmetric).
  - Connected-components clustering via a small union-find: motifs i
    and j are in the same cluster iff Qsym[i,j] <= qvalue. Each
    multi-motif cluster is collapsed via merge_motifs2(); singletons
    pass through unchanged.
  - Returns a list of universalmotif S4 objects shorter than the
    input, or a data.frame of cluster assignments when
    return.clusters = TRUE.

This is the STAMP / TOMTOM-clustering semantics ("group all motifs
that are pairwise significantly similar"), which is more interpretable
than v1's score-threshold hierarchical clustering and removes the
linkage-method / cutoff-unit ambiguity.

merge_motifs() and merge_similar() now emit one-line tips pointing at
the v2 functions when the user's argument set maps cleanly. Silence
with options(universalmotif.suggest.merge_motifs2 = FALSE) /
universalmotif.suggest.merge_similar2 = FALSE.
 

 docs: tighten motif_finder() vignette section
 

 motif_finder: tighten allocations, background, and interruptibility

Four follow-up improvements to the initial motif_finder() landing:

  1. Background from positives. The C++ side was being passed uniform
     (0.25, 0.25, 0.25, 0.25), which diverged from yamtk me's default
     (compute from the actual sequence composition). Now the R wrapper
     computes the bkg from the combined positives + bkg.sequences via
     Biostrings::letterFrequency (with yamtk's 0.001 floor + renorm) and
     forwards it both to the C++ scanner and to each discovered motif's
     @bkg slot.

  2. Allocation churn. Three hot loops were rebuilding scratch buffers
     on every call:
       - fill_cdf()'s tmp_pdf is now a member of the Motif struct,
         grown but never shrunk across CDF builds.
       - enumerate_seeds()'s per-sequence `seen` unordered_set is now
         hoisted out of the inner loop and clear()'d, so the bucket
         capacity is reused across sequences.
       - discover_for_width()'s `covered` bitmask is pre-allocated once
         per width and zeroed in-place between seed attempts (only
         re-shaped if a previous attempt moved it into local_results).

  3. Drop the manual consensus workaround. Now that the upstream
     create_motif() matrix-dispatch bug is fixed (737d8e4),
     create_motif() correctly populates @consensus and motif colnames,
     so the per-column get_consensus() loop in the R wrapper is
     redundant. Twelve lines deleted.

  4. Safety. RcppThread::checkUserInterrupt() now fires at the top of
     each per-width iteration -- the RcppThread variant throws a C++
     exception (vs longjmp), so STL destructors run cleanly and per-
     width buffers are released on Ctrl-C. Also added defensive input
     guards at the motif_finder_cpp entry (empty seqs, bad pvalues,
     non-positive bkg, etc.) so direct C++ callers can't wander into
     runaway allocations or divide-by-zero.

All 21 motif_finder tests still pass. No behavior change on valid
inputs beyond the bkg-from-positives shift (which makes the discovered
PWM log-odds correct for non-uniform sequence sets).
 

 rewrite motif_finder(): yamtk me-aligned de novo discovery

Replaces the 109-line, unexported, never-tested stub with a real
implementation modelled after the yamtk me command-line tool
(https://github.com/bjmt/yamtk). The new function is exported and
returns a universalmotif_df, slotting into the v2 stack alongside
scan_sequences2() / compare_motifs2() / enrich_motifs2().

Pipeline (faithful port of yamtk/src/yamme.c):
  - For each motif width in [min.width, max.width], iterate up to
    nmotifs:
    - enumerate over-represented k-mer seeds (Fisher's exact log-p
      on per-sequence canonical-strand presence)
    - build initial PPM by aligning Hamming-1 neighbours of the
      best seed
    - convert PPM -> integer PWM, build score CDF, choose threshold
      by hit p-value
    - refine PPM twice (re-scan, rebuild, re-threshold)
    - evaluate on positives + negatives; accept if Fisher's per-seq
      presence p-value below stop.pvalue
    - mask covered positions and try the next seed
  - merge per-thread results in (width, discovery_seq) order
  - re-evaluate every motif on pristine positives so cross-width
    dedup compares apples-to-apples
  - drop motifs whose coverage overlap with a lower-p-value motif
    exceeds dedup.overlap (default 0.5)
  - BH q-values across survivors; filter by qvalue cutoff
  - IC-trim low-information flanks

Width-level parallelism via RcppThread::parallelFor. Background is
auto-generated by k=2 euler-shuffle of positives if not supplied.
DNA/RNA only (matches yamtk me).

src/motif_finder.cpp is a fresh ~700-line file; the old C++ helper
(calc_seq_probs_cpp, which only served the stub) is gone.
 

 bump to 1.31.7 and document enrich_motifs2() in NEWS
 

 add enrich_motifs2(): minimalist yamtk-aligned motif enrichment

Companion to scan_sequences2() and compare_motifs2(): a 10-formal
DNA/RNA enrichment function whose default surface mirrors
yamtk enr (https://github.com/bjmt/yamtk). Internally calls
scan_sequences2() on target and (auto-shuffled or user-supplied)
background sets, aggregates per-motif counts, runs
fisher.test(alternative = "greater"), applies BH q-value
adjustment, and returns a plain data.frame whose 13 columns mirror
yamtk's TSV one-for-one (motif, motif.i, consensus,
target.seq.n / .seq.hits / .site.hits, bkg.seq.n / .seq.hits /
.site.hits, enrichment, log2.enrichment, pvalue, qvalue).

Two test modes via the test argument:
  - "seqs"  (default): Fisher on per-sequence hit presence
  - "sites": Fisher on per-position rate
matching v1's mode = "seq.hits" / "total.hits". The "ranksum"
yamtk mode is intentionally out of scope for v2.

enrich_motifs() also gains a one-line suggestion hint that fires
when a call's argument set maps cleanly onto enrich_motifs2().
Silence with options(universalmotif.suggest.enrich_motifs2 = FALSE).
 

 bump to 1.31.5 and document compare_motifs2() in NEWS

DESCRIPTION: 1.31.4 -> 1.31.5 (patch bump, development cycle).

NEWS: under "CHANGES IN VERSION 1.32.0", add a NEW FEATURES entry for
compare_motifs2() (yamtk-aligned motif-vs-motif comparison with both
empirical and parametric null modes) and a MINOR CHANGES entry for
the new compare_motifs() -> compare_motifs2() suggestion hint.
 

 add compare_motifs2(): minimalist yamtk-aligned motif comparison

compare_motifs2() is the motif-comparison counterpart to
scan_sequences2(): a deliberately stripped-down version of
compare_motifs() whose defaults and algorithm mirror the command-line
tool `yamtk cmp` (Tremblay 2026, https://github.com/bjmt/yamtk). It
uses Pearson correlation per column as the only similarity metric and
computes p-values from a TOMTOM-style empirical null PMF built from
the target database, with Bonferroni + Benjamini-Hochberg adjustment.

Signature (eight formals):

  compare_motifs2(motifs, compare.to = NULL,
                  qvalue = 0.1, min.overlap = 5L, RC = TRUE,
                  null = c("empirical", "parametric"),
                  bkg = NULL,
                  matrix.out = c("score", "pvalue", "qvalue"),
                  nthreads = 1)

Two null-PMF modes, both ported verbatim from yamcmp.c:

  - "empirical" (default): per-query column, histogram PCC against
    every target-database column, convolve to overlap length L. Best
    for large databases.
  - "parametric": enumerate 56 simplex-grid columns weighted by a
    Dirichlet-Multinomial PMF (alpha = 4 * bkg). Best for small
    databases or when histogram sparsity would distort the empirical
    estimate.

Two output shapes:

  - compare.to = NULL: square numeric matrix (motif x motif) holding
    the value selected by matrix.out (mean PCC, p-value, or q-value).
  - compare.to provided: long-format data.frame of pairs filtered by
    qvalue, with the same column shape yamtk emits to TSV (subject,
    target, offset, strand, overlap, score, consensus strings,
    pvalue, qvalue).

The reported `score` is mean PCC over the overlap (always in [-1, 1])
rather than the raw sum yamtk reports. The underlying sum is still
what the p-value lookup indexes into.

Output conventions diverge slightly from compare_motifs() to match
yamtk and universalmotif convention: column names are subject/target
(universalmotif), the score is unit-normalised (this package), and
the per-motif p-value asymmetry of the empirical null is preserved
(non-symmetric pvalue/qvalue matrix; the score matrix remains
symmetric).

Performance: 100% hit-set overlap with yamtk in both modes on a
50-motif HOCOMOCO fixture; all score/offset/overlap/strand exactly
identical; p/q-value relative error < 4e-6. Threading scales 3.1-3.3x
at 4 threads. matrix.out = "score" skips the p-value path entirely,
giving a 6-10x speedup when significance isn't needed.

Implementation:

  - src/compare_motifs2.cpp: ~450 lines. Per-pair best alignment
    scanner + yamtk-style p-value path + IUPAC consensus helper.
    Three Rcpp exports, RcppThread-parallelised over queries with
    pre-extract/parallelise/wrap idiom.
  - R/compare_motifs2.R: ~220 lines. Arg validation, motif
    normalisation, mode dispatch, BH adjustment, output reshape.
  - utils-internal.R: suggest_compare_motifs2() hint emitted from
    compare_motifs() when arguments map cleanly onto compare_motifs2().
    Silence with `options(universalmotif.suggest.compare_motifs2 = FALSE)`.

Tests: 16 test blocks / 47 assertions covering matrix symmetry,
diagonal = 1, score range, matrix.out variants, self-comparison
sanity, both null modes, custom bkg behaviour, qvalue filtering,
empty results, AA rejection, character-name compare.to.
 

 Update script
 

 feat: add genome<- method for EnsDb

- Add a `genome<-,EnsDb` method to allow manually changing/setting the Genome
  version (related to issue #159).
 

 Declare NSE globals for check
 

 version bump 1.33.2
 

 un-defunct mirToRanges
 

 Bump version
 

 Silence cyclocomp linter
 

 Refactor .read_consolidate_metadata() with explicit handling of array zarr.json
 

 Add overwrite parameter
 

 Clarify that `zarr_overview()` can apply to groups
 

 Create zarr_consolidate_metadata()
 

 Add deNBI ROR
 

 Update data types support table with ref to grumpy
 

 Simplify .determine_s3_provider()
 

 Avoid renaming vlen-utf8 codec

To simplify code
 

 Used a semi-compressed realign temp bin file
 

 Bioconductor 0.99.19 Vignettes
 

 Added the gesel library.
 

 Bioconductor 0.99.19 Vignettes
 

 Bioconductor 0.99.19 Vignettes
 

 Bioconductor 0.99.19 Vignettes
 

 final version
 

 multiple updates
 

 version 2.3.1
 

 update documentation
 

 correct bug for calculating average
 

 v2.1.3
 

 v2.0.2
 

 handle regulon without idxATAC
 

 v2.0.1
 

 Update Bioconductor workflow for R 4.6 and 3.23 

 Merge branch 'master' of github.com:xiaosaiyao/epiregulon
 

 Merge branch 'devel'
 

 Merge pull request #32 from xiaosaiyao/devel

Devel 

 Merge pull request #29 from xiaosaiyao/devel

Devel 

 Merge branch 'devel'
 

 Merge pull request #28 from xiaosaiyao/devel

Devel 

 Merge branch 'devel'
 

 Merge branch 'master' of ssh.code.roche.com:scwg/gene-transcriptional-network/activity-inference/Epiregulon
 

 Merge pull request #25 from xiaosaiyao/devel

Devel 

 Merge pull request #23 from xiaosaiyao/devel

Devel 

 version bump 1.7.1
 

 incl. email in package doc
 

 add importFrom tags for roxygen2
 

 add returns and examples to azure-class doc
 

 add return value and example for gcp-class doc
 

 formatting and spacing for BiocCheck
 

 update to changes in roxygen2 v8
 

 add avtable_* method descriptions to docs
 

 update R version to 4.6.0
 

 add lifeCycle import to gsutil
 

 version bump 1.3.2
 

 show error message from gcloud if available
 

 version bump 1.7.1
 

 Version bump in devel
 

 GenomeInfoDb 1.49.1: Import data() from BiocGenerics instead of utils
 

 v0.99.4: fix BiocCheck warnings and update installation method
 

 Merge remote-tracking branch 'upstream/devel'
 

 Merge pull request #17 from rformassspectrometry/gabri

fix: issue white spaces in ftp link 

 bug fix: factor-coerce NMI inputs for aricode 1.1.0 in Bioc 3.24

calculate_performance() in R/helper_functions.R was passing
aricode::NMI() a character vector of cell-type labels and a numeric
vector of cluster IDs. aricode 1.1.0 (currently shipping in
Bioconductor 3.24) tightened input handling and now routes mixed-type
arguments through sort_pairs(), which calls as.integer() on the
character argument — turning labels such as "alpha" into NA and
aborting with "NA are not supported." Both grid-search
configurations in the vignette therefore failed, cascading to
"All parameter configurations failed; cannot select best_params"
on the Bioconductor build farm.

Coerce both NMI inputs to factor so labels and clusters share a
well-defined integer encoding. This restores the vignette and
preserves the score semantics (NMI is invariant to label coding).

This is the underlying cause that the previous fix (1.1.1) surfaced
with a clearer error message but did not address.

Verified in bioconductor/bioconductor_docker:devel (R 4.6.0 /
Bioc 3.24):
  R CMD build --keep-empty-dirs --no-resave-data --md5 CellMentor
    → creating vignettes ... OK
  R CMD check --no-manual CellMentor_1.1.2.tar.gz
    → Status: OK
 

 bump to 0.99.1
 

 Update
 

 Biostrings 2.81.2: Move utils from Imports to Suggests
 

 removed a redundant var
 

 BiocGenerics 0.59.3: Define the data() S4 generic

Known issue: Having

    \usage{
      data(x, ...)
    }

in man/data.Rd triggers the following 'R CMD check' warning:

    * checking for code/documentation mismatches ... WARNING
    Data with usage in Rd file 'data.Rd' but not in code:
      ‘x’

which is of course spurious because this man page is **not** about
documenting a dataset but is about documenting a function (the data()
generic function). This is confirmed by the fact that our man page
does **not** contain \docType{data}, so is clearly and unambiguously
documenting a function. Unfortunately, for some reason 'R CMD check'
ignores that. Something to report to the R core folks at some point...

In the mean time, we're putting

    \usage{
      #data(x, ...)
    }

in man/data.Rd to keep 'R CMD check' quiet.
 

 update links based on roxygen changes
 

 use interactive() criteria to assign *h_meta object in .GlobalEnv

- show notification if not running app locally
 

 version bump 1.3.1
 

 use roxygen2 v8
 

 move importFrom lines above examples
 

 shQuote exclude in gsutil_sync rather than double quote
 

 gsutil_is_uri checks if source starts with gs
 

 use encode arg for httr::POST
 

 more test
 

 Bump version
 

 Bump version
 

 bump again
 

 Version bump to 1.39.2
 

 Version bump to 1.39.1 + Date adapted
 

 bug fix: Switching from testthat::expectation to testthat::expect_no_errors

Closes #20
 

 Merge pull request #33 from lamortenera/31-bioconductor-failure-and-risk-of-deprecation

31 bioconductor failure and risk of deprecation 

 Version bump + last touched date added
 

 bug fix: Switching from testthat::expectation to testthat::expect_no_errors

Closes #31
 

 Push to winbuilder again
 

 windows, alsways use srialparam for building vignette
 

 Change url neghyper
 

 bump to 1.31.4 and document scan_sequences2 / dedup_hits in NEWS

DESCRIPTION: 1.31.3 -> 1.31.4 (patch bump, development cycle).

NEWS: under "CHANGES IN VERSION 1.32.0" (the next release version per
Bioconductor's even-minor convention), add NEW FEATURES entries for
scan_sequences2() and dedup_hits(), MINOR CHANGES entries for the
batched dynamic motif_pvalue() path and the scan_sequences ->
scan_sequences2 suggestion hint, and a BUG FIXES entry for the
nthreads = 0 silently-serial regression.
 

 hint at scan_sequences2() from scan_sequences() when applicable

Adds an opt-out message inside scan_sequences() that fires when the
user's argument set maps cleanly onto scan_sequences2()'s feature
surface. The hint suggests trying scan_sequences2() for better
performance.

Conditions for the hint (all must hold):
  threshold.type        == "pvalue"
  use.freq              == 1
  !allow.nonfinite
  !no.overlaps                      (different semantics in scan_sequences2)
  !calc.qvals
  !respect.strand
  motif_pvalue.method   == "dynamic"
  motif alphabet         is DNA or RNA
  !use.gaps || no gapped motifs

The hint is emitted via the existing wmsg()/message() pattern. Silence
with `options(universalmotif.suggest.scan_sequences2 = FALSE)` (default
TRUE; initialised in zzz.R alongside the other universalmotif options).

Implementation is a small helper suggest_scan_sequences2() in
utils-internal.R, called from scan_sequences() right after motif
preprocessing where the alphabet and gap-status are known.

7 new assertions in test_scan_sequences.R confirm the hint fires under
the eligible default case, is suppressed by each disqualifying argument
(no.overlaps, respect.strand, motif_pvalue.method = "exhaustive",
allow.nonfinite via the logodds path, non-pvalue threshold.type), and
is silenced by the option flag. Full regression suite passes.
 

 add scan_sequences2(): minimalist yamtk-aligned scanner

scan_sequences2() is a deliberately stripped-down counterpart to
scan_sequences() whose default surface mirrors `yamtk scan`. Six
formals only: motifs, sequences, pvalue (default 1e-4), RC (TRUE),
nthreads (1), return.granges (NULL = auto). Drops multifreq, gapped
motifs, q-values, exhaustive p-value mode, overlap removal,
respect.strand, allow.nonfinite, and every threshold type other
than p-value.

Reuses the existing C++ scanner (scan_sequences_cpp) and the batched
dynamic-DP p-value path. PWMs are built via universalmotif's own
convert_type_internal() pipeline, so output is not bit-identical to
yamtk (universalmotif rounds counts in the PCM step; yamtk operates
on continuous PPMs throughout) but agrees on ~95% of hits at p=5e-5
on a HOCOMOCOv11 / random-DNA fixture.

Output convention differs from scan_sequences() in two ways:
  (1) coordinates are always start <= end regardless of strand,
      matching BED / GRanges / yamtk;
  (2) the match string on '-' strand hits is reverse-complemented
      (matched orientation), again matching yamtk.

Default output is a GRanges (when GenomicRanges is installed);
data.frame fallback otherwise. data.frame columns:
motif, motif.i, sequence, sequence.i, start, end, strand, score,
score.pct, match, pvalue.

Per-hit p-values are always computed via motif_pvalue(method =
"dynamic"), which now batches across motifs in a single C++ call.
With the batched p-value path this hits ~3.0x of yamtk at 4 threads
on long-sequence workloads -- meeting the Tier-1 roadmap target of
"within 2-3x of yamtk".

Tests: 14 unit tests covering determinism, coord convention, RC
match reversal, score/score.pct internal consistency, GRanges /
data.frame shapes, error cases, and equality with scan_sequences()
on shared hits.
 

 fix nthreads=0 to actually use all available threads

RcppThread::parallelFor(..., nthreads=0) silently runs on a single
worker instead of all cores, despite the documented contract in every
nthreads-aware function. Translate `nthreads = 0` to
parallel::detectCores() at the R level before the value reaches C++,
via a shared resolve_nthreads() helper in utils-internal.R.

Applied to: scan_sequences, shuffle_sequences, compare_motifs,
motif_pvalue, motif_tree, create_sequences, get_bkg, enrich_motifs,
make_DBscores, sequence_complexity, window_string.
 

 Load packages required for the vignette
 

 Mention oversized chunks in NEWS
 

 Fixed tests
 

 Updates to the integrated training in the latest singlepp.

This introduces a top= argument to consolidate the pairwise markers into a
smaller set of per-label markers. The hope is to reduce the runtime of
combineRecomputedResults() by using a smaller feature space.
 

 Merge pull request #109 from Bisaloo/runnable-vignettes

Make vignettes before rendering actually runnable 

 Make vignettes before rendering actually runnable
 

 added sample index and quality into the temp bin
 

 github  action
 

 import clusterProfiler
 

 freed thread context after running
 

 added input binfile buffer
 

 Added junction temp file in cellCounts
 

 temp file tests for junction detection
 

 Fix codefactor issues
 

 Increase coverage
 

 Bioconductor submission changes (#41)

This commit implement the changes suggested during Bioconductor revision:

- It is incorrect to state the edgeR and DEseq only perform different gene abundance analysis.
- The vignette should use Bioconductor styling. Please refer to R Markdown Vignettes.
- You should explain why readcounts contains decimal numbers rather than integers.
- Text has "a q-curve per sample and gene." but Figure 1 has no gene names, unlike Figure 2.
- Don't use for loops where vapply could be used according to Code Syntax and Efficiency.
- Some parameters, such as pi0_method have a set of valid values. Please use match.arg at start of function to check.

In addition, it increases test coverage. 

 version bump 1.31.7
 

 version bump 1.25.3
 

 Merge pull request #121 from LieberInstitute/xenium_vis_functions

Adapt vis_clus() functions for use with Xenium data 

 Merge remote-tracking branch 'upstream/devel' into xenium_vis_functions
 

 Add support on the shiny app for Xenium data
 

 Make vis_grid_gene() and vis_grid_clus() support
`datatype` from vis_clus() and vis_gene()
 

 Update docs
 

 Add datatype checks
 

 Fix Cross-refrences in descriptions
 

 Update date
 

 Adapt vis_gene for xenium data
 

 Add datatype arg to vis_gene
 

 Add vis_gene_c
 

 Fix typo in example
 

 Update doc with fixed example
 

 fix unit
 

 Add xenium example
 

 Add spe_xenium_example to fetch_data (need to update github when ready)
 

 Update docs
 

 Add vis_clus_c() function for handling visulization of Xenium/centroid based data
 

 comply with aricode input expectation
 

 Fix punctuation README.md
 

 Fix format issue
 

 Bioconductor submission review
 

 small fix to unit tests
 

 version bump 1.25.2
 

 use double bracket subsetting instead of $
 

 fixed by roxygen; link refers to topic not Rd name
 

 use multiple lines for importFrom
 

 fixed genome package eval away from `eval(parse`
 

 Bump version
 

 Remove internal function man page
 

 Mention new supported types in features vignette
 

 Merge branch 'devel' of git.bioconductor.org:packages/scDblFinder into devel
 

 Pass on nthreads to xgboost functions
 

 Fix bug for updating
 

 Adapted the code to the R version update to version 4.6 (#139)

- Changed roxygen to 8.0.0
- As consequence all links in documentation were updated
- Ensured all packages required by `COTAN` are in the `Imports` list 

 Fix cyclocomp lint
 

 Mention all grumpy changes in NEWS
 

 Add support and test for v3 structured datatype
 

 Run devtools::document()
 

 Handle new structured data_type name
 

 Simplify .update_fill_value()
 

 Add grumpy to DESCRIPTION
 

 Replace codec_bytes_decode() by grumpy::parse_npy_data()
 

 Update R-CMD-check.yaml

remove ubuntu devel from automated
 

 update for loops
 

 Swithc to snowparam
 

 Mention fixed length string types renames in NEWS
 

 Bump version
 

 Run devtools::document()