Changes in version 0.99.2                        

Reproducibility

  - calibrateThreshold() and runConsensus() take seed, which seeds the
    random positions of the shuffled peaks. It goes to the BiocParallel
    backend built from BPPARAM as its RNGseed, so the same seed returns
    the same threshold on one core or on many, and the random numbers of
    the session are left as they were. The calibration keeps the seed
    and its seedSource, and the object prints it.
  - The documentation said that set.seed() before the call brought the
    same threshold back on a single core. It does not: BiocParallel
    draws from random numbers of its own, which set.seed() does not
    reach, so two identical calls could calibrate two different
    thresholds. The examples, the vignette and the tests use seed now. A
    BiocParallelParam passed as BPPARAM is still used untouched, and
    carries its own RNGseed; seed is ignored with it, with a warning.

                       Changes in version 0.99.1                        

Performance

  - The confirmation step of buildConsensus() is about four times faster
    on genome-scale peak sets. It picked the supporting peak of every
    replicate with a grouped slice_max() and summed the evidence with a
    grouped summarise(), which dplyr evaluates one group at a time, tens
    of thousands of times per iteration. A single sort followed by
    distinct(), and split() with sum(), give the same peaks and the same
    sums to the last bit, so the consensus is unchanged. Three
    replicates of about 12,000 peaks each went from 16 s to 4.4 s. The
    calibration gains as much, since every permutation runs the same
    step.

                       Changes in version 0.99.0                        

First release.

consensusRegions combines the p-values of overlapping peaks across
replicated ChIP-seq, ATAC-seq and CUT&RUN experiments. A region that
convinces two samples and falls short in the third survives, where an
intersection would drop it. The method follows MSPC (Jalili et al.,
2015) and runs in R, with no .NET runtime to install.

Analysis

  - runConsensus() takes the whole workflow in one call. The individual
    functions stay available for the runs you want to inspect partway
    through.
  - Four combination schemes. Fisher reproduces MSPC. Stouffer and
    Lancaster accept per-replicate weights, so a shallow sample counts
    for less rather than being dropped or trusted whole. The rank
    product works on ranks and needs no p-values at all.
  - computeReplicateWeights() reads those weights from BAM files, from
    FRiP values you measured elsewhere, or from the peak sets themselves
    once the alignments are gone.
  - calibrateThreshold() estimates the cut on the combined statistic
    from shuffled peak sets, instead of leaving it at the customary
    1e-8. Give BPPARAM a number of cores to spread the permutations.

Input

  - narrowPeak p-values go straight in. readPeakSets() rank-transforms a
    score column when that is all a file carries, and falls back to a
    weighted k-of-n presence rule for BED3.
  - readPeakSets() puts every replicate on one chromosome naming style.
    Two replicates calling the same chromosome chr1 and 1 share no
    overlap, and the analysis would report that as biology.

Output

  - recentrePeaks() recovers summits and returns fixed-width regions,
    which is what motif analysis and count matrices want.
  - asSummarizedExperiment() hands the consensus to csaw, DiffBind or
    edgeR. exportConsensus() writes BED or narrowPeak.
  - plotRescue(), plotJaccard(), plotUpset() and plotCalibration() cover
    the quality control.

Multiple testing

buildConsensus() corrects within each replicate across every peak it
tested, not across the subset that cleared the combined threshold.
Correcting a family picked for being significant is not
Benjamini-Hochberg, whatever the column is called. Set adjustmentFamily
= "confirmed" to reproduce MSPC.

Neither setting delivers exact false discovery rate control, and the
package does not claim it. Each replicate contributes its most
significant overlapping peak, and a maximum over a selected set does not
behave like an unselected p-value. Overlapping peaks share supporters
too, so their combined values are not independent. Reach for
calibrateThreshold() when you need a defensible error rate rather than a
ranking.