Changes in version 0.99.3 New - ThresholdCalibration class. calibrateThreshold() returned a plain list, which buildConsensus() then had to check by looking for a threshold element. It now returns an S4 object with a validity method, so a malformed calibration is refused where it is made rather than surfacing later as an empty consensus. Accessors are calibrationThreshold(), calibrationThresholdNegLog10(), calibrationCurve(), calibrationTargetFDR(), calibrationAchievedFDR(), calibrationSeed(), nullStatistics() and observedStatistics(), and there is a show() method. $ still reaches the components by name, so code written against the list keeps working. - calibrationResult() returns the calibration of a ConsensusRegions object, so that every slot now has an accessor. A consensus built without calibration carries NULL rather than an empty list. Changes - The components of a ConsensusRegions object are read through their accessors throughout, so @ appears only in the class definition and in the methods themselves. - The stats slot is an S4Vectors::DataFrame rather than a data.frame. - BPPARAM defaults to NULL, which builds a BiocParallel::SerialParam() seeded from seed, so nothing runs in parallel unasked. A BiocParallelParam is used untouched and a number of cores is still accepted as a shorthand. - buildConsensus() hands its argument checking to .resolveBuildSettings(), which can be read and tested on its own. - plotCalibration() labels its axis with a plotmath expression, and ggtext is no longer a dependency. - .pairwiseJaccard() computes the upper triangle only and mirrors it, rather than every ordered pair. A set still scores 1 against itself, and 0 when it covers no bases at all. - The package man page is no longer marked internal, so ?consensusRegions is reachable from the help index, and points at the vignette. - The vignette has an installation section, says how the approach relates to MSPC and to the consensus options already in Bioconductor, and evaluates the parallel and blacklist examples instead of showing them unevaluated. - LICENSE is no longer shipped in the tarball, since GPL-3 needs no copy of its text. - The chromosome-name fallback credits RegionSetDE, by the same author, and links to it. 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.