STACAS v2.2.0

New in this release:

• Enable scale.data option for integration. It may be beneficial to scale the data if batch effects are very large, e.g. when integrating across species, or single cells with single nuclei
• Include gene name conversion table for human genes
• Add min.sample.size parameter to Run.STACAS()
• Fix bug with NA in consistency scores when calculating anchors
• In some datasets dash (-) is replaced by dot (.) in the gene names. Account for this in function for standardizing gene names
• Improve compatibility with Seurat 5 assays

STACAS v2.1.1

New in this release:

• Add compatibility with Seurat v5
• New function for standardizing gene names across datasets (StandardizeGeneSymbols())
• Implement a "reference" parameter to anchor finding function, to allow specifying a seed dataset for integration
• New behavior for large-scale integration: by default, only use the largest N datasets (given by "max.seed.datasets" parameter, by default =20) to calculate integration tree. All remaining datasets are added sequentially.
• Implement "normalization.method" parameter for FindAnchors.STACAS() (enables e.g. using SCT assay)
• Remove future.apply for parallel processing
• Several bug fixes

STACAS version 2.0

New in STACAS-2.0:

• Semi-supervised batch correction:

  • When cell type labels are available, they can be used to guide dataset integration. Briefly, STACAS removes anchors with inconsistent labels, while it retains anchors between cells with the same annotation, or between unannotated cells.
  • New parameters that control the semi-supervised behavior of FindAnchors.STACAS are cell.labels (metadata column with cell labels) and label.confidence (a number between 0 and 1 indicating the confidence on the label assignment)
  • We recommend to prioritize specificity in label assignment, i.e. only include labels for cells that can be confidently annotated, and leave the rest as unlabeled (NA or unknown)

• Downweighting of anchors based on rPCA distance:

  • Compared to STACAS v1, anchors are down-weighted instead of being filtered out based on rPCA distance. This makes the method more robust in cases where most or all anchors between certain pairs of datasets would be removed.
  • New parameters anchor.coverage and correction.scale control how rPCA distances are transformed into scores, using a sigmoid function centered on anchor.coverage with slope correction.scale
  • Anchor scores based on rPCA distance are combined with Seurat anchor confidence scores using a geometric mean, weighted by the new parameter alpha

• Native integration function

  • The function IntegrateData.STACAS is used to perform dataset integration based on the anchors calculated by FindAnchors.STACAS
  • If cell labels were specified to find integration anchors, the flag semisupervised=TRUE tells STACAS to disregard anchors with inconsistent annotation
  • The parameter k.weight controls the number of neighbors used for local anchor reweighting. Compared to Seurat, we check that k.weight is not larger than the actual number of anchors between any dataset pair. k.weight="max" is a shortcut to use all anchors for reweighting, effectively disabling local rescaling of anchor weights

• Streamlined wrapper for STACAS pipeline

  • The Run.STACAS wrapper runs variable feature selection, anchor finding, dataset integration and low dimensional embeddings in one command. This can be useful for a first explorative run on a new dataset using reasonable default parameters.

STACAS version 1.1.0

STACAS release compatible with Seurat 4.0.0

Also new in this version:

• Support for SCTransform normalization
• Throw warning if too few anchors are available after anchor filtering