corto() gains a dpi argument to select
the Data Processing Inequality rule. The default, "target",
is the rule used so far: for each target only the edge from the centroid
with the strongest absolute correlation survives, and results are
identical to previous versions. "triangle" applies the
ARACNe rule, removing an edge only if it is the weakest in a
centroid-centroid-target triangle; it produces larger networks, closer
to those of ARACNe, at a considerably higher computing time. Proposed
and implemented by Hugo Tovar (#16).
New exported dpi_edges(), which applies either DPI
rule to any list of centroid-target edges.
corto()
help page now describe the DPI step as it is actually performed. They
previously described the ARACNe triangle rule, while
corto() applied the per-target rule (#16).corto() uses Pearson correlation on the input matrix
as provided. Rank the matrix beforehand to work on Spearman
correlations.New filter_regulon() and getregulon()
helpers for post-processing a regulon: subset it by centroid, likelihood
and absolute correlation, and flatten it into a data frame of edges with
optional tab-separated export. Proposed by Hugo Tovar (#15).
mra() now performs Signature Master Regulator
Analysis when expmat1 is provided as a named vector, as the
documentation had always described. The null model permutes the
signature values across feature names, and nperm defaults
to 1000 in this mode. Thanks to Hugo Tovar for the detailed diagnosis
and a reference implementation (#13).
mra() (a single expmat1,
no expmat2) now returns the same list as the other modes,
list(nes, pvalue, sig, regulon), instead of a bare matrix.
Use mra(expmat, regulon=regulon)$nes to get the previous
output (#1).corto() failed with
invalid 'row.names' length when a single centroid was
provided. The internal correlation step dropped to a vector and lost its
row names. Thanks to Hualin Wang for the fix (#6, #14).mra() returned an all-NA matrix
whenever any regulon target had zero variance in the input matrix. The
permuted null signatures were not NA-guarded the way the real signature
was.gsea() errored on R >= 4.2 when method
was left at its default, since the default is a length-2 vector. It now
uses match.arg() and defaults to "permutation"
as before.plot_gsea(omit_middle=TRUE) errored on an undefined
legend_position.mraplot() errored on regulons with fewer than 12
targets, which is reachable with the default
minsize=10.ssgsea(scale=TRUE) silently dropped the sample names
from the returned NES matrix.mra() failed on regulons left with a single centroid
after minsize filtering.mraplot() now stops with an informative message when
given sample-by-sample results, which it cannot plot.cnvmat
and inmat had a single sample or a single target in common,
the same dimension-drop problem fixed elsewhere. Guarded with
drop=FALSE.corto() now stops with an informative message when no
edge passes the correlation threshold, instead of failing on an invalid
row name.corto() is roughly 1.2x faster single-threaded and 1.4x
faster on 4 threads. The input matrix is transposed once instead of once
per bootstrap and is no longer shipped twice to the workers, DPI
selection is done by a radix sort rather than a grouped data frame, and
the regulon is assembled in one pass instead of rescanning the edge
table for every centroid. Results are bit-identical to previous
versions.dplyr and gplots are no longer required,
which removes about 20 recursive dependencies and makes installation
considerably faster. knitr and rmarkdown moved
from Imports to Suggests, grDevices and
graphics added.scatter(bgcol=),
val2col(nbreaks=) and plot_gsea2().