stefpeschel/NetCoMi: NetCoMi 1.1.0

This is a major release, which comes with new features, improvements, bugfixes, and reformatted R code.

The commit history has been considerably revised to follow general conventions, which is why the commit hashes have changed and are not in line with the old "master" branch anymore.

New features

• renameTaxa(): New function for renaming taxa in a taxonomic table. It comes with functionality for making unknown and unclassified taxa unique and substituting them by the next higher known taxonomic level. E.g., an unknown genus "g__", where family is the next higher known level, can automatically be renamed to "1_Streptococcaceae(F)". User-defined patterns determine the format of known and substituted names. Unknown names (e.g., NAs) and unclassified taxa can be handled separately. Duplicated names within one or more chosen ranks can also be made unique by numbering them consecutively.

• editLabels(): New function for editing node labels, i.e., shortening to a certain length and removing unwanted characters. It is used by NetCoMi's plot functions plot.microNetProps() and plot.diffnet().

• In netCompare(): The adjusted Rand index is also computed for the largest connected component (LCC). The summary method has been adapted.

• Argument "testRand" added to netCompare(). Performing a permutation test for the adjusted Rand index can now be disabled to save run time.

• Graphlet-based network measures implemented. NetCoMi contains two new exported functions calcGCM() and calcGCD() to compute the Graphlet Correlation Matrix (GCM) of a network and the Graphlet Correlation Distance (GCD) between two networks. Orbits for graphlets with up to four nodes are considered. Furthermore, the GCM is computed with netAnalyze() and the GCD with netCompare() (for the whole network and the largest connected component, respectively). Also the orbit counts are returned. The GCD is added to the summary for class microNetComp objects returned by netCompare().

• Significance test for the GCD: If permutation tests are conducted with netCompare(), the GCD is tested for being significantly different from zero.

• New function testGCM() to test graphlet-based measures for significance. For a single GCM, the correlations are tested for being significantly different from zero. If two GCMs are given, it is tested if the correlations are significantly different between the two groups, that is, the absolute differences between correlations ( $|gc1{ij}-gc2{ij}|$ ) are tested for being different from zero.

• New function plotHeat() for plotting a mixed heatmap where, for instance, values are shown in the upper triangle and corresponding p-values or significance codes in the lower triangle. The function is used for plotting heatmaps of the GCMs, but could also be used for association matrices.

• netAnalyze() now by default returns a heatmap of the GCM(s) with graphlet correlations in the upper triangle and significance codes in the lower triangle.

• Argument "doPlot" added to plot.microNetProps() to suppress the plot if only the return value is of interest.

• New "show" arguments are added to the summary methods for class microNetProps and microNetComp objects. They specify which network properties should be printed in the summary. See the help pages of summary.microNetProps and summary.microNetComp() for details.

• New zero replacement method "pseudoZO" available in netConstruct(). Instead of adding the desired pseudo count to the whole count matrix, it is added to zero counts only if pseudoZO is chosen. The behavior of "pseudo" (a further available method where a pseudo count is added to all counts) has not changed. Adding a pseudo count only to zeros preserves the ratios between non-zero counts, which is desirable.

• createAssoPerm() now accepts objects of class microNet as input (in addition to objects of class microNetProps).

• SPRING's fast version of latent correlation computation (implemented in mixedCCA) is available again. It can be used by setting the netConstruct() parameter measurePar$Rmethod to "approx", which is now the default again.

• The function multAdjust() now has an argument pTrueNull to pre-define the proportion of true null hypotheses for the adaptive BH method.

• netConstruct() has a new argument assoBoot, which enables the computation of bootstrap association matrices outside netConstruct() if bootstrapping is used for sparsification. An example has been added to the help page ?netConstruct. This feature might be useful for very large association matrices (for which the working memory might reach its limit).

Bug fixes

• In netConstruct():

  • Using "bootstrap" as sparsification method in combination with one of the association methods "bicor", "cclasso", "ccrepe", or "gcoda" led to the error: argument "verbose" is missing, with no default, which has been fixed.
  • The "signedPos" transformation did not work properly. Dissimilarities corresponding to negative correlations were set to zero instead of infinity.

• In editLabels(): The function (and thus also plot.microNetProps) threw an error if taxa have been renamed with renameTaxa and the data contain more than 9 taxa with equal names, so that double-digit numbers were added to avoid duplicates.

• Issues in network analysis and plotting if association matrices are used for network construction, but row and/or column names are missing. (issue #65)

• diffnet() threw an error if association matrices are used for network construction instead of count matrices. (issue #66)

• In plot.microNetProps():

  • The function now directly returns an error if x has not the expected class.
  • The cut parameter could not be changed.

• In cclasso(): In rare cases, the function produced complex numbers, which led to an error.

Further changes

• In permutation tests: The permuted group labels must now be different from the original group vector. In other words, the original group vector is strictly avoided in the matrix with permuted group labels. So far, only duplicates were avoided. Only in exact permutation tests (if nPerm equals the possible number of permutations), the original group vector is still included in the permutation matrix. The calculation of p-values has been adapted to the new behavior: p=B/N for exact p-values and p=(B+1)/(N+1) for approximated p-values, where B is the number of permutation test statistics being larger than or equal to the observed one, and N is the number of permutations. So far, p=(B+1)/(N+1) has been used in all cases.

• In plot.microNetProps():

  • The default of shortenLabels is now "none", i.e. the labels are not shortened by default, to avoid confusion about the node labels.
  • The edge filter (specified via edgeFilter and edgeInvisFilter) now refers to the estimated association/dissimilarities instead of edge weights. E.g., setting the threshold to 0.3 for an association network hides edges with a corresponding absolute association below 0.3 even though the edge weight might be different (depending on the transformation used for network construction). (issue #26)
  • If two networks are constructed and the cut parameter is not user-defined, the mean of the two determined cut parameters is now used for both networks so that edge thicknesses are comparable.

• More expressive messages and errors in diffnet and plot.diffnet if no differential associations are detected.

• New function .suppress_warnings() to suppress certain warnings returned by external functions.

• In netConstruct if "multRepl" is used for zero handling: The warning about the proportion of zeros is suppressed by setting the multRepl() parameter "z.warning" to 1.

• The functions makeCluster and stopCluster from parallel package are now used for parallel computation because those from snow package sometimes led to problems on Unix machines.

Style

• The whole R code has been reformatted to follow general conventions.

• The element "clustering_lcc" as part of the netAnalyze output has changed to "clusteringLCC" to be in line with the remaining output.

• Input argument checking of exported function has been revised. New functions .checkArgsXxx() are added to perform argument checking outside the main functions.

• Non-exported functions have been renamed to follow general naming conventions, i.e. that of Bioconductor:

  • Use camelCase for all functions.
  • Non-exported functions have prefix "."
  • See the NEWS file for an overview of the changed function names.