Skip to content

Latest commit

 

History

History
506 lines (388 loc) · 21 KB

NEWS.md

File metadata and controls

506 lines (388 loc) · 21 KB

NetCoMi 1.1.0

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 “.”
    • The following functions have been renamed:
Old names New names
boottest .boottest
calc_association .calcAssociation
calc_diff_props .calcDiffProps
calc_jaccard .calcJaccard
calc_props .calcProps
diff_connect_pairs .diffConnectPairs
diff_connect_variables .diffConnectVariables
diff_connect_network .diffConnectNetwork
filter_edges .filterEdges
filter_nodes .filterNodes
filter_samples .filterSamples
filter_taxa .filterTaxa
first_unequal_element .firstUnequalElement
get_clust_cols .getClustCols
get_node_size .getNodeSize
get_perm_group_mat .getPermGroupMat
get_vec_names .getVecNames
norm_counts .normCounts
permtest_diff_asso .permTestDiffAsso
scale_diss .scaleDiss
sparsify .sparsify
trans_to_diss .transToDiss
trans_to_sim .transToSim
trans_to_adja .transToAdja
zero_treat .zeroTreat

NetCoMi 1.0.3

This is a minor release with some bug fixes and changes in the documentation.

Bug fixes

  • netConstruct() threw an error if the data had no row and/or column names, which is fixed.

  • An edge list is added to the output of netConstruct() (issue #41). See the help page for details.

  • SPRING’s fast version of latent correlation computation (implemented in mixedCCA) is currently not available due to deprecation of the R package chebpol. The issue is fixed by setting the netConstruct() parameter measurePar$Rmethod internally to “original” if SPRING is used for association estimation.

  • In plot.microNetProps(): The xpd parameter is changed to NA so that plotting outside the plot region is possible (useful for legends or additional text).

  • Labels in the network plot can now be suppressed by setting labels = FALSE (issue #43)

  • The netCompare() function threw an error if one of the permutation networks was empty, i.e. had no edges with weight different from zero (issue #38), which is now fixed.

  • Fix issues #29 and #40, where permutation tests did not terminate for small sample sizes. Now, if the possible number of permutations (resulting from the sample size) is smaller than that defined by the user, the function stops and returns an error.

  • Fix a bug in diffnet() (issue #51), where colors in differential networks could not be changed.

  • diffnet() threw an error if the netConstruct() argument jointPrepro was set to TRUE.

NetCoMi 1.0.2

This release includes a range of new features and fixes known bugs and issues.

New features

Improved installation process

Packages that are optionally required in certain settings are not installed together with NetCoMi anymore. Instead, there is a new function installNetCoMiPacks() for installing the remaining packages. If not installed via installNetCoMiPacks(), the required package is installed by the respective NetCoMi function when needed.

installNetCoMiPacks()

New function for installing the R packages used in NetCoMi not listed as dependencies or imports in NetCoMi’s description file.

netConstruct()

  • New argument matchDesign: Implements matched-group (i.e. matched-pair) designs, which are used for permutation tests in netCompare() and diffnet(). c(1,2), for instance, means that one sample in the first group is matched to two samples in the second group. If the argument is not NULL, the matched-group design is kept when generating permuted data.

  • New argument jointPrepro: Specifies whether two data sets (of group one and two) should be preprocessed together. Preprocessing includes sample and taxa filtering, zero treatment, and normalization. Defaults to TRUE if data and group are given, and to FALSE if data and data2 are given, which is similar to the behavior of NetCoMi 1.0.1. For dissimilarity networks, no joint preprocessing is possible.

  • mclr(){SPRING} is now available as normalization method.

  • clr{SpiecEasi} is used for centered log-ratio transformation instead of cenLR(){robCompositions}.

  • "symBetaMode" is accepted as list element of measurePar, which is passed to symBeta(){SpiecEasi}. Only needed for SpiecEasi or SPRING associations.

  • The pseudocount (if zeroMethod = "pseudo") may be freely specified. In v1.0.1, only unit pseudocounts were possible.

netAnalyze()

  • Global network properties are now computed for the whole network as well as for the largest connected component (LCC). The summary of network properties now contains for the whole network only statistics that are not based on shortest paths (or, more generally, also meaningful for disconnected networks). For the LCC, all global properties available in NetCoMi are shown.

  • New global network properties (see the docu of netAnalyze() for definitions):

    • Number of components (only whole network)
    • Relative LCC size (only LCC)
    • Positive edge percentage
    • Natural connectivity
    • Average dissimilarity (only meaningful for the LCC)
    • Average path length (only meaningful for the LCC)
  • New argument centrLCC: Specifies whether to compute centralities only for the LCC. If TRUE, centrality values of disconnected components are zero.

  • New argument avDissIgnoreInf: Indicates whether infinite values should be ignored in the average dissimilarity. If FALSE, infinities are set to 1.

  • New argument sPathAlgo: Algorithm used for computing shortest paths

  • New argument sPathNorm: Indicates whether shortest paths should be normalized by average dissimilarity to improve interpretability.

  • New argument normNatConnect: Indicates whether to normalize natural connectivity values.

  • New argument weightClustCoef: Specifies the algorithm used for computing the global clustering coefficient. If FALSE, transitivity(){igraph} with type = "global" is used (similar to NetCoMi 1.0.1). If TRUE, the local clustering coefficient is computed using transitivity(){igraph} with type = "barrat". The global clustering coefficient is then the arithmetic mean of local values.

  • Argument connect has been changed to connectivity.

  • Documentation extended by definitions of network properties.

summary.microNetProps()

  • New argument clusterLCC: Indicates whether clusters should be shown for the whole network or only for the LCC.

  • The print method for summary.microNetProps was completely revised.

plot.microNetProps()

  • All normalization methods available for network construction can now be used for scaling node sizes (argument nodeSize).

  • New argument normPar: Optional parameters used for normalization.

  • Usage of colorVec changed: Node colors can now be set separately in both groups (colorVec can be a single vector or a list with two vectors). Usage depends on nodeColor (see docu of colorVec).

  • New argument sameFeatCol: If nodeColor = "feature" and colorVec is not given, sameFeatCol indicates whether same features should have same colors in both groups.

  • Argument colorNegAsso has been renamed to negDiffCol. Using the old name leads to a warning.

  • New functionality for using the same layout in both groups (if two networks are plotted). In addition to computing the layout for one group and adopting it for the other group, a union of both layout can be computed and used in both groups so that nodes are placed as optimal as possible equally for both networks. This option is applied via sameLayout = TRUE and layoutGroup = "union". Many thanks to Christian L. Müller and Alice Sommer for providing the idea and R code for this new feature!

netCompare()

  • New arguments for storing association and count matrices of the permuted data into an external file:

    • fileLoadAssoPerm
    • fileLoadCountsPerm
    • storeAssoPerm
    • fileStoreAssoPerm
    • storeCountsPerm
    • fileStoreCountsPerm
  • New argument returnPermProps: If TRUE, global network properties and the respective absolute group differences of the permuted data are returned.

  • New argument returnPermCentr: If TRUE, the computed centrality values and the respective absolute group differences of the permuted data are returned as list with a matrix for each centrality measure.

  • The arguments assoPerm and dissPerm are still existent for compatibility with NetCoMi 1.0.1 but the former elements assoPerm and dissPerm are not returned anymore (matrices are stored in an external file instead).

createAssoPerm()

New function for creating association/dissimilarity matrices for permuted count data. The stored count or association/dissimilarity matrices can then be passed to netCompare() or diffnet() to decrease runtime. The function also allows to generate a matrix permuted group labels without computing associations. Using this matrix, createAssoPerm() furthermore allows to estimate the permutation associations/dissimilarities in blocks (by passing only a subset of the permuted group matrix to createAssoPerm()).

summary.microNetComp()

Summary method has been adapted to the new network properties (analogous to the summary of microNetProps objects, which are returned from netAnalyze())

diffnet()

  • New arguments for storing association and count matrices of the permuted data into an external file:

    • fileLoadAssoPerm
    • fileLoadCountsPerm
    • storeAssoPerm
    • fileStoreAssoPerm
    • storeCountsPerm
    • fileStoreCountsPerm
  • The argument assoPerm is still existent for compatibility with NetCoMi 1.0.1 but the former element assoPerm is not returned anymore (matrices are stored in an external file instead).

  • Changed output: For permutation tests and Fisher’s z-test, a vector and matrix with p-values and the corresponding matrix with group differences are returned for both with and without multiple testing adjustment.

  • Documentation has been revised.

plot.diffnet()

  • New argument adjusted: Indicates whether the adjacency matrix (matrix with group differences) based on adjusted or unadjusted p-values should be plotted.

  • New argument legendPos for positioning the legend.

  • New argument legendArgs for specifying further arguments passed to legend.

colToTransp()

  • The function is now exported and its name has changed from col_to_transp() to colToTransp(). The function expects a color vector as input and adds transparency to each color.

Bug fixes

The major issues fixed in this release are:

  • The following error is solved: Error in update.list(...): argument "new" is missing. The error was caused by a conflict between SpiecEasi and metagenomeSeq, in particular by gplot as a dependency of metagenomeSeq. A former version of gplot was dependend on gdata, which caused the conflict. So, please update gplot and remove the package gdata to fix the error.

  • sparcc() from SpiecEasi package is now used for estimating SparCC associations. For some users, NetCoMi’s Rccp implementation of SparCC caused errors when installing NetCoMi. If these are fixed, the Rcpp implementation will be included again, so that users can decide between the two SparCC versions.

  • VST transformations are now computed correctly.

  • Error when plotting two networks, where one network is empty, has been fixed.