16 May 2014 mcxload 14-137
mcxload — load matrices and tab files from label format
mcxload -abc <fname> (label file) -o <fname> (output file)
[-abc <fname> (label file)] [-123 <fname> (identifier file)] [-o <fname> (output file)] [--stream-mirror (symmetrify, same domain)] [--stream-split (assume different domains)] [-re <mode> (edge deduplication mode)] [-ri <mode> (image symmetrification mode)] [-sif <fname> (SIF label file)] [-etc <fname> ('etc' label file)] [-etc-ai <fname> (leaderless 'etc' label file)] [--expect-values (expect label:weight format)] [-235 <fname> (leader '235' label file)] [-235-ai <fname> (leaderless '235' label file)] [-packed <fname> (file/stream in binary format)] [-pack-cnum <num> (set column range)] [-pack-rnum <num> (set row range)] [-123-max <int> (set domain range)] [-123-maxc <int> (set column range)] [-123-maxr <int> (set row range)] [-write-tab <fname> (save domain tab)] [-write-tabc <fname> (save column tab)] [-write-tabr <fname> (save row tab)] [-strict-tab <fname> (tab universe)] [-strict-tabc <fname> (tabc universe)] [-strict-tabr <fname> (tabr universe)] [-restrict-tab <fname> (tab world)] [-restrict-tabc <fname> (tabc world)] [-restrict-tabr <fname> (tabr world)] [-extend-tab <fname> (tab launch)] [-extend-tabc <fname> (tabc launch)] [-extend-tabr <fname> (tabr launch)] [--stream-log (log transform stream values)] [--stream-neg-log (negative log transform stream values)] [--stream-neg-log10 (negative log-10 transform stream values)] [-stream-tf (transform stream values)] [-tf <tf-spec> (transform (not so) final matrix)] [--transpose (transpose)] [--write-binary (output binary format)] [--debug (debug)] [-h (print synopsis, exit)] [--apropos (print synopsis, exit)] [--version (print version, exit)]
When the output should be an undirected graph it is safest to always use the --stream-mirror option. Edges are stored bidirectionally as two arcs, and this option instructs mcxload to ensure that both arcs are present. In the above examples three different types of format are read. In all formats, the basic unit of specification is that of an arc specified by a source node, a destination node, and optionally a weight. All formats are line based, with -abc specifying a single arc and -etc and -sif specifying multiple arcs corresponding to a shared source node. For -abc the format is
The last field, specifying the arc weight, is optional. If not present the arc weight will be set to the default weight of 1.0.
For -sif the format isThere can be an arbitrary number of destination labels. The relation type field in the second column is required but will be ignored. As an extension it is possible to specify weights, requiring the use of the --expect-values option. Weights are specified by tagging them onto the destination label separated by a colon:
Finally, the format for the -etc option is the same, except that the relation type column is dropped.
mcxload reads label input from a file. The format of the file should be line-based, each line containing two white-space separated strings (labels) and optionally a number separated from the second label by whitespace. In the absence of a value, mcxload will use the default value 1.0. If a tab is present on an input line, mcxload will assume that the tab character is the separator for that line. Lines for which the first non-whitespace character is an octothorpe ('#') are skipped.
mcxload will transform the labels into mcl numerical identifiers and the pairs of labels into graph edges or equivalently matrix entries. The weight of an edge is the value associated with the associated labels. mcxload constructs dictionaries (sometimes just one) that map labels onto mcl identifiers as it goes along. It can optionally write these to file. In MCL (family) parlance, such a dictionary written to file is called a tab file.
It is possible to specify numerical identifiers directly with the -123 option. In this case mcxload assumes a canonical domain (cf mcxio) and will create the minimal canonical domain that supports the data. Also bear in mind the caveat further below.
It is possible to effectively predeclare labels and thus enforce an a-priori known mapping of labels onto numerical identifiers. Labels receive an identifier in the order in which they occur in the input. Predeclaring labels can be achieved by having them appear in the desired order and setting the edge weight to zero.
A major mcxload modality is whether the input refers to a single domain or to two separate domains. An example of the first is where labels are names of people and the value is the extent to which they like one another. This encodes a likability graph where all the nodes represent people. The reasonable thing to do in this case is to create a single dictionary with all names wherever they occur. All tab options (as opposed to tabc and tabr) pertain to this scenario and likewise for the options --graph and --stream-mirror.
An example of the second mode is where the first label is again the name of a person, the second label is the name of an animal species, and the value is the extent to which that person appreciates the species. In this case, the reasonable thing to do is to create two dictionaries, one for persons and one for species. All tabc and tabr options pertain to this scenario. The tabc options always refer to the first label and the tabr options always refer to the second label. The letters c and r refer to column and row respectively. The latter are the names of the matrix domains corresponding to the input domains. Refer to mcxio.
A further mcxload modality is whether it constructs dictionaries on the fly, or whether it proceeds from a tab file already available. By default mcxload will construct dictionaries on the fly. You need to save them with the appropriate -write option(s). All the strict options read a tab file and require any labels in the -abc label input to be present in the corresponding tab file. mcxload will then fail in the face of absent labels. All the restrict options simply ignore labels that are not found in the corresponding tab file. The extend options extend the existing tab file with labels that are not found. It presumably only makes sense to do so if the corresponding -write options are used as well.
The input stream is deduplicated on a per-node neighbourhood basis using the -re option.
mcxload has a few options to transform or select based on the values in the input stream and the values in the constructed matrix. These are --stream-log, --stream-neg-log, --stream-neg-log10, -stream-tf and -tf. Refer to mcxio for a description of the syntax accepted by the latter two options — it is a syntax accepted by a few more mcl siblings. Finally it is possible to transpose the final result using the --transpose option. Keep in mind that mcxload does not accordingly change its idea of row and column domains.
The final matrix can be symmetrified using the -ri option.
The -etc, -235 and -sif options assume a format where all entries for a given column (or equivalently all neighbours for a given node) are joined onto a single line. This can be useful e.g. to read in externally generated clusterings. The -etc and -sif options expect label input, whereas the -235 options expects numbers in the input that are mapped directly onto mcl numerical identifiers. The SIF format expected by -sif requires a relationship type in the second field on each line; this is ignored. As an extension to the SIF format weights may optionally follow the labels, separated from them with a colon character.
CAVEAT
Please note that by feeding the line '1000000000 1' to mcxload with either
of the -235 or -123 options it will try to allocate a
matrix with one billion columns. This is most likely not what is wanted.
Assuming that the input contains fewer than one billion unique labels, one
should use the label options as described above and below.
STAGES
Conceptually, input matrix creation consists of the following stages
Read the input stream, apply -stream-tf transformation specification, and optionally push reverse elements (--stream-mirror).
Deduplicate edges in the context of all edges/arcs originating from a given node according to the -re option.
Apply transpose symmetrification according to the -ri option, if used.
Apply -tf transformation specification.
The file to read label data from. Labels are separated by white-space. The labels may optionally be followed by a value (again separated by white-space), which is taken as the edge weight between the nodes corresponding with the labels. If a tab is present on an input line it is presumed to be the separator for that line, including the value if present. Lines for which the first non-blank character is the octothorpe ('#') are skipped.
The file to read numerical data from. The format is the same as for label data, but the identifiers are directly mapped onto mcl identifiers as described earlier.
The output file where the constructed matrix is written.
Whenever label1 label2 value is encountered in the input, mcxload inserts label2 label1 value in the input stream as well. This option implies that both labels belong to the same domain.
This tells mcxload that the two labels belong to different domains. The program will create two tab files, one for columns and one for rows. This can be used for example to create a logical mapping of gene identifiers to species identifiers.
This specifies how mcxload should collapse repeated entries, that is edges for which a value is specified multiple times. This is done relative to a single node at a time, taking into account all neighbours assembled from the input stream. Note that --stream-mirror will result in duplicated entries if the input contains edge specifications in both ways. Also note that first and last might not result in symmetric input if only --stream-mirror is used.
Write the domain to file. It applies to both label types.
Write the column domain to file. It applies to the first label found on each input line.
Write the column domain to file. It applies to the second label found on each input line.
Read a dictionary from file and require each label to be present in the dictionary. mcxload will exit on absentees.
Read a dictionary from file and require the first label on each line to be present in the dictionary. mcxload will exit on absentees.
Read a dictionary from file and require the second label on each line to be present in the dictionary. mcxload will exit on absentees.
Read a dictionary from file and only accept input lines (edges) for which both labels are present in the dictionary. mcxload will ignore absentees.
Read a dictionary from file and ignore input lines for which the first label is absent from the dictionary.
Read a dictionary from file and ignore input lines for which the second label is absent from the dictionary.
Read a dictionary from file and extend it with any label from the input not yet present in the dictionary.
Read a dictionary from file and extend it with all first labels from the input not yet present in the dictionary.
Read a dictionary from file and extend it with all second labels from the input not yet present in the dictionary.
These options limit the domain ranges accepted by the -123 option. Numbers starting from <int> will be ignored, and the domain(s) will range from zero up to one less than <int>. The first, -123-max governs both domains, and -123-maxc and -123-maxr respectively govern the column and row domain.
Replace each entry by its natural logarithm.
Replace each entry by the negative of its natural logarithm and log-10 representation, respectively. This is for example useful to convert scores that denote probabilities or p-values such as BLAST scores.
Transform the stream values as they are read in according to the syntax described in mcxio.
Transform the matrix values after deduplication and symmetrification according to the syntax described in mcxio.
After the initial matrix has been assembled, it can be symmetrified by either of these options. They indicate the operation used to combine the entries of the transposed matrix and the original matrix. mul is special in that it treats missing entries (which are normally considered zero in mcl matrix operations) as one.
Write the transposed matrix to file. This is obviously not useful when a symmetric matrix has been generated.
The input is read in lines; each line is split on whitespace into labels. For -etc the first label is interpreted as the source node. All other labels are interpreted as destination nodes. Weights may optionally follow the labels, separated from them with a colon character. It is in this case necessary to use the --expect-values option. The SIF (Simple Interaction File) format expected by -sif is similar except that it contains an additional field. In this format the second column denotes the relationship type. It is ignored by mcxload. For -etc-ai (auto-increment) all labels are interpreted as destination nodes and mcxload automatically creates a source node for each line it reads. This option can be useful to read in files encoding a clustering, where each line represents a cluster of white-space separated labels.
The -235 options are similar except that the input is not interpreted as labels but must consist of numbers that explicitly specify the matrix to be built.
The -packed option allows to read machine-readable data directly. The data has to correspond to the data types for indexes and values with with MCL was compiled. The use of -pack-cnum and -pack-rnum is required to set the limits of the ranges of indices that will be read.
The /scripts directory of the MCL software contains scripts packed-example.sh and packed-example2.sh. The first shows the simple binary format that is accepted by -packed. It also documents the required include files and library and the method by which they can be referenced and linked to. The second expands on the first example by multiplexing binary output onto multiple output streams. Each output stream is read and loaded by an independent mcxload instance. The final result is obtained by summing the individual matrices. This can be used to speed up the loading of large data by parallelisation.
The output matrix is written in native binary format — refer to mcxio.
Among other things, this turns on warnings when restrict tab files are used and labels are found to be missing.
Stijn van Dongen.
mcxio, mcxdump, mcl, mclfaq, and mclfamily for an overview of all the documentation and the utilities in the mcl family.