Runs the phenoDisco algorithm.

phenoDisco is a semi-supervised iterative approach to detect new protein clusters.

Usage

phenoDisco(
  object,
  fcol = "markers",
  times = 100,
  GS = 10,
  allIter = FALSE,
  p = 0.05,
  ndims = 2,
  modelNames = mclust.options("emModelNames"),
  G = 1:9,
  BPPARAM,
  tmpfile,
  seed,
  verbose = TRUE,
  dimred = c("PCA", "t-SNE"),
  ...
)

Arguments

object

An instance of class MSnSet.

fcol

A character indicating the organellar markers column name in feature meta-data. Default is markers.

times

Number of runs of tracking. Default is 100.

GS

Group size, i.e how many proteins make a group. Default is 10 (the minimum group size is 4).

allIter

logical, defining if predictions for all iterations should be saved. Default is FALSE.

p

Significance level for outlier detection. Default is 0.05.

ndims

Number of principal components to use as input for the disocvery analysis. Default is 2. Added in version 1.3.9.

modelNames

A vector of characters indicating the models to be fitted in the EM phase of clustering using Mclust. The help file for mclust::mclustModelNames describes the available models. Default model names are c("EII", "VII", "EEI", "VEI", "EVI", "VVI", "EEE", "EEV", "VEV", "VVV"), as returned by mclust.options("emModelNames"). Note that using all these possible models substantially increases the running time. Legacy models are c("EEE","EEV","VEV","VVV"), i.e. only ellipsoidal models.

G

An integer vector specifying the numbers of mixture components (clusters) for which the BIC is to be calculated. The default is G=1:9 (as in Mclust).

BPPARAM

Support for parallel processing using the BiocParallel infrastructure. When missing (default), the default registered BiocParallelParam parameters are used. Alternatively, one can pass a valid BiocParallelParam parameter instance: SnowParam, MulticoreParam, DoparParam, ... see the BiocParallel package for details. To revert to the origianl serial implementation, use NULL.

tmpfile

An optional character to save a temporary MSnSet after each iteration. Ignored if missing. This is useful for long runs to track phenotypes and possibly kill the run when convergence is observed. If the run completes, the temporary file is deleted before returning the final result.

seed

An optional numeric of length 1 specifing the random number generator seed to be used. Only relevant when executed in serialised mode with BPPARAM = NULL. See BPPARAM for details.

verbose

Logical, indicating if messages are to be printed out during execution of the algorithm.

dimred

A characater defining which of Principal Component Analysis ("PCA") or t-Distributed Stochastic Neighbour Embedding ("t-SNE") should be use to reduce dimensions prior to running phenoDisco novelty detection.

...

Additional arguments passed to the dimensionality reduction method. For both PCA and t-SNE, the data is scaled and centred by default, and these parameters (scale and centre for PCA, and pca_scale and pca_center for t-SNE can't be set). When using t-SNE however, it is important to tune the perplexity and max iterations parameters. See the Dimensionality reduction section in the pRoloc vignette for details.

Value

An instance of class MSnSet containing the phenoDisco predictions.

Details

The algorithm performs a phenotype discovery analysis as described in Breckels et al. Using this approach one can identify putative subcellular groupings in organelle proteomics experiments for more comprehensive validation in an unbiased fashion. The method is based on the work of Yin et al. and used iterated rounds of Gaussian Mixture Modelling using the Expectation Maximisation algorithm combined with a non-parametric outlier detection test to identify new phenotype clusters.

One requires 2 or more classes to be labelled in the data and at a very minimum of 6 markers per class to run the algorithm. The function will check and remove features with missing values using the filterNA method.

A parallel implementation, relying on the BiocParallel package, has been added in version 1.3.9. See the BPPARAM arguent for details.

Important: Prior to version 1.1.2 the row order in the output was different from the row order in the input. This has now been fixed and row ordering is now the same in both input and output objects.

References

Yin Z, Zhou X, Bakal C, Li F, Sun Y, Perrimon N, Wong ST. Using iterative cluster merging with improved gap statistics to perform online phenotype discovery in the context of high-throughput RNAi screens. BMC Bioinformatics. 2008 Jun 5;9:264. PubMed PMID: 18534020.

Breckels LM, Gatto L, Christoforou A, Groen AJ, Lilley KS and Trotter MWB. The Effect of Organelle Discovery upon Sub-Cellular Protein Localisation. J Proteomics. 2013 Aug 2;88:129-40. doi: 10.1016/j.jprot.2013.02.019. Epub 2013 Mar 21. PubMed PMID: 23523639.

Author

Lisa M. Breckels <lms79@cam.ac.uk>

Examples

if (FALSE) { # \dontrun{
library(pRolocdata)
data(tan2009r1)
pdres <- phenoDisco(tan2009r1, fcol = "PLSDA")
getPredictions(pdres, fcol = "pd", scol = NULL)
plot2D(pdres, fcol = "pd")

## to pre-process the data with t-SNE instead of PCA
pdres <- phenoDisco(tan2009r1, fcol = "PLSDA", dimred = "t-SNE")
} # }