HME3M Markov pathway classifier.

pathClassifier(
  paths,
  target.class,
  M,
  alpha = 1,
  lambda = 2,
  hme3miter = 100,
  plriter = 1,
  init = "random"
)

Arguments

paths: The training paths computed by pathsToBinary
target.class: he label of the targe class to be classified. This label must be present as a label within the paths\$y object
M: Number of components within the paths to be extracted.
alpha: The PLR learning rate. (between 0 and 1).
lambda: The PLR regularization parameter. (between 0 and 2)
hme3miter: Maximum number of HME3M iterations. It will stop when likelihood change is < 0.001.
plriter: Maximum number of PLR iteractions. It will stop when likelihood change is < 0.001.
init: Specify whether to initialize the HME3M responsibilities with the 3M model - random is recommended.

Value

A list with the following elements. A list with the following values

h: A dataframe with the EM responsibilities.
theta: A dataframe with the Markov parameters for each component.
beta: A dataframe with the PLR coefficients for each component.
proportions: The probability of each HME3M component.
posterior.probs: The HME3M posterior probability.
likelihood: The likelihood convergence history.
plrplr: The posterior predictions from each components PLR model.
path.probabilities: The 3M probabilities for each path belonging to each component.
params: The parameters used to build the model.
y: The binary response variable used by HME3M. A 1 indicates the location of the target.class labels in paths\$y
perf: The training set ROC curve AUC.
label: The HME3M predicted label for each path.
component: The HME3M component assignment for each path.

Details

Take care with selection of lambda and alpha - make sure you check that the likelihood is always increasing.

References

Hancock, Timothy, and Mamitsuka, Hiroshi: A Markov Classification Model for Metabolic Pathways, Workshop on Algorithms in Bioinformatics (WABI) , 2009

Hancock, Timothy, and Mamitsuka, Hiroshi: A Markov Classification Model for Metabolic Pathways, Algorithms for Molecular Biology 2010

Author

Timothy Hancock and Ichigaku Takigawa

Examples

  ## Prepare a weighted reaction network.
  ## Conver a metabolic network to a reaction network.
 data(ex_sbml) # bipartite metabolic network of Carbohydrate metabolism.
 rgraph <- makeReactionNetwork(ex_sbml, simplify=TRUE)
#> This graph was created by an old(er) igraph version.
#> ℹ Call `igraph::upgrade_graph()` on it to use with the current igraph version.
#> For now we convert it on the fly...

  ## Assign edge weights based on Affymetrix attributes and microarray dataset.
 # Calculate Pearson's correlation.
  data(ex_microarray)  # Part of ALL dataset.
  rgraph <- assignEdgeWeights(microarray = ex_microarray, graph = rgraph,
    weight.method = "cor", use.attr="miriam.uniprot",
    y=factor(colnames(ex_microarray)), bootstrap = FALSE)
#> 100 genes were present in the microarray, but not represented in the network.
#> 55 genes were couldn't be found in microarray.
#> Assigning edge weights for label ALL1/AF4 
#> Assigning edge weights for label BCR/ABL 
#> Assigning edge weights for label E2A/PBX1 
#> Assigning edge weights for label NEG 

  ## Get ranked paths using probabilistic shortest paths.
 ranked.p <- pathRanker(rgraph, method="prob.shortest.path",
          K=20, minPathSize=6)
#> Extracting the 20 most probable paths for ALL1/AF4
#> Extracting the 20 most probable paths for BCR/ABL
#> Extracting the 20 most probable paths for E2A/PBX1
#> Extracting the 20 most probable paths for NEG

  ## Convert paths to binary matrix.
  ybinpaths <- pathsToBinary(ranked.p)
  p.class <- pathClassifier(ybinpaths, target.class = "BCR/ABL", M = 3)

  ## Contingency table of classification performance
  table(ybinpaths$y,p.class$label)
#>           
#>             0  1
#>   ALL1/AF4 20  0
#>   BCR/ABL  16  4
#>   E2A/PBX1 19  1
#>   NEG      20  0

  ## Plotting the classifier results.
  plotClassifierROC(p.class)

  plotClusters(ybinpaths, p.class)