mnem: Mixture NEMs - main function.

In cbg-ethz/mnem: Mixture Nested Effects Models

Mixture NEMs - main function.

Description

This function simultaneously learns a mixture of causal networks and clusters of a cell population from single cell perturbation data (e.g. log odds of fold change) with a multi-trait readout. E.g. Pooled CRISPR scRNA-Seq data (Perturb-Seq. Dixit et al., 2016, Crop-Seq. Datlinger et al., 2017).

Usage

mnem(
  D,
  inference = "em",
  search = "greedy",
  phi = NULL,
  theta = NULL,
  mw = NULL,
  method = "llr",
  marginal = FALSE,
  parallel = NULL,
  reduce = FALSE,
  runs = 1,
  starts = 3,
  type = "networks",
  complete = FALSE,
  p = NULL,
  k = NULL,
  kmax = 10,
  verbose = FALSE,
  max_iter = 100,
  parallel2 = NULL,
  converged = -Inf,
  redSpace = NULL,
  affinity = 0,
  evolution = FALSE,
  lambda = 1,
  subtopoX = NULL,
  ratio = TRUE,
  logtype = 2,
  domean = TRUE,
  modulesize = 5,
  compress = FALSE,
  increase = TRUE,
  fpfn = c(0.1, 0.1),
  Rho = NULL,
  ksel = c("kmeans", "silhouette", "cor"),
  nullcomp = FALSE,
  tree = FALSE,
  burnin = 10,
  hastings = TRUE,
  nodeswitch = TRUE,
  postgaps = 10,
  penalized = FALSE,
  accept_range = 1,
  ...
)

Arguments

D	data with cells indexing the columns and features (E-genes) indexing the rows
inference	inference method "em" for expectation maximization or "mcmc" for markov chain monte carlo sampling
search	search method for single network inference "greedy", "exhaustive" or "modules" (also possible: "small", which is greedy with only one edge change per M-step to make for a smooth convergence)
phi	a list of n lists of k networks for n starts of the EM and k components
theta	a list of n lists of k attachment vector for the E-genes for n starts of the EM and k components
mw	mixture weights; if NULL estimated or uniform
method	"llr" for log ratios or foldchanges as input (see ratio)
marginal	logical to compute the marginal likelihood (TRUE)
parallel	number of threads for parallelization of the number of em runs
reduce	logical - reduce search space for exhaustive search to unique networks
runs	number of runs for greedy search
starts	number of starts for the em or mcmc
type	initialize with responsibilities either by "random", "cluster" (each S-gene is clustered and the different S-gene clustered differently combined for several starts), "cluster2" (clustNEM is used to infer reasonable phis, which are then used as a start for one EM run), "cluster3" (global clustering as a start), or "networks" (initialize with random phis), inference='mcmc' only supports 'networks' and 'empty' for unconncected networks phi
complete	if TRUE, optimizes the expected complete log likelihood of the model, otherwise the log likelihood of the observed data
p	initial probabilities as a k (components) times l (cells) matrix
k	number of components
kmax	maximum number of components when k=NULL is inferred
verbose	verbose output
max_iter	maximum iterations (moves for inference='mcmc'. adjust parameter burnin)
parallel2	if parallel=NULL, number of threads for single component optimization
converged	absolute distance for convergence between new and old log likelihood; if set to -Inf, the EM stops if neither the phis nor thetas were changed in the most recent iteration
redSpace	space for "exhaustive" search
affinity	0 is default for soft clustering, 1 is for hard clustering
evolution	logical. If TRUE components are penelized for being different from each other.
lambda	smoothness value for the prior put on the components, if evolution set to TRUE
subtopoX	hard prior on theta as a vector with entry i equal to j, if E-gene i is attached to S-gene j
ratio	logical, if true data is log ratios, if false foldchanges
logtype	logarithm type of the data (e.g. 2 for log2 data or exp(1) for natural)
domean	average the data, when calculating a single NEM (speed improvment)
modulesize	max number of S-genes per module in module search
compress	compress networks after search (warning: penelized likelihood not interpretable)
increase	if set to FALSE, the algorithm will not stop if the likelihood decreases
fpfn	numeric vector of length two with false positive and false negative rates for discrete data
Rho	perturbation matrix with dimensions nxl with n S-genes and l samples; either as probabilities with the sum of probabilities for a sample less or equal to 1 or discrete with 1s and 0s
ksel	character vector of methods for the inference of k; can combine as the first two vlues "hc" (hierarchical clustering) or "kmeans" with "silhouette", "BIC" or "AIC"; the third value is either "cor" for correlation distance or any method accepted by the function 'dist'
nullcomp	if TRUE, adds a null component (k+1)
tree	if TRUE, restrict inference on trees (MCMC not included)
burnin	number of iterations to be discarded prior to analyzing the posterior distribution of the mcmc
hastings	if set to TRUE, the Hastings ratio is calculated
nodeswitch	if set to TRUE, node switching is allowed as a move, additional to the edge moves
postgaps	can be set to numeric. Determines after how many iterations the next Phi mixture is added to the Phi edge Frequency tracker in the mcmc
penalized	if set to TRUE, the penalized likelihood will be used for the mcmc. Per default this is FALSE, since no component learning is involved and sparcity is hence not enforced
accept_range	the random probability the acceptance probability is compared to (default: 1)
...	arguments to function nem

Value

object of class mnem

comp	list of the component with each component being a list of the causal network phi and the E-gene attachment theta
data	input data matrix
limits	list of results for all indpendent searches
ll	log likelihood of the best model
lls	log likelihood ascent of the best model search
mw	vector with mixture weights
probs	kxl matrix containing the cell log likelihoods of the model

Author(s)

Martin Pirkl

Examples

sim <- simData(Sgenes = 3, Egenes = 2, Nems = 2, mw = c(0.4,0.6))
data <- (sim$data - 0.5)/0.5
data <- data + rnorm(length(data), 0, 1)
result <- mnem(data, k = 2, starts = 1)

cbg-ethz/mnem documentation built on June 29, 2024, 1:56 p.m.