cluster_profiles_vb: Cluster methylation profiles using VB
In andreaskapou/BPRMeth: Model higher-order methylation profiles

Description Usage Arguments Value Details Author(s) See Also Examples

General purpose functions for clustering latent profiles for different observation models using Variational Bayes (VB) EM-like algorithm.

cluster_profiles_vb(
  X,
  K = 3,
  model = NULL,
  basis = NULL,
  H = NULL,
  delta_0 = NULL,
  w = NULL,
  gaussian_l = 50,
  alpha_0 = 0.5,
  beta_0 = 0.1,
  vb_max_iter = 100,
  epsilon_conv = 1e-04,
  is_verbose = FALSE,
  ...
)

`X`	The input data, which has to be a `list` of elements of length N, where each element is an `L X C` matrix, where L are the total number of observations. The first column contains the input observations x (i.e. CpG locations). If "binomial" model then C=3, and 2nd and 3rd columns contain total number of trials and number of successes respectively. If "bernoulli" or "gaussian" model, then C=2 containing the output y (e.g. methylation level).
`K`	Integer denoting the total number of clusters K.
`model`	Observation model name as character string. It can be either 'bernoulli', 'binomial', 'beta' or 'gaussian'.
`basis`	A 'basis' object. E.g. see `create_basis`. If NULL, will an RBF object will be created.
`H`	Optional, design matrix of the input data X. If NULL, H will be computed inside the function.
`delta_0`	Parameter vector of the Dirichlet prior on the mixing proportions pi.
`w`	Optional, an (M+1)xK matrix of the initial parameters, where each column consists of the basis function coefficients for each corresponding cluster k. If NULL, will be assigned with default values.
`gaussian_l`	Noise precision parameter, only used when having "gaussian" observation model.
`alpha_0`	Hyperparameter: shape parameter for Gamma distribution. A Gamma distribution is used as prior for the precision parameter tau.
`beta_0`	Hyperparameter: rate parameter for Gamma distribution. A Gamma distribution is used as prior for the precision parameter tau.
`vb_max_iter`	Integer denoting the maximum number of VB iterations.
`epsilon_conv`	Numeric denoting the convergence threshold for VB.
`is_verbose`	Logical, print results during VB iterations.
`...`	Additional parameters.

An object of class cluster_profiles_vb_"obs_model" with the following elements:

W: An (M+1) X K matrix with the optimized parameter values for each cluster, M are the number of basis functions. Each column of the matrix corresponds a different cluster k.
W_Sigma: A list with the covariance matrices of the posterior parmateter W for each cluster k.
r_nk: An (N X K) responsibility matrix of each observations being explained by a specific cluster.
delta: Optimized Dirichlet paramter for the mixing proportions.
alpha: Optimized shape parameter of Gamma distribution.
beta: Optimized rate paramter of the Gamma distribution
basis: The basis object.
lb: The lower bound vector.
labels: Cluster assignment labels.
pi_k: Expected value of mixing proportions.

The modelling and mathematical details for clustering profiles using mean-field variational inference are explained here: http://rpubs.com/cakapourani/ . More specifically:

For Binomial/Bernoulli observation model check: http://rpubs.com/cakapourani/vb-mixture-bpr
For Gaussian observation model check: http://rpubs.com/cakapourani/vb-mixture-lr

C.A.Kapourani C.A.Kapourani@ed.ac.uk

create_basis, cluster_profiles_mle infer_profiles_vb, infer_profiles_mle, infer_profiles_gibbs, create_region_object

# Example of optimizing parameters for synthetic data using 3 RBFs
basis <- create_rbf_object(M=3)
out <- cluster_profiles_vb(X = binomial_data, model = "binomial",
  basis=basis, vb_max_iter = 10)

#-------------------------------------

basis <- create_rbf_object(M=3)
out <- cluster_profiles_vb(X = gaussian_data, model = "gaussian",
  basis=basis, vb_max_iter = 10)