Segmentation and classification of copy number profiles

Share:

Description

Takes in a copy number profile and segments it into predicted regions of equal copy number, and assigns a biologically motivated copy number state to each region using a Hidden Markov Model (HMM). This is an extension to the HMM described in Shah et al., 2006.

Usage

1
2
HMMsegment(correctOut, param = NULL, autosomes = NULL,
    maxiter = 50, getparam = FALSE, verbose = TRUE)

Arguments

correctOut

Output value from correctReadcount

param

If none is provided, will generate a reasonable set of parameters based on the input data, which can optionally be returned for inspection and manual adjustment by setting 'getparam' to TRUE.

See Details for more information on parameters.

A matrix with parameters values in columns for each state in rows:

e

Probability of extending a segment, increase to lengthen segments, decrase to shorten segments. Range: (0, 1)

strength

Strength of initial e suggestion, reducing allows e to change, increasing makes e undefiable. Range: [0, Inf)

mu

Suggested median for copy numbers in state, change to readjust classification of states. Range: (-Inf, Inf)

lambda

Suggested precision (inversed variance) for copy numbers in state, increase to reduce overlap between states. Range: [0, Inf)

nu

Suggested degree of freedom between states, increase to reduce overlap between states. Range: [0, Inf)

kappa

Suggested distribution of states. Should sum to 1.

m

Optimal value for mu, difference from corresponding mu value determines elasticity of the mu value. i.e. Set to identical value as mu if you don't want mu to move much.

eta

Mobility of mu, increase to allow more movement. Range: [0, Inf)

gamma

Prior shape on lambda, gamma distribution. Effects flexibility of lambda.

S

Prior scale on lambda, gamma distribution. Effects flexibility of lambda.

autosomes

Array of LOGICAL values corresponding to the 'chr' argument where an element is TRUE if the chromosome is an autosome, otherwise FALSE. If not provided, will automatically set the following chromosomes to false: "X", "Y", "23", "24", "chrX", chrY", "M", "MT", "chrM".

maxiter

The maximum number of iterations allows for the Maximum-Expectation algorithm, reduce to decrease running time at the expense of robustness.

getparam

If TRUE, generates and returns parameters without running segmentation.

verbose

Set to FALSE if messages are not desired

Details

HMMsegment is a two stage algorithm that first runs an Expectation-Maximization algorithm to find the optimal set of parameters based on suggested parameter inputs, and allowed flexibilities. After iteratively finding the optimal parameters, the actual segmentation of the data is conducted with the Viterbi algorithm, finally output segmented states. This is an extension to the hidden Markov model described in Shah et al., 2006.

Parameters are divided into two main categories:

  • Initial parameters: e, mu, lambda, nu, kappa

  • Flexibility parameters: strength, m, eta, gamma, S

Where initial parameters are treated as starting suggestions for the parameter optimization algorithm, and flexibility parameters (hyperparameters) define how much the initial parameters are allowed to deviate during the search for the optimal parameters.

With a good copy number dataset, in theory, given enough flexibility, the algorithm should always find a similar set of optimal parameters regardless of initial parameters (although running times will vary).

If for some reason you wish to manually set the parameters for the final segmentation process, one should tune all flexibility parameters to minimal values. For example, if you wish to increase the length of segments, you could set:

1
2
3
    param$e <- 0.9999999999999999
    param$strength <- 1e30
  

Which suggests that segments should be very long, and gives minimal to non-existant flexibility to your suggestion.

See vignette for diagrammed example:

1
2
    vignette("HMMcopy")
  

Value

A list object containing multiple values, although in practice only the state assigned to each copy number value in 'states' and the segments of non-overlapping states in 'segs' are of interest.

By default, there are 6 states, which in a diploid sample corresponds to the following chromosomal copies and biological state:

1

<=0 copies, homozogous deletion

2

1 copy, heterozogous deletion

3

2 copies, neutral

4

3 copies, gain

5

4 copies, amplification

6

>=5 copies, high level amplification

The full list of output is as follows:

states

The state assigned to each copy number value

segs

Non-overlapping segments and medians of each segment

mus

Optimal median of of copy numbers in state

lambda

Optimal precision of copy numbers in state

pi

Optimal state distribution

loglik

The likelihood values of each EM algorithm iteration

rho

Posterior marginals (responsibilities) for each position and state

Author(s)

Daniel Lai, Gavin Ha, Sohrab Shah

References

Sohrab P Shah, Xiang Xuan, Ron J DeLeeuw, Mehrnoush Khojasteh, Wan L Lam, Raymond Ng, and Kevin P Murphy. Integrating copy number polymorphisms into array cgh analysis using a robust hmm. Bioinformatics, 22(14):e431-9, Jul 2006.

See Also

correctReadcount, to correct the readcounts prior to segmentation and classification for better results.

Examples

1
2