constructGLM: Constructs GLM to model per-diagonal Hi-C contact dependancy

In rz6/DIADEM: Hi-C differential analysis via dependency modelling of chromatin interactions with generalized linear models

Description

Models Hi-C contacts using (robust) Negative Binomial (or Poisson when the data is underdispersed) regression. Given the fact that Hi-C data suffers from contact decay bias this method is intendent to model each diagonal separately. By default this function uses robust Negative Binomial regression to model interaction dependencies.

Usage

constructGLM(
  df,
  robust.nb = TRUE,
  overdisp.test.pval = 0.01,
  max.nobs = 20000,
  nrep = 10
)

Arguments

df	data frame with predictor, response, outlier columns
robust.nb	logical whther to use robust fitting procedure (see details)
overdisp.test.pval	numeric significance threshold for testing ovedispersion
max.nobs	numeric maximum number of observations (points), i.e. sample size to be taken for robust NB regression estimation (see details)
nrep	numeric number of repetitions for subsampling (see details)

Details

If robust.nb is true then this function uses robust Negative Binomial estimation method developed in \insertCiteaeberhard2014robustDIADEM. This function uses the code of glmrob.nb function written by William Aeberhard, which is available at: https://github.com/williamaeberhard/glmrob.nb.

At first overdispersion test is performed to decide if Negative Binomial or Poisson regression should be used. If robust.nb is true the estimation may consume huge amounts of memory for large sample sizes (like for example 400000 points). In order to prevent that whenever the sample size exceeds max.nobs initial sample is subsampled to max.nobs size and model is estimated on subsample. This procedure is repeated nrep times and final parameter estimate equals average over subsampled estimates.

Value

object of class glm or MASS::glm.nb

References

\insertRef

aeberhard2014robustDIADEM

See Also

glm, glm.nb to see how GLM are constructed, dispersiontest to see how overdispersion is tested

rz6/DIADEM documentation built on Dec. 31, 2019, 3:51 a.m.