HiCglm: An S3 object to represent differential GLM Hi-C model.

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

Description

Models diagonal-wise dependencies between Hi-C data sets with GLM. Model is constructed as follows:

• merge maps1 with maps2

• for each diagonal in diagonal pools

  • take all points from this diagonal, such that they are non zero in map1 (X) and non zero in map2 (Y)

  • remove outliers using robust regression (only performed when robust.nb is false)

  • model response (Y = f(X) or X = f(Y)) using (robust) Negative Binomial distribution

Before fitting the model it's recommended to first inspect correlations between analyzed Hi-C maps before fixing this variable. As the ratio of noise / signal in Hi-C data increases rapidly with decay it's unadvised to use all diagonals for modelling. The number of diagonals to be used will depend on chromosome length, resolution and data quality. One can retain uncorrelated diagonals (pools) by increasing significance threshold (alpha) for PCC significance.

Usage

HiCglm(
  hic.comparator,
  diag.frac = 0.5,
  alpha = 0.05,
  robust.nb = TRUE,
  remove.outliers = TRUE,
  outlier.weight = 0,
  ncores = 1,
  max.nobs = 10000,
  nrep = 20
)

Arguments

hic.comparator	object of type HiCcomparator
diag.frac	fraction of diagonals to use to fit models, by default fraction of chromsome length is used to indicate number of diagonals.
alpha	significance (p-value) threshold to discard diagonals with insignificant Pearson correlation coefficient between X (contacts from contact map 1) and Y (corresponding contacts from contact map 2).
robust.nb	logical whether to use robust Negative Binomial regression as main model (see details), true by default
remove.outliers	logical only relevant if robust.nb is set to false - if true try to remove outliers (potential Differential Interactions) before model fitting (NB regression) using robust regression (IRLS) with bisquare weight function
outlier.weight	numeric weight threshold to remove outliers, 0 by default
ncores	numeric number of cores for parallel processing to speed up computations
max.nobs	numeric maximum number of observations to be used for robust NB regression model estimation (see constructGLM)
nrep	number of repetitions to average parameter estimates during robust NB regression model estimation on large sample sizes (see constructGLM)

Details

When robust.nb is true, the model is fitted using robust procedure developed and published by \insertCiteaeberhard2014robustDIADEM - see constructGLM function for details.

The main assumption behind HiCglm is that given 2 Hi-C datasets (even very different cell lines) their diagonal-wise interaction profiles will be correlated, which should manifest in X,Y plots having funnel like shape (see simulate_null function). In this setting it is further assumed (based on observed X,Y dependencies) that potential Differential Interactions will behave like outliers thereby disobeying the funnel-like pattern. The use of robust Negative Binomial regression diminishes the influence of outliers on model fitting and allows to capture “uncontaminated” (with no DIs) null model.

Value

S3 object of class HiCglm

References

\insertRef

aeberhard2014robustDIADEM

See Also

HiCcomparator on how to construct HiCcomparator object and constructGLM, robustRegBS on how GLM fitting process

Examples

# first create hiccomp (i.e. HiCcomparator object), then
hicglm <- HiCglm(hiccomp)

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