| avg_and_regularize {bakR} | R Documentation |
avg_and_regularize pools and regularizes replicate estimates of kinetic parameters. There are two key steps in this
downstream analysis. 1st, the uncertainty for each feature is used to fit a linear ln(uncertainty) vs. log10(read depth) trend,
and uncertainties for individual features are shrunk towards the regression line. The uncertainty for each feature is a combination of the
Fisher Information asymptotic uncertainty as well as the amount of variability seen between estimates. Regularization of uncertainty
estimates is performed using the analytic results of a Normal distribution likelihood with known mean and unknown variance and conjugate
priors. The prior parameters are estimated from the regression and amount of variability about the regression line. The strength of
regularization can be tuned by adjusting the prior_weight parameter, with larger numbers yielding stronger shrinkage towards
the regression line. The 2nd step is to regularize the average kdeg estimates. This is done using the analytic results of a
Normal distribution likelihood model with unknown mean and known variance and conjugate priors. The prior parameters are estimated from the
population wide kdeg distribution (using its mean and standard deviation as the mean and standard deviation of the normal prior).
In the 1st step, the known mean is assumed to be the average kdeg, averaged across replicates and weighted by the number of reads
mapping to the feature in each replicate. In the 2nd step, the known variance is assumed to be that obtained following regularization
of the uncertainty estimates.
avg_and_regularize(
Mut_data_est,
nreps,
sample_lookup,
feature_lookup,
nbin = NULL,
NSS = FALSE,
Chase = FALSE,
BDA_model = FALSE,
null_cutoff = 0,
Mutrates = NULL,
ztest = FALSE
)
Mut_data_est |
Dataframe with fraction new estimation information. Required columns are:
|
nreps |
Vector of number of replicates in each experimental condition |
sample_lookup |
Dictionary mapping sample names to various experimental details |
feature_lookup |
Dictionary mapping feature IDs to original feature names |
nbin |
Number of bins for mean-variance relationship estimation. If NULL, max of 10 or (number of logit(fn) estimates)/100 is used |
NSS |
Logical; if TRUE, logit(fn)s are compared rather than log(kdeg) so as to avoid steady-state assumption. |
Chase |
Logical; Set to TRUE if analyzing a pulse-chase experiment. If TRUE, kdeg = -ln(fn)/tl where fn is the fraction of reads that are s4U (more properly referred to as the fraction old in the context of a pulse-chase experiment) |
BDA_model |
Logical; if TRUE, variance is regularized with scaled inverse chi-squared model. Otherwise a log-normal model is used. |
null_cutoff |
bakR will test the null hypothesis of |effect size| < |null_cutoff| |
Mutrates |
List containing new and old mutation rate estimates |
ztest |
TRUE; if TRUE, then a z-test is used for p-value calculation rather than the more conservative moderated t-test. |
Effect sizes (changes in kdeg) are obtained as the difference in log(kdeg) means between the reference and experimental sample(s), and the log(kdeg)s are assumed to be independent so that the variance of the effect size is the sum of the log(kdeg) variances. P-values assessing the significance of the effect size are obtained using a moderated t-test with number of degrees of freedom determined from the uncertainty regression hyperparameters and are adjusted for multiple testing using the Benjamini- Hochberg procedure to control false discovery rates (FDRs).
In some cases, the assumed ODE model of RNA metabolism will not accurately model the dynamics of a biological system being analyzed.
In these cases, it is best to compare logit(fraction new)s directly rather than converting fraction new to log(kdeg).
This analysis strategy is implemented when NSS is set to TRUE. Comparing logit(fraction new) is only valid
If a single metabolic label time has been used for all samples. For example, if a label time of 1 hour was used for NR-seq
data from WT cells and a 2 hour label time was used in KO cells, this comparison is no longer valid as differences in
logit(fraction new) could stem from differences in kinetics or label times.
List with dataframes providing information about replicate-specific and pooled analysis results. The output includes:
Fn_Estimates; dataframe with estimates for the fraction new and fraction new uncertainty for each feature in each replicate. The columns of this dataframe are:
Feature_ID; Numerical ID of feature
Exp_ID; Numerical ID for experimental condition (Exp_ID from metadf)
Replicate; Numerical ID for replicate
logit_fn; logit(fraction new) estimate, unregularized
logit_fn_se; logit(fraction new) uncertainty, unregularized and obtained from Fisher Information
nreads; Number of reads mapping to the feature in the sample for which the estimates were obtained
log_kdeg; log of degradation rate constant (kdeg) estimate, unregularized
kdeg; degradation rate constant (kdeg) estimate
log_kd_se; log(kdeg) uncertainty, unregularized and obtained from Fisher Information
sample; Sample name
XF; Original feature name
Regularized_ests; dataframe with average fraction new and kdeg estimates, averaged across the replicates and regularized using priors informed by the entire dataset. The columns of this dataframe are:
Feature_ID; Numerical ID of feature
Exp_ID; Numerical ID for experimental condition (Exp_ID from metadf)
avg_log_kdeg; Weighted average of log(kdeg) from each replicate, weighted by sample and feature-specific read depth
sd_log_kdeg; Standard deviation of the log(kdeg) estimates
nreads; Total number of reads mapping to the feature in that condition
sdp; Prior standard deviation for fraction new estimate regularization
theta_o; Prior mean for fraction new estimate regularization
sd_post; Posterior uncertainty
log_kdeg_post; Posterior mean for log(kdeg) estimate
kdeg; exp(log_kdeg_post)
kdeg_sd; kdeg uncertainty
XF; Original feature name
Effects_df; dataframe with estimates of the effect size (change in logit(fn)) comparing each experimental condition to the reference sample for each feature. This dataframe also includes p-values obtained from a moderated t-test. The columns of this dataframe are:
Feature_ID; Numerical ID of feature
Exp_ID; Numerical ID for experimental condition (Exp_ID from metadf)
L2FC(kdeg); Log2 fold change (L2FC) kdeg estimate or change in logit(fn) if NSS TRUE
effect; LFC(kdeg)
se; Uncertainty in L2FC_kdeg
pval; P-value obtained using effect_size, se, and a z-test
padj; pval adjusted for multiple testing using Benjamini-Hochberg procedure
XF; Original feature name
Mut_rates; list of two elements. The 1st element is a dataframe of s4U induced mutation rate estimates, where the mut column represents the experimental ID and the rep column represents the replicate ID. The 2nd element is the single background mutation rate estimate used
Hyper_Parameters; vector of two elements, named a and b. These are the hyperparameters estimated from the uncertainties for each feature, and represent the two parameters of a Scaled Inverse Chi-Square distribution. Importantly, a is the number of additional degrees of freedom provided by the sharing of uncertainty information across the dataset, to be used in the moderated t-test.
Mean_Variance_lms; linear model objects obtained from the uncertainty vs. read count regression model. One model is run for each Exp_ID