README.md

In CshlSiepelLab/tuSelecter2: Deconvolution of Expression for Nascent RNA Sequencing Data

TITLE: "DENR" AUTHORS: "Noah Dukler; Yixin Zhao"

Quantifying nascent RNA sequencing data with transcript level resolution

Installation

DENR package

Package can be installed with the following line of code:

devtools::install_github("CshlSiepelLab/DENR")

Tensorflow and Keras

The TSS identification method uses the deep learning framework Keras and Tensorflow. Only tensorflow 2.0 or greater is supported. Note that the version of the R tensorflow package is not the same as the python versions it is using under the hood. To check it run.

tensorflow::tf_config()

Instructions on installing tensorflow and keras from R can be found here.

Overview

DENR performs annotation based transcript level quantification on nascent RNA sequencing data. Although it was developed on PRO-seq data it should be generally applicable to any nascent sequencing dataset which has been processed such that each read is represented as a single count representing the 3' end of the transcript being synthesized. For a pipeline to process PRO-seq data in this manner see the Danko Lab's PRO-seq 2.0 pipeline here.

Application tutorial

For details on running the DENR method please see the Introduction vignette. It can be built locally or viewed online here.

Methodology

The DENR method explains the observed polymerase density as a weighted mixture of the underlying transcript annotations. The algorithm is outlined as follows ("*" steps are optional):

1. Construct a unique set of pre-RNA isoform models from annotations
2. Produce non-uniform polymerase density profile*
3. Predict inactive transcript start sites (TSS)*
4. Estimate pre-RNA isoform abundances

Transcript model generation

After being provided annotations, DENR constructs a set of models corresponding to each mature RNA isoform and reduces them to a set of unique pre-RNA isoform models. Isoforms with highly similar 5' and 3' ends are often indistinguishable from each other when looking at nascent RNA sequencing data. The degree to which this occurs depends on the degree of granularity in the isoform models as specified by the user and the length of the 5' and 3' regions that are masked for each isoform. We recommend masking roughly +/- 1kb for all isoforms as initiation and especially termination are messy processes producing highly noisy signals and thus ignoring them generally improves the performance of DENR. DENR then reduces these non-identifiable isoforms to a shared model. This process is outlined in the figure below.

Quantification

pre-RNA isoforms are quantified by minimizing the difference between the polymerase density predicted by isoform models (M) weighted by their abundances (A) vs the observed polymerase density (Y). There are two different cost functions, one based on the raw counts (1), the other based on the log counts (2).

The default model is the one using the log counts as the errors are more gaussian-like and thus a better match the assumption of normally distributed errors implied by the sum-of-squares cost function.

Optional: TSS filtering

TSS filtering is done by providing the fitting step with a list of inactive TSS. The user may create this list however they want, but we do include an deep learning model trained to detect active TSS in the DENR package. In order to run this model keras must be installed on your machine. As this is a fairly simple convolutional model it runs fine on a CPU, with no need for GPU acceleration.

Technical notes

Technical Note 1: If two or more transcripts share the same model but at least one is not marked as inactive, the model will be presumed to be active and allowed to have a non-zero abundance value during fitting.

Technical Note 2: Given that the TSS predictor misses some active TSS with abnormal polymerase density patterns there is an additional heuristic that looks for unexplainable high density polymerase regions. It does this by calculating an upstream polymerase ratio (UPR) using the regions indicated in the schematic below:

If the UPR of an isoform t is greater than or equal to 5, and there are no other active isoforms within 5Kb upstream or 6Kb downstream of the TSS of t, then isoform t is eligible to be assigned a non-zero weight during the fitting step.

Optional: Shape profile

It is know that polymerase density varies as the polymerase moves the body of the gene. Although we already recommend masking out the head and tail regions of the isoforms to avoid the large spikes of polymerase density typically viewed there, we provide another tool to account for the more subtle variation caused by the acceleration. Using heuristic methods to select isoforms that are explained by the vast majority of polymerase density at their locus, DENR constructs an empirical model for relative polymerase densities in the gene body. The profile is then calculated using a loess fit that maps position within the gene on a rescaled [0, 1] range to the observed polymerase density, normalized median density in the gene body. Relative positions within the gene body are calculated such that regions close to the head and tail of the gene where absolute position is important are scaled identically across genes (e.g. head 5kb -> [0, 0.2] and tail 5kb -> [0.8, 1]), and the midsection of each gene is scaled to fit into the remaining space on the [0, 1] interval. Short genes are handled specially to just use the applicable part of the linearly scaled regions. Once this shape profile has been calculated it can be used to adjust the isoform models so that they do not assume uniform polymerase density throughout the gene body.

CshlSiepelLab/tuSelecter2 documentation built on July 18, 2021, 5:09 p.m.