This projects downloads mouse and human ChIP-seq data from the Unibind database, organizing this information and gene-scoring each experiment for various projects in the Pavlab.
These are the main deliverables:
-
A gene score vector is created for each experiment, following the method used in Morin et al., 2023 (https://github.com/PavlidisLab/TR_aggregation). This information is bound into a gene by experiment matrix of binding scores.
Unibind experiments may be "duplicated" in that the same experiment is matched to multiple relevant TF motifs. In all such cases, I average the resulting binding scores so that each experiment is only represented once.
Experiments that have fewer than 100 peaks/regions are discarded. This is a fairly lenient filter.
Finally, I log2 transform and quantile normalize the de-duplicated binding matrix. This is what I use for most downstream analysis.
-
A list of basic metadata is saved out. This only includes the ChIP'd symbol, the experiment/file name, the ID, whether the given experiment was "duplicated", and the count of peaks in the experiment.
The "File" column is really the unique identifier as it includes info on which motif was used. But given that I de-duplicate/average the experiments in the processed data, each experiment is represented by one ID which is used to name the columns of the processed matrix.
-
All experiments are saved out as a list of Genomic Ranges objects. This includes "duplicated" experiments (they are not collapsed). This is to provide convenience for overlap functions.
-
An aggregate binding profile for each available TF, which averages the TF's set of binding vectors. These aggregate profiles were used for the ChIP-seq comparison in my single cell coexpression paper (https://github.com/PavlidisLab/TR_singlecell/).
This list object also contains two further "global" summaries that can be used to identify which genes are commonly bound, regardless of the ChIP'd TF. The first averages binding scores across all experiments -- this is influenced by imbalanced dataset counts across TFs. The second averages within each set of TF experiments and then averages these aggregates, such that each TF has equal weight towards the global profile.
-
A linear model is fit per TF, wherein a one-vs-rest contrast is fit to compare a TF's binding scores versus all other TF's binding scores. The model controls for the number of peaks for each experiment, but does not account for imbalanced coverage (i.e., the "rest" comparison will be influenced by the most commonly represented TFs.)
The processed bind score matrices and associated metadata can also be found in this data repository (09_Scored_Unibind_ChIPseq_data.RDS) https://borealisdata.ca/dataset.xhtml?persistentId=doi:10.5683/SP3/HJ1B24