preciseTAD: A transfer learning framework for 3D domain boundary prediction at base-pair resolution

preciseTAD: A transfer learning framework for 3D domain boundary prediction at base-pair resolution Spiro C. Stilianoudakis, Maggie A. Marshall, Mikhail G. Dozmorov. bioRxiv 2020.09.03.282186; doi: https://doi.org/10.1101/2020.09.03.282186

Predicted preciseTAD boundary points (PTBPs) and regions (PTBRs) for 60 cell lines are available here.

Overview

preciseTAD provides functions to predict the location of boundaries of topologically associated domains (TADs) and chromatin loops at base-level resolution. As an input, it takes BED-formatted genomic coordinates of domain boundaries detected from low-resolution Hi-C data, and coordinates of high-resolution genomic annotations from ENCODE or other consortia. preciseTAD employs several feature engineering strategies and resampling techniques to address class imbalance, and trains an optimized random forest model for predicting low-resolution domain boundaries. Translated on a base-level, preciseTAD predicts the probability for each base to be a boundary. Density-based clustering and scalable partitioning techniques are used to detect precise boundary regions and summit points. Compared with low-resolution boundaries, preciseTAD boundaries are highly enriched for CTCF, RAD21, SMC3, and ZNF143 signal and more conserved across cell lines. The pre-trained model can accurately predict boundaries in another cell line using CTCF, RAD21, SMC3, and ZNF143 annotation data for this cell line.

The main functions (in order of implementation) are:

• extractBoundaries() accepts a 3-column data.frame or matrix with the chromosomal coordinates of user-defined domains and outputs the unique boundaries. The second and third columns are the domain anchor centers.
• bedToGRangesList() accepts a filepath containing BED files representing the coordinates of ChIP-seq defined functional genomic annotations
• createTADdata() accepts a set of unique boundaries and genomic annotations derived from extractBoundaries() and bedToGRangesList(), respectively, to create the data matrix used to build a model to predict domain boundary regions
• TADrandomForest() a wrapper of the randomForest package which implements a random forest binary classification algorithm on domain boundary data
• preciseTAD() which leverages a domain boundary prediction model (i.e., random forest) and density-based clustering to predict TAD boundary coordinates at a base-level resolution

Installation

preciseTAD can be installed from Bioconductor:

# if (!requireNamespace("BiocManager", quietly=TRUE))
#     install.packages("BiocManager")
# BiocManager::install("preciseTAD")
library(preciseTAD)
#> 

The latest version of preciseTAD can be directly installed from Github:

devtools::install_github("dozmorovlab/preciseTAD", build_vignettes = TRUE)
library(preciseTAD)

Usage

Below is a brief workflow of how to implement preciseTAD on binned data from CHR1 to get precise base pair coordinates of TAD boundaries for a 10mb section of CHR 22. For more details, including the example how to use the pre-trained model, see vignette("preciseTAD")

First, you need to obtain called TAD boundaries using an established TAD-caller. As an example, consider the Arrowhead TAD-caller, a part of the juicer suite of tools developed by the Aiden Lab. Arrowhead outputs a .txt file with the chromosomal start and end coordinates of their called TADs. As an example, we have provided Arrowhead TADs for GM12878 at 5kb resolution.

data("arrowhead_gm12878_5kb")
head(arrowhead_gm12878_5kb)
#>   V1        V2        V3
#> 1  1  49375000  50805000
#> 2  1  16830000  17230000
#> 3  1 163355000 164860000
#> 4  1 231935000 233400000
#> 5  1 149035000 149430000
#> 6  1   3995000   5505000

The unique boundaries for CHR1 and CHR22 can be extracted as:

bounds <- extractBoundaries(domains.mat = arrowhead_gm12878_5kb, filter = FALSE, CHR = c("CHR1", "CHR22"), resolution = 5000)
bounds
#> GRanges object with 1901 ranges and 0 metadata columns:
#>         seqnames            ranges strand
#>            <Rle>         <IRanges>  <Rle>
#>     787     chr1     815000-815001      *
#>    9196     chr1     890000-890001      *
#>      37     chr1     915000-915001      *
#>    8923     chr1   1005000-1005001      *
#>     275     chr1   1015000-1015001      *
#>     ...      ...               ...    ...
#>    8379    chr22 50815000-50815001      *
#>   16788    chr22 50935000-50935001      *
#>    8382    chr22 50945000-50945001      *
#>   16791    chr22 51060000-51060001      *
#>   16670    chr22 51235000-51235001      *
#>   -------
#>   seqinfo: 22 sequences from an unspecified genome; no seqlengths

Next, you will need to download cell line-specific ChIP-seq data in the form of BED files from ENCODE. Once, you have downloaded your preferred list of functional genomic annotations, store them in a specific file location. These files can then be converted into a GRangesList object and used for downstream modeling using the following command:

path <- "pathToBEDfiles"
tfbsList <- bedToGRangesList(filepath = path, bedList = NULL, bedNames = NULL, pattern = "*.bed", signal = 4)

As an example, we have already provided a GRangesList object with a variety of transcription factor binding sites specific to the GM12878 cell line. Once you load it in, you can see the list of transcription factors using the following:

data("tfbsList")
names(tfbsList)
#>  [1] "Gm12878-Atf3-Haib"       "Gm12878-Cfos-Sydh"      
#>  [3] "Gm12878-Cmyc-Uta"        "Gm12878-Ctcf-Broad"     
#>  [5] "Gm12878-Egr1-Haib"       "Gm12878-Ets1-Haib"      
#>  [7] "Gm12878-Gabp-Haib"       "Gm12878-Jund-Sydh"      
#>  [9] "Gm12878-Max-Sydh"        "Gm12878-Mazab85725-Sydh"
#> [11] "Gm12878-Mef2a-Haib"      "Gm12878-Mxi1-Sydh"      
#> [13] "Gm12878-P300-Sydh"       "Gm12878-Pol2-Haib"      
#> [15] "Gm12878-Pu1-Haib"        "Gm12878-Rad21-Haib"     
#> [17] "Gm12878-Rfx5-Sydh"       "Gm12878-Sin3a-Sydh"     
#> [19] "Gm12878-Six5-Haib"       "Gm12878-Smc3-Sydh"      
#> [21] "Gm12878-Sp1-Haib"        "Gm12878-Srf-Haib"       
#> [23] "Gm12878-Taf1-Haib"       "Gm12878-Tr4-Sydh"       
#> [25] "Gm12878-Yy1-Sydh"        "Gm12878-Znf143-Sydh"

For the purposes of this example, let’s focus only on CTCF, RAD21, SMC3, and ZNF143 transcription factors.

tfbsList_filt <- tfbsList[names(tfbsList) %in% c("Gm12878-Ctcf-Broad", "Gm12878-Rad21-Haib", "Gm12878-Smc3-Sydh", "Gm12878-Znf143-Sydh")]

Now, using the “ground-truth” boundaries and the following TFBS, we can build the data matrix that will be used for predictive modeling. The following command creates the training data from CHR1 and reserves the testing data from CHR22. We specify 5kb sized genomic bins (to match the resolution used to call the original TADs), a distance-type feature space, and apply random under-sampling (RUS) on the training data only.

set.seed(123)
tadData <- createTADdata(bounds.GR          = bounds,
                         resolution         = 5000,
                         genomicElements.GR = tfbsList_filt,
                         featureType        = "distance",
                         resampling         = "rus",
                         trainCHR           = "CHR1",
                         predictCHR         = "CHR22"
)

We can now implement our machine learning algorithm of choice to predict TAD-boundary regions. Here, we opt for the random forest algorithm.

set.seed(123)
tadModel <- TADrandomForest(trainData    = tadData[[1]],
                            testData     = tadData[[2]],
                            tuneParams   = list(mtry = 2, ntree = 500, nodesize = 1),
                            cvFolds      = 3,
                            cvMetric     = "Accuracy",
                            verbose      = TRUE,
                            model        = TRUE,
                            importances  = TRUE,
                            impMeasure   = "MDA",
                            performances = TRUE)
#> Loading required package: lattice
#> Loading required package: ggplot2
#> + Fold1: mtry=2, ntree=500, nodesize=1 
#> - Fold1: mtry=2, ntree=500, nodesize=1 
#> + Fold2: mtry=2, ntree=500, nodesize=1 
#> - Fold2: mtry=2, ntree=500, nodesize=1 
#> + Fold3: mtry=2, ntree=500, nodesize=1 
#> - Fold3: mtry=2, ntree=500, nodesize=1 
#> Aggregating results
#> Fitting final model on full training set
# The model itself
tadModel[[1]]
#> Random Forest 
#> 
#> 3190 samples
#>    4 predictor
#>    2 classes: 'No', 'Yes' 
#> 
#> No pre-processing
#> Resampling: Cross-Validated (3 fold) 
#> Summary of sample sizes: 2126, 2127, 2127 
#> Resampling results:
#> 
#>   MCC        ROC        Sens    Spec       Pos Pred Value  Neg Pred Value
#>   0.4679235  0.7990197  0.7072  0.7598457  0.747661        0.7224703     
#>   Accuracy   Kappa    
#>   0.7335413  0.4670627
#> 
#> Tuning parameter 'mtry' was held constant at a value of 2
#> Tuning
#>  parameter 'ntree' was held constant at a value of 500
#> Tuning
#>  parameter 'nodesize' was held constant at a value of 1

# Variable importances (mean decrease in accuracy)
tadModel[[2]]
#>                 Feature Importance
#> 4 `Gm12878-Znf143-Sydh`   75.98200
#> 3   `Gm12878-Smc3-Sydh`   62.50681
#> 2  `Gm12878-Rad21-Haib`   62.21715
#> 1  `Gm12878-Ctcf-Broad`   32.61855

# Model performance metrics
tadModel[[3]]
#>              Metric  Performance
#> 1                TN 6.400000e+03
#> 2                FN 6.700000e+01
#> 3                FP 2.954000e+03
#> 4                TP 2.390000e+02
#> 5             Total 9.660000e+03
#> 6       Sensitivity 7.810458e-01
#> 7       Specificity 6.841993e-01
#> 8             Kappa 8.363202e-02
#> 9          Accuracy 6.872671e-01
#> 10 BalancedAccuracy 7.326225e-01
#> 11        Precision 7.485124e-02
#> 12              FPR 3.158007e-01
#> 13              FNR 1.036029e-02
#> 14              NPV 9.896397e-01
#> 15              MCC 1.732169e-01
#> 16               F1 1.366105e-01
#> 17              AUC 7.983352e-01
#> 18           Youden 4.652450e-01
#> 19            AUPRC 9.383824e-02

Lastly, we take our TAD-boundary region predictive model and use it to make predictions on a 10mb section of CHR22:35,000,000-45,000,000.

# Run preciseTAD
set.seed(123)
pt <- preciseTAD( genomicElements.GR = tfbsList_filt,
                  featureType        = "distance",
                  CHR                = "CHR22",
                  chromCoords        = list(35000000, 45000000),
                  tadModel           = tadModel[[1]],
                  threshold          = 1.0,
                  verbose            = FALSE,
                  parallel           = 2,
                  DBSCAN_params      = list(30000, 100),
                  slope              = 5000,
                  genome             = "hg19")

# View preciseTAD predicted boundary coordinates between CHR22:35mb-45mb
pt[[2]]
#> GRanges object with 58 ranges and 0 metadata columns:
#>        seqnames            ranges strand
#>           <Rle>         <IRanges>  <Rle>
#>    [1]    chr22 35337875-35342203      *
#>    [2]    chr22 35413232-35426376      *
#>    [3]    chr22 35543040-35549073      *
#>    [4]    chr22 35620677-35634681      *
#>    [5]    chr22 35740783-35747642      *
#>    ...      ...               ...    ...
#>   [54]    chr22 43261091-43270096      *
#>   [55]    chr22 43425520-43490629      *
#>   [56]    chr22 43774902-43791526      *
#>   [57]    chr22 44272171-44283944      *
#>   [58]    chr22 44382230-44396240      *
#>   -------
#>   seqinfo: 1 sequence from an unspecified genome; no seqlengths