preciseTAD: Precise TAD boundary prediction at base-level resolution...

In dozmorovlab/preciseTAD: preciseTAD: A machine learning framework for precise TAD boundary prediction

Precise TAD boundary prediction at base-level resolution using density-based spatial clustering and partitioning techniques

Description

Precise TAD boundary prediction at base-level resolution using density-based spatial clustering and partitioning techniques

Usage

preciseTAD(
  genomicElements.GR,
  featureType = "distance",
  CHR,
  chromCoords = NULL,
  tadModel,
  threshold = 1,
  verbose = TRUE,
  parallel = NULL,
  DBSCAN_params = list(30000, 100),
  slope = 5000,
  genome = "hg19",
  BaseProbs = FALSE,
  savetobed = FALSE
)

Arguments

genomicElements.GR	GRangesList object containing GRanges from each ChIP-seq BED file that was used to train a predictive model (can be obtained using the bedToGRangesList). Required.
featureType	Controls how the feature space is constructed (one of either "binary", "oc", "op", "signal, or "distance" (log2- transformed). Default and recommended: "distance".
CHR	Controls which chromosome to predict boundaries on at base-level resolution, e.g., CHR22. Required.
chromCoords	List containing the starting bp coordinate and ending bp coordinate that defines the region of the linear genome to make predictions on. If chromCoords is not specified, then predictions will be made on the entire chromosome. Default is NULL.
tadModel	Model object used to obtain predicted probabilities at base-level resolution (examples include gbm, glmnet, svm, glm, etc). For a random forest model, can be obtained using preciseTAD::randomForest). Required.
threshold	Bases with predicted probabilities that are greater than or equal to this value are labeled as potential TAD boundaries. Values in the range of .95-1.0 are suggested. Default is 1. To explore how selection of the 'threshold' parameter affects the results, it is recommended to rerun the function with a different threshold, e.g., 0.99, and compare the results of Normalized Enrichment test (see 'DBSCAN_params' and the 'preciseTADparams' slot).
verbose	Option to print progress. Default is TRUE.
parallel	Option to parallelise the process for obtaining predicted probabilities. Must be number to indicate the number of cores to use in parallel. Default is NULL.
DBSCAN_params	Parameters passed to dbscan in list form containing 1) eps and 2) MinPts. If a vector of different values is passed to either or both eps and MinPts, then each combination of these parameters is evaluated to maximize normalized enrichment (NE) is the provided genomic annotations. Normalized Enrichment is calculated as the number of genomic annotations that overlap with flanked predicted boundary points (see the slope parameter) divided by the total number of predicted boundaries, averaged for all genomic annotations. Parameters yielding maximum NE score are automatically selected for the final prediction. It is advisable to explore results of the NE test, available in the 'preciseTADparams' slot of the returned object (NEmean - mean normalized enrichment, larger the better; k - number of PTBRs), to, potentially, find eps and MinPts parameters providing the number of PTBRs and the NE score better agreeing with the number of boundaries used for training. Default: list(30000, 100). Required.
slope	Controls how much to flank the predicted TAD boundary points for calculating normalized enrichment. Default: 5000 bases. Required.
genome	version of the human genome assembly. Used to filter out bases overlapping centromeric regions. Accepted values - hg19 or hg38. Default: hg19
BaseProbs	Option to include the vector of probabilities for each base-level coordinate. Recommended to be used only when chromCoords is specified. Default: FALSE
savetobed	If true, preciseTAD regions (PTBRs) and preciseTAD points (PTBPs) will be saved as BED-like files into the current folder (as.data.frame(GRanges)). File name convention: <PTBRs/PTBPs>_<threshold>_<MinPts>_<eps>.bed, e.g., PTBR_1_3_30000.bed. If multiple DBSCAN_params are specified, each result will be saved in its own file. Default: FALSE

Value

A list containing 4 elements including: 1) data frame with average (and standard deviation) normalized enrichment (NE) values for each combination of t and eps (only if multiple values are provided for at least paramenter; all subsequent summaries are applied to optimal combination of (t, eps)), 2) the genomic coordinates spanning each preciseTAD predicted region (PTBR), 3) the genomic coordinates of preciseTAD predicted boundaries points (PTBP), 4) a named list including summary statistics of the following: PTBRWidth - PTBR width, PTBRCoverage - the proportion of bases within a PTBR with probabilities that equal to or exceed the threshold (t=1 by default), DistanceBetweenPTBR - the genomic distance between the end of the previous PTBR and the start of the subsequent PTBR, NumSubRegions - the number of the subregions in each PTBR cluster, SubRegionWidth - the width of the subregion forming each PTBR, DistBetweenSubRegions - the genomic distance between the end of the previous PTBR-specific subregion and the start of the subsequent PTBR-specific subregion, NormilizedEnrichment - the normalized enrichment of the genomic annotations used in the model around flanked PTBPs, and BaseProbs - a numeric vector of probabilities for each corresponding base coordinate.

Examples

# Read in ARROWHEAD-called TADs at 5kb
data(arrowhead_gm12878_5kb)

# Extract unique boundaries
bounds.GR <- extractBoundaries(domains.mat = arrowhead_gm12878_5kb,
                               filter = FALSE,
                               CHR = c("CHR21", "CHR22"),
                               resolution = 5000)

# Read in GRangesList of 26 TFBS and filter to include only CTCF, RAD21,
#SMC3, and ZNF143
data(tfbsList)

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

# Create the binned data matrix for CHR1 (training) and CHR22 (testing)
# using 5 kb binning, distance-type predictors from 4 TFBS from
# the GM12878 cell line, and random under-sampling
set.seed(123)
tadData <- createTADdata(bounds.GR = bounds.GR,
                         resolution = 5000,
                         genomicElements.GR = tfbsList_filt,
                         featureType = "distance",
                         resampling = "rus",
                         trainCHR = "CHR21",
                         predictCHR = "CHR22")

# Perform random forest using TADrandomForest by tuning mtry over 10 values
# using 3-fold CV
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)

# Apply preciseTAD on a specific 2mb section of CHR22:17000000-18000000
set.seed(123)
pt <- preciseTAD(genomicElements.GR = tfbsList_filt,
                 featureType = "distance",
                 CHR = "CHR22",
                 chromCoords = list(17000000, 18000000),
                 tadModel = tadModel[[1]],
                 threshold = 1.0,
                 verbose = TRUE,
                 parallel = NULL,
                 DBSCAN_params = list(c(1000, 10000, 30000), c(10, 100, 1000)),
                 slope = 5000,
                 genome = "hg19",
                 BaseProbs = FALSE,
                 savetobed = FALSE)

dozmorovlab/preciseTAD documentation built on April 26, 2022, 9:42 a.m.