F-Seq2: improving the feature density based peak caller with dynamic statistics

Abstract

Genomic and epigenomic features are captured at a genome-wide level by using high-throughput sequencing (HTS) technologies. Peak calling delineates features identified in HTS experiments, such as open chromatin regions and transcription factor binding sites, by comparing the observed read distributions to a random expectation. Since its introduction, F-Seq has been widely used and shown to be the most sensitive and accurate peak caller for DNase I hypersensitive site (DNase-seq) data. However, the first release (F-Seq1) has two key limitations: lack of support for user-input control datasets, and poor test statistic reporting. These constrain its ability to capture systematic and experimental biases inherent to the background distributions in peak prediction, and to subsequently rank predicted peaks by confidence. To address these limitations, we present F-Seq2, which combines kernel density estimation and a dynamic ‘continuous’ Poisson test to account for local biases and accurately rank candidate peaks. The output of F-Seq2 is suitable for irreproducible discovery rate analysis as test statistics are calculated for individual candidate summits, allowing direct comparison of predictions across replicates. These improvements significantly boost the performance of F-Seq2 for ATAC-seq and ChIP-seq datasets, outperforming competing peak callers used by the ENCODE Consortium in terms of precision and recall.

ROCCO: A Robust Method for Detection of Open Chromatin via Convex Optimization

“…However, to gauge performance and ensure viability, some proxy for ground truth is needed. Following [Zhao and Boyle, 2021], we constructed a union set , , of conservative Irreproducible Discovery Rate (IDR) [Li et al, 2011] peaks from ENCODE transcription factor ChIP-seq experiments in the lymphoblast cell line. We assume that the majority of annotated transcription factor binding sites will correspond to open chromatin regions, but we note that variability in binding at a snapshot in time, the incomplete annotation of all transcription factor binding, and cases where factors can bind to non-accessible chromatin introduce notable limitations.…”

“…We then computed precision as and recall as where D X denotes the consensus peaks obtained from method X and set intersections in the numerators are computed using . The F β -score was then calculated as the harmonic mean of precision and recall where recall is weighed β times as much as precision, that is: As in [Zhao and Boyle, 2021], we use the ℱ β -score as the primary metric for comparison of methods since it intuitively combines both precision and recall and is less affected by extreme regions of the precision-recall curve that do not correspond to realistic use-cases.…”

“…However, to gauge performance and ensure viability, some proxy for ground truth is needed. Following (Zhao and Boyle, 2021), we constructed a union set, GT, of conservative IDR peaks from ENCODE transcription factor ChIP-seq experiments in the GM12878 lymphoblast cell line. The F β -score, defined below, was then used to assess the ability to recover the regions in GT using ATAC-seq data from the 56 independent samples.…”

“…As in (Zhao and Boyle, 2021), we use the F β -score as the primary metric for comparison of methods since it intuitively combines both precision and recall and is less affected by extreme regions of the precision-recall curve that do not correspond to realistic use-cases.…”

Tensor Decomposition and Principal Component Analysis-Based Unsupervised Feature Extraction Outperforms State-of-the-Art Methods When Applied to Histone Modification Profiles

“…F-Seq2 [21] was implemented as a Linux command line program. It was downloaded from Github https://github.com/Boyle-Lab/F-Seq2.…”