THiCweed: fast, sensitive detection of sequence features by clustering big datasets

Abstract

We present THiCweed, a new approach to analyzing transcription factor binding data from high-throughput chromatin immunoprecipitation-sequencing (ChIP-seq) experiments. THiCweed clusters bound regions based on sequence similarity using a divisive hierarchical clustering approach based on sequence similarity within sliding windows, while exploring both strands. ThiCweed is specially geared toward data containing mixtures of motifs, which present a challenge to traditional motif-finders. Our implementation is significantly faster than standard motif-finding programs, able to process 30 000 peaks in 1–2 h, on a single CPU core of a desktop computer. On synthetic data containing mixtures of motifs it is as accurate or more accurate than all other tested programs. THiCweed performs best with large ‘window’ sizes (≥50 bp), much longer than typical binding sites (7–15 bp). On real data it successfully recovers literature motifs, but also uncovers complex sequence characteristics in flanking DNA, variant motifs and secondary motifs even when they occur in <5% of the input, all of which appear biologically relevant. We also find recurring sequence patterns across diverse ChIP-seq datasets, possibly related to chromatin architecture and looping. THiCweed thus goes beyond traditional motif finding to give new insights into genomic transcription factor-binding complexity.

Position-specific evolution in transcription factor binding sites, and a fast likelihood calculation for the F81 model

“…Focusing on CTCF, we find that PSSVs inferred for our approach are strongest for genuine CTCF motif instances in CTCF ChIP-seq peak regions (information loss less than ); weaker for scrambled CTCF motifs in CTCF ChIP-seq peak regions, and for genuine CTCF motif instances in random genomic regions (information loss ); and weakest for scrambled CTCF motifs in random genomic regions (information loss ). It is known that there are frequently secondary motifs and sequence features in the neighbourhood of the TF core motif [ 25 , 26 ], so it is plausible that an extended region around the core motif, including chance instances of scrambled motifs, may be under greater selection in ChIP-seq regions compared to random genomic regions. For a truly neutrally evolving region, one expects that the PSSV would be almost flat, but estimates of what fraction of the human genome is functional or under selective constraint range from 7% [ 27 ] to, controversially, 80% [ 28 ], with many authors agreeing that 10%–15% is plausible (see [ 29 ] for a discussion).…”

Position-specific evolution in transcription factor binding sites, and a fast likelihood calculation for the F81 model

“…Focussing on CTCF, we find that PSSVs inferred for our approach are strongest for genuine CTCF motif instances in CTCF ChIP-seq peak regions (information loss < 10%); weaker for scrambled CTCF motifs in CTCF ChIP-seq peak regions, and for genuine CTCF motif instances in random genomic regions (information loss ≈ 15%–18%); and weakest for scrambled CTCF motifs in random genomic regions (information loss ≈ 23%-25%). It is known that there are frequently secondary motifs and sequence features in the neighbourhood of the TF core motif [25, 26], so it is plausible that an extended region around the core motif, including chance instances of scrambled motifs, may be under greater selection in ChIP-seq regions compared to random genomic regions. For a truly neutrally evolving region, one expects that the PSSV would be almost flat, but estimates of what fraction of the human genome is functional or under selective constraint range from 7% [27] to, controversially, 80% [28], with many authors agreeing that 10%–15% is plausible (see [29] for a discussion).…”

A universal framework for detecting cis-regulatory diversity in DNA regulatory regions

“…Motifs may be missed because they are present only in a small set of regions and therefore not statistically overrepresented in the entire set. A few recent methods do account for this by posing this as a clustering problem, with each cluster of sequences being dominated by a potentially di↵erent motif 11,12 . However, these approaches do not take into account combinations of motifs, which may be critical to drive the biochemical activity at the region.…”