Identification of C2H2-ZF binding preferences from ChIP-seq data using RCADE

Abstract

Current methods for motif discovery from chromatin immunoprecipitation followed by sequencing (ChIP-seq) data often identify non-targeted transcription factor (TF) motifs, and are even further limited when peak sequences are similar due to common ancestry rather than common binding factors. The latter aspect particularly affects a large number of proteins from the Cys2His2 zinc finger (C2H2-ZF) class of TFs, as their binding sites are often dominated by endogenous retroelements that have highly similar sequences. Here, we present recognition code-assisted discovery of regulatory elements (RCADE) for motif discovery from C2H2-ZF ChIP-seq data. RCADE combines predictions from a DNA recognition code of C2H2-ZFs with ChIP-seq data to identify models that represent the genuine DNA binding preferences of C2H2-ZF proteins. We show that RCADE is able to identify generalizable binding models even from peaks that are exclusively located within the repeat regions of the genome, where state-of-the-art motif finding approaches largely fail.

Availability and implementation: RCADE is available as a webserver and also for download at http://rcade.ccbr.utoronto.ca/.

Supplementary information: Supplementary data are available at Bioinformatics online.

Contact: t.hughes@utoronto.ca.

Fig. 1.

RCADE workflow and benchmarking results. (A) RCADE starts by predicting a set of motifs from the target C2H2-ZF protein sequence, using a previously published bacterial-one-hybrid assay-based recognition code, or B1H-RC (Najafabadi et al., 2015), which are evaluated against the ChIP-seq peak sequences to identify significantly enriched motifs, and are then iteratively optimized. (B) Benchmarking workflow for evaluation of RCADE. The peak sequences were divided into two sets of ERE-overlapping and non-ERE peaks. The ERE-overlapping peaks for each protein were used for motif discovery using RCADE, and the motifs were validated using non-ERE peaks. (C,D) Validation results for 18 ERE-binding proteins. The arrows show the improvement in the AUROC of RCADE motifs compared with seed B1H-RC motifs. (E) Example motifs for two proteins that show the largest difference between RCADE and MEME validation results. The top-scoring MEME motif is shown for each protein, followed by the top-scoring motif that is directly predicted from protein sequence using the B1H-RC, and the RCADE optimized motif. The Pearson similarity of the B1H-RC and RCADE motifs was calculated as described previously (Najafabadi et al., 2015)