BIOINFORMATICS ORIGINAL PAPER doi:10.1093/bioinformatics/btm080 Sequence analysis

MOTIVATION Position weight matrices (PMWs) are simple models commonly used in motif-finding algorithms to identify short functional elements, such as cis-regulatory motifs, on genes. When few experimentally verified motifs are available, estimation of the PWM may be poor. The resultant PWM may not reliably discriminate a true motif from a false one. While experimentally identifying such motifs remains time-consuming and expensive, low-resolution binding data from techniques such as ChIP-on-chip and ChIP-PET have become available. We propose a novel but simple method to improve a poorly estimated PWM using ChIP data. METHODOLOGY Starting from an existing PWM, a set of ChIP sequences, and a set of background sequences, our method, GAPWM, derives an improved PWM via a genetic algorithm that maximizes the area under the receiver operating characteristic (ROC) curve. GAPWM can easily incorporate prior information such as base conservation. We tested our method on two PMWs (Oct4/Sox2 and p53) using three recently published ChIP data sets (human Oct4, mouse Oct4 and human p53). RESULTS GAPWM substantially increased the sensitivity/specificity of a poorly estimated PWM and further improved the quality of a good PWM. Furthermore, it still functioned when the starting PWM contained a major error. The ROC performance of GAPWM compared favorably with that of MEME and others. With increasing availability of ChIP data, our method provides an alternative for obtaining high-quality PWMs for genome-wide identification of transcription factor binding sites. AVAILABILITY The C source code and all data used in this report are available at http://dir.niehs.nih.gov/dirbb/gapwm. SUPPLEMENTARY INFORMATION Supplementary data are available at Bioinformatics online.