Background The modelling of gene regulation is a major challenge in biomedical research. This process is dominated by transcription factors (TFs) and mutations in their binding sites (TFBSs) may cause the misregulation of genes, eventually leading to disease. The consequences of DNA variants on TF binding are modelled in silico using binding matrices, but it remains unclear whether these are capable of accurately representing in vivo binding. In this study, we present a systematic comparison of binding models for 82 human TFs from three freely available sources: JASPAR matrices, HT-SELEX-generated models and matrices derived from protein binding microarrays (PBMs). We determined their ability to detect experimentally verified “real” in vivo TFBSs derived from ENCODE ChIP-seq data. As negative controls we chose random downstream exonic sequences, which are unlikely to harbour TFBS. All models were assessed by receiver operating characteristics (ROC) analysis. Results While the area- under-curve was low for most of the tested models with only 47 % reaching a score of 0.7 or higher, we noticed strong differences between the various position-specific scoring matrices with JASPAR and HT-SELEX models showing higher success rates than PBM-derived models. In addition, we found that while TFBS sequences showed a higher degree of conservation than randomly chosen sequences, there was a high variability between individual TFBSs. Conclusions Our results show that only few of the matrix-based models used to predict potential TFBS are able to reliably detect experimentally confirmed TFBS. We compiled our findings in a freely accessible web application called ePOSSUM (http:/mutationtaster.charite.de/ePOSSUM/) which uses a Bayes classifier to assess the impact of genetic alterations on TF binding in user-defined sequences. Additionally, ePOSSUM provides information on the reliability of the prediction using our test set of experimentally confirmed binding sites.
