A systematic, large-scale comparison of transcription factor binding site models

Area of Science:

• Genomics and Bioinformatics
• Molecular Biology
• Computational Biology

Background:

• Gene regulation is crucial in biomedical research, with transcription factors (TFs) playing a dominant role.
• Mutations in TF binding sites (TFBSs) can lead to gene misregulation and disease.
• In silico models using binding matrices predict TF binding, but their in vivo accuracy is uncertain.

Purpose of the Study:

• To systematically compare the performance of different in silico binding models for 82 human TFs.
• To assess the ability of these models to detect experimentally verified in vivo TFBSs.
• To evaluate the reliability of current matrix-based models for predicting TF binding.

Main Methods:

• Compared binding models from JASPAR, HT-SELEX, and protein binding microarrays (PBMs) for 82 human TFs.
• Utilized experimentally verified in vivo TFBSs from ENCODE ChIP-seq data as positive examples.
• Employed random downstream exonic sequences as negative controls and assessed models using receiver operating characteristic (ROC) analysis.

Main Results:

• Most models showed low predictive power, with only 47% achieving a score of 0.7 or higher.
• JASPAR and HT-SELEX models demonstrated higher success rates in detecting TFBSs compared to PBM-derived models.
• While TFBS sequences were more conserved than random sequences, significant variability existed among individual TFBSs.

Conclusions:

• Few matrix-based models reliably detect experimentally confirmed TFBSs.
• Developed ePOSSUM, a web application using a Bayes classifier to assess genetic alteration impact on TF binding.
• ePOSSUM provides prediction reliability information based on a validated set of binding sites.