Prediction and validation of protein–protein interactors from genome-wide DNA-binding data using a knowledge-based machine-learning approach

Abstract

The ability to accurately predict the DNA targets and interacting cofactors of transcriptional regulators from genome-wide data can significantly advance our understanding of gene regulatory networks. NKX2-5 is a homeodomain transcription factor that sits high in the cardiac gene regulatory network and is essential for normal heart development. We previously identified genomic targets for NKX2-5 in mouse HL-1 atrial cardiomyocytes using DNA-adenine methyltransferase identification (DamID). Here, we apply machine learning algorithms and propose a knowledge-based feature selection method for predicting NKX2-5 protein : protein interactions based on motif grammar in genome-wide DNA-binding data. We assessed model performance using leave-one-out cross-validation and a completely independent DamID experiment performed with replicates. In addition to identifying previously described NKX2-5-interacting proteins, including GATA, HAND and TBX family members, a number of novel interactors were identified, with direct protein : protein interactions between NKX2-5 and retinoid X receptor (RXR), paired-related homeobox (PRRX) and Ikaros zinc fingers (IKZF) validated using the yeast two-hybrid assay. We also found that the interaction of RXRα with NKX2-5 mutations found in congenital heart disease (Q187H, R189G and R190H) was altered. These findings highlight an intuitive approach to accessing protein–protein interaction information of transcription factors in DNA-binding experiments.

Keywords

• machine learning
• protein–protein interactions
• transcription factors
• gene regulatory networks