Protein Surface Comparison Algorithm for Predicting DNA Binding Proteins - Using Estrogen Response Element (ERE) as a Model

In the post-genome era, the structures of protein domains have been published rapidly. Researches have also gained diversity of interests from the structure data, yet only few have concentrated on protein tertiary structure alignment while alignments focused on protein surfaces are even rarer. However, the surface structure of a protein binding motif should be deemed an important factor to determine the binding strength and resulting biological consequences. DNA-binding proteins include transcription factors which bind to one particular set of DNA sequences via their DNA binding domain and such transcription factor binding could activate or inhibit the transcription of genes that have these sequences close to their promoters. Thus, it is reasonable to assume that proteins that bind to the same DNA sequence are likely to have similar binding domains. In this paper, we developed a protein surface distillation and comparison method to predict the DNA binding proteins which have similar or dis-similar binding domains as compared to the positive and negative controls. We employed the Voronoi diagram and Delaunay triangulation techniques to distill the amino acids which locate on the protein surface. The distilled amino acids forms several triangle plans to represent the surface entities. Also, two matrices Amino Acid Substitution Matrix and Metric SSE Exchange Matrix were combined with RASA (relative solvent accessible area) to be further applied to identify surface features. Finally, a surface comparison method is applied to compare whether there are high structure similarity triangle planes, and to assess if the two proteins exist high resembling surfaces. We used the estrogen response element (ERE) which has a consensus sequence of GGTCAGAGTGACC as a model and 7 binding transcription factors deduced from TRANSFAC including estrogen receptor  (ER) as the positive control and another 7 proteins which were discriminated by TRANSFAC as the negative control to train the system we developed. Experimental data showed that there existed more than 2 times discrepancy of surface information between ERE positive and non-ERE negative binding proteins. Furthermore, as we concentrated on comparing eight major types of proteins prone to bind with ERE, we discovered there was also more than 2 times difference between these two groups. We then used the trained algorithm to quest for ERE binding proteins among 195 proteins in MCF-7 which structure could be deduced from protein databank (PDB). The classifier built with surface information increased 5% of average precision and 5.6% of average recall. Furthermore, we proved that six out of seven predicted ERE binding proteins could be identified from proteomics experiments via a pull-down assay using an affinity ERE probe (Mol. Cell. Proteomics 2009, DOI 10.1074/mcp.M900183-MCP200).