A simple biophysical model for protein-DNA binding data

Transcription factors are critical in the regulation of gene expression. Protein binding microarray (PBM) technology has allowed for efficiently probing transcription factors’ DNA binding specificities [4]. Although PBMs are susceptible to missing transient interactions and are less reliable than advanced microfluidic techniques [3], their analysis is a common method for deriving motifs; in recent years, thousands of eukaryotic proteins have been analyzed with PBM assays. [5] PBM data is typically analyzed with algorithms that maximize sequence information content [2], where the output is a position weight matrix (PWM) that defines positionspecific base frequencies. FeatureREDUCE is a novel algorithm that infers specificity directly in terms of relative affinity. The output is a position-specific affinity matrix (PSAM) which represents the differences in binding free energy Kd(S) = exp(∆G/RT ) that occur with point mutations in the DNA sequence. More precisely, a PSAM is a J × 4 numerical matrix of nucleotide affinities, where each column corresponds to a nucleotide and each row corresponds to a position in the binding site j ∈ J .1 Similar to position weight matrices, PSAMs can be visualized with sequence logos. For this study, we analyzed specificities for the helix-loop-helix family of TFs (HLH) from recent PBM data for eukaryotic TFs [5] because proteins with this structure typically bind to the palindromic E-box CACGTG. In order to capture this motif and the surrounding variation, we chose a binding site of length ten, five positions on each side from the center. Using this data from FeatureREDUCE, we attempt to find a model for predicting a TF’s specificities that performs better than the state-of-the-art. Since the nearest It is common to assume that each nucleotide position contributes independently to the overall binding affinity.