Quantifying sequence readout by transcription factors through principled analysis of high-throughput SELEX data

Transcription factors (TFs) control gene expression by binding to genomic DNA in a sequence-specific manner. In recent years, hundreds of TFs have been characterized using high-throughput in vitro DNA binding assays coupled with deep sequencing. Variations of these assays can characterize binding by TF complexes and by RNA-binding proteins, or binding to chemically modified DNA. However, no unified method for analyzing all these data yet exists. We recently developed an algorithm named No Read Left Behind or NRLB (Rastogi et al., PNAS, 2018), which infers biophysical binding specificity models across the full affinity range from single-round SELEX data. It predicts human MAX homodimer binding in near-perfect agreement with existing low-throughput measurements, captures the specificity of the full-length p53 tetramer, and distinguishes multiple binding modes within a single sample. In addition to the chemical identity of the DNA bases, TF binding affinity is sensitive to the local three-dimensional shape of the DNA double helix. We demonstrate that linear models based on mononucleotide features alone can account for 60–70% of the variance in the DNA shape parameters minor groove width, roll, helix twist, and propeller twist. We also show that NRLB binding models implicitly encode DNA shape readout. Building on these observations, we developed a post hoc analysis method that interprets NRLB models in terms of DNA shape readout (Rube et al., MSB, 2018). Finally, we will discuss our latest algorithm, ProBound, which, unlike NRLB, allows principled modeling of multiple SELEX rounds, chemically modified DNA, and complexes with variable spacing between the DNA binding domains. ProBound works on all currently available data and allows us to build a resource of DNA binding specificity models for hundreds of TFs.