Accounting for structural properties and nucleotide co-variations in the quantitative prediction of binding affinities of protein-DNA interactions.

We describe a quantitative model for predicting the binding affinity of protein-DNA interactions. The described model is based on templates capable of providing a global representation of the modelled transcription factor (TF) binding sites. Templates can capture non independent nucleotide variations and structural properties present in these sites. Tests carried out on the p50p50 and p50p65 variants of the transcription factor NF-kappa B demonstrate a high correlation between the observed binding affinities and the binding affinities predicted by the templates. Only a small subset of training data spanning the space of the binding sites is required to train the templates.