PNNL

Abstract

Residue coupling in protein families is an important indicator for structural and functional conservation. Two residues are coupled if changes of amino acid at one residue location are correlated with changes in the other. Many algorithmic techniques have been proposed to discover couplings in protein families. These approaches discover couplings over amino acid combinations but do not yield mechanistic or other explanations for such couplings. We propose to study couplings in terms of amino acid classes such as polarity, hydrophobicity, size, and reactivity, and present two algorithms for learning probabilistic graphical models of amino acid class-based residue couplings. Our probabilistic graphical models provide a sound basis for predictive, diagnostic, and abductive reasoning. Further, our methods can take optional structural priors into account for building graphical models. The resulting models are useful in assessing the likelihood of a new protein to be a member of a family and for designing new protein sequences by sampling from the graphical model. We apply our approaches to understand couplings in two protein families: Nickel-responsive transcription factors (NikR) and G-protein coupled receptors (GPCRs). The results demonstrate that our graphical models based on sequences, physicochemical properties, and protein structure are capable of detecting amino acid class-based couplings between important residues that play roles in activities of these two families.