Proteins are biomolecules of life. They are composed of one or more chains of amino acids that fold into intricate 3D shapes. The shape of a protein renders it its biological function. The 3D conformation that any protein folds into is a function of the observed backbone and sidechain dihedral angles of its amino acids. Thus, to understand amino acid conformations it becomes necessary to appeal to those dihedral angles and model them in some unified way.
In our recently published research, we have inferred a novel set of statistical mixture models, φΨχal (PhiSiCal), that explain the joint probability distributions of backbone and sidechain dihedral angles [1]. These φΨχal mixture models are derived using the Bayesian and (Shannon) information-theoretic criterion of Minimum Message Length.
The inferred φΨχal models were comprehensively compared with the current state-of-the-art Dunbrack’s backbone-dependant rotamer libraries. Our results clearly demonstrate their accuracy and versatility. Specifically, φΨχal statistical models outperform Dunbrack’s rotamer library both in terms of model complexity (by three orders of magnitude) and its fidelity to explain the observed dihedral angle data (yielding on average 20% more lossless compression) across varying experimental resolutions of structures [1]. Importantly, these statistical models are computationally inexpensive to probabilistically sample from and are geared to support a number of studies, ranging from experimental protein structure validation, refinement, de novo protein design, and protein structure prediction. These applications are currently being actively pursued.