Packpred predicts the functional consequences of point mutations in proteins starting with a 3D structure or model. The underlying algorithm checks for residue environment and packing in making the predictions. While residue depth (http://mspc.bii.a-star.edu.sg) is used to quantitate the environment, a multi-body statistical measure is used to study residue packing. An observed/expected ration was computed for pairs and quadruples of residues at different depths to cluster with a given distance of one another. We tested different neighbourhood distances ranging from 7.0A to 10.5 A in steps of 0.5A. This gives us a measure of preferred neighbourhoods of amino acids, i.e., a parameterization of residue packing.

Optimizing the coefficients of the linear combination

To benchmark our algorithm we used a saturation mutagenesis data set of T4 lysozyme containing 2016 point mutants [Potetee and coworkers, JMB 1991]. A discriminatory threshold score was chosen for each combination such that scores above the threshold were indicative of neutral mutations and scores below the threshold were indicative of deleterious or functionally/structurally compromising mutations. The choice of the threshold was determined by trying to maximize the value of the Matthew's correlation coefficient (MCC)