Investigating 'False Positive' Protein Decoys Based on Core Packing Features

It is well-known that the hydrophobic cores of proteins contribute significantly to their stability. Further, the cores of high-quality experimental protein structures share several important physical features: 1) the cores are densely packed with packing fraction around 0.56, and 2) the interatomic overlap between non-bonded core residues is small. However, many computationally generated decoys do not recapitulate these key features of core packing. We developed a feed-forward neural network based on packing features of cores to predict how well computational models recapitulate real protein structures. While this method achieves high accuracy, there are a number of 'false positive' decoys whose physical features match those expected for experimental structures, but still show low similarity to the true crystal structure. These 'false positive' protein decoys also yield high scores for many other state-of-the-art decoy detection methods. After identifying these 'false positive' decoys, I improved the accuracy of the neural network by including additional features concerning backbone secondary structure. In future studies, we will use the 'false positive' decoys as inputs into molecular dynamics simulations to test their stability in current force fields.