Deep Ranking in Template-free Protein Structure Prediction

The road to the discovery of the biological activities of a protein molecule in the cell goes through knowledge of its three-dimensional, biologically-active structure(s). Current evidence suggests significant regions of the protein universe are inaccessible by either wet-laboratory structure determination or homology modeling. While great progress has been made by computational approaches in elucidating dark regions of the proteome, inherent challenges remain. In this paper, we advance research on addressing one such a challenge known as model (quality) assessment. In essence, the task involves discriminating relevant structure(s) among many computed for a protein of interest. We propose a method based on deep learning and evaluate it on tertiary structures computed by a popular de-novo platform on benchmark datasets. The method uses novel protein residue-residue distance features, improved residue-residue contacts, together with other features, such as energies and model topology similarity, to estimate the quality of protein models. A detailed evaluation shows that the proposed method outperforms related ones and advances the state of the art in model assessment.