Deep convolutional networks for quality assessment of protein folds.

Motivation: The computational prediction of a protein structure from its sequence generally relies on a method to assess the quality of protein models. Most assessment methods rank candidate models using heavily engineered structural features, defined as complex functions of the atomic coordinates. However, very few methods have attempted to learn these features directly from the data. Results: We show that deep convolutional networks can be used to predict the ranking of model structures solely on the basis of their raw three-dimensional atomic densities, without any feature tuning. We develop a deep neural network that performs on par with state-of-the-art algorithms from the literature. The network is trained on decoys from the CASP7 to CASP10 datasets and its performance is tested on the CASP11 dataset. Additional testing on decoys from the CASP12, CAMEO and 3DRobot datasets confirms that the network performs consistently well across a variety of protein structures. While the network learns to assess structural decoys globally and does not rely on any predefined features, it can be analyzed to show that it implicitly identifies regions that deviate from the native structure. Availability and implementation: The code and the datasets are available at https://github.com/lamoureux-lab/3DCNN_MQA. Supplementary information: Supplementary data are available at Bioinformatics online.

Blind prediction of interfacial water positions in CAPRI.

We report the first assessment of blind predictions of water positions at protein-protein interfaces, performed as part of the critical assessment of predicted interactions (CAPRI) community-wide experiment. Groups submitting docking predictions for the complex of the DNase domain of colicin E2 and Im2 immunity protein (CAPRI Target 47), were invited to predict the positions of interfacial water molecules using the method of their choice. The predictions-20 groups submitted a total of 195 models-were assessed by measuring the recall fraction of water-mediated protein contacts. Of the 176 high- or medium-quality docking models-a very good docking performance per se-only 44% had a recall fraction above 0.3, and a mere 6% above 0.5. The actual water positions were in general predicted to an accuracy level no better than 1.5 Å, and even in good models about half of the contacts represented false positives. This notwithstanding, three hotspot interface water positions were quite well predicted, and so was one of the water positions that is believed to stabilize the loop that confers specificity in these complexes. Overall the best interface water predictions was achieved by groups that also produced high-quality docking models, indicating that accurate modelling of the protein portion is a determinant factor. The use of established molecular mechanics force fields, coupled to sampling and optimization procedures also seemed to confer an advantage. Insights gained from this analysis should help improve the prediction of protein-water interactions and their role in stabilizing protein complexes.

HermiteFit: fast-fitting atomic structures into a low-resolution density map using three-dimensional orthogonal Hermite functions.

HermiteFit, a novel algorithm for fitting a protein structure into a low-resolution electron-density map, is presented. The algorithm accelerates the rotation of the Fourier image of the electron density by using three-dimensional orthogonal Hermite functions. As part of the new method, an algorithm for the rotation of the density in the Hermite basis and an algorithm for the conversion of the expansion coefficients into the Fourier basis are presented. HermiteFit was implemented using the cross-correlation or the Laplacian-filtered cross-correlation as the fitting criterion. It is demonstrated that in the Hermite basis the Laplacian filter has a particularly simple form. To assess the quality of density encoding in the Hermite basis, an analytical way of computing the crystallographic R factor is presented. Finally, the algorithm is validated using two examples and its efficiency is compared with two widely used fitting methods, ADP_EM and colores from the Situs package. HermiteFit will be made available at http://nano-d.inrialpes.fr/software/HermiteFit or upon request from the authors.

Dock2D: Synthetic data for the molecular recognition problem.

Predicting the physical interaction of proteins is a cornerstone problem in computational biology. New classes of learning-based algorithms are actively being developed, and are typically trained end-to-end on protein complex structures extracted from the Protein Data Bank. These training datasets tend to be large and difficult to use for prototyping and, unlike image or natural language datasets, they are not easily interpretable by non-experts. We present Dock2D-IP and Dock2DIF, two "toy" datasets that can be used to select algorithms predicting protein-protein interactions-or any other type of molecular interactions. Using two-dimensional shapes as input, each example from Dock2D-IP ("interaction pose") describe the interaction pose of two shapes known to interact and each example from Dock2D-IF ("interaction fact") describes whether two shapes form a stable complex or not, regardless of how they bind. We propose a number of baseline solutions to the problem and show that the same underlying energy function can be learned either by solving the interaction pose task (formulated as an energy-minimization "docking" problem) or the fact-ofinteraction task (formulated as a binding free energy estimation problem).