A novel scoring function to estimate protein-ligand binding affinities has been developed and implemented as the Glide 4.0 XP scoring function and docking protocol. In addition to unique water desolvation energy terms, protein-ligand structural motifs leading to enhanced binding affinity are included: (1) hydrophobic enclosure where groups of lipophilic ligand atoms are enclosed on opposite faces by lipophilic protein atoms, (2) neutral-neutral single or correlated hydrogen bonds in a hydrophobically enclosed environment, and (3) five categories of charged-charged hydrogen bonds. The XP scoring function and docking protocol have been developed to reproduce experimental binding affinities for a set of 198 complexes (RMSDs of 2.26 and 1.73 kcal/mol over all and well-docked ligands, respectively) and to yield quality enrichments for a set of fifteen screens of pharmaceutical importance. Enrichment results demonstrate the importance of the novel XP molecular recognition and water scoring in separating active and inactive ligands and avoiding false positives.

/pdf/extra-precision-glide-docking-and-scoring-incorporating-a-2o3e7pl9pg.pdf

Extra Precision Glide: Docking and Scoring Incorporating a Model of Hydrophobic Enclosure for Protein-Ligand Complexes

The synchronization of chaotic systems

Supervised learning on molecules has incredible potential to be useful in chemistry, drug discovery, and materials science. Luckily, several promising and closely related neural network models invariant to molecular symmetries have already been described in the literature. These models learn a message passing algorithm and aggregation procedure to compute a function of their entire input graph. At this point, the next step is to find a particularly effective variant of this general approach and apply it to chemical prediction benchmarks until we either solve them or reach the limits of the approach. In this paper, we reformulate existing models into a single common framework we call Message Passing Neural Networks (MPNNs) and explore additional novel variations within this framework. Using MPNNs we demonstrate state of the art results on an important molecular property prediction benchmark; these results are strong enough that we believe future work should focus on datasets with larger molecules or more accurate ground truth labels.

Neural Message Passing for Quantum Chemistry

: This thesis applies neural network feature selection techniques to multivariate time series data to improve prediction of a target time series. Two approaches to feature selection are used. First, a subset enumeration method is used to determine which financial indicators are most useful for aiding in prediction of the S&P 500 futures daily price. The candidate indicators evaluated include RSI, Stochastics and several moving averages. Results indicate that the Stochastics and RSI indicators result in better prediction results than the moving averages. The second approach to feature selection is calculation of individual saliency metrics. A new decision boundary-based individual saliency metric, and a classifier independent saliency metric are developed and tested. Ruck's saliency metric, the decision boundary based saliency metric, and the classifier independent saliency metric are compared for a data set consisting of the RSI and Stochastics indicators as well as delayed closing price values. The decision based metric and the Ruck metric results are similar, but the classifier independent metric agrees with neither of the other metrics. The nine most salient features, determined by the decision boundary based metric, are used to train a neural network and the results are presented and compared to other published results. (AN)

/pdf/nonlinear-time-series-analysis-1j3a35n4y3.pdf

Nonlinear Time Series Analysis.

Recently, graph neural networks (GNNs) have revolutionized the field of graph representation learning through effectively learned node embeddings, and achieved state-of-the-art results in tasks such as node classification and link prediction. However, current GNN methods are inherently flat and do not learn hierarchical representations of graphs---a limitation that is especially problematic for the task of graph classification, where the goal is to predict the label associated with an entire graph. Here we propose DiffPool, a differentiable graph pooling module that can generate hierarchical representations of graphs and can be combined with various graph neural network architectures in an end-to-end fashion. DiffPool learns a differentiable soft cluster assignment for nodes at each layer of a deep GNN, mapping nodes to a set of clusters, which then form the coarsened input for the next GNN layer. Our experimental results show that combining existing GNN methods with DiffPool yields an average improvement of 5-10% accuracy on graph classification benchmarks, compared to all existing pooling approaches, achieving a new state-of-the-art on four out of five benchmark data sets.

Hierarchical Graph Representation Learning with Differentiable Pooling

Spatiotemporal time series are analyzed and predicted using reconstructed local states. As numerical examples the evolution of a Kuramoto-Sivashinsky equation and a coupled map lattice are predicted from previously sampled data.

Prediction of spatiotemporal time series based on reconstructed local states

We describe a method for the automatic generation of weakly correlated descriptors for molecular data sets. The method can be regarded as a statistical learning procedure that turns the molecular graph, representing the 2D formula of the compound, into an adaptive whole molecule composite descriptor. By translating the molecular graph structure into a dynamical system, the algorithm can compute an output value that is highly sensitive to the molecular topology. This system can be trained by gradient descent techniques, which rely on the efficient calculation of the gradient by back-propagation. We present computational experiments concerning the classification of the Developmental Therapeutics Program AIDS antiviral screen data set on which the performance of the method compares with that of approaches based on substructure comparison.

Automatic generation of complementary descriptors with molecular graph networks

We describe the application of ensemble methods to binary classification problems on two pharmaceutical compound data sets. Several variants of single and ensembles models of k-nearest neighbors classifiers, support vector machines (SVMs), and single ridge regression models are compared. All methods exhibit robust classification even when more features are given than observations. On two data sets dealing with specific properties of drug-like substances (cytochrome P450 inhibition and “Frequent Hitters”, i.e., unspecific protein inhibition), we achieve classification rates above 90%. We are able to reduce the cross-validated misclassification rate for the Frequent Hitters problem by a factor of 2 compared to previous results obtained for the same data set with different modeling techniques.

/pdf/ensemble-methods-for-classification-in-cheminformatics-2gynvwaxcn.pdf

Ensemble methods for classification in cheminformatics.

A fast algorithm for exact and approximate nearest-neighbor searching is presented that is suitable for tasks encountered in nonlinear signal processing. Empirical benchmarks show that the algorithm's performance depends mainly on the (fractal) dimension D(d) of the data set, which is usually smaller than the dimension D(s) of the vector space in which the data points are embedded. We also compare the running time of our algorithm with those of two previously proposed algorithms for nearest-neighbor searching.

Fast nearest-neighbor searching for nonlinear signal processing

The prediction of transdermal absorption for arbitrary penetrant structures has several important applications in the pharmaceutical industry. We propose a new data-driven, predictive model for skin permeability coefficients k(p) based on an ensemble model using k-nearest-neighbor models and ridge regression. The model was trained and validated with a newly assembled data set containing experimental data and structures for 110 compounds. On the basis of three purely computational descriptors (molecular weight, calculated octanol/water partition coefficient, and solvation free energy), we have developed a model allowing for the reliable, purely computational prediction of skin permeability coefficients. The model is both accurate and robust, as we showed in an extensive validation (correlation coefficient for leave-one-out cross validation: Q = 0.948, mean standard error: 0.2 for log k(p)).

Christian Merkwirth

Papers

Prediction of spatiotemporal time series based on reconstructed local states

Automatic generation of complementary descriptors with molecular graph networks

Ensemble methods for classification in cheminformatics.

Fast nearest-neighbor searching for nonlinear signal processing

A fully computational model for predicting percutaneous drug absorption.