Toward on-the-fly quantum mechanical/molecular mechanical (QM/MM) docking: development and benchmark of a scoring function.

We address the challenges of treating polarization and covalent interactions in docking by developing a hybrid quantum mechanical/molecular mechanical (QM/MM) scoring function based on the semiempirical self-consistent charge density functional tight-binding (SCC-DFTB) method and the CHARMM force field. To benchmark this scoring function within the EADock DSS docking algorithm, we created a publicly available dataset of high-quality X-ray structures of zinc metalloproteins ( http://www.molecular-modelling.ch/resources.php ). For zinc-bound ligands (226 complexes), the QM/MM scoring yielded a substantially improved success rate compared to the classical scoring function (77.0% vs 61.5%), while, for allosteric ligands (55 complexes), the success rate remained constant (49.1%). The QM/MM scoring significantly improved the detection of correct zinc-binding geometries and improved the docking success rate by more than 20% for several important drug targets. The performance of both the classical and the QM/MM scoring functions compare favorably to the performance of AutoDock4, AutoDock4Zn, and AutoDock Vina.

Wurst: a protein threading server with a structural scoring function, sequence profiles and optimized substitution matrices

Wurst is a protein threading program with an emphasis on high quality sequence to structure alignments (http://www.zbh.uni-hamburg.de/wurst). Submitted sequences are aligned to each of about 3000 templates with a conventional dynamic programming algorithm, but using a score function with sophisticated structure and sequence terms. The structure terms are a log-odds probability of sequence to structure fragment compatibility, obtained from a Bayesian classification procedure. A simplex optimization was used to optimize the sequence-based terms for the goal of alignment and model quality and to balance the sequence and structural contributions against each other. Both sequence and structural terms operate with sequence profiles.

Improving drug discovery using a neural networks based parallel scoring function

Virtual Screening (VS) methods can considerably aid clinical research, predicting how ligands interact with drug targets. Most VS methods suppose a unique binding site for the target, but it has been demonstrated that diverse ligands interact with unrelated parts of the target and many VS methods do not take into account this relevant fact. This problem is circumvented by a novel VS methodology named BINDSURF that scans the whole protein surface to find new hotspots, where ligands might potentially interact with, and which is implemented in massively parallel Graphics Processing Units, allowing fast processing of large ligand databases. BINDSURF can thus be used in drug discovery, drug design, drug repurposing and therefore helps considerably in clinical research. However, the accuracy of most VS methods is constrained by limitations in the scoring function that describes biomolecular interactions, and even nowadays these uncertainties are not completely understood. In order to solve this problem, we propose a novel approach where neural networks are trained with databases of known active (drugs) and inactive compounds, and later used to improve VS predictions.

Fast docking using the CHARMM force field with EADock DSS.

The prediction of binding modes (BMs) occurring between a small molecule and a target protein of biological interest has become of great importance for drug development. The overwhelming diversity of needs leaves room for docking approaches addressing specific problems. Nowadays, the universe of docking software ranges from fast and user friendly programs to algorithmically flexible and accurate approaches. EADock2 is an example of the latter. Its multiobjective scoring function was designed around the CHARMM22 force field and the FACTS solvation model. However, the major drawback of such a software design lies in its computational cost. EADock dihedral space sampling (DSS) is built on the most efficient features of EADock2, namely its hybrid sampling engine and multiobjective scoring function. Its performance is equivalent to that of EADock2 for drug-like ligands, while the CPU time required has been reduced by several orders of magnitude. This huge improvement was achieved through a combination of several innovative features including an automatic bias of the sampling toward putative binding sites, and a very efficient tree-based DSS algorithm. When the top-scoring prediction is considered, 57% of BMs of a test set of 251 complexes were reproduced within 2 Å RMSD to the crystal structure. Up to 70% were reproduced when considering the five top scoring predictions. The success rate is lower in cross-docking assays but remains comparable with that of the latest version of AutoDock that accounts for the protein flexibility. © 2011 Wiley Periodicals, Inc. J Comput Chem, 2011.

Docking to heme proteins.

In silico screening has become a valuable tool in drug design, but some drug targets represent real challenges for docking algorithms. This is especially true for metalloproteins, whose interactions with ligands are difficult to parametrize. Our docking algorithm, EADock, is based on the CHARMM force field, which assures a physically sound scoring function and a good transferability to a wide range of systems, but also exhibits difficulties in case of some metalloproteins. Here, we consider the therapeutically important case of heme proteins featuring an iron core at the active site. Using a standard docking protocol, where the iron-ligand interaction is underestimated, we obtained a success rate of 28% for a test set of 50 heme-containing complexes with iron-ligand contact. By introducing Morse-like metal binding potentials (MMBP), which are fitted to reproduce density functional theory calculations, we are able to increase the success rate to 62%. The remaining failures are mainly due to specific ligand-water interactions in the X-ray structures. Testing of the MMBP on a second data set of non iron binders (14 cases) demonstrates that they do not introduce a spurious bias towards metal binding, which suggests that they may reliably be used also for cross-docking studies.

Attracting cavities for docking. Replacing the rough energy landscape of the protein by a smooth attracting landscape.

Molecular docking is a computational approach for predicting the most probable position of ligands in the binding sites of macromolecules and constitutes the cornerstone of structure-based computer-aided drug design. Here, we present a new algorithm called Attracting Cavities that allows molecular docking to be performed by simple energy minimizations only. The approach consists in transiently replacing the rough potential energy hypersurface of the protein by a smooth attracting potential driving the ligands into protein cavities. The actual protein energy landscape is reintroduced in a second step to refine the ligand position. The scoring function of Attracting Cavities is based on the CHARMM force field and the FACTS solvation model. The approach was tested on the 85 experimental ligand-protein structures included in the Astex diverse set and achieved a success rate of 80% in reproducing the experimental binding mode starting from a completely randomized ligand conformer. The algorithm thus compares favorably with current state-of-the-art docking programs. © 2015 The Authors. Journal of Computational Chemistry Published by Wiley Periodicals, Inc.

Prediction of novel and analogous folds using fragment assembly and fold recognition

A number of new and newly improved methods for predicting protein structure developed by the Jones–University College London group were used to make predictions for the CASP6 experiment. Structures were predicted with a combination of fold recognition methods (mGenTHREADER, nFOLD, and THREADER) and a substantially enhanced version of FRAGFOLD, our fragment assembly method. Attempts at automatic domain parsing were made using DomPred and DomSSEA, which are based on a secondary structure parsing algorithm and additionally for DomPred, a simple local sequence alignment scoring function. Disorder prediction was carried out using a new SVM-based version of DISOPRED. Attempts were also made at domain docking and “microdomain” folding in order to build complete chain models for some targets.

Mining for Class-Specific Motifs in Protein Sequence Classification

Background: In protein sequence classification, identification of the sequence motifs or n-grams that can precisely discriminate between classes is a more interesting scientific question than the classification itself. A number of classification methods aim at accurate classification but fail to explain which sequence features indeed contribute to the accuracy. We hypothesize that sequences in lower denominations (n-grams) can be used to explore the sequence landscape and to identify class-specific motifs that discriminate between classes during classification. Discriminative n-grams are short peptide sequences that are highly frequent in one class but are either minimally present or absent in other classes. In this study, we present a new substitution-based scoring function for identifying discriminative n-grams that are highly specific to a class. Results: We present a scoring function based on discriminative n-grams that can effectively discriminate between classes. The scoring function, initially, harvests the entire set of 4- to 8-grams from the protein sequences of different classes in the dataset. Similar n-grams of the same size are combined to form new n-grams, where the similarity is defined by positive amino acid substitution scores in the BLOSUM62 matrix. Substitution has resulted in a large increase in the number of discriminatory n-grams harvested. Due to the unbalanced nature of the dataset, the frequencies of the n-grams are normalized using a dampening factor, which gives more weightage to the n-grams that appear in fewer classes and vice-versa. After the n-grams are normalized, the scoring function identifies discriminative 4- to 8-grams for each class that are frequent enough to be above a selection threshold. By mapping these discriminative n-grams back to the protein sequences, we obtained contiguous n-grams that represent short class-specific motifs in protein sequences. Our method fared well compared to an existing motif finding method known as Wordspy. We have validated our enriched set of class-specific motifs against the functionally important motifs obtained from the NLSdb, Prosite and ELM databases. We demonstrate that this method is very generic; thus can be widely applied to detect class-specific motifs in many protein sequence classification tasks. Conclusion: The proposed scoring function and methodology is able to identify class-specific motifs using discriminative n-grams derived from the protein sequences. The implementation of amino acid substitution scores for similarity detection, and the dampening factor to normalize the unbalanced datasets have significant effect on the performance of the scoring function. Our multipronged validation tests demonstrate that this method can detect class-specific motifs from a wide variety of protein sequence classes with a potential application to detecting proteome-specific motifs of different organisms.

Learning Bayesian networks from data by the incremental compilation of new network polynomials

Probabilistic graphical models are a huge research field in artificial intelligence nowadays. The scope of this work is the study of directed graphical models for the representation of discrete distributions. Two of the main research topics related to this area focus on performing inference over graphical models and on learning graphical models from data. Traditionally, the inference process and the learning process have been treated separately, but given that the learned models structure marks the inference complexity, this kind of strategies will sometimes produce very inefficient models. With the purpose of learning thinner models, in this master thesis we propose a new model for the representation of network polynomials, which we call polynomial trees. Polynomial trees are a complementary representation for Bayesian networks that allows an efficient evaluation of the inference complexity and provides a framework for exact inference. We also propose a set of methods for the incremental compilation of polynomial trees and an algorithm for learning polynomial trees from data using a greedy score+search method that includes the inference complexity as a penalization in the scoring function.

Higher order elicitability and Osband’s principle

A statistical functional, such as the mean or the median, is called elicitable if there is a scoring function or loss function such that the correct forecast of the functional is the unique minimizer of the expected score. Such scoring functions are called strictly consistent for the functional. The elicitability of a functional opens the possibility to compare competing forecasts and to rank them in terms of their realized scores. In this paper, we explore the notion of elicitability for multi-dimensional functionals and give both necessary and sufficient conditions for strictly consistent scoring functions. We cover the case of functionals with elicitable components, but we also show that one-dimensional functionals that are not elicitable can be a component of a higher order elicitable functional. In the case of the variance, this is a known result. However, an important result of this paper is that spectral risk measures with a spectral measure with finite support are jointly elicitable if one adds the “correct” quantiles. A direct consequence of applied interest is that the pair (Value at Risk, Expected Shortfall) is jointly elicitable under mild conditions that are usually fulfilled in risk management applications.

Prediction of HLA-DQ3.2 beta ligands: evidence of multiple registers in class II binding peptides

Motivation: While processing of MHC class II antigens for presentation to helper T-cells is essential for normal immune response, it is also implicated in the pathogenesis of autoimmune disorders and hypersensitivity reactions. Sequence-based computational techniques for predicting HLA-DQ binding peptides have encountered limited success, with few prediction techniques developed using three-dimensional models. Methods: We describe a structure-based prediction model for modeling peptide-DQ3.2 beta complexes. We have developed a rapid and accurate protocol for docking candidate peptides into the DQ3.2 beta receptor and a scoring function to discriminate binders from the background. The scoring function was rigorously trained, tested and validated using experimentally verified DQ3.2 beta binding and non-binding peptides obtained from biochemical and functional studies. Results: Our model predicts DQ3.2 beta binding peptides with high accuracy [area under the receiver operating characteristic (ROC) curve A(ROC) > 0.90], compared with experimental data. We investigated the binding patterns of DQ3.2 beta peptides and illustrate that several registers exist within a candidate binding peptide. Further analysis reveals that peptides with multiple registers occur predominantly for high-affinity binders.

Transporter associated with antigen processing preselection of peptides binding to the MHC:a bioinformatic evaluation

TAP is responsible for the transit of peptides from the cytosol to the lumen of the endoplasmic reticulum. In an immunological context, this event is followed by the binding of peptides to MHC molecules before export to the cell surface and recognition by T cells. Because TAP transport precedes MHC binding, TAP preferences may make a significant contribution to epitope selection. To assess the impact of this preselection, we have developed a scoring function for TAP affinity prediction using the additive method, have used it to analyze and extend the TAP binding motif, and have evaluated how well this model acts as a preselection step in predicting MHC binding peptides. To distinguish between MHC alleles that are exclusively dependent on TAP and those exhibiting only a partial dependence on TAP, two sets of MHC binding peptides were examined: HLA-A*0201 was selected as a representative of partially TAP-dependent HLA alleles, and HLA-A*0301 represented fully TAP-dependent HLA alleles. TAP preselection has a greater impact on TAP-dependent alleles than on TAP-independent alleles. The reduction in the number of nonbinders varied from 10% (TAP-independent) to 33% (TAP-dependent), suggesting that TAP preselection is an important component in the successful in silico prediction of T cell epitopes.

Recomendr-entity recommendation based on ad-hoc dimensions

The growing availability and popularity of opinion rich resources on the online web resources, such as review sites and personal blogs, has made it convenient to find out about the opinions and experiences of layman people. But, simultaneously, this huge eruption of data has made it difficult to reach to a conclusion. In this thesis, I develop a novel recommendation system, Recomendr that can help users digest all the reviews about an entity and compare candidate entities based on ad-hoc dimensions specified by keywords. It expects keyword specified ad-hoc dimensions/features as input from the user and based on those features; it compares the selected range of entities using reviews provided on the related User Generated Contents (UGC) e.g. online reviews. It then rates the textual stream of data using a scoring function and returns the decision based on an aggregate opinion to the user. Evaluation of Recomendr using a data set in the laptop domain shows that it can effectively recommend the best laptop as per user-specified dimensions such as price. Recomendr is a general system that can potentially work for any entities on which online reviews or opinionated text is available.

Improving drug discovery using hybrid softcomputing methods

Virtual screening (VS) methods can considerably aid clinical research, predicting how ligands interact with drug targets. Most VS methods suppose a unique binding site for the target, but it has been demonstrated that diverse ligands interact with unrelated parts of the target and many VS methods do not take into account this relevant fact. This problem is circumvented by a novel VS methodology named BINDSURF that scans the whole protein surface in order to find new hotspots, where ligands might potentially interact with, and which is implemented in last generation massively parallel GPU hardware, allowing fast processing of large ligand databases. BINDSURF can thus be used in drug discovery, drug design, drug repurposing and therefore helps considerably in clinical research. However, the accuracy of most VS methods and concretely BINDSURF is constrained by limitations in the scoring function that describes biomolecular interactions, and even nowadays these uncertainties are not completely understood. In order to improve accuracy of the scoring functions used in BINDSURF we propose a hybrid novel approach where neural networks (NNET) and support vector machines (SVM) methods are trained with databases of known active (drugs) and inactive compounds, being this information exploited afterwards to improve BINDSURF VS predictions.

Improvement of Virtual Screening Predictions using Computational Intelligence Methods

Virtual Screening (VS) methods can considerably aid clinical research, predicting how ligands interact with drug targets. However, the accuracy of most VS methods is constrained by limitations in the scoring function that describes biomolecular interactions, and even nowadays these uncertainties are not completely understood. In order to improve accuracy of scoring functions used in most VS methods we propose a hybrid novel approach where neural networks (NNET) and support vector machines (SVM) methods are trained with databases of known active (drugs) and inactive compounds, this information being exploited afterwards to improve VS predictions.