A support vector machines approach for virtual screening of active compounds of single and multiple mechanisms from large libraries at an improved hit-rate and enrichment factor

Support vector machines (SVM) and other machine-learning (ML) methods have been explored as ligand-based virtual screening (VS) tools for facilitating lead discovery. While exhibiting good hit selection performance, in screening large compound libraries, these methods tend to produce lower hit-rate than those of the best performing VS tools, partly because their training-sets contain limited spectrum of inactive compounds. We tested whether the performance of SVM can be improved by using training-sets of diverse inactive compounds. In retrospective database screening of active compounds of single mechanism (HIV protease inhibitors, DHFR inhibitors, dopamine antagonists) and multiple mechanisms (CNS active agents) from large libraries of 2.986 million compounds, the yields, hit-rates, and enrichment factors of our SVM models are 52.4–78.0%, 4.7–73.8%, and 214–10,543, respectively, compared to those of 62–95%, 0.65–35%, and 20–1200 by structure-based VS and 55–81%, 0.2–0.7%, and 110–795 by other ligand-based VS tools in screening libraries of ≥1 million compounds. The hit-rates are comparable and the enrichment factors are substantially better than the best results of other VS tools. 24.3–87.6% of the predicted hits are outside the known hit families. SVM appears to be potentially useful for facilitating lead discovery in VS of large compound libraries.     

Exploration of DPP-IV inhibitors with a novel scaffold by multistep in silico screening

Dipeptidyl peptidase-IV (DPP-IV), an enzyme that degrades incretins–hormones that promote insulin secretion–is a therapeutic target for type 2 diabetes, with a number of its inhibitors having been launched as therapies for diabetes. Since adverse effects of these inhibitors have recently been reported, the development of novel DPP-IV inhibitors with higher efficacy and safety is required. We, therefore, screened for novel DPP-IV inhibitors using the combination of an in silico drug discovery technique and a DPP-IV assay system. We initially selected seven candidate compounds as DPP-IV inhibitors from a database consisting of four million compounds by a multistep in silico screening procedure combining pharmacophore-based screening, docking calculation and the analysis of three-dimensional quantitative structure-activity relationship. We then measured the inhibitory activity of the selected compounds and identified a hit compound. In addition, we discuss the structure–activity relationship between the binding mode model and inhibitory activity of the hit compound.     

Protein tyrosine phosphatases: Ligand interaction analysis and optimisation of virtual screening

Docking-based virtual screening is an established component of structure-based drug discovery. Nevertheless, scoring and ranking of computationally docked ligand libraries still suffer from many false positives. Identifying optimal docking parameters for a target protein prior to virtual screening can improve experimental hit rates. Here, we examine protocols for virtual screening against the important but challenging class of drug target, protein tyrosine phosphatases. In this study, common interaction features were identified from analysis of protein–ligand binding geometries of more than 50 complexed phosphatase crystal structures. It was found that two interactions were consistently formed across all phosphatase inhibitors: (1) a polar contact with the conserved arginine residue, and (2) at least one interaction with the P-loop backbone amide. In order to investigate the significance of these features on phosphatase-ligand binding, a series of seeded virtual screening experiments were conducted on three phosphatase enzymes, PTP1B, Cdc25b and IF2. It was observed that when the conserved arginine and P-loop amide interactions were used as pharmacophoric constraints during docking, enrichment of the virtual screen significantly increased in the three studied phosphatases, by up to a factor of two in some cases. Additionally, the use of such pharmacophoric constraints considerably improved the ability of docking to predict the inhibitor's bound pose, decreasing RMSD to the crystallographic geometry by 43% on average. Constrained docking improved enrichment of screens against both open and closed conformations of PTP1B. Incorporation of an ordered water molecule in PTP1B screening was also found to generally improve enrichment. The knowledge-based computational strategies explored here can potentially inform structure-based design of new phosphatase inhibitors using docking-based virtual screening.     

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.     

Identification of estrogen receptor α ligands with virtual screening techniques

Utilization of computer-aided molecular discovery methods in virtual screening (VS) is a cost-effective approach to identify novel bioactive small molecules. Unfortunately, no universal VS strategy can guarantee high hit rates for all biological targets, but each target requires distinct, fine-tuned solutions. Here, we have studied in retrospective manner the effectiveness and usefulness of common pharmacophore hypothesis, molecular docking and negative image-based screening as potential VS tools for a widely applied drug discovery target, estrogen receptor α (ERα). The comparison of the methods helps to demonstrate the differences in their ability to identify active molecules. For example, structure-based methods identified an already known active ligand from the widely-used bechmarking decoy molecule set. Although prospective VS against one commercially available database with around 100,000 drug-like molecules did not retrieve many testworthy hits, one novel hit molecule with pIC₅₀ value of 6.6, was identified. Furthermore, our small in-house compound collection of easy-to-synthesize molecules was virtually screened against ERα, yielding to five hit candidates, which were found to be active in vitro having pIC₅₀ values from 5.5 to 6.5.     

Drug/nondrug classification using Support Vector Machines with various feature selection strategies

In conjunction with the advance in computer technology, virtual screening of small molecules has been started to use in drug discovery. Since there are thousands of compounds in early-phase of drug discovery, a fast classification method, which can distinguish between active and inactive molecules, can be used for screening large compound collections. In this study, we used Support Vector Machines (SVM) for this type of classification task. SVM is a powerful classification tool that is becoming increasingly popular in various machine-learning applications. The data sets consist of 631 compounds for training set and 216 compounds for a separate test set. In data pre-processing step, the Pearson's correlation coefficient used as a filter to eliminate redundant features. After application of the correlation filter, a single SVM has been applied to this reduced data set. Moreover, we have investigated the performance of SVM with different feature selection strategies, including SVM–Recursive Feature Elimination, Wrapper Method and Subset Selection. All feature selection methods generally represent better performance than a single SVM while Subset Selection outperforms other feature selection methods. We have tested SVM as a classification tool in a real-life drug discovery problem and our results revealed that it could be a useful method for classification task in early-phase of drug discovery.     

Towards more accurate pharmacophore modeling: Multicomplex-based comprehensive pharmacophore map and most-frequent-feature pharmacophore model of CDK2

Pharmacophore modeling, including ligand- and structure-based approaches, has become an important tool in drug discovery. However, the ligand-based method often strongly depends on the training set selection, and the structure-based pharmacophore model is usually created based on apo structures or a single protein-ligand complex, which might miss some important information. In this study, multicomplex-based method has been suggested to generate a comprehensive pharmacophore map of cyclin-dependent kinase 2 (CDK2) based on a collection of 124 crystal structures of human CDK2-inhibitor complex. Our multicomplex-based comprehensive pharmacophore map contains almost all the chemical features important for CDK2-inhibitor interactions. A comparison with previously reported ligand-based pharmacophores has revealed that the ligand-based models are just a subset of our comprehensive map. Furthermore, one most-frequent-feature pharmacophore model consisting of the most frequent pharmacophore features was constructed based on the statistical frequency information provided by the comprehensive map. Validations to the most-frequent-feature model show that it can not only successfully discriminate between known CDK2 inhibitors and the molecules of focused inactive dataset, but also is capable of correctly predicting the activities of a wide variety of CDK2 inhibitors in an external active dataset. Obviously, this investigation provides some new ideas about how to develop a multicomplex-based pharmacophore model that can be used in virtual screening to discover novel potential lead compounds.     

Discovery of high affinity ligands for β2-adrenergic receptor through pharmacophore-based high-throughput virtual screening and docking

Novel high affinity compounds for human β₂-adrenergic receptor (β₂-AR) were searched among the clean drug-like subset of ZINC database consisting of 9,928,465 molecules that satisfy the Lipinski's rule of five. The screening protocol consisted of a high-throughput pharmacophore screening followed by an extensive amount of docking and rescoring. The pharmacophore model was composed of key features shared by all five inactive states of β₂-AR in complex with inverse agonists and antagonists. To test the discriminatory power of the pharmacophore model, a small-scale screening was initially performed on a database consisting of 117 compounds of which 53 antagonists were taken as active inhibitors and 64 agonists as inactive inhibitors. Accordingly, 7.3% of the ZINC database subset (729,413 compounds) satisfied the pharmacophore requirements, along with 44 antagonists and 17 agonists. Afterwards, all these hit compounds were docked to the inactive apo form of the receptor using various docking and scoring protocols. Following each docking experiment, the best pose was further evaluated based on the existence of key residues for antagonist binding in its vicinity. After final evaluations based on the human intestinal absorption (HIA) and the blood brain barrier (BBB) penetration properties, 62 hit compounds have been clustered based on their structural similarity and as a result four scaffolds were revealed. Two of these scaffolds were also observed in three high affinity compounds with experimentally known K_i values. Moreover, novel chemical compounds with distinct structures have been determined as potential β₂-AR drug candidates.     

PKRank: a novel learning-to-rank method for ligand-based virtual screening using pairwise kernel and RankSVM

The development of a new drug takes over 10 years and costs approximately US $2.6 billion. Virtual compound screening (VS) is a part of efforts to reduce this cost. Learning-to-rank is a machine learning technique in information retrieval that was recently introduced to VS. It works well because the application of VS requires the ranking of compounds. Moreover, learning-to-rank can treat multiple heterogeneous experimental data because it is trained using only the order of activity of compounds. In this study, we propose PKRank, a novel learning-to-rank method for ligand-based VS that uses a pairwise kernel and RankSVM. PKRank is a general case of the method proposed by Zhang et al. with the advantage of extensibility in terms of kernel selection. In comparisons of predictive accuracy, PKRank yielded a more accurate model than the previous method.     

Ensemble docking-based virtual screening yields novel spirocyclic JAK1 inhibitors

Small molecule inhibition of Janus kinases (JAKs) has been demonstrated as a viable strategy for the treatment of various inflammatory conditions and continues to emerge in cancer-related indications. In this study, a large supplier database was screened to identify novel chemistry starting points for JAK1. The docking-based screening was followed up by testing ten hit compounds experimentally, out of which five have displayed single-digit micromolar and submicromolar IC₅₀ values on JAK1. Thus, the study was concluded with the discovery of five novel JAK inhibitors from a tiny screening deck with a remarkable hitrate of 50%. The results have highlighted spirocyclic pyrrolopyrimidines with submicromolar JAK1 IC₅₀ values and a preference for JAK1 over JAK2 as potential starting points in developing a novel class of JAK1 inhibitors.     

Strategies for generating less toxic P-selectin inhibitors: pharmacophore modeling, virtual screening and counter pharmacophore screening to remove toxic hits

This paper describes the generation of ligand-based as well as structure-based models and virtual screening of less toxic P-selectin receptor inhibitors. Ligand-based model, 3D-pharmacophore was generated using 27 quinoline salicylic acid compounds and is used to retrieve the actives of P-selectin. This model contains three hydrogen bond acceptors (HBA), two ring aromatics (RA) and one hydrophobic feature (HY). To remove the toxic hits from the screened molecules, a counter pharmacophore model was generated using inhibitors of dihydrooratate dehydrogenase (DHOD), an important enzyme involved in nucleic acid synthesis, whose inhibition leads to toxic effects. Structure-based models were generated by docking and scoring of inhibitors against P-selectin receptor, to remove the false positives committed by pharmacophore screening. The combination of these ligand-based and structure-based virtual screening models were used to screen a commercial database containing 538,000 compounds.     

Automated Compound Storage and Retrieval System for Microplates and Tubes

As compound libraries continue their exponential growth, assay systems become increasingly complex, and high throughput screening (HTS) technology picks up speed to become ultra-HTS, the need to store more samples and more discrete compound collections in labware of varied dimension, shape, and function, represents a problem of heightening significance for drug discovery laboratories. Adequate storage space is merely the tip of the iceberg. The main challenge comes in finding suitable space- and cost-efficient storage facilities that provide the flexibility, automation, environmental control, and necessary computerization and compound registration and tracking systems that will ensure the utility, security, and integrity of a laboratory's compound and reagent stores. With sample libraries commonly exceeding one million compounds today and companies screening their inventories against an expanding range of targets, the capability for rapid and reliable compound storage and retrieval, cataloging and monitoring, and fast and accurate cherry-picking becomes essential.     

Highly predictive support vector machine (SVM) models for anthrax toxin lethal factor (LF) inhibitors

Anthrax is a highly lethal, acute infectious disease caused by the rod-shaped, Gram-positive bacterium Bacillus anthracis. The anthrax toxin lethal factor (LF), a zinc metalloprotease secreted by the bacilli, plays a key role in anthrax pathogenesis and is chiefly responsible for anthrax-related toxemia and host death, partly via inactivation of mitogen-activated protein kinase kinase (MAPKK) enzymes and consequent disruption of key cellular signaling pathways. Antibiotics such as fluoroquinolones are capable of clearing the bacilli but have no effect on LF-mediated toxemia; LF itself therefore remains the preferred target for toxin inactivation. However, currently no LF inhibitor is available on the market as a therapeutic, partly due to the insufficiency of existing LF inhibitor scaffolds in terms of efficacy, selectivity, and toxicity. In the current work, we present novel support vector machine (SVM) models with high prediction accuracy that are designed to rapidly identify potential novel, structurally diverse LF inhibitor chemical matter from compound libraries. These SVM models were trained and validated using 508 compounds with published LF biological activity data and 847 inactive compounds deposited in the Pub Chem BioAssay database. One model, M1, demonstrated particularly favorable selectivity toward highly active compounds by correctly predicting 39 (95.12%) out of 41 nanomolar-level LF inhibitors, 46 (93.88%) out of 49 inactives, and 844 (99.65%) out of 847 Pub Chem inactives in external, unbiased test sets. These models are expected to facilitate the prediction of LF inhibitory activity for existing molecules, as well as identification of novel potential LF inhibitors from large datasets.     

Multi-step virtual screening to develop selective DYRK1A inhibitors

Developing selective inhibitors for a particular kinase remains a major challenge in kinase-targeted drug discovery. Here we performed a multi-step virtual screening for dual-specificity tyrosine-phosphorylation-regulated kinase 1A (DYRK1A) inhibitors by focusing on the selectivity for DYRK1A over cyclin-dependent kinase 5 (CDK5). To examine the key factors contributing to the selectivity, we constructed logistic regression models to discriminate between actives and inactives for DYRK1A and CDK5, respectively, using residue-based binding free energies. The residue-based parameters were calculated by molecular mechanics-generalized Born surface area (MM-GBSA) decomposition methods for kinase–ligand complexes modeled by computer ligand docking. Based on the findings from the logistic regression models, we built a three-dimensional (3D) pharmacophore model and chose filter criteria for the multi-step virtual screening. The virtual hit compounds obtained from the screening were assessed for their inhibitory activities against DYRK1A and CDK5 by in vitro assay. Our screening identified two novel selective DYRK1A inhibitors with IC₅₀ values of several μM for DYRK1A and >100 μM for CDK5, which can be further optimized to develop more potent selective DYRK1A inhibitors.     

Implementation and development of an automated, ultra-high-capacity, acoustic, flexible dispensing platform for assay-ready plate delivery

Compound management faces the daily challenge of providing high-quality samples to drug discovery. The advent of new screening technologies has seen demand for liquid samples move toward nanoliter ranges, dispensed by contactless acoustic droplet ejection. Within AstraZeneca, a totally integrated assay-ready plate production platform has been created to fully exploit the advantages of this technology. This enables compound management to efficiently deliver large throughputs demanded by high-throughput screening while maintaining regular delivery of smaller numbers of compounds in varying plate formats for cellular or biochemical concentration-response curves in support of hit and lead optimization (structure-activity relationship screening). The automation solution, CODA, has the capability to deliver compounds on demand for single- and multiple-concentration ranges, in batch sizes ranging from 1 sample to 2 million samples, integrating seamlessly into local compound and test management systems. The software handles compound orders intelligently, grouping test requests together dependent on output plate type and serial dilution ranges so that source compound vessels are shared among numerous tests, ensuring conservation of sample, reduced labware and costs, and efficiency of work cell logistics. We describe the development of CODA to address the customer demand, challenges experienced, learning made, and subsequent enhancements.