High-throughput Screening Data Sets Research Articles

GPU Accelerated Support Vector Machines for Mining High-Throughput Screening Data

Support Vector Machine (SVM), one of the most promising tools in chemical informatics, is time-consuming for mining large high-throughput screening (HTS) data sets. Here, we describe a parallelization of SVM-light algorithm on a graphic processor unit (GPU), using molecular fingerprints as descriptors and the Tanimoto index as kernel function. Comparison experiments based on six PubChem Bioassay data sets show that the GPU version is 43-104x faster than SVM-light for building classification models and 112-212x over SVM-light for building regression models.

Systematic Extraction of Structure–Activity Relationship Information from Biological Screening Data

A data mining approach is introduced that automatically extracts SAR information from high-throughput screening data sets and that helps to select active compounds for chemical exploration and hit-to-lead projects. SAR pathways are systematically identified consisting of sequences of similar active compounds with gradual increases in potency. Fully enumerated SAR pathway sets are subjected to pathway scoring, filtering, and mining, and pathways with the most significant SAR information content are prioritized. High-scoring SAR pathways often reveal activity cliffs contained in screening data. Subsets of SAR pathways are analyzed in SAR trees that make it possible to identify microenvironments of significant SAR discontinuity from which hits are preferentially selected. SAR trees of alternative pathways leading to activity cliffs identify key compounds and help to develop chemically intuitive SAR hypotheses.

Elucidation of Structure−Activity Relationship Pathways in Biological Screening Data

A computational molecular network analysis of various high-throughput screening (HTS) data sets including inhibition assays and cell-based screens organizes screening hits according to different local structure-activity relationships (SARs). The resulting network representations make it possible to focus on different local SAR environments in screening data. We have designed a simple scoring function accounting for similarity and potency relationships among hits that identifies SAR pathways leading from active compounds in different SAR contexts to key compounds forming activity cliffs. From these pathways, SAR information can be extracted and utilized to select hits for further analysis. In clusters of hits related by different local SARs, alternative pathways can be systematically explored and ranked according to SAR information content, which makes it possible to prioritize hits in a consistent manner.

Ligand-Target Interaction-Based Weighting of Substructures for Virtual Screening

A methodology is introduced to assign energy-based scores to two-dimensional (2D) structural features based on three-dimensional (3D) ligand-target interaction information and utilize interaction-annotated features in virtual screening. Database molecules containing such fragments are assigned cumulative scores that serve as a measure of similarity to active reference compounds. The Interaction Annotated Structural Features (IASF) method is applied to mine five high-throughput screening (HTS) data sets and often identifies more hits than conventional fragment-based similarity searching or ligand-protein docking.

Formal Concept Analysis for the Identification of Molecular Fragment Combinations Specific for Active and Highly Potent Compounds

We introduce fragment formal concept analysis (FragFCA) to study complex relationships between fragments in active compounds taking potency information into account. Fragment combinations that are unique to active or highly potent compounds or that are shared by molecules having different or overlapping activity profiles are systematically identified using chemically intuitive queries of varying complexity. The methodology is applied to analyze fragment distributions in antagonists of seven G protein coupled receptor targets and identify signature fragments. Pairs or triplets of molecular fragments are found to be most specific for different activity profiles and compound potency levels. In addition, we demonstrate the ability of FragFCA to identify selective hits in high-throughput screening data sets.

Evolving Interpretable Structure−Activity Relationship Models. 2. Using Multiobjective Optimization To Derive Multiple Models

A multiobjective evolutionary algorithm (MOEA) is described for evolving multiple structure-activity relationships (SARs). The SARs are encoded in easy-to-interpret reduced graph queries which describe features that are preferentially present in active compounds compared to inactives. The MOEA addresses a limitation associated with many machine learning methods; that is, the inherent tradeoff that exists in recall and precision which is usually handled by combining the two objectives into a single measure with a consequent loss of control. By simultaneously optimizing recall and precision, the MOEA generates a family of SARs that lie on the precision-recall (PR) curve. The user is then able to select a query with an appropriate balance in the two objectives: for example, a low recall-high precision query may be preferred when establishing the SAR, whereas a high recall-low precision query may be more appropriate in a virtual screening context. Each query on the PR curve aims at capturing the structure-activity information into a single representation, and each can be considered as an alternative (equally valid) solution. We then investigate combining individual queries into teams with the aim of capturing multiple SARs that may exist in a data set, for example, as is commonly seen in high-throughput screening data sets. Team formation is carried out iteratively as a postprocessing step following the evolution of the individual queries. The inclusion of uniqueness as a third objective within the MOEA provides an effective way of ensuring the queries are complementary in the active compounds they describe. Substantial improvements in both recall and precision are seen for some data sets. Furthermore, the resulting queries provide more detailed structure-activity information than is present in a single query.

Automated Quantitative Dose-Response Modeling and Point of Departure Determination for Large Toxicogenomic and High-Throughput Screening Data Sets

Regulatory and homeland security agencies undertake safety and risk assessments to assess the potential hazards of radiation, chemical, biological, and pharmaceutical agents. By law, these assessments must be science-based to ensure public safety and environmental quality. These agencies use dose-response modeling and benchmark dose methods to identify points of departure across single end points elicited by the agent. Regulatory agencies have also begun to examine toxicogenomic data to identify novel biomarkers of exposure and assess potential toxicity. The ToxResponse Modeler streamlines analyses and point of departure (POD) calculations across hundreds of responses (e.g., differential gene expression, changes in metabolite levels) through an automated process capable of large-scale modeling and model selection. The application identifies the best-fit dose-response model utilizing particle swarm optimization and calculates the probabilistic POD. The application analyzed a publicly available 2,3,7,8-tetrachlorodibenzo-p-dioxin dose-response data set of hepatic gene expression data in C57BL/6 mice to identify putative biomarkers. The Gene Ontology mapped these responses to specific functions to differentiate adaptive effects from toxic responses. In principle, safety and risk assessors could use the automated ToxResponse Modeler to analyze any large dose-response data set including outputs from high-throughput screening assays to assist with the ranking and prioritization of compounds that warrant further investigation or development.

Similarity Searching using Compound Class‐Specific Combinations of Substructures Found in Randomly Generated Molecular Fragment Populations

Fine fingerprints. Herein, we show that combinations of randomly generated fragments are signatures of active molecules. Small sets of such fragments are encoded as bit string representations and used for similarity searching. These fingerprints are successfully applied to mine high-throughput screening data sets. Shown are randomly generated substructures encoded as a small fingerprint that were extracted from a fragment pathway specific for cathepsin B inhibitors. Supporting information for this article is available on the WWW under http://www.wiley-vch.de/contents/jc_2452/2008/z700199_s.pdf or from the author. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.

Simulation of Sequential Screening Experiments Using Emerging Chemical Patterns

A method called "Emerging Chemical Patterns" (ECP) has recently been introduced as a novel approach to binary molecular classification (for example, "active" versus "inactive"). The underlying pattern recognition algorithm was first introduced in computer science and then adopted for applications in medicinal chemistry and compound screening. A special feature is its ability to accurately classify molecules on the basis of very small training sets containing only a few compounds. This feature is highly relevant for virtual compound screening when only very few experimental hits are available as templates. Here we adopt ECP calculations to simulate sequential screening using an experimental high-throughput screening (HTS) data set containing inhibitors of dihydrofolate reductase. In doing so, we focus on minimizing the number of database compounds that need to be evaluated in order to identify a substantial fraction of available hits. We demonstrate that iterative ECP calculations recover on average between approximately 19% and approximately 39% of available hits in the data set while dramatically reducing the number of compounds that need to be tested to between approximately 0.002% and approximately 9% of the screening database.

Introduction of a Generally Applicable Method to Estimate Retrieval of Active Molecules for Similarity Searching using Fingerprints

Fingerprints are bit string representations of molecular structure and properties and are among the most widely used computational tools for similarity searching and database screening. Various fingerprint designs are available and their search performance is in general strongly dependent on the compound classes under study and the chemical characteristics of screening databases. Currently, it is not possible to predict the probability of identifying novel hits through fingerprint searching. However, for practical applications, such estimations would be very useful because one might be able, for example, to prioritize fingerprints and compound selection strategies or decide whether or not a similarity search campaign with subsequent experimental evaluation of candidate compounds would be promising at all. We have developed a method that makes it possible to predict the outcome of similarity search calculations using any type of keyed fingerprint. The methodology incorporates bit frequency distributions of reference molecules and the screening database into an information-theoretic function and determines the principally possible recall of active compounds within selection sets of varying size. We calibrate the function on diverse compound classes and accurately predict compound recovery in retrospective virtual screening trials. Furthermore, we correctly predict fingerprint search performance on two experimental high-throughput screening data sets (HTS). Our findings indicate that given a set of reference molecules, a fingerprint, and a screening database, we can readily estimate how likely it will be to retrieve active compounds, without knowledge about the distribution of potential hits in the database.

Analysis of a High-Throughput Screening Data Set Using Potency-Scaled Molecular Similarity Algorithms

Molecular similarity methods for ligand-based virtual screening (VS) generally do not take compound potency as a variable or search parameter into account. We have incorporated a logarithmic potency scaling function into two conceptually distinct VS algorithms to account for relative compound potency during search calculations. A high-throughput screening (HTS) data set containing cathepsin B inhibitors was analyzed to evaluate the effects of potency scaling. Sets of template compounds were randomly selected from the HTS data and used to search for hits having varying potency levels in the presence or absence of potency scaling. Enrichment of potent compounds in small subsets of the HTS data set was observed as a consequence of potency scaling. In part, observed enrichments could be rationalized as a result of recentering chemical reference space on a subspace populated by potent compounds. Our findings suggest that VS calculations using multiple reference compounds can be directed toward the preferential detection of potent database hits by scaling compound contributions according to potency differences.

A Chemogenomic Approach to Discovering Target‐Selective Drugs

An approach is described for identifying scaffolds with predefined selectivity profiles directly from a high-quality high-throughput screening data set, and monitoring that selectivity profile throughout the drug discovery process. This approach results in reduced drug development timelines from "hit" to "lead" stage, and should reduce liabilities for a molecule advancing through the development process.

Enrichment of High-Throughput Screening Data with Increasing Levels of Noise Using Support Vector Machines, Recursive Partitioning, and Laplacian-Modified Naive Bayesian Classifiers

High-throughput screening (HTS) plays a pivotal role in lead discovery for the pharmaceutical industry. In tandem, cheminformatics approaches are employed to increase the probability of the identification of novel biologically active compounds by mining the HTS data. HTS data is notoriously noisy, and therefore, the selection of the optimal data mining method is important for the success of such an analysis. Here, we describe a retrospective analysis of four HTS data sets using three mining approaches: Laplacian-modified naive Bayes, recursive partitioning, and support vector machine (SVM) classifiers with increasing stochastic noise in the form of false positives and false negatives. All three of the data mining methods at hand tolerated increasing levels of false positives even when the ratio of misclassified compounds to true active compounds was 5:1 in the training set. False negatives in the ratio of 1:1 were tolerated as well. SVM outperformed the other two methods in capturing active compounds and scaffolds in the top 1%. A Murcko scaffold analysis could explain the differences in enrichments among the four data sets. This study demonstrates that data mining methods can add a true value to the screen even when the data is contaminated with a high level of stochastic noise.

Sample phenotype clusters in high-density oligonucleotide microarray data sets are revealed using Isomap, a nonlinear algorithm

BackgroundLife processes are determined by the organism's genetic profile and multiple environmental variables. However the interaction between these factors is inherently non-linear [1]. Microarray data is one representation of the nonlinear interactions among genes and genes and environmental factors. Still most microarray studies use linear methods for the interpretation of nonlinear data. In this study, we apply Isomap, a nonlinear method of dimensionality reduction, to analyze three independent large Affymetrix high-density oligonucleotide microarray data sets.ResultsIsomap discovered low-dimensional structures embedded in the Affymetrix microarray data sets. These structures correspond to and help to interpret biological phenomena present in the data. This analysis provides examples of temporal, spatial, and functional processes revealed by the Isomap algorithm. In a spinal cord injury data set, Isomap discovers the three main modalities of the experiment – location and severity of the injury and the time elapsed after the injury. In a multiple tissue data set, Isomap discovers a low-dimensional structure that corresponds to anatomical locations of the source tissues. This model is capable of describing low- and high-resolution differences in the same model, such as kidney-vs.-brain and differences between the nuclei of the amygdala, respectively. In a high-throughput drug screening data set, Isomap discovers the monocytic and granulocytic differentiation of myeloid cells and maps several chemical compounds on the two-dimensional model.ConclusionVisualization of Isomap models provides useful tools for exploratory analysis of microarray data sets. In most instances, Isomap models explain more of the variance present in the microarray data than PCA or MDS. Finally, Isomap is a promising new algorithm for class discovery and class prediction in high-density oligonucleotide data sets.

A 3D Similarity Method for Scaffold Hopping from Known Drugs or Natural Ligands to New Chemotypes

A primary goal of 3D similarity searching is to find compounds with similar bioactivity to a reference ligand but with different chemotypes, i.e., "scaffold hopping". However, an adequate description of chemical structures in 3D conformational space is difficult due to the high-dimensionality of the problem. We present an automated method that simplifies flexible 3D chemical descriptions in which clustering techniques traditionally used in data mining are exploited to create "fuzzy" molecular representations called FEPOPS (feature point pharmacophores). The representations can be used for flexible 3D similarity searching given one or more active compounds without a priori knowledge of bioactive conformations or pharmacophores. We demonstrate that similarity searching with FEPOPS significantly enriches for actives taken from in-house high-throughput screening datasets and from MDDR activity classes COX-2, 5-HT3A, and HIV-RT, while also scaffold or ring-system hopping to new chemical frameworks. Further, inhibitors of target proteins (dopamine 2 and retinoic acid receptor) are recalled by FEPOPS by scaffold hopping from their associated endogenous ligands (dopamine and retinoic acid). Importantly, the method excels in comparison to commonly used 2D similarity methods (DAYLIGHT, MACCS, Pipeline Pilot fingerprints) and a commercial 3D method (Pharmacophore Distance Triplets) at finding novel scaffold classes given a single query molecule.

Dbtop: Topomer similarity searching of conventional structure databases

A new topomer-based method for 3D searching of conventional structural databases is described, according to which 3D molecular structures are compared as sets of fragments or topomers, in single rule-generated conformations oriented by superposition of their fragmentation bonds. A topomer is characterized by its CoMFA-like steric shape and now also by its pharmacophoric features, in some novel ways that are detailed and discussed. To illustrate the behavior of topomer similarity searching, a new dbtop program was used to generate a topomer distance matrix for a diverse set of 26 PDE4 inhibitors and 15 serotonin receptor modulators. With the best of three parameter settings tried, within the 210 shortest topomer distances (of 1460), 94.7% involved pairs of compounds having the same biological activity, and the nearest neighbor to every compound also shared its activity. The standard similarity metric, Tanimoto coefficients of “2D fingerprints”, could achieve a similar selectivity performance only for the 108 shortest distances, and three Tanimoto nearest neighbors had a different biological activity. Topomer similarity also allowed “lead-hopping” among 22 of the 26 PDE4 inhibitors, notably between rolipram and cipamfylline, while “2D fingerprints” Tanimotos recognized similarity only within generally recognized structural classes. In 370 searches of authentic high-throughput screening (HTS) data sets, the typical topomer similarity search rate was about 200 structures per s.

Optimization of focused chemical libraries using recursive partitioning.

A number of methods currently exist for designing chemical libraries. General or universal libraries use a measurement of chemical diversity in their design and seek to cover as much of chemical space as possible in order to maximize the likelihood of discovering a novel lead class of active compounds. Focused chemical libraries are then synthesized to expand on this particular class and thoroughly explore the space about it. Rarely, however, is relevant biological data tightly incorporated in the design of focused libraries. Recursive partitioning is a statistical technique that is used to quickly build SAR models from high-throughput screening data sets and associated chemical descriptors. Using these models in a virtual screening mode significantly increases the probability of finding other active compounds. The predicted activity can be also be used as the fitness function for a genetic algorithm that is designed to select monomer subsets having a higher probability of being active. This dramatically reduces the number of compounds that need to be synthesized in focused libraries thus saving considerable time, effort and expense. This paper describes how recursive partitioning models are used to optimize the design of focused chemical libraries.

Using Recursive Partitioning to Analyze a Large Sar Data Set

Large data sets are becoming much more prevalent in drug discovery as high throughput screening and combinatorial chemistry data sets become available. Conventional linear parametric methods, linear regression, PCR, PLS, etc., often fail in analyzing structurally heterogeneous data sets. This is because the underlying relationships may involve nonlinearities, thresholds and interactions, all of which considerably impede linear additive modeling approaches. Also the conventional assumption that all compounds in data set are acting by the same mechanism is not expected to hold. Recursive partitioning (RP) is able to accommodate all these difficulties; it is also computationally fast; RP can deal with very large data sets - 10 to 100k observations pose no particular problems. Therefore RP invites investigation as a general approach for study of structure activity relationships in large chemistry data sets. The purpose of this paper is to explicate a recursive partitioning procedure, FIRM, through the analysis of a large, structure-activity data set, 1650 compounds with 153 fragment descriptors and monoamine oxidase, MAO, activity. The methodology is successful in uncovering compounds acting through different mechanisms.