MDL Drug Data Report Research Articles

Bioactive Molecules: Perfectly Shaped for Their Target?

In this study, we examined target subsets extracted from the MDL Drug Data Report (MDDR)1 to identify specific molecular shape profiles that are representative for compounds active on those targets. Normalized Principal Moments of Inertia Ratios (NPRs)2 have been used to describe molecular shape of small molecules in a finite triangular descriptor space. The clustering behavior of the MDDR target subsets in a cell-based triangular system shows a significant difference compared to randomly sampled datasets and proves the capability of the NPR descriptor to provide information. For some of the target subsets, certain parts of the descriptor space are unlikely to be occupied by bioactive compounds. All analyzed datasets show a generally biased distribution of molecular shapes: the majority of their compounds exhibit a rod-like character. The influence of the employed 3D conformer generators on this distribution has been assessed as well as the capability of multiple conformations of compounds to increase the shape space covered.

Predicting drug polypharmacology using a novel surface property similarity-based approach

In recent years, polypharmacology is becoming anincreasingly important aspect in drug design. For exam-ple, pharmaceutical companies are discovering more andmore cases in which multiple drugs bind to a given tar-get (promiscuous targets) and in which a given drugbinds to more than one target (promiscuous ligands).Both of these phenomena are clearly of great impor-tance when considering drug side-effects. Given thatscreening drugs against all the proteins expressed by thehuman genome is infeasible, several computational tech-niques for predicting the pharmacological profiles ofdrugs have been developed, ranging from statistical ana-lyses of chemical fingerprints and biological activities [1]to 3D docking of ligand structures into protein pockets.Here we present a novel shape-based approach whichuses spherical harmonic (SH) representations [2,3] tocompare molecular surfaces and key surface propertiesvery efficiently. This approach compares targets by theSH similarity of their ligands and also of their bindingpockets. This allows promiscuous ligands and targets tobe identified and characterized.In this contribution, we present details of ourapproach applied to a subset of the MDL Drug DataReport (MDDR) database containing 65367 compoundsdistributed over 249 diverse pharmacological targets forwhich experimental binding information is known. Thesimilarity of each ligand to each target’s ligand set isquantified and used to predict promiscuity. To ourknowledge, this is the largest all-against-all polypharma-cological study to have been carried out using shape-based techniques. We compare our promiscuity predic-tions with computational and experimental resultsobtained by Keiser et al.[4]. We also analyse thecorrelation between binding pocket shapes and ligand-based promiscuity predictions using the ligand andpocket shape similarity matrices.

An Enhancement of Bayesian Inference Network for Ligand-Based Virtual Screening using Features Selection

Problem statement: Similarity based Virtual Screening (VS) deals with a large amount of data containing irrelevant and/or redundant fragments or features. Recent use of Bayesian network as an alternative for existing tools for similarity based VS has received noticeable attention of the researchers in the field of chemoinformatics. Approach: To this end, different models of Bayesian network have been developed. In this study, we enhance the Bayesian Inference Network (BIN) using a subset of selected molecule's features. Results: In this approach, a few features were filtered from the molecular fingerprint features based on a features selection approach. Conclusion: Simulated virtual screening experiments with MDL Drug Data Report (MDDR) data sets showed that the proposed method provides simple ways of enhancing the cost effectiveness of ligand-based virtual screening searches, especially for higher diversity data set.

New Fragment Weighting Scheme for the Bayesian Inference Network in Ligand-Based Virtual Screening

Many of the conventional similarity methods assume that molecular fragments that do not relate to biological activity carry the same weight as the important ones. One possible approach to this problem is to use the Bayesian inference network (BIN), which models molecules and reference structures as probabilistic inference networks. The relationships between molecules and reference structures in the Bayesian network are encoded using a set of conditional probability distributions, which can be estimated by the fragment weighting function, a function of the frequencies of the fragments in the molecule or the reference structure as well as throughout the collection. The weighting function combines one or more fragment weighting schemes. In this paper, we have investigated five different weighting functions and present a new fragment weighting scheme. Later on, these functions were modified to combine the new weighting scheme. Simulated virtual screening experiments with the MDL Drug Data Report (23) and maximum unbiased validation data sets show that the use of new weighting scheme can provide significantly more effective screening when compared with the use of current weighting schemes.

Brainstorming: weighted voting prediction of inhibitors for protein targets

The “Brainstorming” approach presented in this paper is a weighted voting method that can improve the quality of predictions generated by several machine learning (ML) methods. First, an ensemble of heterogeneous ML algorithms is trained on available experimental data, then all solutions are gathered and a consensus is built between them. The final prediction is performed using a voting procedure, whereby the vote of each method is weighted according to a quality coefficient calculated using multivariable linear regression (MLR). The MLR optimization procedure is very fast, therefore no additional computational cost is introduced by using this jury approach. Here, brainstorming is applied to selecting actives from large collections of compounds relating to five diverse biological targets of medicinal interest, namely HIV-reverse transcriptase, cyclooxygenase-2, dihydrofolate reductase, estrogen receptor, and thrombin. The MDL Drug Data Report (MDDR) database was used for selecting known inhibitors for these protein targets, and experimental data was then used to train a set of machine learning methods. The benchmark dataset (available at http://bio.icm.edu.pl/∼darman/chemoinfo/benchmark.tar.gz) can be used for further testing of various clustering and machine learning methods when predicting the biological activity of compounds. Depending on the protein target, the overall recall value is raised by at least 20% in comparison to any single machine learning method (including ensemble methods like random forest) and unweighted simple majority voting procedures.

In silico guided search for anti-viral properties of a natural methylenecyclopropane glucoside

(2E)-2-(6-hydroxyhexyliden)cyclopropyl-β-glucopyranoside (1) was isolated from the rhizomes of the tree fern Metaxya rostrata, Metaxyaceae, a Costa Rican traditional herbal remedy against intestinal diseases [1]. Literature research for similarity of the structural characteristics resulted in different related compounds with the very uncommon structural part of methylenecyclopropane [2,3,4], which showed anti-viral activitiy against Epstein-Barr-, Human Herpes-, Hepatitis B-, Human Immunodeficiency- and Human Cytomegalo-virus in vitro. Thus, an in silico guided search for possible anti-viral activities of (1) was realized. Pharmacophoric profiling against over 2200 models representing over 280 putative pharmacological targets [5] suggested α-glucosidase as potential target for (1). A structural similarity search in the MDL Drug Data Report (MDDR) database [6] confirmed this result making α-glucosidase a promising target for biological evaluation of compound (1).

Model-Free Drug-Likeness from Fragments

We developed a drug-likeness filter (DLF), starting from molecular fragments and molecular weight (MW), a key property relevant in drug design. The molecular fragments were selected from extended connectivity atom environments based on their occurrence ratio in our collection of drugs and "nondrugs". The DLF recalls 87.05% of compounds from DRUGS (N = 3823) and 40.25% of compounds from the Available Chemicals Directory, (ACD, N = 178 0 11), using molecular fragments only. By adding MW (under 600) as an additional filter, 78.81% of DRUGS and 40.17% of ACD are recalled. The DLF procedure was externally validated using the MDL Drug Data Report (MDDR) data set (N = 169 277): 78.45% of compounds were recalled using the molecular fragments only, while 65.64% pass the DLF-MW filter. Over 50% of a pesticides collection (N = 1482) passed the DLF, as these chemicals share molecular fragments with known drugs. Developed as a model-free filter, DLF is perhaps less useful in discriminating drugs from nondrugs but more likely to rapidly eliminate those chemicals rich in nondrug-like fragments. Since almost 40% of ACD, the standard reference set for nondrugs, contain drug-like molecules, by using a rule-based system such as DLF, one is less likely to mislabel nondrugs due to overfitting. Reliable benchmarks for nondrugs are not likely to exist since medicinal chemistry catalogs tend to be biased toward existing drugs.

Virtual High Throughput Screening Using Combined Random Forest and Flexible Docking

We present here the random forest supervised machine learning algorithm applied to flexible docking results from five typical virtual high throughput screening (HTS) studies. Our approach is aimed at: i) reducing the number of compounds to be tested experimentally against the given protein target and ii) extending results of flexible docking experiments performed only on a subset of a chemical library in order to select promising inhibitors from the whole dataset. The random forest (RF) method is applied and tested here on compounds from the MDL drug data report (MDDR). The recall values for selected five diverse protein targets are over 90% and the performance reaches 100%. This machine learning method combined with flexible docking is capable to find 60% of the active compounds for most protein targets by docking only 10% of screened ligands. Therefore our in silico approach is able to scan very large databases rapidly in order to predict biological activity of small molecule inhibitors and provides an effective alternative for more computationally demanding methods in virtual HTS.

KNNsim: k-Nearest neighbors similarity with genetic algorithm features optimization enhances the prediction of activity classes for small molecules

Protein targets specificity classification is an important step in computational drug development and design efforts. The enhanced classification models of small chemical molecules enable the rapid scanning of large compounds databases. Here, we present the k-nearest neighbors with genetic algorithm feature optimization approach for selection of small molecule protein inhibitors. The method is trained on selected, diverse activity classes of the MDL drug data report (MDDR) with ligands described using simple atom pairs two dimensional chemical descriptors. The accuracy of inhibitors identification is presented in confusion tables with calculated recall and precision values. The precision for selected types of targets exceeded 70%, and the recall reaches 40%. As a consequence, the method can be easily applied to large commercial compounds collections in a drug development campaign in order to significantly reduce the number of ligands for further costly experimental validation.

A Combination Method of the Tanimoto Coefficient and Proximity Measure of Random Forest for Compound Activity Prediction

Chemical and biological activities of compounds provide valuable information for discovering new drugs. The compound fingerprint that is represented by structural information of the activities is used for candidates for investigating similarity. However, there are several problems with predicting accuracy from the requirement in the compound structural similarity. Although the amount of compound data is growing rapidly, the number of well-annotated compounds, e.g., those in the MDL Drug Data Report (MDDR)database, has not increased quickly. Since the compounds that are known to have some activities of a biological class of the target are rare in the drug discovery process, the accuracy of the prediction should be increased as the activity decreases or the false positive rate should be maintained in databases that have a large number of un-annotated compounds and a small number of annotated compounds of the biological activity. In this paper, we propose a new similarity scoring method composed of a combination of the Tanimoto coefficient and the proximity measure of random forest. The score contains two properties that are derived from unsupervised and supervised methods of partial dependence for compounds. Thus, the proposed method is expected to indicate compounds that have accurate activities. By evaluating the performance of the prediction compared with the two scores of the Tanimoto coefficient and the proximity measure, we demonstrate that the prediction result of the proposed scoring method is better than those of the two methods by using the Linear Discriminant Analysis (LDA) method. We estimate the prediction accuracy of compound datasets extracted from MDDR using the proposed method. It is also shown that the proposed method can identify active compounds in datasets including several un-annotated compounds.

Target Specific Compound Identification Using a Support Vector Machine

In many cases at the beginning of an HTS-campaign, some information about active molecules is already available. Often known active compounds (such as substrate analogues, natural products, inhibitors of a related protein or ligands published by a pharmaceutical company) are identified in low-throughput validation studies of the biochemical target. In this study we evaluate the effectiveness of a support vector machine applied for those compounds and used to classify a collection with unknown activity. This approach was aimed at reducing the number of compounds to be tested against the given target. Our method predicts the biological activity of chemical compounds based on only the atom pairs (AP) two dimensional topological descriptors. The supervised support vector machine (SVM) method herein is trained on compounds from the MDL drug data report (MDDR) known to be active for specific protein target. For detailed analysis, five different biological targets were selected including cyclooxygenase-2, dihydrofolate reductase, thrombin, HIV-reverse transcriptase and antagonists of the estrogen receptor. The accuracy of compound identification was estimated using the recall and precision values. The sensitivities for all protein targets exceeded 80% and the classification performance reached 100% for selected targets. In another application of the method, we addressed the absence of an initial set of active compounds for a selected protein target at the beginning of an HTS-campaign. In such a case, virtual high-throughput screening (vHTS) is usually applied by using a flexible docking procedure. However, the vHTS experiment typically contains a large percentage of false positives that should be verified by costly and time-consuming experimental follow-up assays. The subsequent use of our machine learning method was found to improve the speed (since the docking procedure was not required for all compounds from the database) and also the accuracy of the HTS hit lists (the enrichment factor).

Genetic Algorithm-Optimized QSPR Models for Bioavailability, Protein Binding, and Urinary Excretion

In this work, a genetic algorithm (GA) was applied to build up a set of QSPR (quantitative structure-property relationship) models for human absolute oral bioavailability, plasma protein binding, and urinary excretion using the counts of molecular fragments as descriptors. For a pharmacokinetic property, the consensus score of a set of models (20 or 30) was found to improve the correlation coefficient and reduce the standard error significantly. Key fragments that may boost or reduce pharmacokinetic properties were also identified. Databases searches were performed for a set of key fragments identified by bioavailability models. The percentage of hit rates of bioavailability-boosting fragments were significantly higher than those of bioavailability-reducing fragments for MDDR (MDL Drug Data Report), a database of drugs and drug leads entered or entering development. On the other hand, the opposite trend was observed for ACD (Available Chemicals Directory), a database of all kinds of available compounds.

Toward automated biochemotype annotation for large compound libraries

Combinatorial chemistry allows scientists to probe large synthetically accessible chemical space. However, identifying the sub-space which is selectively associated with an interested biological target, is crucial to drug discovery and life sciences. This paper describes a process to automatically annotate biochemotypes of compounds in a library and thus to identify bioactivity related chemotypes (biochemotypes) from a large library of compounds. The process consists of two steps: (1) predicting all possible bioactivities for each compound in a library, and (2) deriving possible biochemotypes based on predictions. The Prediction of Activity Spectra for Substances program (PASS) was used in the first step. In second step, structural similarity and scaffold-hopping technologies are employed. These technologies are used to derive biochemotypes from bioactivity predictions and the corresponding annotated biochemotypes from MDL Drug Data Report (MDDR) database. About a one million (982,889) commercially available compound library (CACL) has been tested using this process. This paper demonstrates the feasibility of automatically annotating biochemotypes for large libraries of compounds. Nevertheless, some issues need to be considered in order to improve the process. First, the prediction accuracy of PASS program has no significant correlation with the number of compounds in a training set. Larger training sets do not necessarily increase the maximal error of prediction (MEP), nor do they increase the hit structural diversity. Smaller training sets do not necessarily decrease MEP, nor do they decrease the hit structural diversity. Second, the success of systematic bioactivity prediction relies on modeling, training data, and the definition of bioactivities (biochemotype ontology). Unfortunately, the biochemotype ontology was not well developed in the PASS program. Consequently, "ill-defined" bioactivities can reduce the quality of predictions. This paper suggests the ways in which the systematic bioactivities prediction program should be improved.

NIPALSTREE: A New Hierarchical Clustering Approach for Large Compound Libraries and Its Application to Virtual Screening

A hierarchical clustering algorithm--NIPALSTREE--was developed that is able to analyze large data sets in high-dimensional space. The result can be displayed as a dendrogram. At each tree level the algorithm projects a data set via principle component analysis onto one dimension. The data set is sorted according to this one dimension and split at the median position. To avoid distortion of clusters at the median position, the algorithm identifies a potentially more suited split point left or right of the median. The procedure is recursively applied on the resulting subsets until the maximal distance between cluster members exceeds a user-defined threshold. The approach was validated in a retrospective screening study for angiotensin converting enzyme (ACE) inhibitors. The resulting clusters were assessed for their purity and enrichment in actives belonging to this ligand class. Enrichment was observed in individual branches of the dendrogram. In further retrospective virtual screening studies employing the MDL Drug Data Report (MDDR), COBRA, and the SPECS catalog, NIPALSTREE was compared with the hierarchical k-means clustering approach. Results show that both algorithms can be used in the context of virtual screening. Intersecting the result lists obtained with both algorithms improved enrichment factors while losing only few chemotypes.

Virtual Screening against Metalloenzymes for Inhibitors and Substrates

Molecular docking uses the three-dimensional structure of a receptor to screen databases of small molecules for potential ligands, often based on energetic complementarity. For many docking scoring functions, which calculate nonbonded interactions, metalloenzymes are challenging because of the partial covalent nature of metal-ligand interactions. To investigate how well molecular docking can identify potential ligands of metalloenzymes using a "standard" scoring function, we have docked the MDL Drug Data Report (MDDR), a functionally annotated database of 95,000 small molecules, against the X-ray crystal structures of five metalloenzymes. These enzymes included three zinc proteases, the nickel analogue of an iron enzyme, and a molybdenum metalloenzyme. The ability of the docking program to retrospectively enrich the annotated ligands as high-scoring hits for each enzyme and to calculate proper geometries was evaluated. In all five systems, the annotated ligands within the MDDR were enriched at least 20 times over random. To test the approach prospectively, a sixth target, the zinc beta-lactamase from Bacteroides fragilis, was screened against the fragment-like subset of the ZINC database. We purchased and tested 15 compounds from among the top 50 top-ranked ligands from docking, and found 5 inhibitors with apparent K(i) values less than 120 microM, the best of which was 2 microM. A more ambitious test still was predicting actual substrates for a seventh target, a Zn-dependent phosphotriesterase from Pseudomonas diminuta. Screening the Available Chemicals Directory (ACD) identified 25 thiophosphate esters as potential substrates within the top 100 ranked compounds. Eight of these, all previously uncharacterized for this enzyme, were acquired and tested, and all were confirmed experimentally as substrates. These results suggest that a simple, noncovalent scoring function may be used to identify inhibitors of at least some metalloenzymes.

Hierarchical Docking of Databases of Multiple Ligand Conformations

Ligand flexibility is an important problem in molecular docking and virtual screening. To address this challenge, we investigate a hierarchical pre-organization of multiple conformations of small molecules. Such organization of pre-calculated conformations removes the exploration of ligand conformational space from the docking calculation and allows for concise representation of what can be thousands of conformations. The hierarchy also recognizes and prunes incompatible conformations early in the calculation, eliminating redundant calculations of fit. We investigate the method by docking the MDL Drug Data Report (MDDR), an annotated database of 100,000 molecules, into apo and holo forms of seven unrelated targets. This annotated database allows us to track the ranking of tens to hundreds of annotated ligands in each of the docking systems. The binding sites and database are prepared in an automated fashion in an attempt to remove some human bias from the calculations. Many thousands of explicit and implicit ligand conformations may be docked in calculations not much longer than required for single conformer docking. As long as internal energies are not considered, recombination with the hierarchy is additive as the number of degrees of freedom is increased. Molecules with even millions of conformations can be docked in a few minutes on a single desktop computer.

A Hierarchical Clustering Approach for Large Compound Libraries

A modified version of the k-means clustering algorithm was developed that is able to analyze large compound libraries. A distance threshold determined by plotting the sum of radii of leaf clusters was used as a termination criterion for the clustering process. Hierarchical trees were constructed that can be used to obtain an overview of the data distribution and inherent cluster structure. The approach is also applicable to ligand-based virtual screening with the aim to generate preferred screening collections or focused compound libraries. Retrospective analysis of two activity classes was performed: inhibitors of caspase 1 [interleukin 1 (IL1) cleaving enzyme, ICE] and glucocorticoid receptor ligands. The MDL Drug Data Report (MDDR) and Collection of Bioactive Reference Analogues (COBRA) databases served as the compound pool, for which binary trees were produced. Molecules were encoded by all Molecular Operating Environment 2D descriptors and topological pharmacophore atom types. Individual clusters were assessed for their purity and enrichment of actives belonging to the two ligand classes. Significant enrichment was observed in individual branches of the cluster tree. After clustering a combined database of MDDR, COBRA, and the SPECS catalog, it was possible to retrieve MDDR ICE inhibitors with new scaffolds using COBRA ICE inhibitors as seeds. A Java implementation of the clustering method is available via the Internet (http://www.modlab.de).

Calculating Similarities Between Biological Activities in the MDL Drug Data Report Database.

There are a number of licensed databases that assign biological activities to druglike compounds. The MDL Drug Data Report (MDDR), compiled from the patent literature, is a popular example. It contains several hundred distinct activities, some of which are therapeutic areas (e.g., Antihypertensive) and some of which are related to specific enzymes or receptors (e.g., ACE inhibitor). There are several data mining applications where it would be useful to calculate a similarity between any two activities. Two distinct activity labels can have a significant similarity for a number of reasons: two activities can be nearly synonymous (e.g., CCK B antagonist vs Gastrin antagonist), one activity may be a subset of another (e.g., Dopamine (D2) agonist vs Dopamine agonist), or an activity can be the mechanism by which another activity works (e.g., ACE inhibitor vs Antihypertensive), etc. In an ideal world, similarities for two activities could be calculated simply by comparing the compounds they have in common, but in hand-curated databases such as the MDDR the assignment of activities to compounds are inevitably inconsistent and incomplete. We propose a number of methods of calculating activity-activity similarities that hopefully compensate for errors in hand-curation. Two of these, TIMI and trend vector, show promise. Soft clustering of the activities using a union of similarity methods shows a reasonable association of therapeutic areas with their mechanisms.

Calculating similarities between biological activities in the MDL Drug Data Report database.

There are a number of licensed databases that assign biological activities to druglike compounds. The MDL Drug Data Report (MDDR), compiled from the patent literature, is a popular example. It contains several hundred distinct activities, some of which are therapeutic areas (e.g., Antihypertensive) and some of which are related to specific enzymes or receptors (e.g., ACE inhibitor). There are several data mining applications where it would be useful to calculate a similarity between any two activities. Two distinct activity labels can have a significant similarity for a number of reasons: two activities can be nearly synonymous (e.g., CCK B antagonist vs Gastrin antagonist), one activity may be a subset of another (e.g., Dopamine (D2) agonist vs Dopamine agonist), or an activity can be the mechanism by which another activity works (e.g., ACE inhibitor vs Antihypertensive), etc. In an ideal world, similarities for two activities could be calculated simply by comparing the compounds they have in common, but in hand-curated databases such as the MDDR the assignment of activities to compounds are inevitably inconsistent and incomplete. We propose a number of methods of calculating activity-activity similarities that hopefully compensate for errors in hand-curation. Two of these, TIMI and trend vector, show promise. Soft clustering of the activities using a union of similarity methods shows a reasonable association of therapeutic areas with their mechanisms.

Molecular similarity searching using atom environments, information-based feature selection, and a naïve Bayesian classifier.

A novel technique for similarity searching is introduced. Molecules are represented by atom environments, which are fed into an information-gain-based feature selection. A naïve Bayesian classifier is then employed for compound classification. The new method is tested by its ability to retrieve five sets of active molecules seeded in the MDL Drug Data Report (MDDR). In comparison experiments, the algorithm outperforms all current retrieval methods assessed here using two- and three-dimensional descriptors and offers insight into the significance of structural components for binding.