Search

Reaction classification has often been considered an important task for many different applications, and has traditionally been accomplished using hand-coded rule-based approaches. However, the availability of large collections of reactions enables data-driven approaches to be developed. We present the development and validation of a 336-class machine learning-based classification model integrated within a Conformal Prediction (CP) framework to associate reaction class predictions with confidence estimations. We also propose a data-driven approach for "dynamic" reaction fingerprinting to maximize the effectiveness of reaction encoding, as well as developing a novel reaction classification system that organizes labels into four hierarchical levels (SHREC: Sheffield Hierarchical REaction Classification). We show that the performance of the CP augmented model can be improved by defining confidence thresholds to detect predictions that are less likely to be false. For example, the external validation of the model reports 95% of predictions as correct by filtering out less than 15% of the uncertain classifications. The application of the model is demonstrated by classifying two reaction data sets: one extracted from an industrial ELN and the other from the medicinal chemistry literature. We show how confidence estimations and class compositions across different levels of information can be used to gain immediate insights on the nature of reaction collections and hidden relationships between reaction classes.

There has been a growing interest in multitask prediction in chemoinformatics, helped by the increasing use of deep neural networks in this field. This technique is applied to multitarget data sets, where compounds have been tested against different targets, with the aim of developing models to predict a profile of biological activities for a given compound. However, multitarget data sets tend to be sparse; i.e., not all compound-target combinations have experimental values. There has been little research on the effect of missing data on the performance of multitask methods. We have used two complete data sets to simulate sparseness by removing data from the training set. Different models to remove the data were compared. These sparse sets were used to train two different multitask methods, deep neural networks and Macau, which is a Bayesian probabilistic matrix factorization technique. Results from both methods were remarkably similar and showed that the performance decrease because of missing data is at first small before accelerating after large amounts of data are removed. This work provides a first approximation to assess how much data is required to produce good performance in multitask prediction exercises. Electronic supplementary material The online version of this article (10.1186/s13321-018-0281-z) contains supplementary material, which is available to authorized users.

This paper summarises work in chemoinformatics carried out in the Information School of the University of Sheffield during the period 2002-2014. Research studies are described on fingerprint-based similarity searching, data fusion, applications of reduced graphs and pharmacophore mapping, and on the School's teaching in chemoinformatics.

Spectral clustering involves placing objects into clusters based on the eigenvectors and eigenvalues of an associated matrix. The technique was first applied to molecular data by Brewer [J. Chem. Inf. Model. 2007, 47, 1727-1733] who demonstrated its use on a very small dataset of 125 COX-2 inhibitors. We have determined suitable parameters for spectral clustering using a wide variety of molecular descriptors and several datasets of a few thousand compounds and compared the results of clustering using a nonoverlapping version of Brewer's use of Sarker and Boyer's algorithm with that of Ward's and k-means clustering. We then replaced the exact eigendecomposition method with two different approximate methods and concluded that Singular Value Decomposition is the most appropriate method for clustering larger compound collections of up to 100,000 compounds. We have also used spectral clustering with the Tversky coefficient to generate two sets of clusters linked by a common set of eigenvalues and have used this novel approach to cluster sets of fragments such as those used in fragment-based drug design.

The design of new alerts, that is, collections of structural features observed to result in toxicological activity, can be a slow process and may require significant input from toxicology and chemistry experts. A method has therefore been developed to help automate alert identification by mining descriptions of activating structural features directly from toxicity data sets. The method is based on jumping emerging pattern mining which is applied to a set of toxic and nontoxic compounds that are represented using atom pair descriptors. Using the resulting jumping emerging patterns, it is possible to cluster toxic compounds into groups defined by the presence of shared structural features and to arrange the clusters into hierarchies. The methodology has been tested on a number of data sets for Ames mutagenicity, oestrogenicity, and hERG channel inhibition end points. These tests have shown the method to be effective at clustering the data sets around minimal jumping-emerging structural patterns and finding descriptions of potentially activating structural features. Furthermore, the mined structural features have been shown to be related to some of the known alerts for all three tested end points.

Molecular interaction fields provide a useful description of ligand binding propensity and have found widespread use in computer-aided drug design, for example, to characterize protein binding sites and in small molecular applications, such as three-dimensional quantitative structure-activity relationships, physicochemical property prediction, and virtual screening. However, the grids on which the field data are stored are typically very large, consisting of thousands of data points, which make them cumbersome to store and manipulate. The wavelet transform is a commonly used data compression technique, for example, in signal processing and image compression. Here we use the wavelet transform to encode molecular interaction fields as wavelet thumbnails, which represent the original grid data in significantly reduced volumes. We describe a method for aligning wavelet thumbnails based on extracting extrema from the thumbnails and subsequently use them for virtual screening. We demonstrate that wavelet thumbnails provide an effective method of capturing the three-dimensional information encoded in a molecular interaction field.

This paper describes the extension of our earlier multi-objective method for generating plausible pharmacophore hypotheses to incorporate partial matches. Diverse sets of molecules rarely adopt exactly the same binding mode, and so allowing the identification of partial matches allows our program to be applied to larger and more diverse datasets. The method explores the conformational space of a series of ligands simultaneously with their alignment using a multi-objective genetic algorithm (MOGA). The principles of Pareto ranking are used to evolve a diverse set of pharmacophore hypotheses that are optimised on conformational energy of the ligands, the goodness of the overlay and the volume of the overlay. A partial match is defined as a pharmacophoric feature that is present in at least two, but not all, of the ligands in the set. The number of ligands that map to a given pharmacophore point is taken into account when evaluating an overlay. The method is applied to a number of test cases extracted from the Protein Data Bank (PDB) where the true overlay is known.

Knowledge Discovery in Databases (KDD) refers to the use of methodologies from machine learning, pattern recognition, statistics, and other fields to extract knowledge from large collections of data, where the knowledge is not explicitly available as part of the database structure. In this paper, we describe four modern data mining techniques, Rough Set Theory (RST), Association Rule Mining (ARM), Emerging Pattern Mining (EP), and Formal Concept Analysis (FCA), and we have attempted to give an exhaustive list of their chemoinformatics applications. One of the main strengths of these methods is their descriptive ability. When used to derive rules, for example, in structure-activity relationships, the rules have clear physical meaning. This review has shown that there are close relationships between the methods. Often apparent differences lie in the way in which the problem under investigation has been formulated which can lead to the natural adoption of one or other method. For example, the idea of a structural alert, as a structure which is present in toxic and absent in nontoxic compounds, leads to the natural formulation of an Emerging Pattern search. Despite the similarities between the methods, each has its strengths. RST is useful for dealing with uncertain and noisy data. Its main chemoinformatics applications so far have been in feature extraction and feature reduction, the latter often as input to another data mining method, such as an Support Vector Machine (SVM). ARM has mostly been used for frequent subgraph mining. EP and FCA have both been used to mine both structural and nonstructural patterns for classification of both active and inactive molecules. Since their introduction in the 1980s and 1990s, RST, ARM, EP, and FCA have found wide-ranging applications, with many thousands of citations in Web of Science, but their adoption by the chemoinformatics community has been relatively slow. Advances, both in computer power and in algorithm development, mean that there is the potential to apply these techniques to larger data sets and thus to different problems in the future.

Pharmacophore methods provide a way of establishing a structure activity relationship for a series of known active ligands. Often, there are several plausible hypotheses that could explain the same set of ligands and, in such cases, it is important that the chemist is presented with alternatives that can be tested with different synthetic compounds. Existing pharmacophore methods involve either generating an ensemble of conformers and considering each conformer of each ligand in turn or exploring conformational space on-the-fly. The ensemble methods tend to produce a large number of hypotheses and require considerable effort to analyse the results, whereas methods that vary conformation on-the-fly typically generate a single solution that represents one possible hypothesis, even though several might exist. We describe a new method for generating multiple pharmacophore hypotheses with full conformational flexibility being explored on-the-fly. The method is based on multiobjective evolutionary algorithm techniques and is designed to search for an ensemble of diverse yet plausible overlays which can then be presented to the chemist for further investigation.

A knowledge-based method for calculating the similarity of functional groups is described and validated. The method is based on experimental information derived from small molecule crystal structures. These data are used in the form of scatterplots that show the likelihood of a non-bonded interaction being formed between functional group A (the `central group') and functional group B (the `contact group' or `probe'). The scatterplots are converted into three-dimensional maps that show the propensity of the probe at different positions around the central group. Here we describe how to calculate the similarity of a pair of central groups based on these maps. The similarity method is validated using bioisosteric functional group pairs identified in the Bioster database and Relibase. The Bioster database is a critical compilation of thousands of bioisosteric molecule pairs, including drugs, enzyme inhibitors and agrochemicals. Relibase is an object-oriented database containing structural data about protein-ligand interactions. The distributions of the similarities of the bioisosteric functional group pairs are compared with similarities for all the possible pairs in IsoStar, and are found to be significantly different. Enrichment factors are also calculated showing the similarity method is statistically significantly better than random in predicting bioisosteric functional group pairs.

Predicting the risk of toxic and environmental effects of chemical compounds is of great importance to all chemical industries [1]. Expert systems have shown success in predicting toxic risk by applying established knowledge of toxicology encoded as a knowledge base of structural alerts and a reasoning model. A disadvantage of expert systems is that developing new structural alerts requires considerable time and effort from domain experts. In order to expedite this process a software tool has been developed that can automatically mine representations of activating features directly from toxicity datasets and present them in an interpretable form. Our knowledge discovery tool applies emerging pattern (EP) mining [2]: a form of association rule mining [3] that is well known to computer science, but is relatively new to chemistry [4]. The EP mining algorithm accepts any data expressed as a series of binary properties, which is divided into two classes, and extracts patterns of those properties that are frequent within the data and are more frequent in one data class compared to the other. By mining emerging patterns from toxicity datasets, encoded as fingerprints of binary descriptors, the tool generates patterns of features that distinguish toxicants from innocuous compounds. These patterns represent potentially activating features of the toxic compounds that may then be used to define new alerts. The knowledge discovery tool has been tested using a public dataset of 3489 mutagens and 2981 non-mutagens, encoded as fingerprints of approximately 2000 functional groups and ring descriptors. EPs were produced and grouped into a number of hierarchical families. Six of the EPs that represented distinct chemical classes were selected for manual inspection by a toxicology expert. Relevant literature was analysed to find a mechanistic rationale for the mined features, which resulted in four new structural alerts for in vitro mutagenicity.

A program for overlaying multiple flexible molecules has been developed. Candidate overlays are generated by a novel fingerprint algorithm, scored on three objective functions (union volume, hydrogen-bond match, and hydrophobic match), and ranked by constrained Pareto ranking. A diverse subset of the best ranked solutions is chosen using an overlay-dissimilarity metric. If necessary, the solutions can be optimised. A multi-objective genetic algorithm can be used to find additional overlays with a given mapping of chemical features but different ligand conformations. The fingerprint algorithm may also be used to produce constrained overlays, in which user-specified chemical groups are forced to be superimposed. The program has been tested on several sets of ligands, for each of which the true overlay is known from protein–ligand crystal structures. Both objective and subjective success criteria indicate that good results are obtained on the majority of these sets. Electronic supplementary material The online version of this article (doi:10.1007/s10822-012-9573-y) contains supplementary material, which is available to authorized users.

SPROUT is a computer program for constrained structure generation. It is designed to generate molecules for a range of applications in molecular recognition. The program uses a number of approximations that enable a wide variety of diverse structures to be generated. Practical use of the program is demonstrated in two examples. The first demonstrates the ability of the program to generate candidate inhibitors for a receptor site of known 3D structure, specifically the GDP binding site of p21. In the second example, structures are generated to fit a pharmacophore hypothesis that models morphine agonists.

This chapter analyses the use of similarity fusion in similarity searching of chemical databases. The ranked retrieval of molecules from a database can be modelled using both analytical and simulation approaches describing the similarities between an active reference structure and both the active and the non-active molecules in the database. The simulation model described here has the advantage that it can handle unmatched molecules, i.e., those that occur in one of the ranked similarity lists that are to be fused but that do not occur in the other. Our analyses provide insights into why the results of similarity fusion are often inconsistent when used for virtual screening.

A knowledge-based approach to the de novo design of synthetically feasible molecules is described. The method is based on reaction vectors which represent the structural changes that take place at the reaction center along with the environment in which the reaction occurs. The reaction vectors are derived automatically from a database of reactions which is not restricted by size or reaction complexity. A structure generation algorithm has been developed whereby reaction vectors can be applied to previously unseen starting materials in order to suggest novel syntheses. The approach has been implemented in KNIME and is validated by reproducing known synthetic routes. We then present applications of the method in different drug design scenarios including lead optimization and library enumeration. The method offers great potential for capturing and using the growing body of data on reactions that is becoming available through electronic laboratory notebooks.

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.

Three field-based similarity methods are compared in retrospective virtual screening experiments. The methods are the CatShape module of CATALYST, ROCS, and an in-house program developed at the University of Sheffield called FBSS. The programs are used in both rigid and flexible searches carried out in the MDL Drug Data Report. UNITY 2D fingerprints are also used to provide a comparison with a more traditional approach to similarity searching, and similarity based on simple whole-molecule properties is used to provide a baseline for the more sophisticated searches. Overall, UNITY 2D fingerprints and ROCS with the chemical force field option gave comparable performance and were superior to the shape-only 3D methods. When the flexible methods were compared with the rigid methods, it was generally found that the flexible methods gave slightly better results than their respective rigid methods; however, the increased performance did not justify the additional computational cost required.

A system is being developed, based on structure generation methods, for the design of molecules to fit a variety of constraints. The system is modular and uses techniques from artificial intelligence to solve general problems in molecule design. Heuristics are being developed based, in part, on the different kinds of constraints that might be available to the system, e.g., the shape and electrostatic nature of a cavity. The heuristics are then used to restrict the combinatorial explosion that is inherent in structure generation. The program has been tested using the appa binding site of Trypsin and a range of different structures that fit the site have been generated. Some of these structures are very closely related to appa and the predictive value of the program has also been demonstrated.

An evolutionary statistical learning method was applied to classify drugs according to their biological target and also to discriminate between a compilation of oral and nonoral drugs. The emphasis was placed not only on how well the models predict but also on their interpretability. In an enhancement to previous studies, the consistency of the model weights over several runs of the genetic algorithm was considered with the goal of producing comprehensible models. Via this approach, the descriptors and their ranges that contribute most to class discrimination were identified. Selecting a bin step size that enables the average descriptor properties of the class being trained to be captured improves the interpretability and discriminatory power of a model. The performance, consistency, and robustness of such models were further enhanced by using two novel approaches that reduce the variability between individual solutions: consensus and splice modeling. Finally, the ability of the genetic algorithm to discriminate between activity classes was compared with a similarity searching method, while naive Bayes classifiers and support vector machines were applied in discriminating the oral and nonoral drugs.

The large numbers of compounds that are now available in drug discovery programmes have resulted in the need for methods to select compounds, both from external suppliers and in combinatorial library design procedures. In this article we describe a method that has been developed for scoring and ranking compounds according to their likelihood of exhibiting activity. The method can be used to determine the order in which compounds should be screened as well as to guide compound acquisition programmes. We then describe a series of experiments we have conducted that explore the benefits of designing combinatorial libraries through an analysis of product space rather than reactant space. The experiments are based on several different diversity metrics and molecular descriptors. We also show how product-based selection allows multi-objectives to be optimised simultaneously, for example, diversity and physicochemical properties, allowing the design of diverse and drug-like libraries.

This paper evaluates the effectiveness of various similarity coefficients for 2D similarity searching when multiple bioactive target structures are available. Similarity searches using several different activity classes within the MDL Drug Data Report and the Dictionary of Natural Products databases are performed using BCI 2D fingerprints. Using data fusion techniques to combine the resulting nearest neighbor lists we obtain group recall results which, in many cases, are a considerable improvement on standard average recall values obtained for individual structures. It is shown that the degree of improvement can be related to the structural diversity of the activity class that is searched for, the best results being found for the most diverse groups. The group recall of active compounds using subsets of the class is also investigated: for highly self-similar activity classes, the group recall improvement saturates well before the full activity class size is reached. A rough correlation is found between the relative improvement using the group recall and the square of the number of unique compounds available in all of the merged lists. The Tanimoto coefficient is found unambiguously to be the best coefficient to use for the recovery of active compounds using multiple targets. Furthermore, when using the Tanimoto coefficient, the "MAX" fusion rule is found to be more effective than the "SUM" rule for the combination of similarity searches from multiple targets. The use of group recall can lead to improved enrichment in database searches and virtual screening.

Reduced graphs provide summary representations of chemical structures. Here, a variety of different types of reduced graphs are compared in similarity searches. The reduced graphs are found to give comparable performance to Daylight fingerprints in terms of the number of active compounds retrieved. However, no one type of reduced graph is found to be consistently superior across a variety of different data sets. Consequently, a representative set of reduced graphs was chosen and used together with Daylight fingerprints in data fusion experiments. The results show improved performance in 10 out of 11 data sets compared to using Daylight fingerprints alone. Finally, the potential of using reduced graphs to build SAR models is demonstrated using recursive partitioning. An SAR model consistent with a published model is found following just two splits in the decision tree.

Reduced graphs provide summary representations of chemical structures. In this work, the effectiveness of reduced graphs for similarity searching is investigated. Different types of reduced graphs are introduced that aim to summarize features of structures that have the potential to form interactions with receptors while retaining the topology between the features. Similarity searches have been carried out across a variety of different activity classes. The effectiveness of the reduced graphs at retrieving compounds with the same activity as known target compounds is compared with searching using Daylight fingerprints. The reduced graphs are shown to be effective for similarity searching and to retrieve more diverse active compounds than those found using Daylight fingerprints; they thus represent a complementary similarity searching tool.

The last deacde has seen a revolutionary change in the processes used to discover novel bioactive compounds in the pharmaceutical and agrochemical industries. This change is due to the introduction of automation techniques which allow tens or hundreds of thousands compounds to be synthesised simultaneously and then to be screened for activity rapidly. These techniques of combinatorial synthesis and high throughput screening have vastly increased the throughput of the traditional structure-activity cycle. Despite the initial enthusiasm for the methods, early results have been disappointing, producing fewer hits than were expected or hits that have undesirable properties to be suitable as new drugs or agrochemicals. It is now realised that the number of compounds that could potentially be considered as new bioactive compounds is enormous compared to the numbers that can be handled in practise, even using automated techniques. Thus, efficient and effective methods are required for designing the sets of compounds to be used in combinatorial syntheses and to be screened in high throughput screening experiments. It is not possible to explore such large search spaces systematically and hence many methods have been developed for designing combinatorial libraries. Evolutionary algorithms are well suited to search for solutions to large combinatorial problems and this chapter reviews the application of genetic algorithms, a sub-branch of evolutionary algorithms, to combinatorial library design.

This paper reports how Zamora's smallest set of smallest rings algorithm has been modified and extended to provide an algorithm that will find the extended set of smallest rings (ESSR) for specific and structurally explicit generic structures. Modifications are necessary to find the ESSR rings within a partial structure connection table, while extensions are required to perceive the ESSR rings that span several such connection tables. Particular care is involved when structures containing primary cut faces or alternative embedments are processed

The rings perceived within and across the partial structures of structurally explicit generics are analyzed to produce ring-screen information that complements the existing fragment screens. This paper presents and discusses the format of these ring screens. The resultant bit vectors for each partial structure are accumulated to retain their logical relationships by means of a «bubble-up» of the vectors. The principles of the bubble-up are outlined and explained by means of several examples

The achievements of the project are reviewed briefly. They include: a) the design, implementation and testing of a representation which closely mirrors the characteristics of generic structures in patents, and which enables an internal representation to be created, b) the application of formal grammar theory to some of the problems caused by generic radical terms, and c) the development of a number of search representations including both fragments and ring descriptors, as well as reduced graph representations, together with appropriate search algorithms, for screening and atom- and bond-level searching.

The technology for storing and searching large databases of specific chemical substances is well established; public and private systems and services have operated successfully for a decade or more. Attention has now turned to more complex database types — to 3-D chemical structures, to biological macro- molecules, and to generic structures in patents. These all pose formidable problems in terms of the complexity of the data structures, the search algorithms, and the processing power required.