Advancing Cheminformatics-A Theme Issue in Honor of Professor Jürgen Bajorath.

While cheminformatics problems have been actively researched since the early 1960s, as witnessed by the QSAR approaches developed by Toshio Fujita and Corwin Hansch [...].

Prediction of activity cliffs on the basis of images using convolutional neural networks.

An activity cliff (AC) is formed by a pair of structurally similar compounds with a large difference in potency. Accordingly, ACs reveal structure-activity relationship (SAR) discontinuity and provide SAR information for compound optimization. Herein, we have investigated the question if ACs could be predicted from image data. Therefore, pairs of structural analogs were extracted from different compound activity classes that formed or did not form ACs. From these compound pairs, consistently formatted images were generated. Image sets were used to train and test convolutional neural network (CNN) models to systematically distinguish between ACs and non-ACs. The CNN models were found to predict ACs with overall high accuracy, as assessed using alternative performance measures, hence establishing proof-of-principle. Moreover, gradient weights from convolutional layers were mapped to test compounds and identified characteristic structural features that contributed to successful predictions. Weight-based feature visualization revealed the ability of CNN models to learn chemistry from images at a high level of resolution and aided in the interpretation of model decisions with intrinsic black box character.

Set-Theoretic Formalism for Treating Ligand-Target Datasets.

Data on ligand-target (LT) interactions has played a growing role in drug research for several decades. Even though the amount of data has grown significantly in size and coverage during this period, most datasets remain difficult to analyze because of their extreme sparsity, as there is no activity data whatsoever for many LT pairs. Even within clusters of data there tends to be a lack of data completeness, making the analysis of LT datasets problematic. The current effort extends earlier works on the development of set-theoretic formalisms for treating thresholded LT datasets. Unlike many approaches that do not address pairs of unknown interaction, the current work specifically takes account of their presence in addition to that of active and inactive pairs. Because a given LT pair can be in any one of three states, the binary logic of classical set-theoretic methods does not strictly apply. The current work develops a formalism, based on ternary set-theoretic relations, for treating thresholded LT datasets. It also describes an extension of the concept of data completeness, which is typically applied to sets of ligands and targets, to the local data completeness of individual ligands and targets. The set-theoretic formalism is applied to the analysis of simple and joint polypharmacologies based on LT activity profiles, and it is shown that null pairs provide a means for determining bounds to these values. The methodology is applied to a dataset of protein kinase inhibitors as an illustration of the method. Although not dealt with here, work is currently underway on a more refined treatment of activity values that is based on increasing the number of activity classes.

Automatic Identification of Analogue Series from Large Compound Data Sets: Methods and Applications.

Analogue series play a key role in drug discovery. They arise naturally in lead optimization efforts where analogues are explored based on one or a few core structures. However, it is much harder to accurately identify and extract pairs or series of analogue molecules in large compound databases with no predefined core structures. This methodological review outlines the most common and recent methodological developments to automatically identify analogue series in large libraries. Initial approaches focused on using predefined rules to extract scaffold structures, such as the popular Bemis-Murcko scaffold. Later on, the matched molecular pair concept led to efficient algorithms to identify similar compounds sharing a common core structure by exploring many putative scaffolds for each compound. Further developments of these ideas yielded, on the one hand, approaches for hierarchical scaffold decomposition and, on the other hand, algorithms for the extraction of analogue series based on single-site modifications (so-called matched molecular series) by exploring potential scaffold structures based on systematic molecule fragmentation. Eventually, further development of these approaches resulted in methods for extracting analogue series defined by a single core structure with several substitution sites that allow convenient representations, such as R-group tables. These methods enable the efficient analysis of large data sets with hundreds of thousands or even millions of compounds and have spawned many related methodological developments.

Role of Formulation Parameters on Intravitreal Dosing Accuracy Using 1 mL Hypodermic Syringes.

PURPOSE: Evaluation of product viscosity, density and aeration on the dose delivery and accuracy for intravitreal injections with commonly used commercially available hypodermic 1 mL syringes. METHODS: Six commercially available hypodermic 1 mL syringes with different specifications were used for the study. Syringes were filled with the test solutions with different densities and viscosities. Syringes were also subjected to shaking stress to introduce aeration in the test solutions in the presence of different surfactant concentrations with and without high antibody concentration. Target intravitreal volumes of 100 µL, 50 µL and 30 µL were tested to assess dosing accuracy in a controlled simulated administration setup using DIN ISO 11040-4 guidelines and Zwick/Roell Z010 TN instrument. RESULTS: With increasing product viscosity, higher volumes and hence doses were delivered especially for very low volumes like 50 µL and 30 µL. No impact of increasing product density was found on the delivered dose. The presence of surfactants or high protein concentration can lead to aeration, which also negatively affects the dose accuracy and precision. CONCLUSION: Formulation parameters like viscosity can have an impact on dose delivery using hypodermic syringes for intravitreal injections and on the resulting glide force.

Evaluation of Different Quality-Relevant Aspects of Closed System Transfer Devices (CSTDs).

PURPOSE: Health care professionals can be exposed to hazardous drugs such as cytostatics during preparation of drugs for administration. Closed sytem transfer devices (CSTDs) were introduced to provide protection for healthcare professional against unintended exposure to hazardous drugs. The interest in CSTDs has significantly increased after USP <800> monograph was issued. The majority of the studies published so far on CSTDs have focused on their "containment" function. However, other important attributes for CSTDs with potential importance for product quality impact are not yet fully evaluated. METHODS: In the current study, we evaluated four sytems from different suppliers, in combination with different container closure systems (CCS), using solutions of different viscosity and surface tension. The different CSTD / CCS combinations were tested for (a) containment (integrity) using a highly sensitive helium leak test, (b) the force required for mounting the vial adaptor, (c) contribution to visible and subvisible particles as well as (d) the hold-up volume. RESULTS: Results show that the majority of CSTDs may have leaks varying in size, and that some of them generated visible particles due to stopper coring and subvisible particles, both due to silicon oil and particulate contaminations of the Devices. Finally, the holdup volume was up to 1 mL depending on the CSTD type, vial size and solution viscosity. CONCLUSION: These results show that there is a need to evaluate the compatibility of CSTD systems to select the best system for the intended use and that CSTDs may adversely impact product quality and delivered dose.

From Qualitative to Quantitative Analysis of Activity and Property Landscapes.

Activity or, more generally, property landscapes (PLs) have been considered as an attractive way to visualize and explore structure-property relationships (SPRs) contained in large data sets of chemical compounds. For graphical analysis, three-dimensional representations reminiscent of natural landscapes are particularly intuitive. So far, the use of such landscape models has essentially been confined to qualitative assessment. We describe recent efforts to analyze PLs in a more quantitative manner, which make it possible to calculate topographical similarity values for comparison of landscape models as a measure of relative SPR information content.

Computational Method for Quantitative Comparison of Activity Landscapes on the Basis of Image Data.

Activity landscape (AL) models are used for visualizing and interpreting structure-activity relationships (SARs) in compound datasets. Therefore, ALs are designed to present chemical similarity and compound potency information in context. Different two- or three-dimensional (2D or 3D) AL representations have been introduced. For SAR analysis, 3D AL models are particularly intuitive. In these models, an interpolated potency surface is added as a third dimension to a 2D projection of chemical space. Accordingly, AL topology can be associated with characteristic SAR features. Going beyond visualization and a qualitative assessment of SARs, it would be very helpful to compare 3D ALs of different datasets in more quantitative terms. However, quantitative AL analysis is still in its infancy. Recently, it has been shown that 3D AL models with pre-defined topologies can be correctly classified using machine learning. Classification was facilitated on the basis of AL image feature representations learned with convolutional neural networks. Therefore, we have further investigated image analysis for quantitative comparison of 3D ALs and devised an approach to determine (dis)similarity relationships for ALs representing different compound datasets. Herein, we report this approach and demonstrate proof-of-principle. The methodology makes it possible to computationally compare 3D ALs and quantify topological differences reflecting varying SAR information content. For SAR exploration in drug design, this adds a quantitative measure of AL (dis)similarity to graphical analysis.

Systematic computational identification of promiscuity cliff pathways formed by inhibitors of the human kinome.

The ability of a small molecule to interact with multiple target proteins provides the molecular basis of polypharmacology. So-defined compound promiscuity is intensely investigated in drug discovery. For example, for kinase inhibitors, the interplay between target selectivity and promiscuity plays a decisive role for different therapeutic applications. The "promiscuity cliff" (PC) concept was introduced previously to aid in promiscuity analysis. A PC is defined as a pair of structurally similar compounds with a large difference in promiscuity. Accordingly, PCs can reveal small structural modifications that might be responsible for selectivity or multi-target activity. In network representations, PCs form clusters of varying size and complexity that are difficult to analyze interactively. Herein, we introduce a computational method to systematically identify PC pathways, which are particularly rich in structure-promiscuity information, and extract them from PC clusters. PC pathways provide informative templates for experimental design. In a proof-of-concept investigation, we have applied the new computational approach to systematically identify pathways in more than 600 PC clusters formed by inhibitors of the human kinome, demonstrating the utility of the method and revealing many interesting promiscuity patterns.

Influence of Varying Training Set Composition and Size on Support Vector Machine-Based Prediction of Active Compounds.

Support vector machine (SVM) modeling is one of the most popular machine learning approaches in chemoinformatics and drug design. The influence of training set composition and size on predictions currently is an underinvestigated issue in SVM modeling. In this study, we have derived SVM classification and ranking models for a variety of compound activity classes under systematic variation of the number of positive and negative training examples. With increasing numbers of negative training compounds, SVM classification calculations became increasingly accurate and stable. However, this was only the case if a required threshold of positive training examples was also reached. In addition, consideration of class weights and optimization of cost factors substantially aided in balancing the calculations for increasing numbers of negative training examples. Taken together, the results of our analysis have practical implications for SVM learning and the prediction of active compounds. For all compound classes under study, top recall performance and independence of compound recall of training set composition was achieved when 250-500 active and 500-1000 randomly selected inactive training instances were used. However, as long as ∼50 known active compounds were available for training, increasing numbers of 500-1000 randomly selected negative training examples significantly improved model performance and gave very similar results for different training sets.

Maximum common substructure-based Tversky index: an asymmetric hybrid similarity measure.

Current approaches for the assessment of molecular similarity can generally be divided into descriptor-based and substructure-based methods. The former require the application of similarity metrics that yield continuous similarity values, whereas the readout of the latter is binary (i.e. similar vs. not similar). However, it is also possible to combine descriptor-based and substructure-based methods to exploit advantages of individual methods in context and generate similarity measures for special applications. Herein we present a hybrid measure for asymmetric similarity calculations on the basis of maximum common core structures. This similarity function can be effectively applied to compare small reference compounds with larger test molecules, which is difficult using conventional metrics.

Lessons learned from the design of chemical space networks and opportunities for new applications.

The concept of chemical space is of fundamental relevance in chemical informatics and computer-aided drug discovery. In a series of articles published in the Journal of Computer-Aided Molecular Design, principles of chemical space design were evaluated, molecular networks proposed as an alternative to conventional coordinate-based chemical reference spaces, and different types of chemical space networks (CSNs) constructed and analyzed. Central to the generation of CSNs was the way in which molecular similarity relationships were assessed and a primary focal point was the network-based representation of biologically relevant chemical space. The design and comparison of CSNs based upon alternative similarity measures can be viewed as an evolutionary path with interesting lessons learned along the way. CSN design has matured to the point that such chemical space representations can be used in practice. In this contribution, highlights from the sequence of CSN design efforts are discussed in context, providing a perspective for future practical applications.

Design of chemical space networks on the basis of Tversky similarity.

Chemical space networks (CSNs) have been introduced as a coordinate-free representation of chemical space. In CSNs, nodes represent compounds and edges pairwise similarity relationships. These network representations are mostly used to navigate sections of biologically relevant chemical space. Different types of CSNs have been designed on the basis of alternative similarity measures including continuous numerical similarity values or substructure-based similarity criteria. CSNs can be characterized and compared on the basis of statistical concepts from network science. Herein, a new CSN design is introduced that is based upon asymmetric similarity assessment using the Tversky coefficient and termed TV-CSN. Compared to other CSNs, TV-CSNs have unique features. While CSNs typically contain separate compound communities and exhibit small world character, many TV-CSNs are also scale-free in nature and contain hubs, i.e., extensively connected central compounds. Compared to other CSNs, these hubs are a characteristic of TV-CSN topology. Hub-containing compound communities are of particular interest for the exploration of structure-activity relationships.

Design of chemical space networks using a Tanimoto similarity variant based upon maximum common substructures.

Chemical space networks (CSNs) have recently been introduced as an alternative to other coordinate-free and coordinate-based chemical space representations. In CSNs, nodes represent compounds and edges pairwise similarity relationships. In addition, nodes are annotated with compound property information such as biological activity. CSNs have been applied to view biologically relevant chemical space in comparison to random chemical space samples and found to display well-resolved topologies at low edge density levels. The way in which molecular similarity relationships are assessed is an important determinant of CSN topology. Previous CSN versions were based on numerical similarity functions or the assessment of substructure-based similarity. Herein, we report a new CSN design that is based upon combined numerical and substructure similarity evaluation. This has been facilitated by calculating numerical similarity values on the basis of maximum common substructures (MCSs) of compounds, leading to the introduction of MCS-based CSNs (MCS-CSNs). This CSN design combines advantages of continuous numerical similarity functions with a robust and chemically intuitive substructure-based assessment. Compared to earlier version of CSNs, MCS-CSNs are characterized by a further improved organization of local compound communities as exemplified by the delineation of drug-like subspaces in regions of biologically relevant chemical space.

Design and characterization of chemical space networks for different compound data sets.

Chemical Space Networks (CSNs) are generated for different compound data sets on the basis of pairwise similarity relationships. Such networks are thought to complement and further extend traditional coordinate-based views of chemical space. Our proof-of-concept study focuses on CSNs based upon fingerprint similarity relationships calculated using the conventional Tanimoto similarity metric. The resulting CSNs are characterized with statistical measures from network science and compared in different ways. We show that the homophily principle, which is widely considered in the context of social networks, is a major determinant of the topology of CSNs of bioactive compounds, designed as threshold networks, typically giving rise to community structures. Many properties of CSNs are influenced by numerical features of the conventional Tanimoto similarity metric and largely dominated by the edge density of the networks, which depends on chosen similarity threshold values. However, properties of different CSNs with constant edge density can be directly compared, revealing systematic differences between CSNs generated from randomly collected or bioactive compounds.

Comparison of bioactive chemical space networks generated using substructure- and fingerprint-based measures of molecular similarity.

Chemical space networks (CSNs) have recently been introduced as a conceptual alternative to coordinate-based representations of chemical space. CSNs were initially designed as threshold networks using the Tanimoto coefficient as a continuous similarity measure. The analysis of CSNs generated from sets of bioactive compounds revealed that many statistical properties were strongly dependent on their edge density. While it was difficult to compare CSNs at pre-defined similarity threshold values, CSNs with constant edge density were directly comparable. In the current study, alternative CSN representations were constructed by applying the matched molecular pair (MMP) formalism as a substructure-based similarity criterion. For more than 150 compound activity classes, MMP-based CSNs (MMP-CSNs) were compared to corresponding threshold CSNs (THR-CSNs) at a constant edge density by applying different parameters from network science, measures of community structure distributions, and indicators of structure-activity relationship (SAR) information content. MMP-CSNs were found to be an attractive alternative to THR-CSNs, yielding low edge densities and well-resolved topologies. MMP-CSNs and corresponding THR-CSNs often had similar topology and closely corresponding community structures, although there was only limited overlap in similarity relationships. The homophily principle from network science was shown to affect MMP-CSNs and THR-CSNs in different ways, despite the presence of conserved topological features. Moreover, activity cliff distributions in alternative CSN designs markedly differed, which has important implications for SAR analysis.

Introducing the 'active search' method for iterative virtual screening.

A method is introduced for sequential similarity searching for active compounds. Given a set of known actives and a screening database, a strategy is devised to optimally rank test compounds by observing the outcome of each iteration before selecting the next compound. This 'active search' approach is based upon Bayesian decision theory. A typical ranking procedure used in virtual compound screening corresponds to a myopic approximation to the optimal strategy. Exploratory active search represents a less-myopic approach and is shown to accurately identify a variety of active compounds in iterative virtual screening trials on 120 compound classes. Source code and data for the active search approach presented herein is made freely available.

Large-scale assessment of activity landscape feature probabilities of bioactive compounds.

Activity landscape representations integrate pairwise compound similarity and potency relationships and provide direct access to characteristic structure-activity relationship features in compound data sets. Because pairwise compound comparisons provide the foundation of activity landscape design, the assessment of specific landscape features such as activity cliffs has generally been confined to the level of compound pairs. A conditional probability-based approach has been applied herein to assign most probable activity landscape features to individual compounds. For example, for a given data set compound, it was determined if it would preferentially engage in the formation of activity cliffs or other landscape features. In a large-scale effort, we have determined conditional activity landscape feature probabilities for more than 160,000 compounds with well-defined activity annotations contained in 427 different target-based data sets. These landscape feature probabilities provide a detailed view of how different activity landscape features are distributed over currently available bioactive compounds.

Design of an activity landscape view taking compound-based feature probabilities into account.

Activity landscapes (ALs) of compound data sets are rationalized as graphical representations that integrate similarity and potency relationships between active compounds. ALs enable the visualization of structure-activity relationship (SAR) information and are thus computational tools of interest for medicinal chemistry. For AL generation, similarity and potency relationships are typically evaluated in a pairwise manner and major AL features are assessed at the level of compound pairs. In this study, we add a conditional probability formalism to AL design that makes it possible to quantify the probability of individual compounds to contribute to characteristic AL features. Making this information graphically accessible in a molecular network-based AL representation is shown to further increase AL information content and helps to quickly focus on SAR-informative compound subsets. This feature probability-based AL variant extends the current spectrum of AL representations for medicinal chemistry applications.

Chemoinformatic approaches for navigating large chemical spaces.

INTRODUCTION: Large chemical spaces (CSs) include traditional large compound collections, combinatorial libraries covering billions to trillions of molecules, DNA-encoded chemical libraries comprising complete combinatorial CSs in a single mixture, and virtual CSs explored by generative models. The diverse nature of these types of CSs require different chemoinformatic approaches for navigation. AREAS COVERED: An overview of different types of large CSs is provided. Molecular representations and similarity metrics suitable for large CS exploration are discussed. A summary of navigation of CSs in generative models is provided. Methods for characterizing and comparing CSs are discussed. EXPERT OPINION: The size of large CSs might restrict navigation to specialized algorithms and limit it to considering neighborhoods of structurally similar molecules. Efficient navigation of large CSs not only requires methods that scale with size but also requires smart approaches that focus on better but not necessarily larger molecule selections. Deep generative models aim to provide such approaches by implicitly learning features relevant for targeted biological properties. It is unclear whether these models can fulfill this ideal as validation is difficult as long as the covered CSs remain mainly virtual without experimental verification.