Large-scale similarity search profiling of ChEMBL compound data sets

Exploring Chemical Reaction Space with Machine Learning Models: Representation and Feature Perspective

Chemical reactions serve as foundational building blocks for organic chemistry and drug design. In the era of large AI models, data-driven approaches have emerged to innovate the design of novel reactions, optimize existing ones for higher yields, and discover new pathways for synthesizing chemical structures comprehensively. To effectively address these challenges with machine learning models, it is imperative to derive robust and informative representations or engage in feature engineering using extensive data sets of reactions. This work aims to provide a comprehensive review of established reaction featurization approaches, offering insights into the selection of representations and the design of features for a wide array of tasks. The advantages and limitations of employing SMILES, molecular fingerprints, molecular graphs, and physics-based properties are meticulously elaborated. Solutions to bridge the gap between different representations will also be critically evaluated. Additionally, we introduce a new frontier in chemical reaction pretraining, holding promise as an innovative yet unexplored avenue.

Effectiveness of molecular fingerprints for exploring the chemical space of natural products

Natural products are a diverse class of compounds with promising biological properties, such as high potency and excellent selectivity. However, they have different structural motifs than typical drug-like compounds, e.g., a wider range of molecular weight, multiple stereocenters and higher fraction of sp3-hybridized carbons. This makes the encoding of natural products via molecular fingerprints difficult, thus restricting their use in cheminformatics studies. To tackle this issue, we explored over 30 years of research to systematically evaluate which molecular fingerprint provides the best performance on the natural product chemical space. We considered 20 molecular fingerprints from four different sources, which we then benchmarked on over 100,000 unique natural products from the COCONUT (COlleCtion of Open Natural prodUcTs) and CMNPD (Comprehensive Marine Natural Products Database) databases. Our analysis focused on the correlation between different fingerprints and their classification performance on 12 bioactivity prediction datasets. Our results show that different encodings can provide fundamentally different views of the natural product chemical space, leading to substantial differences in pairwise similarity and performance. While Extended Connectivity Fingerprints are the de-facto option to encoding drug-like compounds, other fingerprints resulted to match or outperform them for bioactivity prediction of natural products. These results highlight the need to evaluate multiple fingerprinting algorithms for optimal performance and suggest new areas of research. Finally, we provide an open-source Python package for computing all molecular fingerprints considered in the study, as well as data and scripts necessary to reproduce the results, at https://github.com/dahvida/NP_Fingerprints .

FGFR1Pred: an artificial intelligence-based model for predicting fibroblast growth factor receptor 1 inhibitor

Fibroblast growth factor receptors (FGFRs) are a family of cell surface receptors that bind to fibroblast growth factor (FGF) and mediate various cellular functions (translocating proteins, tissue repair, cell proliferation, development, and differentiation) through complex signaling pathways. The FGFR1 growth receptor is essential in the pathogenesis of numerous malignancies, including but not limited to breast cancer, bladder cancer, hepatocellular carcinoma (HCC), and cholangiocarcinoma. The higher levels of FGFR1 expression on the surface of cancer cells cause overly active signaling, which leads to rapid cell proliferation, resulting in a high spread of cancer cells. The kinases that FGFR1 activates migrate across the cell nucleus, activating genes and kinase proteins necessary for the growth and survival of cancerous cells. Therefore, FGFR1 targeting shows therapeutic promise in some diseases, including cancer. Inhibitors of FGFR1s are being developed and studied for their potential to block aberrant FGFR1 signaling and inhibit cancer growth. Since the discovery of new FGFR1 inhibitors in the laboratory is difficult, expensive, time-consuming, and labor-intensive, only a small number of FGFR1 inhibitors have been approved by the FDA for use in the treatment of cancer. To accelerate drug discovery by efficiently exploring the vast chemical space, and identifying potential candidates with higher accuracy and reduced cost, we developed artificial intelligence (AI)-based prediction models for FGFR1 inhibitors using a dataset of 2356 chemical compounds. Four machine learning (ML) algorithms (SVM, RF, k-NN, and ANN) were used to train different prediction models based on molecular descriptors (1D and 2D, with and without molecular fingerprints). Among all trained models, the random forest (RF)-based prediction model achieved the highest accuracy on the training (98.9%), test (89.8%), and external test (90.3%) datasets. The developed inhibitor prediction model (FGFR1Pred) provides a valuable tool for identifying potential FGFR1 inhibitors, expediting the drug discovery process and ultimately facilitating the development of new therapeutics. The model is made available at https://github.com/PGlab-NIPER/FGFR1Pred.git.

Recent Advances in Computer-Aided Structure-Based Drug Design on Ion Channels

Ion channels play important roles in fundamental biological processes, such as electric signaling in cells, muscle contraction, hormone secretion, and regulation of the immune response. Targeting ion channels with drugs represents a treatment option for neurological and cardiovascular diseases, muscular degradation disorders, and pathologies related to disturbed pain sensation. While there are more than 300 different ion channels in the human organism, drugs have been developed only for some of them and currently available drugs lack selectivity. Computational approaches are an indispensable tool for drug discovery and can speed up, especially, the early development stages of lead identification and optimization. The number of molecular structures of ion channels has considerably increased over the last ten years, providing new opportunities for structure-based drug development. This review summarizes important knowledge about ion channel classification, structure, mechanisms, and pathology with the main focus on recent developments in the field of computer-aided, structure-based drug design on ion channels. We highlight studies that link structural data with modeling and chemoinformatic approaches for the identification and characterization of new molecules targeting ion channels. These approaches hold great potential to advance research on ion channel drugs in the future.

Maximizing the Performance of Similarity-Based Virtual Screening Methods by Generating Synergy from the Integration of 2D and 3D Approaches

Methods for the pairwise comparison of 2D and 3D molecular structures are established approaches in virtual screening. In this work, we explored three strategies for maximizing the virtual screening performance of these methods: (i) the merging of hit lists obtained from multi-compound screening using a single screening method, (ii) the merging of the hit lists obtained from 2D and 3D screening by parallel selection, and (iii) the combination of both of these strategies in an integrated approach. We found that any of these strategies led to a boost in virtual screening performance, with the clearest advantages observed for the integrated approach. On test sets for virtual screening, covering 50 pharmaceutically relevant proteins, the integrated approach, using sets of five query molecules, yielded, on average, an area under the receiver operating characteristic curve (AUC) of 0.84, an early enrichment among the top 1% of ranked compounds (EF1%) of 53.82 and a scaffold recovery rate among the top 1% of ranked compounds (SRR1%) of 0.50. In comparison, the 2D and 3D methods on their own (when using a single query molecule) yielded AUC values of 0.68 and 0.54, EF1% values of 19.96 and 17.52, and SRR1% values of 0.20 and 0.17, respectively. In conclusion, based on these results, the integration of 2D and 3D methods, via a (balanced) parallel selection strategy, is recommended, and, in particular, when combined with multi-query screening.

Differential Multimolecule Fingerprint for Similarity Search─Making Use of Active and Inactive Compound Sets in Virtual Screening

In conventional fingerprint methods, the similarity between two molecules is calculated using the Tanimoto index as a numerical criterion. Thus, the query molecules in virtual screening should be most representative of the wanted compound class at hand. In the concept introduced here, all available active molecules form a multimolecule fingerprint in which the appearing features are weighted according to their respective frequency. The features of inactive molecules are treated likewise and the resulting values are subtracted from those of the active ones. The obtained differential multimolecule fingerprint (DMMFP) is thus specific for the respective class of compounds. To account for the noninteger representation within this fingerprint, a modified Sørensen-Dice coefficient is used to compute the similarity. Potentially active molecules yield positive scores, whereas presumably inactive ones are denoted by negative values. The concept was applied to Angiotensin-converting enzyme (ACE) inhibitors, β2-adrenoceptor ligands, leukotriene A4 hydrolase inhibitors, dopamine D3 antagonists, and cytochrome CYP2C9 substrates, for which experimental binding affinities are known and was tested against decoys from DUD-E and a further background database consisting of molecules from the dark chemical matter, which comprises compounds that appear as frequent hitters across multiple assays. Using the 166 publicly available keys of the MACCS fingerprint and the larger PubChem fingerprint, actives were recovered with very high sensitivity. Furthermore, three marketed ACE inhibitors as well as the carbonic anhydrase II inhibitor dorzolamide were detected in the dark chemical matter data set. For comparison, the DMMFP was also used with a Bayesian classifier, for which the specificity (correctly classified inactives) and likewise the accuracy was superior. Conversely, the similarity score produced by the Sørensen-Dice coefficient showed its potential for the early recognition of (potentially) active molecules.

Introducing a Chemically Intuitive Core-Substituent Fingerprint Designed to Explore Structural Requirements for Effective Similarity Searching and Machine Learning

Fingerprint (FP) representations of chemical structure continue to be one of the most widely used types of molecular descriptors in chemoinformatics and computational medicinal chemistry. One often distinguishes between two- and three-dimensional (2D and 3D) FPs depending on whether they are derived from molecular graphs or conformations, respectively. Primary application areas for FPs include similarity searching and compound classification via machine learning, especially for hit identification. For these applications, 2D FPs are particularly popular, given their robustness and for the most part comparable (or better) performance to 3D FPs. While a variety of FP prototypes has been designed and evaluated during earlier times of chemoinformatics research, new developments have been rare over the past decade. At least in part, this has been due to the situation that topological (atom environment) FPs derived from molecular graphs have evolved as a gold standard in the field. We were interested in exploring the question of whether the amount of structural information captured by state-of-the-art 2D FPs is indeed required for effective similarity searching and compound classification or whether accounting for fewer structural features might be sufficient. Therefore, pursuing a "structural minimalist" approach, we designed and implemented a new 2D FP based upon ring and substituent fragments obtained by systematically decomposing large numbers of compounds from medicinal chemistry. The resulting FP termed core-substituent FP (CSFP) captures much smaller numbers of structural features than state-of-the-art 2D FPs. However, CSFP achieves high performance in similarity searching and machine learning, demonstrating that less structural information is required for establishing molecular similarity relationships than is often believed. Given its high performance and chemical tangibility, CSFP is also relevant for practical applications in medicinal chemistry.

Turbo Similarity Searching: Effect of Partial Ranking and Fusion Rules on ChEMBL Database

Turbo Similarity Searching (TSS) is the simplest and most recent chemical similarity searching (SS) approach, which improves the effectiveness of SS by performing a multi-target searching. TSS has four important elements, namely structural representation, similarity coefficient, number of nearest neighbours (NNs), and fusion rule, and any changes in these elements could affect the TSS results. A previous study suggested the advantage of using large numbers of reference compounds with small fractions of the database structures to obtain a better recall in group fusion. Therefore, this study aims to investigate the effect of partial ranking on TSS utilising different fusion rules and different numbers of NNs on the ChEMBL database and to evaluate whether these observations hold in TSS. Furthermore, the objective is to observe the effect of the indirect relationship feature of TSS on the partial ranking investigation. The results showed that the effect of using partial ranking on TSS was significant. This study also found that the performance of TSS improved as the database proportions used in the fusion process decreased and by using a small number of NNs. In addition, fusion rules based on reciprocal rank positions (RKP), maximum similarity score (sMAX), and sMNZ were superior to all the other fusion rules.

Prediction of Molecular Properties Using Molecular Topographic Map

Prediction of molecular properties plays a critical role towards rational drug design. In this study, the Molecular Topographic Map (MTM) is proposed, which is a two-dimensional (2D) map that can be used to represent a molecule. An MTM is generated from the atomic features set of a molecule using generative topographic mapping and is then used as input data for analyzing structure-property/activity relationships. In the visualization and classification of 20 amino acids, differences of the amino acids can be visually confirmed from and revealed by hierarchical clustering with a similarity matrix of their MTMs. The prediction of molecular properties was performed on the basis of convolutional neural networks using MTMs as input data. The performance of the predictive models using MTM was found to be equal to or better than that using Morgan fingerprint or MACCS keys. Furthermore, data augmentation of MTMs using mixup has improved the prediction performance. Since molecules converted to MTMs can be treated like 2D images, they can be easily used with existing neural networks for image recognition and related technologies. MTM can be effectively utilized to predict molecular properties of small molecules to aid drug discovery research.

Machine Learning in Discovery of New Antivirals and Optimization of Viral Infections Therapy

Nowadays, computational approaches play an important role in the design of new drug-like compounds and optimization of pharmacotherapeutic treatment of diseases. The emerging growth of viral infections, including those caused by the Human Immunodeficiency Virus (HIV), Ebola virus, recently detected coronavirus, and some others, leads to many newly infected people with a high risk of death or severe complications. A huge amount of chemical, biological, clinical data is at the disposal of the researchers. Therefore, there are many opportunities to find the relationships between the particular features of chemical data and the antiviral activity of biologically active compounds based on machine learning approaches. Biological and clinical data can also be used for building models to predict relationships between viral genotype and drug resistance, which might help determine the clinical outcome of treatment. In the current study, we consider machine-learning approaches in the antiviral research carried out during the past decade. We overview in detail the application of machine-learning methods for the design of new potential antiviral agents and vaccines, drug resistance prediction, and analysis of virus-host interactions. Our review also covers the perspectives of using the machine-learning approaches for antiviral research, including Dengue, Ebola viruses, Influenza A, Human Immunodeficiency Virus, coronaviruses, and some others.

Harnessing In Silico, In Vitro, and In Vivo Data to Understand the Toxicity Landscape of Polycyclic Aromatic Compounds (PACs)

Polycyclic aromatic compounds (PACs) are compounds with a minimum of two six-atom aromatic fused rings. PACs arise from incomplete combustion or thermal decomposition of organic matter and are ubiquitous in the environment. Within PACs, carcinogenicity is generally regarded to be the most important public health concern. However, toxicity in other systems (reproductive and developmental toxicity, immunotoxicity) has also been reported. Despite the large number of PACs identified in the environment, research attention to understand exposure and health effects of PACs has focused on a relatively limited subset, namely polycyclic aromatic hydrocarbons (PAHs), the PACs with only carbon and hydrogen atoms. To triage the rest of the vast number of PACs for more resource-intensive testing, we developed a data-driven approach to contextualize hazard characterization of PACs, by leveraging the available data from various data streams (in silico toxicity, in vitro activity, structural fingerprints, and in vivo data availability). The PACs were clustered on the basis of their in silico toxicity profiles containing predictions from 8 different categories (carcinogenicity, cardiotoxicity, developmental toxicity, genotoxicity, hepatotoxicity, neurotoxicity, reproductive toxicity, and urinary toxicity). We found that PACs with the same parent structure (e.g., fluorene) could have diverse in silico toxicity profiles. In contrast, PACs with similar substituted groups (e.g., alkylated-PAHs) or heterocyclics (e.g., N-PACs) with varying ring sizes could have similar in silico toxicity profiles, suggesting that these groups are better candidates for toxicity read-across analysis. The clusters/regions associated with certain in silico toxicity, in vitro activity, and structural fingerprints were identified. We found that genotoxicity/carcinogenicity (in silico toxicity) and xenobiotic homeostasis and stress response (in vitro activity), respectively, dominate the toxicity/activity variation seen in the PACs. The "hot spots" with enriched toxicity/activity in conjunction with availability of in vivo carcinogenicity data revealed regions of either data-poor (hydroxylated-PAHs) or data-rich (unsubstituted, parent PAHs) PACs. These regions offer potential targets for prioritization of further in vivo assessment and for chemical read-across efforts. The analysis results are searchable through an interactive web application (https://ntp.niehs.nih.gov/go/pacs_tableau), allowing for alternative hypothesis generation.

Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and Highlight Important Binding Interactions

Current deep learning methods for structure-based virtual screening take the structures of both the protein and the ligand as input but make little or no use of the protein structure when predicting ligand binding. Here, we show how a relatively simple method of data set augmentation forces such deep learning methods to take into account information from the protein. Models trained in this way are more generalizable (make better predictions on protein/ligand complexes from a different distribution to the training data). They also assign more meaningful importance to the protein and ligand atoms involved in binding. Overall, our results show that data set augmentation can help deep learning-based virtual screening to learn physical interactions rather than data set biases.

QSAR-derived affinity fingerprints (part 1): fingerprint construction and modeling performance for similarity searching, bioactivity classification and scaffold hopping

An affinity fingerprint is the vector consisting of compound’s affinity or potency against the reference panel of protein targets. Here, we present the QAFFP fingerprint, 440 elements long in silico QSAR-based affinity fingerprint, components of which are predicted by Random Forest regression models trained on bioactivity data from the ChEMBL database. Both real-valued (rv-QAFFP) and binary (b-QAFFP) versions of the QAFFP fingerprint were implemented and their performance in similarity searching, biological activity classification and scaffold hopping was assessed and compared to that of the 1024 bits long Morgan2 fingerprint (the RDKit implementation of the ECFP4 fingerprint). In both similarity searching and biological activity classification, the QAFFP fingerprint yields retrieval rates, measured by AUC (~ 0.65 and ~ 0.70 for similarity searching depending on data sets, and ~ 0.85 for classification) and EF5 (~ 4.67 and ~ 5.82 for similarity searching depending on data sets, and ~ 2.10 for classification), comparable to that of the Morgan2 fingerprint (similarity searching AUC of ~ 0.57 and ~ 0.66, and EF5 of ~ 4.09 and ~ 6.41, depending on data sets, classification AUC of ~ 0.87, and EF5 of ~ 2.16). However, the QAFFP fingerprint outperforms the Morgan2 fingerprint in scaffold hopping as it is able to retrieve 1146 out of existing 1749 scaffolds, while the Morgan2 fingerprint reveals only 864 scaffolds.

Identifying the Structural Features of Diphenyl Ether Analogues for InhA Inhibition: A 2D and 3D QSAR Based Study

Background:

Enoyl acyl carrier protein reductase (InhA) is a validated target for Mycobacterium. It is an enzyme which is associated with the biosynthesis of mycolic acids in type II fatty acid synthase system. Mycobacterial cell wall majorly comprises mycolic acids, which are responsible for virulence of the microorganism. Several diphenyl ether derivatives have been known to be direct inhibitors of InhA.

Objective:

In the present work, a Quantitative Structure Activity Relationship (QSAR) study was performed to identify the structural features of diphenyl ether analogues which contribute to InhA inhibitory activity in a favourable way.

Method:

Both 2D and 3D QSAR models were built and compared. Several fingerprint based 2D QSAR models were generated and their relationship with the structural features was studied. Models which corroborated the inhibitory activity of the molecules with their structural features were selected and studied in detail.

Results:

A 2D-QSAR model, with dendritic fingerprints having regression coefficient, for test set molecules Q2 =0.8132 and for the training set molecules, R2 =0.9607 was obtained. Additionally, an atom-based 3D-QSAR model with Q2 =0.7697 and R2 =0.9159 was also constructed.

Conclusion:

The data reported by various models provides guidance for the designing of structurally new diphenyl ether inhibitors with potential activity against InhA of M. tuberculosis.

Identification of Target Associations for Polypharmacology from Analysis of Crystallographic Ligands of the Protein Data Bank

The design of a chemical entity that potently and selectively binds to a biological target of therapeutic relevance has dominated the scene of drug discovery so far. However, recent findings suggest that multitarget ligands may be endowed with superior efficacy and be less prone to drug resistance. The Protein Data Bank (PDB) provides experimentally validated structural information about targets and bound ligands. Therefore, it represents a valuable source of information to help identifying active sites, understanding pharmacophore requirements, designing novel ligands, and inferring structure-activity relationships. In this study, we performed a large-scale analysis of the PDB by integrating different ligand-based and structure-based approaches, with the aim of identifying promising target associations for polypharmacology based on reported crystal structure information. First, the 2D and 3D similarity profiles of the crystallographic ligands were evaluated using different ligand-based methods. Then, activity data of pairs of similar ligands binding to different targets were inspected by comparing structural information with bioactivity annotations reported in the ChEMBL, BindingDB, BindingMOAD, and PDBbind databases. Afterward, extensive docking screenings of ligands in the identified cross-targets were made in order to validate and refine the ligand-based results. Finally, the therapeutic relevance of the identified target combinations for polypharmacology was evaluated from comparison with information on therapeutic targets reported in the Therapeutic Target Database (TTD). The results led to the identification of several target associations with high therapeutic potential for polypharmacology.

Perturbation Theory Machine Learning Modeling of Immunotoxicity for Drugs Targeting Inflammatory Cytokines and Study of the Antimicrobial G1 Using Cytometric Bead Arrays

ChEMBL biological activities prediction for 1-5-bromofur-2-il-2-bromo-2-nitroethene (G1) is a difficult task for cytokine immunotoxicity. The current study presents experimental results for G1 interaction with mouse Th1/Th2 and pro-inflammatory cytokines using a cytometry bead array (CBA). In the in vitro test of CBA, the results show no significant differences between the mean values of the Th1/Th2 cytokines for the samples treated with G1 with respect to the negative control, but there are moderate differences for cytokine values between different periods (24/48 h). The experiments show no significant differences between the mean values of the pro-inflammatory cytokines for the samples treated with G1, regarding the negative control, except for the values of tumor necrosis factor (TNF) and Interleukin (IL6) between the group treated with G1 and the negative control at 48 h. Differences occur for these cytokines in the periods (24/48 h). The study confirmed that the antimicrobial G1 did not alter the Th1/Th2 cytokines concentration in vitro in different periods, but it can alter TNF and IL6. G1 promotes free radicals production and activates damage processes in macrophages culture. In order to predict all ChEMBL activities for drugs in other experimental conditions, a ChEMBL data set was constructed using 25 biological activities, 1366 assays, 2 assay types, 4 assay organisms, 2 organisms, and 12 cytokine targets. Molecular descriptors calculated with Rcpi and 15 machine learning methods were used to find the best model able to predict if a drug could be active or not against a specific cytokine, in specific experimental conditions. The best model is based on 120 selected molecular descriptors and a deep neural network with area under the curve of the receiver operating characteristic of 0.904 and accuracy of 0.832. This model predicted 1384 G1 biological activities against cytokines in all ChEMBL data set experimental conditions.

Evaluation of Cross-Validation Strategies in Sequence-Based Binding Prediction Using Deep Learning

Binding prediction between targets and drug-like compounds through deep neural networks has generated promising results in recent years, outperforming traditional machine learning-based methods. However, the generalization capability of these classification models is still an issue to be addressed. In this work, we explored how different cross-validation strategies applied to data from different molecular databases affect to the performance of binding prediction proteochemometrics models. These strategies are (1) random splitting, (2) splitting based on K-means clustering (both of actives and inactives), (3) splitting based on source database, and (4) splitting based both in the clustering and in the source database. These schemas are applied to a deep learning proteochemometrics model and to a simple logistic regression model to be used as baseline. Additionally, two different ways of describing molecules in the model are tested: (1) by their SMILES and (2) by three fingerprints. The classification performance of our deep learning-based proteochemometrics model is comparable to the state of the art. Our results show that the lack of generalization of these models is due to a bias in public molecular databases and that a restrictive cross-validation schema based on compound clustering leads to worse but more robust and credible results. Our results also show better performance when representing molecules by their fingerprints.

Statistical-based database fingerprint: chemical space dependent representation of compound databases

Background

Simplified representation of compound databases has several applications in cheminformatics. Herein, we introduce an alternative and general method to build single fingerprint representations of compound databases. The approach is inspired on the previously published modal fingerprints that are aimed to capture the most significant bits of a fingerprint representation for a compound data set. The novelty of the herein proposed statistical-based database fingerprint (SB-DFP) is that it is generated based on binomial proportions comparisons taking as reference the distribution of “1” bits on a large representative set of the chemical space.

Results

To illustrate the Method, SB-DFPs were constructed for 28 epigenetic target data sets retrieved from a recently published epigenomics database of interest in probe and drug discovery. For each target data set, the SB-DFPs were built based on two representative fingerprints of different design using as reference a data set with more than 15 million compounds from ZINC. The application of SB-DFP was illustrated and compared to other methods through association relationships of the 28 epigenetic data sets and similarity searching. It was found that SB-DFPs captured overall, the common features between data sets and the distinct features of each set. In similarity searching SB-DFP equaled or outperformed other approaches for at least 20 out of the 28 sets.

Conclusions

SB-DFP is a general approach based on binomial proportion comparisons to represent a compound data set with a single fingerprint. SB-DFP can be developed, at least in principle, based on any fingerprint and reference data set. SB-DFP is a good alternative for exploration of relationships between targets through its associated compound data sets and performing similarity searching.

Electronic supplementary material

The online version of this article (10.1186/s13321-018-0311-x) contains supplementary material, which is available to authorized users.

Machine learning in chemoinformatics and drug discovery

Chemoinformatics is an established discipline focusing on extracting, processing and extrapolating meaningful data from chemical structures. With the rapid explosion of chemical 'big' data from HTS and combinatorial synthesis, machine learning has become an indispensable tool for drug designers to mine chemical information from large compound databases to design drugs with important biological properties. To process the chemical data, we first reviewed multiple processing layers in the chemoinformatics pipeline followed by the introduction of commonly used machine learning models in drug discovery and QSAR analysis. Here, we present basic principles and recent case studies to demonstrate the utility of machine learning techniques in chemoinformatics analyses; and we discuss limitations and future directions to guide further development in this evolving field.

Cross-Classified Multilevel Modelling of the Effectiveness of Similarity-Based Virtual Screening

The screening effectiveness of a chemical similarity search depends on a range of factors, including the bioactivity of interest, the types of similarity coefficient and fingerprint that comprise the similarity measure, and the nature of the reference structure that is being searched against a database. This study introduces the use of cross-classified multilevel modelling as a way to investigate the relative importance of these four factors when carrying out similarity searches on the ChEMBL database. Two principal conclusions can be drawn from the analyses: that the fingerprint plays a more important role than the similarity coefficient in determining the effectiveness of a similarity search, and that comparative studies of similarity measures should involve many more reference structures than has been the case in much previous work.

Shallow Representation Learning via Kernel PCA Improves QSAR Modelability

Linear models offer a robust, flexible, and computationally efficient set of tools for modeling quantitative structure-activity relationships (QSARs) but have been eclipsed in performance by nonlinear methods. Support vector machines (SVMs) and neural networks are currently among the most popular and accurate QSAR methods because they learn new representations of the data that greatly improve modelability. In this work, we use shallow representation learning to improve the accuracy of L1 regularized logistic regression (LASSO) and meet the performance of Tanimoto SVM. We embedded chemical fingerprints in Euclidean space using Tanimoto (a.k.a. Jaccard) similarity kernel principal component analysis (KPCA) and compared the effects on LASSO and SVM model performance for predicting the binding activities of chemical compounds against 102 virtual screening targets. We observed similar performance and patterns of improvement for LASSO and SVM. We also empirically measured model training and cross-validation times to show that KPCA used in concert with LASSO classification is significantly faster than linear SVM over a wide range of training set sizes. Our work shows that powerful linear QSAR methods can match nonlinear methods and demonstrates a modular approach to nonlinear classification that greatly enhances QSAR model prototyping facility, flexibility, and transferability.

Protein-Ligand Scoring with Convolutional Neural Networks

Computational approaches to drug discovery can reduce the time and cost associated with experimental assays and enable the screening of novel chemotypes. Structure-based drug design methods rely on scoring functions to rank and predict binding affinities and poses. The ever-expanding amount of protein-ligand binding and structural data enables the use of deep machine learning techniques for protein-ligand scoring. We describe convolutional neural network (CNN) scoring functions that take as input a comprehensive three-dimensional (3D) representation of a protein-ligand interaction. A CNN scoring function automatically learns the key features of protein-ligand interactions that correlate with binding. We train and optimize our CNN scoring functions to discriminate between correct and incorrect binding poses and known binders and nonbinders. We find that our CNN scoring function outperforms the AutoDock Vina scoring function when ranking poses both for pose prediction and virtual screening.

Flexible Analog Search with Kernel PCA Embedded Molecule Vectors

Studying analog series to find structural transformations that enhance the activity and ADME properties of lead compounds is an important part of drug development. Matched molecular pair (MMP) search is a powerful tool for analog analysis that imitates researchers' ability to select pairs of compounds that differ only by small well-defined transformations. Abstraction is a challenge for existing MMP search algorithms, which can result in the omission of relevant, inexact MMPs, and inclusion of irrelevant, contextually dissimilar MMPs. In this work, we present a new method for MMP search that returns approximate results and enables flexible control over abstraction of contextual information. We illustrate the concepts and mechanics of our method with a series of exemplar MMP queries, and then benchmark search accuracy using MMPs found by fragment indexing. We show that we can search for MMPs in a context dependent manner, and accurately approximate context independent fragment index based MMP search over a range of fingerprint and dataset conditions. Our method can be used to search for pairwise correspondences among analog sets and bolster MMP datasets where data is missing or incomplete.

Molecular Similarity Concepts for Informatics Applications

The assessment of small molecule similarity is a central task in chemoinformatics and medicinal chemistry. A variety of molecular representations and metrics are applied to computationally evaluate and quantify molecular similarity. A critically important aspect of molecular similarity analysis in chemoinformatics and pharmaceutical research is that one is typically not interested in quantifying the degree of structural or chemical similarity between compounds per se, but rather in extrapolating from molecular similarity to property similarity. In other words, one assumes that there is a correlation between calculated similarity and specific properties of small molecules including, first and foremost, biological activities. Although similarity is a priori a subjective concept, and difficult to quantify, it must computationally be assessed in a formally consistent manner. Otherwise, there is little utility of similarity calculations. Consistent treatment requires approximations to be made and the consideration of alternative computational similarity concepts, as discussed herein.

Filtered circular fingerprints improve either prediction or runtime performance while retaining interpretability

BACKGROUND

Even though circular fingerprints have been first introduced more than 50 years ago, they are still widely used for building highly predictive, state-of-the-art (Q)SAR models. Historically, these structural fragments were designed to search large molecular databases. Hence, to derive a compact representation, circular fingerprint fragments are often folded to comparatively short bit-strings. However, folding fingerprints introduces bit collisions, and therefore adds noise to the encoded structural information and removes its interpretability. Both representations, folded as well as unprocessed fingerprints, are often used for (Q)SAR modeling.

RESULTS

We show that it can be preferable to build (Q)SAR models with circular fingerprint fragments that have been filtered by supervised feature selection, instead of applying folded or all fragments. Compared to folded fingerprints, filtered fingerprints significantly increase predictive performance and remain unambiguous and interpretable. Compared to unprocessed fingerprints, filtered fingerprints reduce the computational effort and are a more compact and less redundant feature representation. Depending on the selected learning algorithm filtering yields about equally predictive (Q)SAR models. We demonstrate the suitability of filtered fingerprints for (Q)SAR modeling by presenting our freely available web service Collision-free Filtered Circular Fingerprints that provides rationales for predictions by highlighting important structural features in the query compound (see http://coffer.informatik.uni-mainz.de).

CONCLUSIONS

Circular fingerprints are potent structural features that yield highly predictive models and encode interpretable structural information. However, to not lose interpretability, circular fingerprints should not be folded when building prediction models. Our experiments show that filtering is a suitable option to reduce the high computational effort when working with all fingerprint fragments. Additionally, our experiments suggest that the area under precision recall curve is a more sensible statistic for validating (Q)SAR models for virtual screening than the area under ROC or other measures for early recognition.

GRAPHICAL ABSTRACT

Comparing structural fingerprints using a literature-based similarity benchmark

Background

The concept of molecular similarity is one of the central ideas in cheminformatics, despite the fact that it is ill-defined and rather difficult to assess objectively. Here we propose a practical definition of molecular similarity in the context of drug discovery: molecules A and B are similar if a medicinal chemist would be likely to synthesise and test them around the same time as part of the same medicinal chemistry program. The attraction of such a definition is that it matches one of the key uses of similarity measures in early-stage drug discovery. If we make the assumption that molecules in the same compound activity table in a medicinal chemistry paper were considered similar by the authors of the paper, we can create a dataset of similar molecules from the medicinal chemistry literature. Furthermore, molecules with decreasing levels of similarity to a reference can be found by either ordering molecules in an activity table by their activity, or by considering activity tables in different papers which have at least one molecule in common.

Results

Using this procedure with activity data from ChEMBL, we have created two benchmark datasets for structural similarity that can be used to guide the development of improved measures. Compared to similar results from a virtual screen, these benchmarks are an order of magnitude more sensitive to differences between fingerprints both because of their size and because they avoid loss of statistical power due to the use of mean scores or ranks. We measure the performance of 28 different fingerprints on the benchmark sets and compare the results to those from the Riniker and Landrum (J Cheminf 5:26, 2013. doi:10.1186/1758-2946-5-26) ligand-based virtual screening benchmark.

Conclusions

Extended-connectivity fingerprints of diameter 4 and 6 are among the best performing fingerprints when ranking diverse structures by similarity, as is the topological torsion fingerprint. However, when ranking very close analogues, the atom pair fingerprint outperforms the others tested. When ranking diverse structures or carrying out a virtual screen, we find that the performance of the ECFP fingerprints significantly improves if the bit-vector length is increased from 1024 to 16,384.Graphical abstract

Activity-relevant similarity values for fingerprints and implications for similarity searching

A largely unsolved problem in chemoinformatics is the issue of how calculated compound similarity relates to activity similarity, which is central to many applications. In general, activity relationships are predicted from calculated similarity values. However, there is no solid scientific foundation to bridge between calculated molecular and observed activity similarity. Accordingly, the success rate of identifying new active compounds by similarity searching is limited. Although various attempts have been made to establish relationships between calculated fingerprint similarity values and biological activities, none of these has yielded generally applicable rules for similarity searching. In this study, we have addressed the question of molecular versus activity similarity in a more fundamental way. First, we have evaluated if activity-relevant similarity value ranges could in principle be identified for standard fingerprints and distinguished from similarity resulting from random compound comparisons. Then, we have analyzed if activity-relevant similarity values could be used to guide typical similarity search calculations aiming to identify active compounds in databases. It was found that activity-relevant similarity values can be identified as a characteristic feature of fingerprints. However, it was also shown that such values cannot be reliably used as thresholds for practical similarity search calculations. In addition, the analysis presented herein helped to rationalize differences in fingerprint search performance.

Activity-relevant similarity values for fingerprints and implications for similarity searching

A largely unsolved problem in chemoinformatics is the issue of how calculated compound similarity relates to activity similarity, which is central to many applications. In general, activity relationships are predicted from calculated similarity values. However, there is no solid scientific foundation to bridge between calculated molecular and observed activity similarity. Accordingly, the success rate of identifying new active compounds by similarity searching is limited. Although various attempts have been made to establish relationships between calculated fingerprint similarity values and biological activities, none of these has yielded generally applicable rules for similarity searching. In this study, we have addressed the question of molecular versus activity similarity in a more fundamental way. First, we have evaluated if activity-relevant similarity value ranges could in principle be identified for standard fingerprints and distinguished from similarity resulting from random compound comparisons. Then, we have analyzed if activity-relevant similarity values could be used to guide typical similarity search calculations aiming to identify active compounds in databases. It was found that activity-relevant similarity values can be identified as a characteristic feature of fingerprints. However, it was also shown that such values cannot be reliably used as thresholds for practical similarity search calculations. In addition, the analysis presented herein helped to rationalize differences in fingerprint search performance.

Accurate and efficient target prediction using a potency-sensitive influence-relevance voter

BACKGROUND

A number of algorithms have been proposed to predict the biological targets of diverse molecules. Some are structure-based, but the most common are ligand-based and use chemical fingerprints and the notion of chemical similarity. These methods tend to be computationally faster than others, making them particularly attractive tools as the amount of available data grows.

RESULTS

Using a ChEMBL-derived database covering 490,760 molecule-protein interactions and 3236 protein targets, we conduct a large-scale assessment of the performance of several target-prediction algorithms at predicting drug-target activity. We assess algorithm performance using three validation procedures: standard tenfold cross-validation, tenfold cross-validation in a simulated screen that includes random inactive molecules, and validation on an external test set composed of molecules not present in our database.

CONCLUSIONS

We present two improvements over current practice. First, using a modified version of the influence-relevance voter (IRV), we show that using molecule potency data can improve target prediction. Second, we demonstrate that random inactive molecules added during training can boost the accuracy of several algorithms in realistic target-prediction experiments. Our potency-sensitive version of the IRV (PS-IRV) obtains the best results on large test sets in most of the experiments. Models and software are publicly accessible through the chemoinformatics portal at http://chemdb.ics.uci.edu/.

Target enhanced 2D similarity search by using explicit biological activity annotations and profiles

BACKGROUND

The enriched biological activity information of compounds in large and freely-accessible chemical databases like the PubChem Bioassay Database has become a powerful research resource for the scientific research community. Currently, 2D fingerprint based conventional similarity search (CSS) is the most common widely used approach for database screening, but it does not typically incorporate the relative importance of fingerprint bits to biological activity.

RESULTS

In this study, a large-scale similarity search investigation has been carried out on 208 well-defined compound activity classes extracted from PubChem Bioassay Database. An analysis was performed to compare the search performance of three types of 2D similarity search approaches: 2D fingerprint based conventional similarity search approach (CSS), iterative similarity search approach with multiple active compounds as references (ISS), and fingerprint based iterative similarity search with classification (ISC), which can be regarded as the combination of iterative similarity search with active references and a reversed iterative similarity search with inactive references. Compared to the search results returned by CSS, ISS improves recall but not precision. Although ISC causes the false rejection of active hits, it improves the precision with statistical significance, and outperforms both ISS and CSS. In a second part of this study, we introduce the profile concept into the three types of searches. We find that the profile based non-iterative search can significantly improve the search performance by increasing the recall rate. We also find that profile based ISS (PBISS) and profile based ISC (PBISC) significantly decreases ISS search time without sacrificing search performance.

CONCLUSIONS

On the basis of our large-scale investigation directed against a wide spectrum of pharmaceutical targets, we conclude that ISC and ISS searches perform better than 2D fingerprint similarity searching and that profile based versions of these algorithms do nearly as well in less time. We also suggest that the profile version of the iterative similarity searches are both better performing and potentially quicker than the standard algorithm.

Mapping chemical structure-activity information of HAART-drug cocktails over complex networks of AIDS epidemiology and socioeconomic data of U.S. counties

Using computational algorithms to design tailored drug cocktails for highly active antiretroviral therapy (HAART) on specific populations is a goal of major importance for both pharmaceutical industry and public health policy institutions. New combinations of compounds need to be predicted in order to design HAART cocktails. On the one hand, there are the biomolecular factors related to the drugs in the cocktail (experimental measure, chemical structure, drug target, assay organisms, etc.); on the other hand, there are the socioeconomic factors of the specific population (income inequalities, employment levels, fiscal pressure, education, migration, population structure, etc.) to study the relationship between the socioeconomic status and the disease. In this context, machine learning algorithms, able to seek models for problems with multi-source data, have to be used. In this work, the first artificial neural network (ANN) model is proposed for the prediction of HAART cocktails, to halt AIDS on epidemic networks of U.S. counties using information indices that codify both biomolecular and several socioeconomic factors. The data was obtained from at least three major sources. The first dataset included assays of anti-HIV chemical compounds released to ChEMBL. The second dataset is the AIDSVu database of Emory University. AIDSVu compiled AIDS prevalence for >2300 U.S. counties. The third data set included socioeconomic data from the U.S. Census Bureau. Three scales or levels were employed to group the counties according to the location or population structure codes: state, rural urban continuum code (RUCC) and urban influence code (UIC). An analysis of >130,000 pairs (network links) was performed, corresponding to AIDS prevalence in 2310 counties in U.S. vs. drug cocktails made up of combinations of ChEMBL results for 21,582 unique drugs, 9 viral or human protein targets, 4856 protocols, and 10 possible experimental measures. The best model found with the original data was a linear neural network (LNN) with AUROC>0.80 and accuracy, specificity, and sensitivity≈77% in training and external validation series. The change of the spatial and population structure scale (State, UIC, or RUCC codes) does not affect the quality of the model. Unbalance was detected in all the models found comparing positive/negative cases and linear/non-linear model accuracy ratios. Using synthetic minority over-sampling technique (SMOTE), data pre-processing and machine-learning algorithms implemented into the WEKA software, more balanced models were found. In particular, a multilayer perceptron (MLP) with AUROC=97.4% and precision, recall, and F-measure >90% was found.

Ligand-based target prediction with signature fingerprints

When evaluating a potential drug candidate it is desirable to predict target interactions in silico prior to synthesis in order to assess, e.g., secondary pharmacology. This can be done by looking at known target binding profiles of similar compounds using chemical similarity searching. The purpose of this study was to construct and evaluate the performance of chemical fingerprints based on the molecular signature descriptor for performing target binding predictions. For the comparison we used the area under the receiver operating characteristics curve (AUC) complemented with net reclassification improvement (NRI). We created two open source signature fingerprints, a bit and a count version, and evaluated their performance compared to a set of established fingerprints with regards to predictions of binding targets using Tanimoto-based similarity searching on publicly available data sets extracted from ChEMBL. The results showed that the count version of the signature fingerprint performed on par with well-established fingerprints such as ECFP. The count version outperformed the bit version slightly; however, the count version is more complex and takes more computing time and memory to run so its usage should probably be evaluated on a case-by-case basis. The NRI based tests complemented the AUC based ones and showed signs of higher power.

Prediction of multi-target networks of neuroprotective compounds with entropy indices and synthesis, assay, and theoretical study of new asymmetric 1,2-rasagiline carbamates

In a multi-target complex network, the links (L_ij) represent the interactions between the drug (d_i) and the target (t_j), characterized by different experimental measures (K_i, K_m, IC₅₀, etc.) obtained in pharmacological assays under diverse boundary conditions (c_j). In this work, we handle Shannon entropy measures for developing a model encompassing a multi-target network of neuroprotective/neurotoxic compounds reported in the CHEMBL database. The model predicts correctly >8300 experimental outcomes with Accuracy, Specificity, and Sensitivity above 80%–90% on training and external validation series. Indeed, the model can calculate different outcomes for >30 experimental measures in >400 different experimental protocolsin relation with >150 molecular and cellular targets on 11 different organisms (including human). Hereafter, we reported by the first time the synthesis, characterization, and experimental assays of a new series of chiral 1,2-rasagiline carbamate derivatives not reported in previous works. The experimental tests included: (1) assay in absence of neurotoxic agents; (2) in the presence of glutamate; and (3) in the presence of H₂O₂. Lastly, we used the new Assessing Links with Moving Averages (ALMA)-entropy model to predict possible outcomes for the new compounds in a high number of pharmacological tests not carried out experimentally.

Asymmetric clustering index in a case study of 5-HT1A receptor ligands

The automatic clustering of chemical compounds is an important branch of chemoinformatics. In this paper the Asymmetric Clustering Index (Aci) is proposed to assess how well an automatically created partition reflects the reference. The asymmetry allows for a distinction between the fixed reference and the numerically constructed partition. The introduced index is applied to evaluate the quality of hierarchical clustering procedures for 5-HT_1A receptor ligands. We find that the most appropriate combination of parameters for the hierarchical clustering of compounds with a determined activity for this biological target is the Klekota Roth fingerprint combined with the complete linkage function and the Buser similarity metric.

Follow up: Compound data sets and software tools for chemoinformatics and medicinal chemistry applications: update and data transfer

In 2012, we reported 30 compound data sets and/or programs developed in our laboratory in a data article and made them freely available to the scientific community to support chemoinformatics and computational medicinal chemistry applications. These data sets and computational tools were provided for download from our website. Since publication of this data article, we have generated 13 new data sets with which we further extend our collection of publicly available data and tools. Due to changes in web servers and website architectures, data accessibility has recently been limited at times. Therefore, we have also transferred our data sets and tools to a public repository to ensure full and stable accessibility. To aid in data selection, we have classified the data sets according to scientific subject areas. Herein, we describe new data sets, introduce the data organization scheme, summarize the database content and provide detailed access information in ZENODO (doi: 10.5281/zenodo.8451 and doi:10.5281/zenodo.8455).

Improving structural similarity based virtual screening using background knowledge

BACKGROUND

Virtual screening in the form of similarity rankings is often applied in the early drug discovery process to rank and prioritize compounds from a database. This similarity ranking can be achieved with structural similarity measures. However, their general nature can lead to insufficient performance in some application cases. In this paper, we provide a link between ranking-based virtual screening and fragment-based data mining methods. The inclusion of binding-relevant background knowledge into a structural similarity measure improves the quality of the similarity rankings. This background knowledge in the form of binding relevant substructures can either be derived by hand selection or by automated fragment-based data mining methods.

RESULTS

In virtual screening experiments we show that our approach clearly improves enrichment factors with both applied variants of our approach: the extension of the structural similarity measure with background knowledge in the form of a hand-selected relevant substructure or the extension of the similarity measure with background knowledge derived with data mining methods.

CONCLUSION

Our study shows that adding binding relevant background knowledge can lead to significantly improved similarity rankings in virtual screening and that even basic data mining approaches can lead to competitive results making hand-selection of the background knowledge less crucial. This is especially important in drug discovery and development projects where no receptor structure is available or more frequently no verified binding mode is known and mostly ligand based approaches can be applied to generate hit compounds.

Molecular crime scene investigation - dusting for fingerprints

In chemoinformatics and drug design, fingerprints (FPs) are defined as string representations of molecular structure and properties and are popular descriptors for similarity searching. FPs are generally characterized by the simplicity of their design and ease of use. Despite a long history in chemoinformatics, the potential and limitations of FP searching are often not well under- stood. Standard FPs can also be subjected to engineering techniques to tune them for specific search applications.

S. Cordeiro MN, Yañez M, González-Díaz H, García-Mera X. Model for high-throughput screening of multitarget drugs in chemical neurosciences: synthesis, assay, and theoretic study of rasagiline carbamates

The disappointing results obtained in recent clinical trials renew the interest in experimental/computational techniques for the discovery of neuroprotective drugs. In this context, multitarget or multiplexing QSAR models (mt-QSAR/mx-QSAR) may help to predict neurotoxicity/neuroprotective effects of drugs in multiple assays, on drug targets, and in model organisms. In this work, we study a data set downloaded from CHEMBL; each data point (>8000) contains the values of one out of 37 possible measures of activity, 493 assays, 169 molecular or cellular targets, and 11 different organisms (including human) for a given compound. In this work, we introduce the first mx-QSAR model for neurotoxicity/neuroprotective effects of drugs based on the MARCH-INSIDE (MI) method. First, we used MI to calculate the stochastic spectral moments (structural descriptors) of all compounds. Next, we found a model that classified correctly 2955 out of 3548 total cases in the training and validation series with Accuracy, Sensitivity, and Specificity values>80%. The model also showed excellent results in Computational-Chemistry simulations of High-Throughput Screening (CCHTS) experiments, with accuracy=90.6% for 4671 positive cases. Next, we reported the synthesis, characterization, and experimental assays of new rasagiline derivatives. We carried out three different experimental tests: assay (1) in the absence of neurotoxic agents, assay (2) in the presence of glutamate, and assay (3) in the presence of H2O2. Compounds 11 with 27.4%, 8 with 11.6%, and 9 with 15.4% showed the highest neuroprotective effects in assays (1), (2), and (3), respectively. After that, we used the mx-QSAR model to carry out a CCHTS of the new compounds in >400 unique pharmacological tests not carried out experimentally. Consequently, this model may become a promising auxiliary tool for the discovery of new drugs for the treatment of neurodegenerative diseases.

Noncontiguous atom matching structural similarity function

Measuring similarity between molecules is a fundamental problem in cheminformatics. Given that similar molecules tend to have similar physical, chemical, and biological properties, the notion of molecular similarity plays an important role in the exploration of molecular data sets, query-retrieval in molecular databases, and in structure-property/activity modeling. Various methods to define structural similarity between molecules are available in the literature, but so far none has been used with consistent and reliable results for all situations. We propose a new similarity method based on atom alignment for the analysis of structural similarity between molecules. This method is based on the comparison of the bonding profiles of atoms on comparable molecules, including features that are seldom found in other structural or graph matching approaches like chirality or double bond stereoisomerism. The similarity measure is then defined on the annotated molecular graph, based on an iterative directed graph similarity procedure and optimal atom alignment between atoms using a pairwise matching algorithm. With the proposed approach the similarities detected are more intuitively understood because similar atoms in the molecules are explicitly shown. This noncontiguous atom matching structural similarity method (NAMS) was tested and compared with one of the most widely used similarity methods (fingerprint-based similarity) using three difficult data sets with different characteristics. Despite having a higher computational cost, the method performed well being able to distinguish either different or very similar hydrocarbons that were indistinguishable using a fingerprint-based approach. NAMS also verified the similarity principle using a data set of structurally similar steroids with differences in the binding affinity to the corticosteroid binding globulin receptor by showing that pairs of steroids with a high degree of similarity (>80%) tend to have smaller differences in the absolute value of binding activity. Using a highly diverse set of compounds with information about the monoamine oxidase inhibition level, the method was also able to recover a significantly higher average fraction of active compounds when the seed is active for different cutoff threshold values of similarity. Particularly, for the cutoff threshold values of 86%, 93%, and 96.5%, NAMS was able to recover a fraction of actives of 0.57, 0.63, and 0.83, respectively, while the fingerprint-based approach was able to recover a fraction of actives of 0.41, 0.40, and 0.39, respectively. NAMS is made available freely for the whole community in a simple Web based tool as well as the Python source code at http://nams.lasige.di.fc.ul.pt/.