Machine learning methods in chemoinformatics

Network-based piecewise linear regression for QSAR modelling

Quantitative Structure-Activity Relationship (QSAR) models are critical in various areas of drug discovery, for example in lead optimisation and virtual screening. Recently, the need for models that are not only predictive but also interpretable has been highlighted. In this paper, a new methodology is proposed to build interpretable QSAR models by combining elements of network analysis and piecewise linear regression. The algorithm presented, modSAR, splits data using a two-step procedure. First, compounds associated with a common target are represented as a network in terms of their structural similarity, revealing modules of similar chemical properties. Second, each module is subdivided into subsets (regions), each of which is modelled by an independent linear equation. Comparative analysis of QSAR models across five data sets of protein inhibitors obtained from ChEMBL is reported and it is shown that modSAR offers similar predictive accuracy to popular algorithms, such as Random Forest and Support Vector Machine. Moreover, we show that models built by modSAR are interpretatable, capable of evaluating the applicability domain of the compounds and serve well tasks such as virtual screening and the development of new drug leads.

A Drug Decision Support System for Developing a Successful Drug Candidate Using Machine Learning Techniques

BACKGROUND

Virtual screening of candidate drug molecules using machine learning techniques plays a key role in pharmaceutical industry to design and discovery of new drugs. Computational classification methods can determine drug types according to the disease groups and distinguish approved drugs from withdrawn ones.

INTRODUCTION

Classification models developed in this study can be used as a simple filter in drug modelling to eliminate potentially inappropriate molecules in the early stages. In this work, we developed a Drug Decision Support System (DDSS) to classify each drug candidate molecule as potentially drug or non-drug and to predict its disease group.

METHODS

Molecular descriptors were identified for the determination of a number of rules in drug molecules. They were derived using ADRIANA.Code program and Lipinski's rule of five. We used Artificial Neural Network (ANN) to classify drug molecules correctly according to the types of diseases. Closed frequent molecular structures in the form of subgraph fragments were also obtained with Gaston algorithm included in ParMol Package to find common molecular fragments for withdrawn drugs.

RESULTS

We observed that TPSA, XlogP Natoms, HDon_O and TPSA are the most distinctive features in the pool of the molecular descriptors and evaluated the performances of classifiers on all datasets and found that classification accuracies are very high on all the datasets. Neural network models achieved 84.6% and 83.3% accuracies on test sets including cardiac therapy, anti-epileptics and anti-parkinson drugs with approved and withdrawn drugs for drug classification problems.

CONCLUSION

The experimental evaluation shows that the system is promising at determination of potential drug molecules to classify drug molecules correctly according to the types of diseases.

Improving prediction of phenotypic drug response on cancer cell lines using deep convolutional network

Background

Understanding the phenotypic drug response on cancer cell lines plays a vital role in anti-cancer drug discovery and re-purposing. The Genomics of Drug Sensitivity in Cancer (GDSC) database provides open data for researchers in phenotypic screening to build and test their models. Previously, most research in these areas starts from the molecular fingerprints or physiochemical features of drugs, instead of their structures.

Results

In this paper, a model called twin Convolutional Neural Network for drugs in SMILES format (tCNNS) is introduced for phenotypic screening. tCNNS uses a convolutional network to extract features for drugs from their simplified molecular input line entry specification (SMILES) format and uses another convolutional network to extract features for cancer cell lines from the genetic feature vectors respectively. After that, a fully connected network is used to predict the interaction between the drugs and the cancer cell lines. When the training set and the testing set are divided based on the interaction pairs between drugs and cell lines, tCNNS achieves 0.826, 0.831 for the mean and top quartile of the coefficient of determinant (R²) respectively and 0.909, 0.912 for the mean and top quartile of the Pearson correlation (R_p) respectively, which are significantly better than those of the previous works (Ammad-Ud-Din et al., J Chem Inf Model 54:2347–9, 2014), (Haider et al., PLoS ONE 10:0144490, 2015), (Menden et al., PLoS ONE 8:61318, 2013). However, when the training set and the testing set are divided exclusively based on drugs or cell lines, the performance of tCNNS decreases significantly and R_p and R² drop to barely above 0.

Conclusions

Our approach is able to predict the drug effects on cancer cell lines with high accuracy, and its performance remains stable with less but high-quality data, and with fewer features for the cancer cell lines. tCNNS can also solve the problem of outliers in other feature space. Besides achieving high scores in these statistical metrics, tCNNS also provides some insights into the phenotypic screening. However, the performance of tCNNS drops in the blind test.

Electronic supplementary material

The online version of this article (10.1186/s12859-019-2910-6) contains supplementary material, which is available to authorized users.

Error Tolerance of Machine Learning Algorithms across Contemporary Biological Targets

Machine learning continues to make strident advances in the prediction of desired properties concerning drug development. Problematically, the efficacy of machine learning in these arenas is reliant upon highly accurate and abundant data. These two limitations, high accuracy and abundance, are often taken together; however, insight into the dataset accuracy limitation of contemporary machine learning algorithms may yield insight into whether non-bench experimental sources of data may be used to generate useful machine learning models where there is a paucity of experimental data. We took highly accurate data across six kinase types, one GPCR, one polymerase, a human protease, and HIV protease, and intentionally introduced error at varying population proportions in the datasets for each target. With the generated error in the data, we explored how the retrospective accuracy of a Naïve Bayes Network, a Random Forest Model, and a Probabilistic Neural Network model decayed as a function of error. Additionally, we explored the ability of a training dataset with an error profile resembling that produced by the Free Energy Perturbation method (FEP+) to generate machine learning models with useful retrospective capabilities. The categorical error tolerance was quite high for a Naïve Bayes Network algorithm averaging 39% error in the training set required to lose predictivity on the test set. Additionally, a Random Forest tolerated a significant degree of categorical error introduced into the training set with an average error of 29% required to lose predictivity. However, we found the Probabilistic Neural Network algorithm did not tolerate as much categorical error requiring an average of 20% error to lose predictivity. Finally, we found that a Naïve Bayes Network and a Random Forest could both use datasets with an error profile resembling that of FEP+. This work demonstrates that computational methods of known error distribution like FEP+ may be useful in generating machine learning models not based on extensive and expensive in vitro-generated datasets.

A primer on natural product-based virtual screening

Databases play an important role in various computational techniques, including virtual screening (VS) and molecular modeling in general. These collections of molecules can contain a large amount of information, making them suitable for several drug discovery applications. For example, vendor, bioactivity data or target type can be found when searching a database. The introduction of these data resources and their characteristics is used for the design of an experiment. The description of the construction of a database can also be a good advisor for the creation of a new one. There are free available databases and commercial virtual libraries of molecules. Furthermore, a computational chemist can find databases for a general purpose or a specific subset such as natural products (NPs). In this chapter, NP database resources are presented, along with some guidelines when preparing an NP database for drug discovery purposes.

Exploiting machine learning for end-to-end drug discovery and development

A variety of machine learning methods such as naive Bayesian, support vector machines and more recently deep neural networks are demonstrating their utility for drug discovery and development. These leverage the generally bigger datasets created from high-throughput screening data and allow prediction of bioactivities for targets and molecular properties with increased levels of accuracy. We have only just begun to exploit the potential of these techniques but they may already be fundamentally changing the research process for identifying new molecules and/or repurposing old drugs. The integrated application of such machine learning models for end-to-end (E2E) application is broadly relevant and has considerable implications for developing future therapies and their targeting.

Alternative data mining/machine learning methods for the analytical evaluation of food quality and authenticity - A review

In recent years, the variety and volume of data acquired by modern analytical instruments in order to conduct a better authentication of food has dramatically increased. Several pattern recognition tools have been developed to deal with the large volume and complexity of available trial data. The most widely used methods are principal component analysis (PCA), partial least squares-discriminant analysis (PLS-DA), soft independent modelling by class analogy (SIMCA), k-nearest neighbours (kNN), parallel factor analysis (PARAFAC), and multivariate curve resolution-alternating least squares (MCR-ALS). Nevertheless, there are alternative data treatment methods, such as support vector machine (SVM), classification and regression tree (CART) and random forest (RF), that show a great potential and more advantages compared to conventional ones. In this paper, we explain the background of these methods and review and discuss the reported studies in which these three methods have been applied in the area of food quality and authenticity. In addition, we clarify the technical terminology used in this particular area of research.

Deep learning for predicting toxicity of chemicals: a mini review

Humans and wildlife inhabit a world with panoply of natural and synthetic chemicals. Alarmingly, only a limited number of chemicals have undergone comprehensive toxicological evaluation due to limitations of traditional toxicity testing. High-throughput screening assays provide a higher-speed alternative for conventional toxicity testing. Advancement of high-throughput bioassay technology has greatly increased chemical toxicity data volumes in the past decade, pushing toxicology research into a "big data" era. However, traditional data analysis methods fail to effectively process large data volumes, presenting both a challenge and an opportunity for toxicologists. Deep learning, a machine learning method leveraging deep neural networks (DNNs), is a proven useful tool for building quantitative structure-activity relationship (QSAR) models for toxicity prediction utilizing these new large datasets. In this mini review, a brief technical background on DNNs is provided, and the current state of chemical toxicity prediction models built with DNNs is reviewed. In addition, relevant toxicity data sources are summarized, possible limitations are discussed, and perspectives on DNN utilization in chemical toxicity prediction are given.

Comparative Study of Multitask Toxicity Modeling on a Broad Chemical Space

Acute toxicity is one of the most challenging properties to predict purely with computational methods due to its direct relationship to biological interactions. Moreover, toxicity can be represented by different end points: it can be measured for different species using different types of administration, etc., and it is questionable if the knowledge transfer between end points is possible. We performed a comparative study of prediction multitask toxicity for a broad chemical space using different descriptors and modeling algorithms and applied multitask learning for a large toxicity data set extracted from the Registry of Toxic Effects of Chemical Substances (RTECS). We demonstrated that multitask modeling provides significant improvement over single-output models and other machine learning methods. Our research reveals that multitask learning can be very useful to improve the quality of acute toxicity modeling and raises a discussion about the usage of multitask approaches for regulation purposes. Our MultiTox models are freely available in OCHEM platform ( ochem.eu/multitox ) under CC-BY-NC license.

Prediction of higher-selectivity catalysts by computer-driven workflow and machine learning

Catalyst design in asymmetric reaction development has traditionally been driven by empiricism, wherein experimentalists attempt to qualitatively recognize structural patterns to improve selectivity. Machine learning algorithms and chemoinformatics can potentially accelerate this process by recognizing otherwise inscrutable patterns in large datasets. Herein we report a computationally guided workflow for chiral catalyst selection using chemoinformatics at every stage of development. Robust molecular descriptors that are agnostic to the catalyst scaffold allow for selection of a universal training set on the basis of steric and electronic properties. This set can be used to train machine learning methods to make highly accurate predictive models over a broad range of selectivity space. Using support vector machines and deep feed-forward neural networks, we demonstrate accurate predictive modeling in the chiral phosphoric acid-catalyzed thiol addition to N-acylimines.

ChemSuite: A package for chemoinformatics calculations and machine learning

Prediction of biological and toxicological properties of small molecules using in silico approaches has become a wide practice in pharmaceutical research to lessen the cost and enhance productivity. The development of a tool "ChemSuite," a stand-alone application for chemoinformatics calculations and machine-learning model development, is reported. Availability of multi-functional features makes it widely acceptable in various fields. Force field such as UFF is incorporated in tool for optimization of molecules. Packages like RDKit, PyDPI and PaDEL help to calculate 1D, 2D and 3D descriptors and more than 10 types of fingerprints. MinMax Scaler and Z-Score algorithms are available to normalize descriptor values. Varied descriptor selection and machine-learning algorithms are available for model development. It allows the user to add their own algorithm or extend the software for various scientific purposes. It is free, open source and has user-friendly graphical interface, and it can work on all major platforms.

Web-Based Tools for Polypharmacology Prediction

Drug promiscuity or polypharmacology is the ability of small molecules to interact with multiple protein targets simultaneously. In drug discovery, understanding the polypharmacology of potential drug molecules is crucial to improve their efficacy and safety, and to discover the new therapeutic potentials of existing drugs. Over the past decade, several computational methods have been developed to study the polypharmacology of small molecules, many of which are available as Web services. In this chapter, we review some of these Web tools focusing on ligand based approaches. We highlight in particular our recently developed polypharmacology browser (PPB) and its application for finding the side targets of a new inhibitor of the TRPV6 calcium channel.

In Silico Target Prediction for Small Molecules

Drugs modulate disease states through their actions on targets in the body. Determining these targets aids the focused development of new treatments, and helps to better characterize those already employed. One means of accomplishing this is through the deployment of in silico methodologies, harnessing computational analytical and predictive power to produce educated hypotheses for experimental verification. Here, we provide an overview of the current state of the art, describe some of the well-established methods in detail, and reflect on how they, and emerging technologies promoting the incorporation of complex and heterogeneous data-sets, can be employed to improve our understanding of (poly)pharmacology.

Practical Model Selection for Prospective Virtual Screening

Virtual (computational) high-throughput screening provides a strategy for prioritizing compounds for experimental screens, but the choice of virtual screening algorithm depends on the data set and evaluation strategy. We consider a wide range of ligand-based machine learning and docking-based approaches for virtual screening on two protein–protein interactions, PriA-SSB and RMI-FANCM, and present a strategy for choosing which algorithm is best for prospective compound prioritization. Our workflow identifies a random forest as the best algorithm for these targets over more sophisticated neural network-based models. The top 250 predictions from our selected random forest recover 37 of the 54 active compounds from a library of 22,434 new molecules assayed on PriA-SSB. We show that virtual screening methods that perform well on public data sets and synthetic benchmarks, like multi-task neural networks, may not always translate to prospective screening performance on a specific assay of interest.

Deep Convolutional Neural Networks for the Prediction of Molecular Properties: Challenges and Opportunities Connected to the Data

We present a flexible deep convolutional neural network method for the analysis of arbitrary sized graph structures representing molecules. This method, which makes use of the Lipinski RDKit module, an open-source cheminformatics software, enables the incorporation of any global molecular (such as molecular charge and molecular weight) and local (such as atom hybridization and bond orders) information. In this paper, we show that this method significantly outperforms another recently proposed method based on deep convolutional neural networks on several datasets that are studied. Several best practices for training deep convolutional neural networks on chemical datasets are also highlighted within the article, such as how to select the information to be included in the model, how to prevent overfitting and how unbalanced classes in the data can be handled.

Brain Theranostics and Radiotheranostics: Exosomes and Graphenes In Vivo as Novel Brain Theranostics

Brain disease is one of the greatest threats to public health. Brain theranostics is recently taking shape, indicating the treatments of stroke, inflammatory brain disorders, psychiatric diseases, neurodevelopmental disease, and neurodegenerative disease. However, several factors, such as lack of endophenotype classification, blood-brain barrier (BBB), target determination, ignorance of biodistribution after administration, and complex intercellular communication between brain cells, make brain theranostics application difficult, especially when it comes to clinical application. So, a more thorough understanding of each aspect is needed. In this review, we focus on recent studies regarding the role of exosomes in intercellular communication of brain cells, therapeutic effect of graphene quantum dots, transcriptomics/epitranscriptomics approach for target selection, and in vitro/in vivo considerations.

Three-dimensional descriptors for aminergic GPCRs: dependence on docking conformation and crystal structure

Three-dimensional descriptors are often used to search for new biologically active compounds, in both ligand- and structure-based approaches, capturing the spatial orientation of molecules. They frequently constitute an input for machine learning-based predictions of compound activity or quantitative structure-activity relationship modeling; however, the distribution of their values and the accuracy of depicting compound orientations might have an impact on the power of the obtained predictive models. In this study, we analyzed the distribution of three-dimensional descriptors calculated for docking poses of active and inactive compounds for all aminergic G protein-coupled receptors with available crystal structures, focusing on the variation in conformations for different receptors and crystals. We demonstrated that the consistency in compound orientation in the binding site is rather not correlated with the affinity itself, but is more influenced by other factors, such as the number of rotatable bonds and crystal structure used for docking studies. The visualizations of the descriptors distributions were prepared and made available online at http://chem.gmum.net/vischem_stability , which enables the investigation of chemical structures referring to particular data points depicted in the figures. Moreover, the performed analysis can assist in choosing crystal structure for docking studies, helping in selection of conditions providing the best discrimination between active and inactive compounds in machine learning-based experiments.

QSAR-Based Virtual Screening: Advances and Applications in Drug Discovery

Virtual screening (VS) has emerged in drug discovery as a powerful computational approach to screen large libraries of small molecules for new hits with desired properties that can then be tested experimentally. Similar to other computational approaches, VS intention is not to replace in vitro or in vivo assays, but to speed up the discovery process, to reduce the number of candidates to be tested experimentally, and to rationalize their choice. Moreover, VS has become very popular in pharmaceutical companies and academic organizations due to its time-, cost-, resources-, and labor-saving. Among the VS approaches, quantitative structure–activity relationship (QSAR) analysis is the most powerful method due to its high and fast throughput and good hit rate. As the first preliminary step of a QSAR model development, relevant chemogenomics data are collected from databases and the literature. Then, chemical descriptors are calculated on different levels of representation of molecular structure, ranging from 1D to nD, and then correlated with the biological property using machine learning techniques. Once developed and validated, QSAR models are applied to predict the biological property of novel compounds. Although the experimental testing of computational hits is not an inherent part of QSAR methodology, it is highly desired and should be performed as an ultimate validation of developed models. In this mini-review, we summarize and critically analyze the recent trends of QSAR-based VS in drug discovery and demonstrate successful applications in identifying perspective compounds with desired properties. Moreover, we provide some recommendations about the best practices for QSAR-based VS along with the future perspectives of this approach.

Finding the structural requirements of diverse HIV-1 protease inhibitors using multiple QSAR modelling for lead identification

Multiple Quantitative Structure-Activity Relationship (QSAR) analysis is widely used in drug discovery for lead identification. Human Immunodeficiency Virus (HIV) protease is one of the key targets for the treatment of Acquired Immunodeficiency Syndrome (AIDS). One of the major challenges for the design of HIV-1 protease inhibitors (HIV PRIs) is to increase the inhibitory activities against the enzyme to a level where the problem associated to drug resistance may be considerably delayed. Herein, chemometric analyses were performed with 346 structurally diverse HIV PRIs with experimental bioactivities against a sub-type B mutant to develop highly predictable QSAR models and also to identify the effective structural determinants for higher affinity against HIV PR. The QSAR models were developed using OCHEM-based machine learning tools (ASNN, FSMLR, KNN, RF, MANN and XGBoost), with descriptors calculated by eight different software packages. Simultaneously, a Monte Carlo optimization-based QSAR modelling was performed using SMILES and graph-based descriptors to understand fragment and topochemical contributions. To validate the actual predictability of all these models, an additional set of 104 compounds (also with known experimental activities) with slightly different chemical space were employed. This ligand-based study serves as a crucial benchmark for further development of the HIV protease inhibitors with improved activities.

A Predictive Tool for Electrophilic Aromatic Substitutions Using Machine Learning

At the early stages of the drug development process, thousands of compounds are synthesized in order to attain the best possible potency and pharmacokinetic properties. Once successful scaffolds are identified, large libraries of analogues are made, which is a challenging and time-consuming task. Recently, late stage functionalization (LSF) has become increasingly prominent since these reactions selectively functionalize C-H bonds, allowing to quickly produce analogues. Classical electrophilic aromatic halogenations are a powerful type of reaction in the LSF toolkit. However, the introduction of an electrophile in a regioselective manner on a drug-like molecule is a challenging task. Herein we present a machine learning model able to predict the reactive site of an electrophilic aromatic substitution with an accuracy of 93% (internal validation set). The model takes as input a SMILES of a compound and uses six quantum mechanics descriptors to identify its reactive site(s). On an external validation set, 90% of all molecules were correctly predicted.

The hepatotoxic potential of protein kinase inhibitors predicted with Random Forest and Artificial Neural Networks

Protein kinases (PKs) play a role in many pivotal aspects of cellular function. Dysregulation and mutations of protein kinases are involved in the development of different diseases, which might be treated by inhibition of the corresponding kinase. Protein kinase inhibitors (PKIs) are generally well tolerated, but unexpected and serious adverse events on the heart, lung, kidney and liver were observed clinically. In this study, the structure-activity relationship of PKIs in relation to hepatotoxicity was investigated. A dataset of 165 PKIs was compiled and the probability of human hepatotoxicity with two different machine learning algorithms (Random Forest and Artificial Neural Networks) was analysed. The estimated probability of hepatotoxicity was generally high for single PKIs. However, depending on the target kinase of the PKI, a difference in hepatotoxic potential could be observed. The similarity of the PKIs to each other is caused by the conserved site of action of the protein kinases. Hepatotoxicity may therefore always be an issue in PKIs.

In Silico Studies in Drug Research Against Neurodegenerative Diseases

Background

Neurodegenerative diseases such as Alzheimer's disease (AD), amyotrophic lateral sclerosis, Parkinson's disease (PD), spinal cerebellar ataxias, and spinal and bulbar muscular atrophy are described by slow and selective degeneration of neurons and axons in the central nervous system (CNS) and constitute one of the major challenges of modern medicine. Computer-aided or in silico drug design methods have matured into powerful tools for reducing the number of ligands that should be screened in experimental assays.

Methods

In the present review, the authors provide a basic background about neurodegenerative diseases and in silico techniques in the drug research. Furthermore, they review the various in silico studies reported against various targets in neurodegenerative diseases, including homology modeling, molecular docking, virtual high-throughput screening, quantitative structure activity relationship (QSAR), hologram quantitative structure activity relationship (HQSAR), 3D pharmacophore mapping, proteochemometrics modeling (PCM), fingerprints, fragment-based drug discovery, Monte Carlo simulation, molecular dynamic (MD) simulation, quantum-mechanical methods for drug design, support vector machines, and machine learning approaches.

Results

Detailed analysis of the recently reported case studies revealed that the majority of them use a sequential combination of ligand and structure-based virtual screening techniques, with particular focus on pharmacophore models and the docking approach.

Conclusion

Neurodegenerative diseases have a multifactorial pathoetiological origin, so scientists have become persuaded that a multi-target therapeutic strategy aimed at the simultaneous targeting of multiple proteins (and therefore etiologies) involved in the development of a disease is recommended in future.

Comparing and Validating Machine Learning Models for Mycobacterium tuberculosis Drug Discovery

Tuberculosis is a global health dilemma. In 2016, the WHO reported 10.4 million incidences and 1.7 million deaths. The need to develop new treatments for those infected with Mycobacterium tuberculosis ( Mtb) has led to many large-scale phenotypic screens and many thousands of new active compounds identified in vitro. However, with limited funding, efforts to discover new active molecules against Mtb needs to be more efficient. Several computational machine learning approaches have been shown to have good enrichment and hit rates. We have curated small molecule Mtb data and developed new models with a total of 18,886 molecules with activity cutoffs of 10 μM, 1 μM, and 100 nM. These data sets were used to evaluate different machine learning methods (including deep learning) and metrics and to generate predictions for additional molecules published in 2017. One Mtb model, a combined in vitro and in vivo data Bayesian model at a 100 nM activity yielded the following metrics for 5-fold cross validation: accuracy = 0.88, precision = 0.22, recall = 0.91, specificity = 0.88, kappa = 0.31, and MCC = 0.41. We have also curated an evaluation set ( n = 153 compounds) published in 2017, and when used to test our model, it showed the comparable statistics (accuracy = 0.83, precision = 0.27, recall = 1.00, specificity = 0.81, kappa = 0.36, and MCC = 0.47). We have also compared these models with additional machine learning algorithms showing Bayesian machine learning models constructed with literature Mtb data generated by different laboratories generally were equivalent to or outperformed deep neural networks with external test sets. Finally, we have also compared our training and test sets to show they were suitably diverse and different in order to represent useful evaluation sets. Such Mtb machine learning models could help prioritize compounds for testing in vitro and in vivo.

Prediction Is a Balancing Act: Importance of Sampling Methods to Balance Sensitivity and Specificity of Predictive Models Based on Imbalanced Chemical Data Sets

Increase in the number of new chemicals synthesized in past decades has resulted in constant growth in the development and application of computational models for prediction of activity as well as safety profiles of the chemicals. Most of the time, such computational models and its application must deal with imbalanced chemical data. It is indeed a challenge to construct a classifier using imbalanced data set. In this study, we analyzed and validated the importance of different sampling methods over non-sampling method, to achieve a well-balanced sensitivity and specificity of a machine learning model trained on imbalanced chemical data. Additionally, this study has achieved an accuracy of 93.00%, an AUC of 0.94, F1 measure of 0.90, sensitivity of 96.00% and specificity of 91.00% using SMOTE sampling and Random Forest classifier for the prediction of Drug Induced Liver Injury (DILI). Our results suggest that, irrespective of data set used, sampling methods can have major influence on reducing the gap between sensitivity and specificity of a model. This study demonstrates the efficacy of different sampling methods for class imbalanced problem using binary chemical data sets.

Mathematical structural descriptors and mutagenicity assessment: a study with congeneric and diverse datasets$

Quantitative bioactivity and toxicity assessment of chemical compounds plays a central role in drug discovery as it saves a substantial amount of resources. To this end, high-performance computing has enabled researchers and practitioners to leverage hundreds, or even thousands, of computed molecular descriptors for the activity prediction of candidate compounds. In this paper, we evaluate the utility of two large groups of chemical descriptors by such predictive modelling, as well as chemical structure discovery, through empirical analysis. We use a suite of commercially available and in-house software to calculate molecular descriptors for two sets of chemical mutagens - a homogeneous set of 95 amines, and a diverse set of 508 chemicals. Using calculated descriptors, we model the mutagenic activity of these compounds using a number of methods from the statistics and machine-learning literature, and use robust principal component analysis to investigate the low-dimensional subspaces that characterize these chemicals. Our results suggest that combining different sets of descriptors is likely to result in a better predictive model - but that depends on the compounds being modelled and the modelling technique being used.

Editor's Highlight: Identification of Any Structure-Specific Hepatotoxic Potential of Different Pyrrolizidine Alkaloids Using Random Forests and Artificial Neural Networks

Pyrrolizidine alkaloids (PAs) are characteristic metabolites of some plant families and form a powerful defense mechanism against herbivores. More than 600 different PAs are known. PAs are ester alkaloids composed of a necine base and a necic acid, which can be used to divide PAs in different structural subcategories. The main target organs for PA metabolism and toxicity are liver and lungs. Additionally, PAs are potentially genotoxic, carcinogenic and exhibit developmental toxicity. Only for very few PAs, in vitro and in vivo investigations have characterized their toxic potential. However, these investigations suggest that structural differences have an influence on the toxicity of single PAs. To investigate this structural relationship for a large number of PAs, a quantitative structural-activity relationship (QSAR) analysis for hepatotoxicity of over 600 different PAs was performed, using Random Forest- and artificial Neural Networks-algorithms. These models were trained with a recently established dataset specific for acute hepatotoxicity in humans. Using this dataset, a set of molecular predictors was identified to predict the hepatotoxic potential of each compound in validated QSAR models. Based on these models, the hepatotoxic potential of the 602 PAs was predicted and the following hepatotoxic rank order in 3 main categories defined (1) for necine base: otonecine > retronecine > platynecine; (2) for necine base modification: dehydropyrrolizidine ≫ tertiary PA = N-oxide; and (3) for necic acid: macrocyclic diester ≥ open-ring diester > monoester. A further analysis with combined structural features revealed that necic acid has a higher influence on the acute hepatotoxicity than the necine base.

A Machine Learning Approach for Predicting HIV Reverse Transcriptase Mutation Susceptibility of Biologically Active Compounds

HIV resistance emerging against antiretroviral drugs represents a great threat to the continued prolongation of the lifespans of HIV-infected patients. Therefore, methods capable of predicting resistance susceptibility in the development of compounds are in great need. By targeting the major reverse transcription residues Y181, K103, and L100, we used the biological activities of compounds against these enzymes and the wild-type reverse transcriptase to create Naïve Bayes Networks. Through this machine learning approach, we could predict, with high accuracy, whether a compound would be susceptible to a loss of potency due to resistance. Also, we could perfectly predict retrospectively whether compounds would be susceptible to both a K103 mutant RT and a Y181 mutant RT. In the study presented here, our method outperformed a traditional molecular mechanics approach. This method should be of broad interest beyond drug discovery efforts, and serves to expand the utility of machine learning for the prediction of physical, chemical, or biological properties using the vast information available in the literature.

Machine Learning in Computer-Aided Synthesis Planning

Computer-aided synthesis planning (CASP) is focused on the goal of accelerating the process by which chemists decide how to synthesize small molecule compounds. The ideal CASP program would take a molecular structure as input and output a sorted list of detailed reaction schemes that each connect that target to purchasable starting materials via a series of chemically feasible reaction steps. Early work in this field relied on expert-crafted reaction rules and heuristics to describe possible retrosynthetic disconnections and selectivity rules but suffered from incompleteness, infeasible suggestions, and human bias. With the relatively recent availability of large reaction corpora (such as the United States Patent and Trademark Office (USPTO), Reaxys, and SciFinder databases), consisting of millions of tabulated reaction examples, it is now possible to construct and validate purely data-driven approaches to synthesis planning. As a result, synthesis planning has been opened to machine learning techniques, and the field is advancing rapidly. In this Account, we focus on two critical aspects of CASP and recent machine learning approaches to both challenges. First, we discuss the problem of retrosynthetic planning, which requires a recommender system to propose synthetic disconnections starting from a target molecule. We describe how the search strategy, necessary to overcome the exponential growth of the search space with increasing number of reaction steps, can be assisted through a learned synthetic complexity metric. We also describe how the recursive expansion can be performed by a straightforward nearest neighbor model that makes clever use of reaction data to generate high quality retrosynthetic disconnections. Second, we discuss the problem of anticipating the products of chemical reactions, which can be used to validate proposed reactions in a computer-generated synthesis plan (i.e., reduce false positives) to increase the likelihood of experimental success. While we introduce this task in the context of reaction validation, its utility extends to the prediction of side products and impurities, among other applications. We describe neural network-based approaches that we and others have developed for this forward prediction task that can be trained on previously published experimental data. Machine learning and artificial intelligence have revolutionized a number of disciplines, not limited to image recognition, dictation, translation, content recommendation, advertising, and autonomous driving. While there is a rich history of using machine learning for structure-activity models in chemistry, it is only now that it is being successfully applied more broadly to organic synthesis and synthesis design. As reported in this Account, machine learning is rapidly transforming CASP, but there are several remaining challenges and opportunities, many pertaining to the availability and standardization of both data and evaluation metrics, which must be addressed by the community at large.

Exploring G Protein-Coupled Receptors (GPCRs) Ligand Space via Cheminformatics Approaches: Impact on Rational Drug Design

The primary goal of rational drug discovery is the identification of selective ligands which act on single or multiple drug targets to achieve the desired clinical outcome through the exploration of total chemical space. To identify such desired compounds, computational approaches are necessary in predicting their drug-like properties. G Protein-Coupled Receptors (GPCRs) represent one of the largest and most important integral membrane protein families. These receptors serve as increasingly attractive drug targets due to their relevance in the treatment of various diseases, such as inflammatory disorders, metabolic imbalances, cardiac disorders, cancer, monogenic disorders, etc. In the last decade, multitudes of three-dimensional (3D) structures were solved for diverse GPCRs, thus referring to this period as the "golden age for GPCR structural biology." Moreover, accumulation of data about the chemical properties of GPCR ligands has garnered much interest toward the exploration of GPCR chemical space. Due to the steady increase in the structural, ligand, and functional data of GPCRs, several cheminformatics approaches have been implemented in its drug discovery pipeline. In this review, we mainly focus on the cheminformatics-based paradigms in GPCR drug discovery. We provide a comprehensive view on the ligand- and structure-based cheminformatics approaches which are best illustrated via GPCR case studies. Furthermore, an appropriate combination of ligand-based knowledge with structure-based ones, i.e., integrated approach, which is emerging as a promising strategy for cheminformatics-based GPCR drug design is also discussed.

Modeling and predicting chiral stationary phase enantioselectivity: An efficient random forest classifier using an optimally balanced training dataset and an aggregation strategy

Predicting whether a chiral column will be effective is a daily task for many analysts. Moreover, finding the best chiral column for separating a particular racemic compound is mostly a matter of trial and error that may take up to a week in some cases. In this study we have developed a novel prediction approach based on combining a random forest classifier and an optimized discretization method for dealing with enantioselectivity as a continuous variable. Using the optimization results, models were trained on data sets divided into four enantioselectivity classes. The best model performances were achieved by over-sampling the minority classes (α ≤ 1.10 and α ≥ 2.00), down-sampling the majority class (1.2 ≤ α < 2.0), and aggregating multicategory predictions into binary classifications. We tested our method on 41 chiral stationary phases using layered fingerprints as descriptors. Experimental results show that this learning methodology was successful in terms of average area under the Receiver Operating Characteristic curve, Kappa indices and F-measure for structure-based prediction of the enantioselective behavior of 34 chiral columns.

Efficient iterative virtual screening with Apache Spark and conformal prediction

Background

Docking and scoring large libraries of ligands against target proteins forms the basis of structure-based virtual screening. The problem is trivially parallelizable, and calculations are generally carried out on computer clusters or on large workstations in a brute force manner, by docking and scoring all available ligands.

Contribution

In this study we propose a strategy that is based on iteratively docking a set of ligands to form a training set, training a ligand-based model on this set, and predicting the remainder of the ligands to exclude those predicted as ‘low-scoring’ ligands. Then, another set of ligands are docked, the model is retrained and the process is repeated until a certain model efficiency level is reached. Thereafter, the remaining ligands are docked or excluded based on this model. We use SVM and conformal prediction to deliver valid prediction intervals for ranking the predicted ligands, and Apache Spark to parallelize both the docking and the modeling.

Results

We show on 4 different targets that conformal prediction based virtual screening (CPVS) is able to reduce the number of docked molecules by 62.61% while retaining an accuracy for the top 30 hits of 94% on average and a speedup of 3.7. The implementation is available as open source via GitHub (https://github.com/laeeq80/spark-cpvs) and can be run on high-performance computers as well as on cloud resources.

Development of models for classification of action between heat-clearing herbs and blood-activating stasis-resolving herbs based on theory of traditional Chinese medicine

Background

Action (“gongxiao” in Chinese) of traditional Chinese medicine (TCM) is the high recapitulation for therapeutic and health-preserving effects under the guidance of TCM theory. TCM-defined herbal properties (“yaoxing” in Chinese) had been used in this research. TCM herbal property (TCM-HP) is the high generalization and summary for actions, both of which come from long-term effective clinical practice in two thousands of years in China. However, the specific relationship between TCM-HP and action of TCM is complex and unclear from a scientific perspective. The research about this is conducive to expound the connotation of TCM-HP theory and is of important significance for the development of the TCM-HP theory.

Methods

One hundred and thirty-three herbs including 88 heat-clearing herbs (HCHs) and 45 blood-activating stasis-resolving herbs (BAHRHs) were collected from reputable TCM literatures, and their corresponding TCM-HPs/actions information were collected from Chinese pharmacopoeia (2015 edition). The Kennard–Stone (K–S) algorithm was used to split 133 herbs into 100 calibration samples and 33 validation samples. Then, machine learning methods including supported vector machine (SVM), k-nearest neighbor (kNN) and deep learning methods including deep belief network (DBN), convolutional neutral network (CNN) were adopted to develop action classification models based on TCM-HP theory, respectively. In order to ensure robustness, these four classification methods were evaluated by using the method of tenfold cross validation and 20 external validation samples for prediction.

Results

As results, 72.7–100% of 33 validation samples including 17 HCHs and 16 BASRHs were correctly predicted by these four types of methods. Both of the DBN and CNN methods gave out the best results and their sensitivity, specificity, precision, accuracy were all 100.00%. Especially, the predicted results of external validation set showed that the performance of deep learning methods (DBN, CNN) were better than traditional machine learning methods (kNN, SVM) in terms of their sensitivity, specificity, precision, accuracy. Moreover, the distribution patterns of TCM-HPs of HCHs and BASRHs were also analyzed to detect the featured TCM-HPs of these two types of herbs. The result showed that the featured TCM-HPs of HCHs were cold, bitter, liver and stomach meridians entered, while those of BASRHs were warm, bitter and pungent, liver meridian entered.

Conclusions

The performance on validation set and external validation set of deep learning methods (DBN, CNN) were better than machine learning models (kNN, SVM) in sensitivity, specificity, precision, accuracy when predicting the actions of heat-clearing and blood-activating stasis-resolving based on TCM-HP theory. The deep learning classification methods owned better generalization ability and accuracy when predicting the actions of heat-clearing and blood-activating stasis-resolving based on TCM-HP theory. Besides, the methods of deep learning would help us to improve our understanding about the relationship between herbal property and action, as well as to enrich and develop the theory of TCM-HP scientifically.

Electronic supplementary material

The online version of this article (10.1186/s13020-018-0169-x) contains supplementary material, which is available to authorized users.

Transfer and Multi-task Learning in QSAR Modeling: Advances and Challenges

Medicinal chemistry projects involve some steps aiming to develop a new drug, such as the analysis of biological targets related to a given disease, the discovery and the development of drug candidates for these targets, performing parallel biological tests to validate the drug effectiveness and side effects. Approaches as quantitative study of activity-structure relationships (QSAR) involve the construction of predictive models that relate a set of descriptors of a chemical compound series and its biological activities with respect to one or more targets in the human body. Datasets used to perform QSAR analyses are generally characterized by a small number of samples and this makes them more complex to build accurate predictive models. In this context, transfer and multi-task learning techniques are very suitable since they take information from other QSAR models to the same biological target, reducing efforts and costs for generating new chemical compounds. Therefore, this review will present the main features of transfer and multi-task learning studies, as well as some applications and its potentiality in drug design projects.

Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks

In de novo drug design, computational strategies are used to generate novel molecules with good affinity to the desired biological target. In this work, we show that recurrent neural networks can be trained as generative models for molecular structures, similar to statistical language models in natural language processing. We demonstrate that the properties of the generated molecules correlate very well with the properties of the molecules used to train the model. In order to enrich libraries with molecules active toward a given biological target, we propose to fine-tune the model with small sets of molecules, which are known to be active against that target. Against Staphylococcus aureus, the model reproduced 14% of 6051 hold-out test molecules that medicinal chemists designed, whereas against Plasmodium falciparum (Malaria), it reproduced 28% of 1240 test molecules. When coupled with a scoring function, our model can perform the complete de novo drug design cycle to generate large sets of novel molecules for drug discovery.

Combining machine learning models of in vitro and in vivo bioassays improves rat carcinogenicity prediction

In vitro genotoxicity bioassays are cost-efficient methods of assessing potential carcinogens. However, many genotoxicity bioassays are inappropriate for detecting chemicals eliciting non-genotoxic mechanisms, such as tumour promotion, this necessitates the use of in vivo rodent carcinogenicity (IVRC) assays. In silico IVRC modelling could potentially address the low throughput and high cost of this assay. We aimed to develop and combine computational QSAR models of novel bioassays for the prediction of IVRC results and compare with existing software. QSAR models were generated from existing Ames (n = 6512), Syrian Hamster Embryonic (SHE, n = 410), ISSCAN rodent carcinogenicity (ISC, n = 834) and GreenScreen GADD45a-GFP (n = 1415) chemical datasets. These models mapped the molecular descriptors of each compound to their respective assay result using machine learning algorithms (adaboost, k-Nearest Neighbours, C.45 Decision Tree, Multilayer Perceptron, Random Forest). The best performing models were combined with k-Nearest Neighbours to create a cascade model for IVRC prediction. High QSAR model performance was observed from ten time 10-fold cross-validation with above 80% accuracy and 0.85 AUC for each assay dataset. The cascade model predicted rat carcinogenicity with 69.3% accuracy and 0.700 AUC. This study demonstrates the novelty of a combined approach for IVRC prediction, with higher performance than existing software.

Computational Toxicology and Drug Discovery

The use of computational toxicology methods within drug discovery began in the early 2000s with applications such as predicting bacterial mutagenicity and hERG inhibition. The field has been continuously expanding ever since and the tasks at hand have become more complex. These approaches are now strategically integrated into the risk assessment process, as a complement to in vitro and in vivo methods. Today, computational toxicology can be used in every phase of drug discovery and development, from profiling large libraries early on, to predicting off-target effects in the mid-discovery phase, to assessing potential mutagenic impurities in development and degradants as part of life-cycle management. This chapter provides an overview of the field and describes the application of computational toxicology throughout the entire discovery and development process.

Synthesis meets theory: Past, present and future of rational chemistry

Chemical synthesis has its roots in the empirical approach of alchemy. Nonetheless, the birth of the scientific method, the technical and technological advances (exploiting revolutionary discoveries in physics) and the improved management and sharing of growing databases greatly contributed to the evolution of chemistry from an esoteric ground into a mature scientific discipline during these last 400 years. Furthermore, thanks to the evolution of computational resources, platforms and media in the last 40 years, theoretical chemistry has added to the puzzle the final missing tile in the process of “rationalizing” chemistry. The use of mathematical models of chemical properties, behaviors and reactivities is nowadays ubiquitous in literature. Theoretical chemistry has been successful in the difficult task of complementing and explaining synthetic results and providing rigorous insights when these are otherwise unattainable by experiment. The first part of this review walks the reader through a concise historical overview on the evolution of the “model” in chemistry. Salient milestones have been highlighted and briefly discussed. The second part focuses more on the general description of recent state-of-the-art computational techniques currently used worldwide by chemists to produce synergistic models between theory and experiment. Each section is complemented by key-examples taken from the literature that illustrate the application of the technique discussed therein.