Maximum unbiased validation (MUV) data sets for virtual screening based on PubChem bioactivity data

Insights into performance evaluation of compound-protein interaction prediction methods

MOTIVATION

Machine-learning-based prediction of compound-protein interactions (CPIs) is important for drug design, screening and repurposing. Despite numerous recent publication with increasing methodological sophistication claiming consistent improvements in predictive accuracy, we have observed a number of fundamental issues in experiment design that produce overoptimistic estimates of model performance.

RESULTS

We systematically analyze the impact of several factors affecting generalization performance of CPI predictors that are overlooked in existing work: (i) similarity between training and test examples in cross-validation; (ii) synthesizing negative examples in absence of experimentally verified negative examples and (iii) alignment of evaluation protocol and performance metrics with real-world use of CPI predictors in screening large compound libraries. Using both state-of-the-art approaches by other researchers as well as a simple kernel-based baseline, we have found that effective assessment of generalization performance of CPI predictors requires careful control over similarity between training and test examples. We show that, under stringent performance assessment protocols, a simple kernel-based approach can exceed the predictive performance of existing state-of-the-art methods. We also show that random pairing for generating synthetic negative examples for training and performance evaluation results in models with better generalization in comparison to more sophisticated strategies used in existing studies. Our analyses indicate that using proposed experiment design strategies can offer significant improvements for CPI prediction leading to effective target compound screening for drug repurposing and discovery of putative chemical ligands of SARS-CoV-2-Spike and Human-ACE2 proteins.

AVAILABILITY AND IMPLEMENTATION

Code and supplementary material available at https://github.com/adibayaseen/HKRCPI.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

On modeling and utilizing chemical compound information with deep learning technologies: A task-oriented approach

A large number of chemical compounds are available in databases such as PubChem and ZINC. However, currently known compounds, though large, represent only a fraction of possible compounds, which is known as chemical space. Many of these compounds in the databases are annotated with properties and assay data that can be used for drug discovery efforts. For this goal, a number of machine learning algorithms have been developed and recent deep learning technologies can be effectively used to navigate chemical space, especially for unknown chemical compounds, in terms of drug-related tasks. In this article, we survey how deep learning technologies can model and utilize chemical compound information in a task-oriented way by exploiting annotated properties and assay data in the chemical compounds databases. We first compile what kind of tasks are trying to be accomplished by machine learning methods. Then, we survey deep learning technologies to show their modeling power and current applications for accomplishing drug related tasks. Next, we survey deep learning techniques to address the insufficiency issue of annotated data for more effective navigation of chemical space. Chemical compound information alone may not be powerful enough for drug related tasks, thus we survey what kind of information, such as assay and gene expression data, can be used to improve the prediction power of deep learning models. Finally, we conclude this survey with four important newly developed technologies that are yet to be fully incorporated into computational analysis of chemical information.

Confidence bands and hypothesis tests for hit enrichment curves

In virtual screening for drug discovery, hit enrichment curves are widely used to assess the performance of ranking algorithms with regard to their ability to identify early enrichment. Unfortunately, researchers almost never consider the uncertainty associated with estimating such curves before declaring differences between performance of competing algorithms. Uncertainty is often large because the testing fractions of interest to researchers are small. Appropriate inference is complicated by two sources of correlation that are often overlooked: correlation across different testing fractions within a single algorithm, and correlation between competing algorithms. Additionally, researchers are often interested in making comparisons along the entire curve, not only at a few testing fractions. We develop inferential procedures to address both the needs of those interested in a few testing fractions, as well as those interested in the entire curve. For the former, four hypothesis testing and (pointwise) confidence intervals are investigated, and a newly developed EmProc approach is found to be most effective. For inference along entire curves, EmProc-based confidence bands are recommended for simultaneous coverage and minimal width. While we focus on the hit enrichment curve, this work is also appropriate for lift curves that are used throughout the machine learning community. Our inferential procedures trivially extend to enrichment factors, as well.

Protein-Ligand Docking in the Machine-Learning Era

Molecular docking plays a significant role in early-stage drug discovery, from structure-based virtual screening (VS) to hit-to-lead optimization, and its capability and predictive power is critically dependent on the protein-ligand scoring function. In this review, we give a broad overview of recent scoring function development, as well as the docking-based applications in drug discovery. We outline the strategies and resources available for structure-based VS and discuss the assessment and development of classical and machine learning protein-ligand scoring functions. In particular, we highlight the recent progress of machine learning scoring function ranging from descriptor-based models to deep learning approaches. We also discuss the general workflow and docking protocols of structure-based VS, such as structure preparation, binding site detection, docking strategies, and post-docking filter/re-scoring, as well as a case study on the large-scale docking-based VS test on the LIT-PCBA data set.

Ligand-Enhanced Negative Images Optimized for Docking Rescoring

Despite the pivotal role of molecular docking in modern drug discovery, the default docking scoring functions often fail to recognize active ligands in virtual screening campaigns. Negative image-based rescoring improves docking enrichment by comparing the shape/electrostatic potential (ESP) of the flexible docking poses against the target protein’s inverted cavity volume. By optimizing these negative image-based (NIB) models using a greedy search, the docking rescoring yield can be improved massively and consistently. Here, a fundamental modification is implemented to this shape-focused pharmacophore modelling approach—actual ligand 3D coordinates are incorporated into the NIB models for the optimization. This hybrid approach, labelled as ligand-enhanced brute-force negative image-based optimization (LBR-NiB), takes the best from both worlds, i.e., the all-roundedness of the NIB models and the difficult to emulate atomic arrangements of actual protein-bound small-molecule ligands. Thorough benchmarking, focused on proinflammatory targets, shows that the LBR-NiB routinely improves the docking enrichment over prior iterations of the R-NiB methodology. This boost can be massive, if the added ligand information provides truly essential binding information that was lacking or completely missing from the cavity-based NIB model. On a practical level, the results indicate that the LBR-NiB typically works well when the added ligand 3D data originates from a high-quality source, such as X-ray crystallography, and, yet, the NIB model compositions can also sometimes be improved by fusing into them, for example, with flexibly docked solvent molecules. In short, the study demonstrates that the protein-bound ligands can be used to improve the shape/ESP features of the negative images for effective docking rescoring use in virtual screening.

TocoDecoy: A New Approach to Design Unbiased Datasets for Training and Benchmarking Machine-Learning Scoring Functions

Development of accurate machine-learning-based scoring functions (MLSFs) for structure-based virtual screening against a given target requires a large unbiased dataset with structurally diverse actives and decoys. However, most datasets for the development of MLSFs were designed for traditional SFs and may suffer from hidden biases and data insufficiency. Hereby, we developed a new approach named Topology-based and Conformation-based decoys generation (TocoDecoy), which integrates two strategies to generate decoys by tweaking the actives for a specific target, to generate unbiased and expandable datasets for training and benchmarking MLSFs. For hidden bias evaluation, the performance of InteractionGraphNet (IGN) trained on the TocoDecoy, LIT-PCBA, and DUD-E-like datasets was assessed. The results illustrate that the IGN model trained on the TocoDecoy dataset is competitive with that trained on the LIT-PCBA dataset but remarkably outperforms that trained on the DUD-E dataset, suggesting that the decoys in TocoDecoy are unbiased for training and benchmarking MLSFs.

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.

Improved drug-target interaction prediction with intermolecular graph transformer

The identification of active binding drugs for target proteins (referred to as drug-target interaction prediction) is the key challenge in virtual screening, which plays an essential role in drug discovery. Although recent deep learning-based approaches achieve better performance than molecular docking, existing models often neglect topological or spatial of intermolecular information, hindering prediction performance. We recognize this problem and propose a novel approach called the Intermolecular Graph Transformer (IGT) that employs a dedicated attention mechanism to model intermolecular information with a three-way Transformer-based architecture. IGT outperforms state-of-the-art (SoTA) approaches by 9.1% and 20.5% over the second best option for binding activity and binding pose prediction, respectively, and exhibits superior generalization ability to unseen receptor proteins than SoTA approaches. Furthermore, IGT exhibits promising drug screening ability against severe acute respiratory syndrome coronavirus 2 by identifying 83.1% active drugs that have been validated by wet-lab experiments with near-native predicted binding poses. Source code and datasets are available at https://github.com/microsoft/IGT-Intermolecular-Graph-Transformer.

Machine Learning in Chemoinformatics and Medicinal Chemistry

In chemoinformatics and medicinal chemistry, machine learning has evolved into an important approach. In recent years, increasing computational resources and new deep learning algorithms have put machine learning onto a new level, addressing previously unmet challenges in pharmaceutical research. In silico approaches for compound activity predictions, de novo design, and reaction modeling have been further advanced by new algorithmic developments and the emergence of big data in the field. Herein, novel applications of machine learning and deep learning in chemoinformatics and medicinal chemistry are reviewed. Opportunities and challenges for new methods and applications are discussed, placing emphasis on proper baseline comparisons, robust validation methodologies, and new applicability domains. Expected final online publication date for the Annual Review of Biomedical Data Science, Volume 5 is August 2022. Please see http://www.annualreviews.org/page/journal/pubdates for revised estimates.

Application of deep metric learning to molecular graph similarity

Graph based methods are increasingly important in chemistry and drug discovery, with applications ranging from QSAR to molecular generation. Combining graph neural networks and deep metric learning concepts, we expose a framework for quantifying molecular graph similarity based on distance between learned embeddings separate from any endpoint. Using a minimal definition of similarity, and data from the ZINC database of public compounds, this work demonstrate the properties of the embedding and its suitability for a range of applications, among them a novel reconstruction loss method for training deep molecular auto-encoders. Finally, we compare the applications of the embedding to standard practices, with a focus on known failure points and edge cases; concluding that our approach can be used in conjunction to existing methods.

Identification of Sclareol As a Natural Neuroprotective Cav 1.3-Antagonist Using Synthetic Parkinson-Mimetic Gene Circuits and Computer-Aided Drug Discovery

Parkinson's disease (PD) results from selective loss of substantia nigra dopaminergic (SNc DA) neurons, and is primarily caused by excessive activity-related Ca²⁺ oscillations. Although L-type voltage-gated calcium channel blockers (CCBs) selectively inhibiting Ca_v 1.3 are considered promising candidates for PD treatment, drug discovery is hampered by the lack of high-throughput screening technologies permitting isoform-specific assessment of Cav-antagonistic activities. Here, a synthetic-biology-inspired drug-discovery platform enables identification of PD-relevant drug candidates. By deflecting Cav-dependent activation of nuclear factor of activated T-cells (NFAT)-signaling to repression of reporter gene translation, they engineered a cell-based assay where reporter gene expression is activated by putative CCBs. By using this platform in combination with in silico virtual screening and a trained deep-learning neural network, sclareol is identified from a essential oils library as a structurally distinctive compound that can be used for PD pharmacotherapy. In vitro studies, biochemical assays and whole-cell patch-clamp recordings confirmed that sclareol inhibits Ca_v 1.3 more strongly than Ca_v 1.2 and decreases firing responses of SNc DA neurons. In a mouse model of PD, sclareol treatment reduced DA neuronal loss and protected striatal network dynamics as well as motor performance. Thus, sclareol appears to be a promising drug candidate for neuroprotection in PD patients.

Similarity-Based Virtual Screen Using Enhanced Siamese Deep Learning Methods

Traditional drug production is a long and complex process that leads to new drug production. The virtual screening technique is a computational method that allows chemical compounds to be screened at an acceptable time and cost. Several databases contain information on various aspects of biologically active substances. Simple statistical tools are difficult to use because of the enormous amount of information and complex data samples of molecules that are structurally heterogeneous recorded in these databases. Many techniques for capturing the biological similarity between a test compound and a known target ligand in LBVS have been established. However, despite the good performances of the above methods compared to their prior, especially when dealing with molecules that have homogeneous active structural elements, they are not satisfied when dealing with molecules that are structurally heterogeneous. Deep learning models have recently achieved considerable success in a variety of disciplines due to their powerful generalization and feature extraction capabilities. Also, the Siamese network has been used in similarity models for more complicated data samples, especially with heterogeneous data samples. The main aim of this study is to enhance the performance of similarity searching, especially with molecules that are structurally heterogeneous. The Siamese architecture will be enhanced using two similarity distance layers with one fusion layer to further improve the similarity measurements between molecules and then adding many layers after the fusion layer for some models to improve the retrieval recall. In this architecture, several methods of deep learning have been used, which are long short-term memory (LSTM), gated recurrent unit (GRU), convolutional neural network-one dimension (CNN1D), and convolutional neural network-two dimensions (CNN2D). A series of experiments have been carried out on real-world data sets, and the results have shown that the proposed methods outperformed the existing methods.

Optimization of Cavity-Based Negative Images to Boost Docking Enrichment in Virtual Screening

Molecular docking is a key in silico method used routinely in modern drug discovery projects. Although docking provides high-quality ligand binding predictions, it regularly fails to separate the active compounds from the inactive ones. In negative image-based rescoring (R-NiB), the shape/electrostatic potential (ESP) of docking poses is compared to the negative image of the protein's ligand binding cavity. While R-NiB often improves the docking yield considerably, the cavity-based models do not reach their full potential without expert editing. Accordingly, a greedy search-driven methodology, brute force negative image-based optimization (BR-NiB), is presented for optimizing the models via iterative editing and benchmarking. Thorough and unbiased training, testing and stringent validation with a multitude of drug targets, and alternative docking software show that BR-NiB ensures excellent docking efficacy. BR-NiB can be considered as a new type of shape-focused pharmacophore modeling, where the optimized models contain only the most vital cavity information needed for effectively filtering docked actives from the inactive or decoy compounds. Finally, the BR-NiB code for performing the automated optimization is provided free-of-charge under MIT license via GitHub (https://github.com/jvlehtonen/brutenib) for boosting the success rates of docking-based virtual screening campaigns.

AI in 3D compound design

The success of Artificial Intelligence (AI) across a wide range of domains has fuelled significant interest in its application to designing novel compounds and screening compounds against a specific target. However, many existing AI methods either do not account for the 3D structure of the target at all or struggle to capture meaningful spatial information from the target. In this Opinion, we highlight a range of recent structure-aware approaches which utilise deep learning for compound design and virtual screening. We discuss how such methods can be better integrated into existing drug discovery pipelines by facilitating the design of compounds which conform to a specified design hypothesis and by uncovering key protein-ligand interactions which can be used to aid molecule design.

Virtual Screening with Gnina 1.0

Virtual screening-predicting which compounds within a specified compound library bind to a target molecule, typically a protein-is a fundamental task in the field of drug discovery. Doing virtual screening well provides tangible practical benefits, including reduced drug development costs, faster time to therapeutic viability, and fewer unforeseen side effects. As with most applied computational tasks, the algorithms currently used to perform virtual screening feature inherent tradeoffs between speed and accuracy. Furthermore, even theoretically rigorous, computationally intensive methods may fail to account for important effects relevant to whether a given compound will ultimately be usable as a drug. Here we investigate the virtual screening performance of the recently released Gnina molecular docking software, which uses deep convolutional networks to score protein-ligand structures. We find, on average, that Gnina outperforms conventional empirical scoring. The default scoring in Gnina outperforms the empirical AutoDock Vina scoring function on 89 of the 117 targets of the DUD-E and LIT-PCBA virtual screening benchmarks with a median 1% early enrichment factor that is more than twice that of Vina. However, we also find that issues of bias linger in these sets, even when not used directly to train models, and this bias obfuscates to what extent machine learning models are achieving their performance through a sophisticated interpretation of molecular interactions versus fitting to non-informative simplistic property distributions.

Ligand-Based Virtual Screening Based on the Graph Edit Distance

Chemical compounds can be represented as attributed graphs. An attributed graph is a mathematical model of an object composed of two types of representations: nodes and edges. Nodes are individual components, and edges are relations between these components. In this case, pharmacophore-type node descriptions are represented by nodes and chemical bounds by edges. If we want to obtain the bioactivity dissimilarity between two chemical compounds, a distance between attributed graphs can be used. The Graph Edit Distance allows computing this distance, and it is defined as the cost of transforming one graph into another. Nevertheless, to define this dissimilarity, the transformation cost must be properly tuned. The aim of this paper is to analyse the structural-based screening methods to verify the quality of the Harper transformation costs proposal and to present an algorithm to learn these transformation costs such that the bioactivity dissimilarity is properly defined in a ligand-based virtual screening application. The goodness of the dissimilarity is represented by the classification accuracy. Six publicly available datasets-CAPST, DUD-E, GLL&GDD, NRLiSt-BDB, MUV and ULS-UDS-have been used to validate our methodology and show that with our learned costs, we obtain the highest ratios in identifying the bioactivity similarity in a structurally diverse group of molecules.

Applications of Artificial Intelligence in Drug Design: Opportunities and Challenges

Artificial intelligence (AI) has undergone rapid development in recent years and has been successfully applied to real-world problems such as drug design. In this chapter, we review recent applications of AI to problems in drug design including virtual screening, computer-aided synthesis planning, and de novo molecule generation, with a focus on the limitations of the application of AI therein and opportunities for improvement. Furthermore, we discuss the broader challenges imposed by AI in translating theoretical practice to real-world drug design; including quantifying prediction uncertainty and explaining model behavior.

Extracting Predictive Representations from Hundreds of Millions of Molecules

The construction of appropriate representations remains essential for molecular predictions due to intricate molecular complexity. Additionally, it is often expensive and ethically constrained to generate labeled data for supervised learning in molecular sciences, leading to challenging small and diverse data sets. In this work, we develop a self-supervised learning approach to pretrain models from over 700 million unlabeled molecules in multiple databases. The intrinsic chemical logic learned from this approach enables the extraction of predictive representations from task-specific molecular sequences in a fine-tuned process. To understand the importance of self-supervised learning from unlabeled molecules, we assemble three models with different combinations of databases. Moreover, we propose a protocol based on data traits to automatically select the optimal model for a specific task. To validate the proposed method, we consider 10 benchmarks and 38 virtual screening data sets. Extensive validation indicates that the proposed method shows superb performance.

Mol2Context-vec: learning molecular representation from context awareness for drug discovery

With the rapid development of proteomics and the rapid increase of target molecules for drug action, computer-aided drug design (CADD) has become a basic task in drug discovery. One of the key challenges in CADD is molecular representation. High-quality molecular expression with chemical intuition helps to promote many boundary problems of drug discovery. At present, molecular representation still faces several urgent problems, such as the polysemy of substructures and unsmooth information flow between atomic groups. In this research, we propose a deep contextualized Bi-LSTM architecture, Mol2Context-vec, which can integrate different levels of internal states to bring dynamic representations of molecular substructures. And the obtained molecular context representation can capture the interactions between any atomic groups, especially a pair of atomic groups that are topologically distant. Experiments show that Mol2Context-vec achieves state-of-the-art performance on multiple benchmark datasets. In addition, the visual interpretation of Mol2Context-vec is very close to the structural properties of chemical molecules as understood by humans. These advantages indicate that Mol2Context-vec can be used as a reliable and effective tool for molecular expression. Availability: The source code is available for download in https://github.com/lol88/Mol2Context-vec.

Similarity-Based Virtual Screen Using Enhanced Siamese Multi-Layer Perceptron

Traditional drug development is a slow and costly process that leads to the production of new drugs. Virtual screening (VS) is a computational procedure that measures the similarity of molecules as one of its primary tasks. Many techniques for capturing the biological similarity between a test compound and a known target ligand have been established in ligand-based virtual screens (LBVSs). However, despite the good performances of the above methods compared to their predecessors, especially when dealing with molecules that have structurally homogenous active elements, they are not satisfied when dealing with molecules that are structurally heterogeneous. The main aim of this study is to improve the performance of similarity searching, especially with molecules that are structurally heterogeneous. The Siamese network will be used due to its capability to deal with complicated data samples in many fields. The Siamese multi-layer perceptron architecture will be enhanced by using two similarity distance layers with one fused layer, then multiple layers will be added after the fusion layer, and then the nodes of the model that contribute less or nothing during inference according to their signal-to-noise ratio values will be pruned. Several benchmark datasets will be used, which are: the MDL Drug Data Report (MDDR-DS1, MDDR-DS2, and MDDR-DS3), the Maximum Unbiased Validation (MUV), and the Directory of Useful Decoys (DUD). The results show the outperformance of the proposed method on standard Tanimoto coefficient (TAN) and other methods. Additionally, it is possible to reduce the number of nodes in the Siamese multilayer perceptron model while still keeping the effectiveness of recall on the same level.

End-to-end learning for compound activity prediction based on binding pocket information

BACKGROUND

Recently, machine learning-based ligand activity prediction methods have been greatly improved. However, if known active compounds of a target protein are unavailable, the machine learning-based method cannot be applied. In such cases, docking simulation is generally applied because it only requires a tertiary structure of the target protein. However, the conformation search and the evaluation of binding energy of docking simulation are computationally heavy and thus docking simulation needs huge computational resources. Thus, if we can apply a machine learning-based activity prediction method for a novel target protein, such methods would be highly useful. Recently, Tsubaki et al. proposed an end-to-end learning method to predict the activity of compounds for novel target proteins. However, the prediction accuracy of the method was still insufficient because it only used amino acid sequence information of a protein as the input.

RESULTS

In this research, we proposed an end-to-end learning-based compound activity prediction using structure information of a binding pocket of a target protein. The proposed method learns the important features by end-to-end learning using a graph neural network both for a compound structure and a protein binding pocket structure. As a result of the evaluation experiments, the proposed method has shown higher accuracy than an existing method using amino acid sequence information.

CONCLUSIONS

The proposed method achieved equivalent accuracy to docking simulation using AutoDock Vina with much shorter computing time. This indicated that a machine learning-based approach would be promising even for novel target proteins in activity prediction.

Machine-learning methods for ligand-protein molecular docking

Artificial intelligence (AI) is often presented as a new Industrial Revolution. Many domains use AI, including molecular simulation for drug discovery. In this review, we provide an overview of ligand-protein molecular docking and how machine learning (ML), especially deep learning (DL), a subset of ML, is transforming the field by tackling the associated challenges.

Comprehensive Survey of Recent Drug Discovery Using Deep Learning

Drug discovery based on artificial intelligence has been in the spotlight recently as it significantly reduces the time and cost required for developing novel drugs. With the advancement of deep learning (DL) technology and the growth of drug-related data, numerous deep-learning-based methodologies are emerging at all steps of drug development processes. In particular, pharmaceutical chemists have faced significant issues with regard to selecting and designing potential drugs for a target of interest to enter preclinical testing. The two major challenges are prediction of interactions between drugs and druggable targets and generation of novel molecular structures suitable for a target of interest. Therefore, we reviewed recent deep-learning applications in drug-target interaction (DTI) prediction and de novo drug design. In addition, we introduce a comprehensive summary of a variety of drug and protein representations, DL models, and commonly used benchmark datasets or tools for model training and testing. Finally, we present the remaining challenges for the promising future of DL-based DTI prediction and de novo drug design.

Recent trends in artificial intelligence-driven identification and development of anti-neurodegenerative therapeutic agents

Neurological disorders affect various aspects of life. Finding drugs for the central nervous system is a very challenging and complex task due to the involvement of the blood-brain barrier, P-glycoprotein, and the drug's high attrition rates. The availability of big data present in online databases and resources has enabled the emergence of artificial intelligence techniques including machine learning to analyze, process the data, and predict the unknown data with high efficiency. The use of these modern techniques has revolutionized the whole drug development paradigm, with an unprecedented acceleration in the central nervous system drug discovery programs. Also, the new deep learning architectures proposed in many recent works have given a better understanding of how artificial intelligence can tackle big complex problems that arose due to central nervous system disorders. Therefore, the present review provides comprehensive and up-to-date information on machine learning/artificial intelligence-triggered effort in the brain care domain. In addition, a brief overview is presented on machine learning algorithms and their uses in structure-based drug design, ligand-based drug design, ADMET prediction, de novo drug design, and drug repurposing. Lastly, we conclude by discussing the major challenges and limitations posed and how they can be tackled in the future by using these modern machine learning/artificial intelligence approaches.

Predicting Drug-Target Interactions with Deep-Embedding Learning of Graphs and Sequences

Computational approaches for predicting drug-target interactions (DTIs) play an important role in drug discovery since conventional screening experiments are time-consuming and expensive. In this study, we proposed end-to-end representation learning of a graph neural network with an attention mechanism and an attentive bidirectional long short-term memory (BiLSTM) to predict DTIs. For efficient training, we introduced a bidirectional encoder representations from transformers (BERT) pretrained method to extract substructure features from protein sequences and a local breadth-first search (BFS) to learn subgraph information from molecular graphs. Integrating both models, we developed a DTI prediction system. As a result, the proposed method achieved high performances with increases of 2.4% and 9.4% for AUC and recall, respectively, on unbalanced datasets compared with other methods. Extensive experiments showed that our model can relatively screen potential drugs for specific protein. Furthermore, visualizing the attention weights provides biological insight.

Deep Scoring Neural Network Replacing the Scoring Function Components to Improve the Performance of Structure-Based Molecular Docking

Accurate prediction of protein-ligand interactions can greatly promote drug development. Recently, a number of deep-learning-based methods have been proposed to predict protein-ligand binding affinities. However, these methods independently extract the feature representations of proteins and ligands but ignore the relative spatial positions and interaction pairs between them. Here, we propose a virtual screening method based on deep learning, called Deep Scoring, which directly extracts the relative position information and atomic attribute information on proteins and ligands from the docking poses. Furthermore, we use two Resnets to extract the features of ligand atoms and protein residues, respectively, and generate an atom-residue interaction matrix to learn the underlying principles of the interactions between proteins and ligands. This is then followed by a dual attention network (DAN) to generate the attention for two related entities (i.e., proteins and ligands) and to weigh the contributions of each atom and residue to binding affinity prediction. As a result, Deep Scoring outperforms other structure-based deep learning methods in terms of screening performance (area under the receiver operating characteristic curve (AUC) of 0.901 for an unbiased DUD-E version), pose prediction (AUC of 0.935 for PDBbind test set), and generalization ability (AUC of 0.803 for the CHEMBL data set). Finally, Deep Scoring was used to select novel ERK2 inhibitor, and two compounds (D264-0698 and D483-1785) were obtained with potential inhibitory activity on ERK2 through the biological experiments.

GHOST: Adjusting the Decision Threshold to Handle Imbalanced Data in Machine Learning

Machine learning classifiers trained on class imbalanced data are prone to overpredict the majority class. This leads to a larger misclassification rate for the minority class, which in many real-world applications is the class of interest. For binary data, the classification threshold is set by default to 0.5 which, however, is often not ideal for imbalanced data. Adjusting the decision threshold is a good strategy to deal with the class imbalance problem. In this work, we present two different automated procedures for the selection of the optimal decision threshold for imbalanced classification. A major advantage of our procedures is that they do not require retraining of the machine learning models or resampling of the training data. The first approach is specific for random forest (RF), while the second approach, named GHOST, can be potentially applied to any machine learning classifier. We tested these procedures on 138 public drug discovery data sets containing structure-activity data for a variety of pharmaceutical targets. We show that both thresholding methods improve significantly the performance of RF. We tested the use of GHOST with four different classifiers in combination with two molecular descriptors, and we found that most classifiers benefit from threshold optimization. GHOST also outperformed other strategies, including random undersampling and conformal prediction. Finally, we show that our thresholding procedures can be effectively applied to real-world drug discovery projects, where the imbalance and characteristics of the data vary greatly between the training and test sets.

Prediction of chemical compounds properties using a deep learning model

The discovery of new medications in a cost-effective manner has become the top priority for many pharmaceutical companies. Despite decades of innovation, many of their processes arguably remain relatively inefficient. One such process is the prediction of biological activity. This paper describes a new deep learning model, capable of conducting a preliminary screening of chemical compounds in-silico. The model has been constructed using a variation autoencoder to generate chemical compound fingerprints, which have been used to create a regression model to predict their LogD property and a classification model to predict binding in selected assays from the ChEMBL dataset. The conducted experiments demonstrate accurate prediction of the properties of chemical compounds only using structural definitions and also provide several opportunities to improve upon this model in the future.

Convolutional Neural Network-based Virtual Screening

Virtual screening is an important means for lead compound discovery. The scoring function is the key to selecting hit compounds. Many scoring functions are currently available; however, there are no all-purpose scoring functions because different scoring functions tend to have conflicting results. Recently, neural networks, especially convolutional neural networks, have constantly been penetrating drug design and most CNN-based virtual screening methods are superior to traditional docking methods, such as Dock and AutoDock. CNNbased virtual screening is expected to improve the previous model of overreliance on computational chemical screening. Utilizing the powerful learning ability of neural networks provides us with a new method for evaluating compounds. We review the latest progress of CNN-based virtual screening and propose prospects.

Deep Learning in Virtual Screening: Recent Applications and Developments

Drug discovery is a cost and time-intensive process that is often assisted by computational methods, such as virtual screening, to speed up and guide the design of new compounds. For many years, machine learning methods have been successfully applied in the context of computer-aided drug discovery. Recently, thanks to the rise of novel technologies as well as the increasing amount of available chemical and bioactivity data, deep learning has gained a tremendous impact in rational active compound discovery. Herein, recent applications and developments of machine learning, with a focus on deep learning, in virtual screening for active compound design are reviewed. This includes introducing different compound and protein encodings, deep learning techniques as well as frequently used bioactivity and benchmark data sets for model training and testing. Finally, the present state-of-the-art, including the current challenges and emerging problems, are examined and discussed.

LINGO-DL: a text-based approach for molecular similarity searching

The line notations of chemical structures are more compact than those of graphs and connection tables, so they can be useful for storing and transferring a large number of molecular structures. The simplified molecular input line system (SMILES) representation is the most extensively used, as it is much easier to utilise and comprehend than others, and it can be generated automatically from connection tables. A SMILES represents and encodes the molecule structure. It has been used by an existing method, LINGO, to calculate the molecular similarities and predict the structure-related properties. The LINGO method decomposes a canonical SMILES into a set of substrings of four characters referred to as LINGOs. The purpose of LINGO method is to measure the similarity between a pair of molecules by comparing the LINGOs that occur in each molecule. This paper aims to introduce an alternative version of the LINGO method using LINGOs of different lengths, called LINGO-DL. LINGO-DL is based on the fragmentation of canonical SMILES into substrings of three different lengths rather than one in LINGO method. Retrospective virtual screening experiments with MDDR, DUD, and MUV datasets show that the LINGO-DL outperforms the LINGO method, especially when the active molecules being sought have a high degree of structural heterogeneity.

A review on compound-protein interaction prediction methods: Data, format, representation and model

There has recently been a rapid progress in computational methods for determining protein targets of small molecule drugs, which will be termed as compound protein interaction (CPI). In this review, we comprehensively review topics related to computational prediction of CPI. Data for CPI has been accumulated and curated significantly both in quantity and quality. Computational methods have become powerful ever to analyze such complex the data. Thus, recent successes in the improved quality of CPI prediction are due to use of both sophisticated computational techniques and higher quality information in the databases. The goal of this article is to provide reviews of topics related to CPI, such as data, format, representation, to computational models, so that researchers can take full advantages of these resources to develop novel prediction methods. Chemical compounds and protein data from various resources were discussed in terms of data formats and encoding schemes. For the CPI methods, we grouped prediction methods into five categories from traditional machine learning techniques to state-of-the-art deep learning techniques. In closing, we discussed emerging machine learning topics to help both experimental and computational scientists leverage the current knowledge and strategies to develop more powerful and accurate CPI prediction methods.

Generating Property-Matched Decoy Molecules Using Deep Learning

MOTIVATION

An essential step in the development of virtual screening methods is the use of established sets of actives and decoys for benchmarking and training. However, the decoy molecules in commonly used sets are biased meaning that methods often exploit these biases to separate actives and decoys, and do not necessarily learn to perform molecular recognition. This fundamental issue prevents generalisation and hinders virtual screening method development.

RESULTS

We have developed a deep learning method (DeepCoy) that generates decoys to a user's preferred specification in order to remove such biases or construct sets with a defined bias.We validated DeepCoy using two established benchmarks, DUD-E and DEKOIS 2.0. For all 102 DUD-E targets and 80 of the 81 DEKOIS 2.0 targets, our generated decoy molecules more closely matched the active molecules' physicochemical properties while introducing no discernible additional risk of false negatives. The DeepCoy decoys improved the Deviation from Optimal Embedding (DOE) score by an average of 81% and 66%, respectively, decreasing from 0.166 to 0.032 for DUD-E and from 0.109 to 0.038 for DEKOIS 2.0. Further, the generated decoys are harder to distinguish than the original decoy molecules via docking with Autodock Vina, with virtual screening performance falling from an AUC ROC of 0.70 to 0.63.

AVAILABILITY AND IMPLEMENTATION

The code is available at https://github.com/oxpig/DeepCoy. Generated molecules can be downloaded from http://opig.stats.ox.ac.uk/resources.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Machine Learning Models to Predict Inhibition of the Bile Salt Export Pump

Cholestatic liver injury is frequently associated with drug inhibition of bile salt transporters, such as the bile salt export pump (BSEP). Reliable in silico models to predict BSEP inhibition directly from chemical structures would significantly reduce costs during drug discovery and could help avoid injury to patients. We report our development of classification and regression models for BSEP inhibition with substantially improved performance over previously published models. We assessed the performance effects of different methods of chemical featurization, data set partitioning, and class labeling and identified the methods producing models that generalized best to novel chemical entities.

A comprehensive comparison of molecular feature representations for use in predictive modeling

Machine learning methods are commonly used for predicting molecular properties to accelerate material and drug design. An important part of this process is deciding how to represent the molecules. Typically, machine learning methods expect examples represented by vectors of values, and many methods for calculating molecular feature representations have been proposed. In this paper, we perform a comprehensive comparison of different molecular features, including traditional methods such as fingerprints and molecular descriptors, and recently proposed learnable representations based on neural networks. Feature representations are evaluated on 11 benchmark datasets, used for predicting properties and measures such as mutagenicity, melting points, activity, solubility, and IC50. Our experiments show that several molecular features work similarly well over all benchmark datasets. The ones that stand out most are Spectrophores, which give significantly worse performance than other features on most datasets. Molecular descriptors from the PaDEL library seem very well suited for predicting physical properties of molecules. Despite their simplicity, MACCS fingerprints performed very well overall. The results show that learnable representations achieve competitive performance compared to expert based representations. However, task-specific representations (graph convolutions and Weave methods) rarely offer any benefits, even though they are computationally more demanding. Lastly, combining different molecular feature representations typically does not give a noticeable improvement in performance compared to individual feature representations.

Artificial Intelligence in Drug Discovery: A Comprehensive Review of Data-driven and Machine Learning Approaches

As expenditure on drug development increases exponentially, the overall drug discovery process requires a sustainable revolution. Since artificial intelligence (AI) is leading the fourth industrial revolution, AI can be considered as a viable solution for unstable drug research and development. Generally, AI is applied to fields with sufficient data such as computer vision and natural language processing, but there are many efforts to revolutionize the existing drug discovery process by applying AI. This review provides a comprehensive, organized summary of the recent research trends in AI-guided drug discovery process including target identification, hit identification, ADMET prediction, lead optimization, and drug repositioning. The main data sources in each field are also summarized in this review. In addition, an in-depth analysis of the remaining challenges and limitations will be provided, and proposals for promising future directions in each of the aforementioned areas.

Saagar-A New, Extensible Set of Molecular Substructures for QSAR/QSPR and Read-Across Predictions

Molecular structure-based predictive models provide a proven alternative to costly and inefficient animal testing. However, due to a lack of interpretability of predictive models built with abstract molecular descriptors they have earned the notoriety of being black boxes. Interpretable models require interpretable descriptors to provide chemistry-backed predictive reasoning and facilitate intelligent molecular design. We developed a novel set of extensible chemistry-aware substructures, Saagar, to support interpretable predictive models and read-across protocols. Performance of Saagar in chemical characterization and search for structurally similar actives for read-across applications was compared with four publicly available fingerprint sets (MACCS (166), PubChem (881), ECFP4 (1024), ToxPrint (729)) in three benchmark sets (MUV, ULS, and Tox21) spanning ∼145 000 compounds and 78 molecular targets at 1%, 2%, 5%, and 10% false discovery rates. In 18 of the 20 comparisons, interpretable Saagar features performed better than the publicly available, but less interpretable and fixed-bit length, fingerprints. Examples are provided to show the enhanced capability of Saagar in extracting compounds with higher scaffold similarity. Saagar features are interpretable and efficiently characterize diverse chemical collections, thus making them a better choice for building interpretable predictive in silico models and read-across protocols.

Virtual screening and drug repurposing experiments to identify potential novel selective MAO-B inhibitors for Parkinson's disease treatment

The main study's purpose is to detect novel natural products (NPs) that are potentially selective MAO-B inhibitors and, additionally, to computationally reposition the marketed drugs with a new therapeutic role for Parkinson's disease. To reach the goals, 3D similarity search, docking, ADMETox, and drug repurposing approaches were employed. Thus, an unbiased benchmarking dataset was built including selective and nonselective inhibitors for MAO-B compliant with both ligand- and structure-based virtual screening approaches. A retrospective and prospective mining scenario was applied to SPECS NP and DrugBank databases to detect novel scaffolds with potential benefits for Parkinson's disease patients. Out of the three best selected natural products, cardamomin showed excellently predicted drug-like properties, superior pharmacological profile, and specific interactions with MAO-B active site, indicating a potential selectivity over MAO-B. Two marketed drugs, fenamisal and monobenzone, were proposed as promising candidates repurposed for Parkinson's disease. The application of shape, physicochemical, and electrostatic similarity searches protocol emerged as a plausible solution to explore MAO-B inhibitors selectivity. This protocol might serve as a rewarding tool in early drug discovery and can be extended to other protein targets.

Comparing Fingerprints for Ligand-Based Virtual Screening: A Fast and Scalable Approach for Unbiased Evaluation

Ligand-based virtual screening is a useful tool for drug and probe discovery due to its high accessibility and scalability. The recent identification of bias in many data sets that were used in performance evaluation, quantified by the asymmetric validation embedding (AVE) score, has prompted the reanalysis of models to determine which performs best. Based on the understanding that ligand data are made up of blocks of highly correlated instances, we introduce a technique that quickly generates splits with AVE distributed close to zero using a combination of clustering and removal of the most biased data. We used our technique to compare the performance of the Morgan and CATS fingerprints and show that, after debiasing, the implementation of the CATS fingerprint performs significantly better. The code to replicate these results and perform low-bias splits is available at https://github.com/ljmartin/fp_low_ave.

Selecting machine-learning scoring functions for structure-based virtual screening

Interest in docking technologies has grown parallel to the ever increasing number and diversity of 3D models for macromolecular therapeutic targets. Structure-Based Virtual Screening (SBVS) aims at leveraging these experimental structures to discover the necessary starting points for the drug discovery process. It is now established that Machine Learning (ML) can strongly enhance the predictive accuracy of scoring functions for SBVS by exploiting large datasets from targets, molecules and their associations. However, with greater choice, the question of which ML-based scoring function is the most suitable for prospective use on a given target has gained importance. Here we analyse two approaches to select an existing scoring function for the target along with a third approach consisting in generating a scoring function tailored to the target. These analyses required discussing the limitations of popular SBVS benchmarks, the alternatives to benchmark scoring functions for SBVS and how to generate them or use them using freely-available software.

Transfer Learning for Drug Discovery

The data sets available to train models for in silico drug discovery efforts are often small. Indeed, the sparse availability of labeled data is a major barrier to artificial-intelligence-assisted drug discovery. One solution to this problem is to develop algorithms that can cope with relatively heterogeneous and scarce data. Transfer learning is a type of machine learning that can leverage existing, generalizable knowledge from other related tasks to enable learning of a separate task with a small set of data. Deep transfer learning is the most commonly used type of transfer learning in the field of drug discovery. This Perspective provides an overview of transfer learning and related applications to drug discovery to date. Furthermore, it provides outlooks on the future development of transfer learning for drug discovery.

Benchmarking Data Sets from PubChem BioAssay Data: Current Scenario and Room for Improvement

Developing realistic data sets for evaluating virtual screening methods is a task that has been tackled by the cheminformatics community for many years. Numerous artificially constructed data collections were developed, such as DUD, DUD-E, or DEKOIS. However, they all suffer from multiple drawbacks, one of which is the absence of experimental results confirming the impotence of presumably inactive molecules, leading to possible false negatives in the ligand sets. In light of this problem, the PubChem BioAssay database, an open-access repository providing the bioactivity information of compounds that were already tested on a biological target, is now a recommended source for data set construction. Nevertheless, there exist several issues with the use of such data that need to be properly addressed. In this article, an overview of benchmarking data collections built upon experimental PubChem BioAssay input is provided, along with a thorough discussion of noteworthy issues that one must consider during the design of new ligand sets from this database. The points raised in this review are expected to guide future developments in this regard, in hopes of offering better evaluation tools for novel in silico screening procedures.

One molecular fingerprint to rule them all: drugs, biomolecules, and the metabolome

BACKGROUND

Molecular fingerprints are essential cheminformatics tools for virtual screening and mapping chemical space. Among the different types of fingerprints, substructure fingerprints perform best for small molecules such as drugs, while atom-pair fingerprints are preferable for large molecules such as peptides. However, no available fingerprint achieves good performance on both classes of molecules.

RESULTS

Here we set out to design a new fingerprint suitable for both small and large molecules by combining substructure and atom-pair concepts. Our quest resulted in a new fingerprint called MinHashed atom-pair fingerprint up to a diameter of four bonds (MAP4). In this fingerprint the circular substructures with radii of r = 1 and r = 2 bonds around each atom in an atom-pair are written as two pairs of SMILES, each pair being combined with the topological distance separating the two central atoms. These so-called atom-pair molecular shingles are hashed, and the resulting set of hashes is MinHashed to form the MAP4 fingerprint. MAP4 significantly outperforms all other fingerprints on an extended benchmark that combines the Riniker and Landrum small molecule benchmark with a peptide benchmark recovering BLAST analogs from either scrambled or point mutation analogs. MAP4 furthermore produces well-organized chemical space tree-maps (TMAPs) for databases as diverse as DrugBank, ChEMBL, SwissProt and the Human Metabolome Database (HMBD), and differentiates between all metabolites in HMBD, over 70% of which are indistinguishable from their nearest neighbor using substructure fingerprints.

CONCLUSION

MAP4 is a new molecular fingerprint suitable for drugs, biomolecules, and the metabolome and can be adopted as a universal fingerprint to describe and search chemical space. The source code is available at https://github.com/reymond-group/map4 and interactive MAP4 similarity search tools and TMAPs for various databases are accessible at http://map-search.gdb.tools/ and http://tm.gdb.tools/map4/.