Profile-QSAR 2.0: Kinase Virtual Screening Accuracy Comparable to Four-Concentration IC50s for Realistically Novel Compounds

Kinase Drug Discovery: Impact of Open Science and Artificial Intelligence

Given their central role in signal transduction, protein kinases (PKs) were first implicated in cancer development, caused by aberrant intracellular signaling events. Since then, PKs have become major targets in different therapeutic areas. The preferred approach to therapeutic intervention of PK-dependent diseases is the use of small molecules to inhibit their catalytic phosphate group transfer activity. PK inhibitors (PKIs) are among the most intensely pursued drug candidates, with currently 80 approved compounds and several hundred in clinical trials. Following the elucidation of the human kinome and development of robust PK expression systems and high-throughput assays, large volumes of PK/PKI data have been produced in industrial and academic environments, more so than for many other pharmaceutical targets. In addition, hundreds of X-ray structures of PKs and their complexes with PKIs have been reported. Substantial amounts of PK/PKI data have been made publicly available in part as a result of open science initiatives. PK drug discovery is further supported through the incorporation of data science approaches, including the development of various specialized databases and online resources. Compound and activity data wealth compared to other targets has also made PKs a focal point for the application of artificial intelligence (AI) in pharmaceutical research. Herein, we discuss the interplay of open and data science in PK drug discovery and review exemplary studies that have substantially contributed to its development, including kinome profiling or the analysis of PKI promiscuity versus selectivity. We also take a close look at how AI approaches are beginning to impact PK drug discovery in light of their increasing data orientation.

An Ensemble Spectral Prediction (ESP) model for metabolite annotation

MOTIVATION

A key challenge in metabolomics is annotating measured spectra from a biological sample with chemical identities. Currently, only a small fraction of measurements can be assigned identities. Two complementary computational approaches have emerged to address the annotation problem: mapping candidate molecules to spectra, and mapping query spectra to molecular candidates. In essence, the candidate molecule with the spectrum that best explains the query spectrum is recommended as the target molecule. Despite candidate ranking being fundamental in both approaches, limited prior works incorporated rank learning tasks in determining the target molecule.

RESULTS

We propose a novel machine learning model, Ensemble Spectral Prediction (ESP), for metabolite annotation. ESP takes advantage of prior neural network-based annotation models that utilize multilayer perceptron (MLP) networks and Graph Neural Networks (GNNs). Based on the ranking results of the MLP- and GNN-based models, ESP learns a weighting for the outputs of MLP and GNN spectral predictors to generate a spectral prediction for a query molecule. Importantly, training data is stratified by molecular formula to provide candidate sets during model training. Further, baseline MLP and GNN models are enhanced by considering peak dependencies through label mixing and multi-tasking on spectral topic distributions. When trained on the NIST 2020 dataset and evaluated on the relevant candidate sets from PubChem, ESP improves average rank by 23.7% and 37.2% over the MLP and GNN baselines, respectively, demonstrating performance gain over state-of-the-art neural network approaches. However, MLP approaches remain strong contenders when considering top five ranks. Importantly, we show that annotation performance is dependent on the training dataset, the number of molecules in the candidate set and candidate similarity to the target molecule.

AVAILABILITY AND IMPLEMENTATION

The ESP code, a trained model, and a Jupyter notebook that guide users on using the ESP tool is available at https://github.com/HassounLab/ESP.

Advances in machine intelligence-driven virtual screening approaches for big-data

Virtual screening (VS) is an integral and ever-evolving domain of drug discovery framework. The VS is traditionally classified into ligand-based (LB) and structure-based (SB) approaches. Machine intelligence or artificial intelligence has wide applications in the drug discovery domain to reduce time and resource consumption. In combination with machine intelligence algorithms, VS has emerged into revolutionarily progressive technology that learns within robust decision orders for data curation and hit molecule screening from large VS libraries in minutes or hours. The exponential growth of chemical and biological data has evolved as "big-data" in the public domain demands modern and advanced machine intelligence-driven VS approaches to screen hit molecules from ultra-large VS libraries. VS has evolved from an individual approach (LB and SB) to integrated LB and SB techniques to explore various ligand and target protein aspects for the enhanced rate of appropriate hit molecule prediction. Current trends demand advanced and intelligent solutions to handle enormous data in drug discovery domain for screening and optimizing hits or lead with fewer or no false positive hits. Following the big-data drift and tremendous growth in computational architecture, we presented this review. Here, the article categorized and emphasized individual VS techniques, detailed literature presented for machine learning implementation, modern machine intelligence approaches, and limitations and deliberated the future prospects.

AI is a viable alternative to high throughput screening: a 318-target study

High throughput screening (HTS) is routinely used to identify bioactive small molecules. This requires physical compounds, which limits coverage of accessible chemical space. Computational approaches combined with vast on-demand chemical libraries can access far greater chemical space, provided that the predictive accuracy is sufficient to identify useful molecules. Through the largest and most diverse virtual HTS campaign reported to date, comprising 318 individual projects, we demonstrate that our AtomNet® convolutional neural network successfully finds novel hits across every major therapeutic area and protein class. We address historical limitations of computational screening by demonstrating success for target proteins without known binders, high-quality X-ray crystal structures, or manual cherry-picking of compounds. We show that the molecules selected by the AtomNet® model are novel drug-like scaffolds rather than minor modifications to known bioactive compounds. Our empirical results suggest that computational methods can substantially replace HTS as the first step of small-molecule drug discovery.

Predicting drug activity against cancer through genomic profiles and SMILES

Due to the constant increase in cancer rates, the disease has become a leading cause of death worldwide, enhancing the need for its detection and treatment. In the era of personalized medicine, the main goal is to incorporate individual variability in order to choose more precisely which therapy and prevention strategies suit each person. However, predicting the sensitivity of tumors to anticancer treatments remains a challenge. In this work, we propose two deep neural network models to predict the impact of anticancer drugs in tumors through the half-maximal inhibitory concentration (IC50). These models join biological and chemical data to apprehend relevant features of the genetic profile and the drug compounds, respectively. In order to predict the drug response in cancer cell lines, this study employed different DL methods, resorting to Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs). In the first stage, two autoencoders were pre-trained with high-dimensional gene expression and mutation data of tumors. Afterward, this genetic background is transferred to the prediction models that return the IC50 value that portrays the potency of a substance in inhibiting a cancer cell line. When comparing RSEM Expected counts and TPM as methods for displaying gene expression data, RSEM has been shown to perform better in deep models and CNNs model can obtain better insight in these types of data. Moreover, the obtained results reflect the effectiveness of the extracted deep representations in the prediction of the IC50 value that portrays the potency of a substance in inhibiting a tumor, achieving a performance of a mean squared error of 1.06 and surpassing previous state-of-the-art models.

Poor Generalization by Current Deep Learning Models for Predicting Binding Affinities of Kinase Inhibitors

The extreme surge of interest over the past decade surrounding the use of neural networks has inspired many groups to deploy them for predicting binding affinities of drug-like molecules to their receptors. A model that can accurately make such predictions has the potential to screen large chemical libraries and help streamline the drug discovery process. However, despite reports of models that accurately predict quantitative inhibition using protein kinase sequences and inhibitors' SMILES strings, it is still unclear whether these models can generalize to previously unseen data. Here, we build a Convolutional Neural Network (CNN) analogous to those previously reported and evaluate the model over four datasets commonly used for inhibitor/kinase predictions. We find that the model performs comparably to those previously reported, provided that the individual data points are randomly split between the training set and the test set. However, model performance is dramatically deteriorated when all data for a given inhibitor is placed together in the same training/testing fold, implying that information leakage underlies the models' performance. Through comparison to simple models in which the SMILES strings are tokenized, or in which test set predictions are simply copied from the closest training set data points, we demonstrate that there is essentially no generalization whatsoever in this model. In other words, the model has not learned anything about molecular interactions, and does not provide any benefit over much simpler and more transparent models. These observations strongly point to the need for richer structure-based encodings, to obtain useful prospective predictions of not-yet-synthesized candidate inhibitors.

Large-Scale Modeling of Sparse Protein Kinase Activity Data

Protein kinases are a protein family that plays an important role in several complex diseases such as cancer and cardiovascular and immunological diseases. Protein kinases have conserved ATP binding sites, which when targeted can lead to similar activities of inhibitors against different kinases. This can be exploited to create multitarget drugs. On the other hand, selectivity (lack of similar activities) is desirable in order to avoid toxicity issues. There is a vast amount of protein kinase activity data in the public domain, which can be used in many different ways. Multitask machine learning models are expected to excel for these kinds of data sets because they can learn from implicit correlations between tasks (in this case activities against a variety of kinases). However, multitask modeling of sparse data poses two major challenges: (i) creating a balanced train-test split without data leakage and (ii) handling missing data. In this work, we construct a protein kinase benchmark set composed of two balanced splits without data leakage, using random and dissimilarity-driven cluster-based mechanisms, respectively. This data set can be used for benchmarking and developing protein kinase activity prediction models. Overall, the performance on the dissimilarity-driven cluster-based split is lower than on random split-based sets for all models, indicating poor generalizability of models. Nevertheless, we show that multitask deep learning models, on this very sparse data set, outperform single-task deep learning and tree-based models. Finally, we demonstrate that data imputation does not improve the performance of (multitask) models on this benchmark set.

Building Machine Learning Small Molecule Melting Points and Solubility Models Using CCDC Melting Points Dataset

Predicting solubility of small molecules is a very difficult undertaking due to the lack of reliable and consistent experimental solubility data. It is well known that for a molecule in a crystal lattice to be dissolved, it must, first, dissociate from the lattice and then, second, be solvated. The melting point of a compound is proportional to the lattice energy, and the octanol-water partition coefficient (log P) is a measure of the compound's solvation efficiency. The CCDC's melting point dataset of almost one hundred thousand compounds was utilized to create widely applicable machine learning models of small molecule melting points. Using the general solubility equation, the aqueous thermodynamic solubilities of the same compounds can be predicted. The global model could be easily localized by adding additional melting point measurements for a chemical series of interest.

Improvement of multi-task learning by data enrichment: application for drug discovery

Multi-task learning in deep neural networks has become a topic of growing importance in many research fields, including drug discovery. However, applying multi-task learning poses new challenges in improving prediction performance. This study investigated the potential of training data enrichment to enhance multi-task model prediction quality in drug discovery. The study evaluated four scenarios with varying degrees of information capacity of the training data and applied two types of test data to evaluate prediction performance. We used three datasets: ViralChEMBL, which consisted of binary activities of compounds against viral species, was applied for the classification task; pQSAR(159) and pQSAR(4267), which consisted of bio-activities of compounds and assays from the research of the profile-QSAR method, were applied for regression tasks. We built multi-task models based on the feed-forward DNNs using the PyTorch framework. Our findings showed that training data enrichment could be an effective means of enhancing prediction performance in multi-task learning, but the degree of improvement depends on the quality of the training data. The more unique compounds and targets the training data included, the more new compound-target interactions are required for prediction improvement. Also, we found out that even using multi-task learning, one could not predict the interactions of compounds that are highly dissimilar from those used for model training. The study provides some recommendations for effectively employing multi-task learning in drug discovery to improve prediction accuracy and facilitate the discovery of novel drug candidates.

Computational Prediction of Drug Phenotypic Effects Based on Substructure-Phenotype Associations

Identifying drug phenotypic effects, including therapeutic effects and adverse drug reactions (ADRs), is an inseparable part for evaluating the potentiality of new drug candidates (NDCs). However, current computational methods for predicting phenotypic effects of NDCs are mainly based on the overall structure of an NDC or a related target. These approaches often lead to inconsistencies between the structures and functions and limit the prediction space of NDCs. In this study, first, we constructed quantitative associations of substructure-domain, domain-ADR, and domain-ATC (Anatomical Therapeutic Chemical Classification System code) through L1LOG and L1SVM machine learning models. These associations represent relationships between phenotypes (ADRs and ATCs) and local structures of drugs and proteins. Then, based on these established associations, substructure-phenotype relationships were constructed which were utilized to quantify drug-phenotype relationships. Thus, this approach could achieve high-throughput and effective evaluations of the druggability of NDCs by referring to the established substructure-phenotype relationships and structural information of NDCs without additional prior knowledge. Using this computational pipeline, 83,205 drug-ATC relationships (including 1,479 drugs and 178 ATCs) and 306,421 drug-ADR relationships (including 1,752 drugs and 454 ADRs) were predicted in total. The prediction results were validated at four levels: five-fold cross validation, public databases, literature, and molecular docking. Furthermore, three case studies demonstrated the feasibility of our method. 79 ATCs and 269 ADRs were predicted to be related to Maraviroc, an approved drug, including the existing antiviral effect in clinical use. Additionally, we also found risk substructures of severe ADRs, for example, SUB215 (>= 1, saturated or only aromatic carbon ring size 7) can result in shock. And we analyzed the mechanism of action (MOA) of interested drugs based on the established drug-substructure-domain-protein associations. In a word, this approach through establishing drug-substructure-phenotype relationships can achieve quantitative prediction of phenotypes for a given NDC or drug without any prior knowledge except its structure information. Using that way, we can directly obtain the relationships between substructure and phenotype of a compound, which is more convenient to analyze the phenotypic mechanism of drugs and accelerate the process of rational drug design.

Large-scale comparison of machine learning algorithms for target prediction of natural products

Natural products (NPs) and their derivatives are important resources for drug discovery. There are many in silico target prediction methods that have been reported, however, very few of them distinguish NPs from synthetic molecules. Considering the fact that NPs and synthetic molecules are very different in many characteristics, it is necessary to build specific target prediction models of NPs. Therefore, we collected the activity data of NPs and their derivatives from the public databases and constructed four datasets, including the NP dataset, the NPs and its first-class derivatives dataset, the NPs and all its derivatives and the ChEMBL26 compounds dataset. Conditions, including activity thresholds and input features, were explored to access the performance of eight machine learning methods of target prediction of NPs, including support vector machines (SVM), extreme gradient boosting, random forests, K-nearest neighbor, naive Bayes, feedforward neural networks (FNN), convolutional neural networks and recurrent neural networks. As a result, the NPs and all their derivatives datasets were selected to build the best NP-specific models. Furthermore, the consensus models, as well as the voting models, were additionally applied to improve the prediction performance. More evaluations were made on the external validation set and the results demonstrated that (1) the NP-specific model performed better on the target prediction of NPs than the traditional models training on the whole compounds of ChEMBL26. (2) The consensus model of FNN + SVM possessed the best overall performance, and the voting model can significantly improve recall and specificity.

Artificial intelligence and machine-learning approaches in structure and ligand-based discovery of drugs affecting central nervous system

CNS disorders are indications with a very high unmet medical needs, relatively smaller number of available drugs, and a subpar satisfaction level among patients and caregiver. Discovery of CNS drugs is extremely expensive affair with its own unique challenges leading to extremely high attrition rates and low efficiency. With explosion of data in information age, there is hardly any aspect of life that has not been touched by data driven technologies such as artificial intelligence (AI) and machine learning (ML). Drug discovery is no exception, emergence of big data via genomic, proteomic, biological, and chemical technologies has driven pharmaceutical giants to collaborate with AI oriented companies to revolutionise drug discovery, with the goal of increasing the efficiency of the process. In recent years many examples of innovative applications of AI and ML techniques in CNS drug discovery has been reported. Research on therapeutics for diseases such as schizophrenia, Alzheimer's and Parkinsonism has been provided with a new direction and thrust from these developments. AI and ML has been applied to both ligand-based and structure-based drug discovery and design of CNS therapeutics. In this review, we have summarised the general aspects of AI and ML from the perspective of drug discovery followed by a comprehensive coverage of the recent developments in the applications of AI/ML techniques in CNS drug discovery.

Analysis of the benefits of imputation models over traditional QSAR models for toxicity prediction

Recently, imputation techniques have been adapted to predict activity values among sparse bioactivity matrices, showing improvements in predictive performance over traditional QSAR models. These models are able to use experimental activity values for auxiliary assays when predicting the activity of a test compound on a specific assay. In this study, we tested three different multi-task imputation techniques on three classification-based toxicity datasets: two of small scale (12 assays each) and one large scale with 417 assays. Moreover, we analyzed in detail the improvements shown by the imputation models. We found that test compounds that were dissimilar to training compounds, as well as test compounds with a large number of experimental values for other assays, showed the largest improvements. We also investigated the impact of sparsity on the improvements seen as well as the relatedness of the assays being considered. Our results show that even a small amount of additional information can provide imputation methods with a strong boost in predictive performance over traditional single task and multi-task predictive models.

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.

Analysis of Training and Seed Bias in Small Molecules Generated with a Conditional Graph-Based Variational Autoencoder─Insights for Practical AI-Driven Molecule Generation

The application of deep learning to generative molecule design has shown early promise for accelerating lead series development. However, questions remain concerning how factors like training, data set, and seed bias impact the technology's utility to medicinal and computational chemists. In this work, we analyze the impact of seed and training bias on the output of an activity-conditioned graph-based variational autoencoder (VAE). Leveraging a massive, labeled data set corresponding to the dopamine D2 receptor, our graph-based generative model is shown to excel in producing desired conditioned activities and favorable unconditioned physical properties in generated molecules. We implement an activity-swapping method that allows for the activation, deactivation, or retention of activity of molecular seeds, and we apply independent deep learning classifiers to verify the generative results. Overall, we uncover relationships between noise, molecular seeds, and training set selection across a range of latent-space sampling procedures, providing important insights for practical AI-driven molecule generation.

Active Site Sequence Representations of Human Kinases Outperform Full Sequence Representations for Affinity Prediction and Inhibitor Generation: 3D Effects in a 1D Model

Recent advances in deep learning have enabled the development of large-scale multimodal models for virtual screening and de novo molecular design. The human kinome with its abundant sequence and inhibitor data presents an attractive opportunity to develop proteochemometric models that exploit the size and internal diversity of this family of targets. Here, we challenge a standard practice in sequence-based affinity prediction models: instead of leveraging the full primary structure of proteins, each target is represented by a sequence of 29 discontiguous residues defining the ATP binding site. In kinase-ligand binding affinity prediction, our results show that the reduced active site sequence representation is not only computationally more efficient but consistently yields significantly higher performance than the full primary structure. This trend persists across different models, data sets, and performance metrics and holds true when predicting pIC₅₀ for both unseen ligands and kinases. Our interpretability analysis reveals a potential explanation for the superiority of the active site models: whereas only mild statistical effects about the extraction of three-dimensional (3D) interaction sites take place in the full sequence models, the active site models are equipped with an implicit but strong inductive bias about the 3D structure stemming from the discontiguity of the active sites. Moreover, in direct comparisons, our models perform similarly or better than previous state-of-the-art approaches in affinity prediction. We then investigate a de novo molecular design task and find that the active site provides benefits in the computational efficiency, but otherwise, both kinase representations yield similar optimized affinities (for both SMILES- and SELFIES-based molecular generators). Our work challenges the assumption that the full primary structure is indispensable for modeling human kinases.

Splitting chemical structure data sets for federated privacy-preserving machine learning

With the increase in applications of machine learning methods in drug design and related fields, the challenge of designing sound test sets becomes more and more prominent. The goal of this challenge is to have a realistic split of chemical structures (compounds) between training, validation and test set such that the performance on the test set is meaningful to infer the performance in a prospective application. This challenge is by its own very interesting and relevant, but is even more complex in a federated machine learning approach where multiple partners jointly train a model under privacy-preserving conditions where chemical structures must not be shared between the different participating parties. In this work we discuss three methods which provide a splitting of a data set and are applicable in a federated privacy-preserving setting, namely: a. locality-sensitive hashing (LSH), b. sphere exclusion clustering, c. scaffold-based binning (scaffold network). For evaluation of these splitting methods we consider the following quality criteria (compared to random splitting): bias in prediction performance, classification label and data imbalance, similarity distance between the test and training set compounds. The main findings of the paper are a. both sphere exclusion clustering and scaffold-based binning result in high quality splitting of the data sets, b. in terms of compute costs sphere exclusion clustering is very expensive in the case of federated privacy-preserving setting.

Validation strategies for target prediction methods

Computational methods for target prediction, based on molecular similarity and network-based approaches, machine learning, docking and others, have evolved as valuable and powerful tools to aid the challenging task of mode of action identification for bioactive small molecules such as drugs and drug-like compounds. Critical to discerning the scope and limitations of a target prediction method is understanding how its performance was evaluated and reported. Ideally, large-scale prospective experiments are conducted to validate the performance of a model; however, this expensive and time-consuming endeavor is often not feasible. Therefore, to estimate the predictive power of a method, statistical validation based on retrospective knowledge is commonly used. There are multiple statistical validation techniques that vary in rigor. In this review we discuss the validation strategies employed, highlighting the usefulness and constraints of the validation schemes and metrics that are employed to measure and describe performance. We address the limitations of measuring only generalized performance, given that the underlying bioactivity and structural data are biased towards certain small-molecule scaffolds and target families, and suggest additional aspects of performance to consider in order to produce more detailed and realistic estimates of predictive power. Finally, we describe the validation strategies that were employed by some of the most thoroughly validated and accessible target prediction methods.

SWnet: a deep learning model for drug response prediction from cancer genomic signatures and compound chemical structures

Background

One of the major challenges in precision medicine is accurate prediction of individual patient’s response to drugs. A great number of computational methods have been developed to predict compounds activity using genomic profiles or chemical structures, but more exploration is yet to be done to combine genetic mutation, gene expression, and cheminformatics in one machine learning model.

Results

We presented here a novel deep-learning model that integrates gene expression, genetic mutation, and chemical structure of compounds in a multi-task convolutional architecture. We applied our model to the Genomics of Drug Sensitivity in Cancer (GDSC) and Cancer Cell Line Encyclopedia (CCLE) datasets. We selected relevant cancer-related genes based on oncology genetics database and L1000 landmark genes, and used their expression and mutations as genomic features in model training. We obtain the cheminformatics features for compounds from PubChem or ChEMBL. Our finding is that combining gene expression, genetic mutation, and cheminformatics features greatly enhances the predictive performance.

Conclusion

We implemented an extended Graph Neural Network for molecular graphs and Convolutional Neural Network for gene features. With the employment of multi-tasking and self-attention functions to monitor the similarity between compounds, our model outperforms recently published methods using the same training and testing datasets.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-021-04352-9.

Enzyme Promiscuity Prediction Using Hierarchy-Informed Multi-Label Classification

MOTIVATION

As experimental efforts are costly and time consuming, computational characterization of enzyme capabilities is an attractive alternative. We present and evaluate several machine-learning models to predict which of 983 distinct enzymes, as defined via the Enzyme Commission (EC) numbers, are likely to interact with a given query molecule. Our data consists of enzyme-substrate interactions from the BRENDA database. Some interactions are attributed to natural selection and involve the enzyme's natural substrates. The majority of the interactions however involve non-natural substrates, thus reflecting promiscuous enzymatic activities.

RESULTS

We frame this "enzyme promiscuity prediction" problem as a multi-label classification task. We maximally utilize inhibitor and unlabelled data to train prediction models that can take advantage of known hierarchical relationships between enzyme classes. We report that a hierarchical multi-label neural network, EPP-HMCNF, is the best model for solving this problem, outperforming k-nearest neighbours similarity-based and other machine learning models. We show that inhibitor information during training consistently improves predictive power, particularly for EPP-HMCNF. We also show that all promiscuity prediction models perform worse under a realistic data split when compared to a random data split, and when evaluating performance on non-natural substrates compared to natural substrates.

AVAILABILITY AND IMPLEMENTATION

We provide Python code for EPP-HMCNF and other models in a repository termed EPP (Enzyme Promiscuity Prediction) at https://github.com/hassounlab/EPP.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Artificial intelligence and machine learning-aided drug discovery in central nervous system diseases: State-of-the-arts and future directions

Neurological disorders significantly outnumber diseases in other therapeutic areas. However, developing drugs for central nervous system (CNS) disorders remains the most challenging area in drug discovery, accompanied with the long timelines and high attrition rates. With the rapid growth of biomedical data enabled by advanced experimental technologies, artificial intelligence (AI) and machine learning (ML) have emerged as an indispensable tool to draw meaningful insights and improve decision making in drug discovery. Thanks to the advancements in AI and ML algorithms, now the AI/ML-driven solutions have an unprecedented potential to accelerate the process of CNS drug discovery with better success rate. In this review, we comprehensively summarize AI/ML-powered pharmaceutical discovery efforts and their implementations in the CNS area. After introducing the AI/ML models as well as the conceptualization and data preparation, we outline the applications of AI/ML technologies to several key procedures in drug discovery, including target identification, compound screening, hit/lead generation and optimization, drug response and synergy prediction, de novo drug design, and drug repurposing. We review the current state-of-the-art of AI/ML-guided CNS drug discovery, focusing on blood-brain barrier permeability prediction and implementation into therapeutic discovery for neurological diseases. Finally, we discuss the major challenges and limitations of current approaches and possible future directions that may provide resolutions to these difficulties.

Collaborative Profile-QSAR: A Natural Platform for Building Collaborative Models among Competing Companies

Massively multitask bioactivity models that transfer learning between thousands of assays have been shown to work dramatically better than separate models trained on each individual assay. In particular, the applicability domain for a given model can expand from compounds similar to those tested in that specific assay to those tested across the full complement of contributing assays. If many large companies would share their assay data and train models on the superset, predictions should be better than what each company can do alone. However, a company's compounds, targets, and activities are among their most guarded trade secrets. Strategies have been proposed to share just the individual collaborators' models, without exposing any of the training data. Profile-QSAR (pQSAR) is a two-level, multitask, stacked model. It uses profiles of level-1 predictions from single-task models for thousands of assays as compound descriptors for level-2 models. This work describes its simple and natural adaptation to safe collaboration by model sharing. Broad model sharing has not yet been implemented across multiple large companies, so there are numerous unanswered questions. Novartis was formed from several mergers and acquisitions. In principle, this should allow an internal simulation of model sharing. In practice, the lack of metadata about the origins of compounds and assays made this difficult. Nevertheless, we have attempted to simulate this process and propose some findings: multitask pQSAR is always an improvement over single-task models; collaborative multitask modeling did not improve predictions on internal compounds; collaboration did improve predictions for external compounds but far less than the purely internal multitask modeling for internal compounds; collaborative models for external compounds increasingly improve as overlap between compound collections increases; combining profiles from inside and outside the company is not best, with internal predictions better using only the inside profile and external using only the outside profile, but a consensus of models using all three profiles is best on external compounds and a good compromise on internal compounds. We anticipate similar results from other model-sharing approaches. Indeed, since collaborative pQSAR through model sharing is mathematically identical to pQSAR using actual shared data, we believe our conclusions should apply to collaborative modeling by any current method even including the unlikely scenario of directly sharing all chemical structures and assay data.

Evolution of Novartis' Small Molecule Screening Deck Design

This article summarizes the evolution of the screening deck at the Novartis Institutes for BioMedical Research (NIBR). Historically, the screening deck was an assembly of all available compounds. In 2015, we designed a first deck to facilitate access to diverse subsets with optimized properties. We allocated the compounds as plated subsets on a 2D grid with property based ranking in one dimension and increasing structural redundancy in the other. The learnings from the 2015 screening deck were applied to the design of a next generation in 2019. We found that using traditional leadlikeness criteria (mainly MW, clogP) reduces the hit rates of attractive chemical starting points in subset screening. Consequently, the 2019 deck relies on solubility and permeability to select preferred compounds. The 2019 design also uses NIBR's experimental assay data and inferred biological activity profiles in addition to structural diversity to define redundancy across the compound sets.

Practical Applications of Deep Learning To Impute Heterogeneous Drug Discovery Data

Contemporary deep learning approaches still struggle to bring a useful improvement in the field of drug discovery because of the challenges of sparse, noisy, and heterogeneous data that are typically encountered in this context. We use a state-of-the-art deep learning method, Alchemite, to impute data from drug discovery projects, including multitarget biochemical activities, phenotypic activities in cell-based assays, and a variety of absorption, distribution, metabolism, and excretion (ADME) endpoints. The resulting model gives excellent predictions for activity and ADME endpoints, offering an average increase in R² of 0.22 versus quantitative structure-activity relationship methods. The model accuracy is robust to combining data across uncorrelated endpoints and projects with different chemical spaces, enabling a single model to be trained for all compounds and endpoints. We demonstrate improvements in accuracy on the latest chemistry and data when updating models with new data as an ongoing medicinal chemistry project progresses.

QSAR-derived affinity fingerprints (part 2): modeling performance for potency prediction

Affinity fingerprints report the activity of small molecules across a set of assays, and thus permit to gather information about the bioactivities of structurally dissimilar compounds, where models based on chemical structure alone are often limited, and model complex biological endpoints, such as human toxicity and in vitro cancer cell line sensitivity. Here, we propose to model in vitro compound activity using computationally predicted bioactivity profiles as compound descriptors. To this aim, we apply and validate a framework for the calculation of QSAR-derived affinity fingerprints (QAFFP) using a set of 1360 QSAR models generated using K_i, K_d, IC₅₀ and EC₅₀ data from ChEMBL database. QAFFP thus represent a method to encode and relate compounds on the basis of their similarity in bioactivity space. To benchmark the predictive power of QAFFP we assembled IC₅₀ data from ChEMBL database for 18 diverse cancer cell lines widely used in preclinical drug discovery, and 25 diverse protein target data sets. This study complements part 1 where the performance of QAFFP in similarity searching, scaffold hopping, and bioactivity classification is evaluated. Despite being inherently noisy, we show that using QAFFP as descriptors leads to errors in prediction on the test set in the ~ 0.65-0.95 pIC₅₀ units range, which are comparable to the estimated uncertainty of bioactivity data in ChEMBL (0.76-1.00 pIC₅₀ units). We find that the predictive power of QAFFP is slightly worse than that of Morgan2 fingerprints and 1D and 2D physicochemical descriptors, with an effect size in the 0.02-0.08 pIC₅₀ units range. Including QSAR models with low predictive power in the generation of QAFFP does not lead to improved predictive power. Given that the QSAR models we used to compute the QAFFP were selected on the basis of data availability alone, we anticipate better modeling results for QAFFP generated using more diverse and biologically meaningful targets. Data sets and Python code are publicly available at https://github.com/isidroc/QAFFP_regression .

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

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

Using Predicted Bioactivity Profiles to Improve Predictive Modeling

Predictive modeling is a cornerstone in early drug development. Using information for multiple domains or across prediction tasks has the potential to improve the performance of predictive modeling. However, aggregating data often leads to incomplete data matrices that might be limiting for modeling. In line with previous studies, we show that by generating predicted bioactivity profiles, and using these as additional features, prediction accuracy of biological endpoints can be improved. Using conformal prediction, a type of confidence predictor, we present a robust framework for the calculation of these profiles and the evaluation of their impact. We report on the outcomes from several approaches to generate the predicted profiles on 16 datasets in cytotoxicity and bioactivity and show that efficiency is improved the most when including the p-values from conformal prediction as bioactivity profiles.

Imputation versus prediction: applications in machine learning for drug discovery

Imputation is a powerful statistical method that is distinct from the predictive modelling techniques more commonly used in drug discovery. Imputation uses sparse experimental data in an incomplete dataset to predict missing values by leveraging correlations between experimental assays. This contrasts with quantitative structure–activity relationship methods that use only descriptor – assay correlations. We summarize three recent imputation strategies – heterogeneous deep imputation, assay profile methods and matrix factorization – and compare these with quantitative structure–activity relationship methods, including deep learning, in drug discovery settings. We comment on the value added by imputation methods when used in an ongoing project and find that imputation produces stronger models, earlier in the project, over activity and absorption, distribution, metabolism and elimination end points.

Biased Diversity for Effective Virtual Screening

Virtual screening is no longer merely a matter of identifying the subset of compounds from a large collection likely to be active against a particular endpoint. This viewpoint shares some distinctive practices at Novartis, where virtual screening combines multiple computational tools that marry the competing goals of biasing the selection of compounds toward multiple desired properties, while diversifying the selection to sample the available chemistry space, identifying quality compounds that inform drug discovery. Topics include the various considerations needed for a successful virtual screening practice: triaging, compound quality, accuracy and test sets, activity prediction including multitask modeling, virtual profiling, automation, multiproperty bias, diversity and property spaces, and biased-diversity designs.

Current methods and challenges for deep learning in drug discovery

Driven by rapid advances in computer hardware and publicly available datasets over the past decade, deep learning has achieved tremendous success in the transformation of many computational disciplines. These novel technologies have had considerable impact on computer-aided drug design as well, throughout all stages of the development pipeline. A flexible toolbox of neural architectures has been developed that are well-suited to represent the sequential, topological, or geometrical concepts of chemistry and biology; and that are able to either discriminate existing molecules or to generate new ones from scratch. For some biochemical prediction tasks, the state of the art has been advanced; however, for complex and practically relevant projects, the outcomes are less clear-cut. Current deep learning methods rely on massive amounts of labeled examples, but drug discovery data is comparatively limited in quantity and quality. These problems need to be resolved and existing sources used more effectively to demonstrate that deep learning can revolutionize the field in general.

Novel Consensus Architecture To Improve Performance of Large-Scale Multitask Deep Learning QSAR Models

Advances in the development of high-throughput screening and automated chemistry have rapidly accelerated the production of chemical and biological data, much of them freely accessible through literature aggregator services such as ChEMBL and PubChem. Here, we explore how to use this comprehensive mapping of chemical biology space to support the development of large-scale quantitative structure-activity relationship (QSAR) models. We propose a new deep learning consensus architecture (DLCA) that combines consensus and multitask deep learning approaches together to generate large-scale QSAR models. This method improves knowledge transfer across different target/assays while also integrating contributions from models based on different descriptors. The proposed approach was validated and compared with proteochemometrics, multitask deep learning, and Random Forest methods paired with various descriptors types. DLCA models demonstrated improved prediction accuracy for both regression and classification tasks. The best models together with their modeling sets are provided through publicly available web services at https://predictor.ncats.io .

All-Assay-Max2 pQSAR: Activity Predictions as Accurate as Four-Concentration IC50s for 8558 Novartis Assays

Profile-quantitative structure-activity relationship (pQSAR) is a massively multitask, two-step machine learning method with unprecedented scope, accuracy, and applicability domain. In step one, a "profile" of conventional single-assay random forest regression models are trained on a very large number of biochemical and cellular pIC₅₀ assays using Morgan 2 substructural fingerprints as compound descriptors. In step two, a panel of partial least squares (PLS) models are built using the profile of pIC₅₀ predictions from those random forest regression models as compound descriptors (hence the name). Previously described for a panel of 728 biochemical and cellular kinase assays, we have now built an enormous pQSAR from 11 805 diverse Novartis (NVS) IC₅₀ and EC₅₀ assays. This large number of assays, and hence of compound descriptors for PLS, dictated reducing the profile by only including random forest regression models whose predictions correlate with the assay being modeled. The random forest regression and pQSAR models were evaluated with our "realistically novel" held-out test set, whose median average similarity to the nearest training set member across the 11 805 assays was only 0.34, comparable to the novelty of compounds actually selected from virtual screens. For the 11 805 single-assay random forest regression models, the median correlation of prediction with the experiment was only r_ext² = 0.05, virtually random, and only 8% of the models achieved our standard success threshold of r_ext² = 0.30. For pQSAR, the median correlation was r_ext² = 0.53, comparable to four-concentration experimental IC₅₀s, and 72% of the models met our r_ext² > 0.30 standard, totaling 8558 successful models. The successful models included assays from all of the 51 annotated target subclasses, as well as 4196 phenotypic assays, indicating that pQSAR can be applied to virtually any disease area. Every month, all models are updated to include new measurements, and predictions are made for 5.5 million NVS compounds, totaling 50 billion predictions. Common uses have included virtual screening, selectivity design, toxicity and promiscuity prediction, mechanism-of-action prediction, and others. Several such actual applications are described.

Concepts of Artificial Intelligence for Computer-Assisted Drug Discovery

Artificial intelligence (AI), and, in particular, deep learning as a subcategory of AI, provides opportunities for the discovery and development of innovative drugs. Various machine learning approaches have recently (re)emerged, some of which may be considered instances of domain-specific AI which have been successfully employed for drug discovery and design. This review provides a comprehensive portrayal of these machine learning techniques and of their applications in medicinal chemistry. After introducing the basic principles, alongside some application notes, of the various machine learning algorithms, the current state-of-the art of AI-assisted pharmaceutical discovery is discussed, including applications in structure- and ligand-based virtual screening, de novo drug design, physicochemical and pharmacokinetic property prediction, drug repurposing, and related aspects. Finally, several challenges and limitations of the current methods are summarized, with a view to potential future directions for AI-assisted drug discovery and design.

Opportunities and challenges using artificial intelligence in ADME/Tox

A recent conference organized a panel of scientists representing small and big pharma companies, who work at the interface of machine learning (ML) and absorption, distribution, metabolism, excretion, and toxicology (ADME/Tox). With the recent rebirth of AI related to pharma, it is timely to present this collaborative commentary to capture the diverging opinions on the past, present and future role of AI for ADME/Tox and how it can be applied in newer areas such as nanomaterials.

Imputation of Assay Bioactivity Data Using Deep Learning

We describe a novel deep learning neural network method and its application to impute assay pIC₅₀ values. Unlike conventional machine learning approaches, this method is trained on sparse bioactivity data as input, typical of that found in public and commercial databases, enabling it to learn directly from correlations between activities measured in different assays. In two case studies on public domain data sets we show that the neural network method outperforms traditional quantitative structure-activity relationship (QSAR) models and other leading approaches. Furthermore, by focusing on only the most confident predictions the accuracy is increased to R² > 0.9 using our method, as compared to R² = 0.44 when reporting all predictions.

PotentialNet for Molecular Property Prediction

The arc of drug discovery entails a multiparameter optimization problem spanning vast length scales. The key parameters range from solubility (angstroms) to protein-ligand binding (nanometers) to in vivo toxicity (meters). Through feature learning-instead of feature engineering-deep neural networks promise to outperform both traditional physics-based and knowledge-based machine learning models for predicting molecular properties pertinent to drug discovery. To this end, we present the PotentialNet family of graph convolutions. These models are specifically designed for and achieve state-of-the-art performance for protein-ligand binding affinity. We further validate these deep neural networks by setting new standards of performance in several ligand-based tasks. In parallel, we introduce a new metric, the Regression Enrichment Factor EF_χ ^(R), to measure the early enrichment of computational models for chemical data. Finally, we introduce a cross-validation strategy based on structural homology clustering that can more accurately measure model generalizability, which crucially distinguishes the aims of machine learning for drug discovery from standard machine learning tasks.

Discovering Highly Potent Molecules from an Initial Set of Inactives Using Iterative Screening

The versatility of similarity searching and quantitative structure-activity relationships to model the activity of compound sets within given bioactivity ranges (i.e., interpolation) is well established. However, their relative performance in the common scenario in early stage drug discovery where lots of inactive data but no active data points are available (i.e., extrapolation from the low-activity to the high-activity range) has not been thoroughly examined yet. To this aim, we have designed an iterative virtual screening strategy which was evaluated on 25 diverse bioactivity data sets from ChEMBL. We benchmark the efficiency of random forest (RF), multiple linear regression, ridge regression, similarity searching, and random selection of compounds to identify a highly active molecule in the test set among a large number of low-potency compounds. We use the number of iterations required to find this active molecule to evaluate the performance of each experimental setup. We show that linear and ridge regression often outperform RF and similarity searching, reducing the number of iterations to find an active compound by a factor of 2 or more. Even simple regression methods seem better able to extrapolate to high-bioactivity ranges than RF, which only provides output values in the range covered by the training set. In addition, examination of the scaffold diversity in the data sets used shows that in some cases similarity searching and RF require two times as many iterations as random selection depending on the chemical space covered in the initial training data. Lastly, we show using bioactivity data for COX-1 and COX-2 that our framework can be extended to multitarget drug discovery, where compounds are selected by concomitantly considering their activity against multiple targets. Overall, this study provides an approach for iterative screening where only inactive data are present in early stages of drug discovery in order to discover highly potent compounds and the best experimental set up in which to do so.

X-ray Structures and Feasibility Assessment of CLK2 Inhibitors for Phelan-McDermid Syndrome

CLK2 inhibition has been proposed as a potential mechanism to improve autism and neuronal functions in Phelan-McDermid syndrome (PMDS). Herein, the discovery of a very potent indazole CLK inhibitor series and the CLK2 X-ray structure of the most potent analogue are reported. This new indazole series was identified through a biochemical CLK2 Caliper assay screen with 30k compounds selected by an in silico approach. Novel high-resolution X-ray structures of all CLKs, including the first CLK4 X-ray structure, bound to known CLK2 inhibitor tool compounds (e.g., TG003, CX-4945), are also shown and yield insight into inhibitor selectivity in the CLK family. The efficacy of the new CLK2 inhibitors from the indazole series was demonstrated in the mouse brain slice assay, and potential safety concerns were investigated. Genotoxicity findings in the human lymphocyte micronucleus test (MNT) assay are shown by using two structurally different CLK inhibitors to reveal a major concern for pan-CLK inhibition in PMDS.

Most Ligand-Based Classification Benchmarks Reward Memorization Rather than Generalization

Undetected overfitting can occur when there are significant redundancies between training and validation data. We describe AVE, a new measure of training-validation redundancy for ligand-based classification problems, that accounts for the similarity among inactive molecules as well as active ones. We investigated seven widely used benchmarks for virtual screening and classification, and we show that the amount of AVE bias strongly correlates with the performance of ligand-based predictive methods irrespective of the predicted property, chemical fingerprint, similarity measure, or previously applied unbiasing techniques. Therefore, it may be the case that the previously reported performance of most ligand-based methods can be explained by overfitting to benchmarks rather than good prospective accuracy.

3D-e-Chem: Structural Cheminformatics Workflows for Computer-Aided Drug Discovery

eScience technologies are needed to process the information available in many heterogeneous types of protein-ligand interaction data and to capture these data into models that enable the design of efficacious and safe medicines. Here we present scientific KNIME tools and workflows that enable the integration of chemical, pharmacological, and structural information for: i) structure-based bioactivity data mapping, ii) structure-based identification of scaffold replacement strategies for ligand design, iii) ligand-based target prediction, iv) protein sequence-based binding site identification and ligand repurposing, and v) structure-based pharmacophore comparison for ligand repurposing across protein families. The modular setup of the workflows and the use of well-established standards allows the re-use of these protocols and facilitates the design of customized computer-aided drug discovery workflows.