Big Data and Artificial Intelligence Modeling for Drug Discovery

Abstract

Due to the massive data sets available for drug candidates, modern drug discovery has advanced to the big data era. Central to this shift is the development of artificial intelligence approaches to implementing innovative modeling based on the dynamic, heterogeneous, and large nature of drug data sets. As a result, recently developed artificial intelligence approaches such as deep learning and relevant modeling studies provide new solutions to efficacy and safety evaluations of drug candidates based on big data modeling and analysis. The resulting models provided deep insights into the continuum from chemical structure to in vitro, in vivo, and clinical outcomes. The relevant novel data mining, curation, and management techniques provided critical support to recent modeling studies. In summary, the new advancement of artificial intelligence in the big data era has paved the road to future rational drug development and optimization, which will have a significant impact on drug discovery procedures and, eventually, public health.

INTRODUCTION

Drug research and development is a complex, expensive, time-consuming procedure and has a high attrition rate (1). Drug attritions that happen in clinical studies induce great resource loss, and currently, nine out of ten drug candidates fail between phase I clinical trials and regulatory approval (2). Compared to traditional animal models, both in vitro and in silico approaches have great potential to lower the cost of drug discovery. The application of in vitro and in silico protocols in the early stages of the drug research and development procedure can reduce the number of drug attritions by identifying drug candidates with suitable therapeutic activities and excluding unsuitable compounds with undesirable side effects (3–6). However, the results of in vitro and in silico testing normally have low correlations to drug activities in vivo, especially for efficacy and complex side effects (7, 8).

Artificial intelligence (AI), which is sometimes presented as machine intelligence, refers to the ability of computers to learn from existing data. Computational modeling based on AI is a promising method to evaluate compounds for their potential biological activities and toxicities. Existing computational models, such as those based on quantitative structure-activity relationship (QSAR) approaches (9), can be used to quickly predict large numbers of new compounds for various biological end points. The existing models (e.g., those available in commercial drug discovery software) can make predictions of simple physicochemical properties (e.g., logP and solubility) and thus are relatively precise in predicting the pharmacokinetic properties of new compounds with simple mechanisms; however, the models for complex biological properties (e.g., drug efficacy and side effects) are far from optimal (8, 10) (Figure 1). Critical issues existed in previous QSAR modeling studies such as the use of small training sets (11), experimental data errors in training sets (12, 13), and a lack of experimental validations (14). The resulting QSAR model predictions of new compounds were questionable due to their coverage of a limited chemical space (15), existing activity cliffs (16), and overfitting (17, 18). The primary hypothesis of QSAR modeling (i.e., similar compounds will have similar activities) sometimes proved to be flawed (10, 19–21), indicating that training sets with only chemical structure information and target activity are not enough to answer the above challenges.

Figure 1.

Challenges of data-driven artificial intelligence modeling in modern, computer-aided drug discovery.

With the great progress of combinatorial chemistry since the 1990s, large chemical libraries have become the major source of new chemical development procedures (22, 23). Over the past ten years, this effort has also stimulated the development of high-throughput screening (HTS) techniques (24–26). HTS is a process that screens thousands to millions of compounds using a rapid and standardized protocol. Current HTS techniques are usually combined with robotic methods and require few resources to test a chemical library. Parallel HTS data processing and assay miniaturization have become increasingly popular in pharmaceutical industries and regulatory agencies as they greatly reduce the cost of experimental testing (27, 28). The chemical-response data obtained from HTS keep growing daily and contribute to the current big data environment. Facilitated by the combined efforts of HTS and combinatorial chemical synthesis, modern screening programs produce enormous amounts of biological data, especially regarding drug responses on specific targets (29, 30).

The challenges raised by big data are known as the “four Vs”: volume (scale of data), velocity (growth of data), variety (diversity of sources), and veracity (uncertainty of data) (31, 32). The data sets available for drug development, especially in pharmaceutical industries, may involve many compounds (e.g., from 100,000 to several million) that were tested against many targets (33), and traditional QSAR modeling and machine learning approaches are not always suited to dealing with these types of data under these conditions. Furthermore, the uncertainty of available data (or data sparsity) is one of the major obstacles of using big data (32). Unfortunately, when coupled with more complex biological mechanisms such as drug responses, the sparsity and variety of the resulting data increased dramatically from in vitro to in vivo studies (Figure 1). This big data scenario necessitated the development of new computational approaches to deal with high-volume, multidimensional, and high-sparsity data sources to predict drug efficacy and side effects in animals and/or humans.

The challenges to using big data discussed above and the involvement of new types of data (e.g., images) have demanded the recent development of novel AI approaches to advance predictive modeling in modern drug discovery (34–36). The popular AI approaches in the current big data era are based on deep learning (3, 4, 37). One of the early efforts of applying deep learning in the drug discovery process in pharmaceutical industries was the 2012 QSAR machine learning challenge supported by Merck (38). In this challenge, deep learning models showed significantly better predictivity than traditional machine learning approaches for 15 absorption, distribution, metabolism, and excretion (ADME) and toxicity data sets for drug candidates developed at Merck. Since then, and with the development of neural network approaches [e.g., convolutional neural networks (CNNs)], deep learning has been widely applied to drug discovery approaches. Although still viewed as a black box algorithm (39, 40), the current progress of AI supported by deep learning has shown great promise in rational drug discovery in this era of big data. The big data challenges; relevant AI developments; and modeling for drugs and drug candidates, especially those studies using deep learning and other new techniques, are the primary focus of this review.

BIG DATA IN DRUG DISCOVERY

The term big data describes a collection of data sets that are so large and complex that they are too difficult to process with traditional data analysis tools (41). Big data is gaining increasing recognition in clinical studies and other research areas driven by biological data (42, 43). As one of the fields generating a massive amount of data, modern drug discovery has moved into the big data era. The need for novel computational techniques, including data mining/generation, curation, storage, and management, brings new challenges and opportunities to the research community.

Several data-sharing projects, in parallel with the developments of HTS techniques in various screening centers, were also initiated in the past ten years. For example, PubChem is a public repository for chemical structures and their biological properties (44–46). In ten years, the number of PubChem compounds increased from 25 million in 2008 (46) to 96 million in 2018 (47). During the same period, the number of bioassays that were deposited into PubChem increased from 1,197 in 2008 (46) to over a million in 2018 (47). The current statistics of PubChem indicate that the repository contains 97.3 million compounds and 1.1 million bioassays (https://pubchem.ncbi.nlm.nih.gov). The tremendous amount of PubChem bioassay data that are updated daily constitutes a publicly accessible big data resource for compounds, including most drugs and drug candidates, with a variety of target response information. Similar to PubChem, ChEMBL is a database containing binding, functional, ADME, and toxicity data for numerous compounds (48). Compared to PubChem, ChEMBL contains a large amount of manually curated data from the literature. Currently, the ChEMBL database consists of over 2.2 million compounds tested against over 12,000 targets, resulting in activity data for 15 million compound-target pairs (https://www.ebi.ac.uk/chembl/).

Several other data sources are specifically designed for drugs and drug candidates. For example, DrugBank (https://www.drugbank.ca) is a publicly available database containing all approved drugs with their mechanisms, interactions, and relevant targets (49). The latest release of DrugBank (version 5.1.2, released December 20, 2018) contains 12,110 drug entries, including 2,553 approved small-molecule drugs, 1,280 approved biotech (protein/peptide) drugs, 130 nutraceuticals, and over 5,842 experimental drugs. DrugMatrix (https://ntp.niehs.nih.gov/results/drugmatrix/index.html), on the other hand, focuses on the toxicogenomic data of drugs to reduce the time to formulate a xenobiotic’s potential for toxicity. The current DrugMatrix database contains large-scale gene expression data from tissues of rats administered over 600 drugs, mostly targeting several major organs (e.g., liver). The Binding Database (BindingDB) is a public, web-accessible resource of drug-target binding data, shown as measured binding affinities (50). The targets included in BindingDB are proteins/enzymes that are considered drug targets. BindingDB currently contains 1,587,753 binding data for 7,235 protein targets and 710,301 small molecules (https://www.bindingdb.org/bind/index.jsp).

The public big data sources can also be characterized by the size of electronic files for these data sets. For example, the current PubChem bioassay database has around 240 million bioactivities, which are contained in 30 GB of XML files. Instead of using personal computers with central processing units, the use of new hardware techniques such as cloud computation (41, 51) and graphics processing units (GPUs) (52) is necessary to process and analyze these available big data.

BIG DATA MODELING CHALLENGES: MISSING DATA AND BIASED DATA

The response profiles of 2,118 approved drugs tested against 531 PubChem assays (each assay having at least 25 active responses among these drug molecules) are shown in Figure 2. The results were generated using an in-house automatic data profiling tool (http://ciipro.rutgers.edu/) (53). There are more than a million data points in this response profile. Nevertheless, many responses in this profile were shown as missing data (Figure 2). Furthermore, the ratio of active versus inactive responses is also biased (approximately 1:6 in this profile). For example, two well-known drugs were included in this profile: acetaminophen (CAS 103-90-2), which has 16 active and 213 inactive responses, and acetylsalicylic acid (aspirin, CAS 50-78-2), which has 14 active and 237 inactive responses. Due to the nature of the HTS techniques, the HTS data normally consist of much fewer active than inactive responses (21, 54), especially for the drugs. In an early review of pharmacological space based on 4.8 million unique compounds, only 275,000 of them showed one (or more) active response when tested against 1,036 targets (55), indicating that most of the testing results were negative. Notably, the drugs that showed the most active responses in public big data sets are for chemotherapy purposes, which normally have critical side effects and other off-target interactions. For example, bortezomib (CAS 179324-69-7) is a chemotherapy drug used to treat multiple myeloma and mantle cell lymphoma. It has the most active responses (258 actives and 49 inactives) in the response profile of Figure 2.

Figure 2.

Bioprofile of 2,118 approved drugs from DrugBank (x axis) represented by the response data obtained from 531 PubChem assays(y axis). Each assay against all drug molecules (one column) has at least 25 active responses (red spots). Data from DrugBank (https://www.drugbank.ca) and PubChem (https://pubchem.ncbi.nlm.nih.gov).

The missing data issue is a common problem of big data modeling (56). In previous studies, a common solution was to develop QSAR models for individual assays and use the resulting models to predict target compounds that were not tested against these assays (19, 20, 57). This approach was applicable only when the predicted data used for model development had simple biological mechanisms (e.g., logPs or structural rigid target bindings). However, this process still introduced uncertainty into the modeling process due to the prediction errors from QSAR models (57). When dealing with heterogeneous and complex data (e.g., clinical data), advanced statistical methods such as multiple imputations are needed (58, 59). To reflect the biased nature of HTS data, emphasis should be given to active rather than inactive results during modeling procedures (53). Early-stage computational studies normally used pharmacophore modeling to identify chemical features that were responsible for relevant bioactivities (60–62). The later modeling projects using machine learning approaches needed the biased training sets to be preprocessed by using various methods such as downsampling to balance active and inactive results (63–65).

ADVANCING ARTIFICIAL INTELLIGENCE FROM MACHINE LEARNING TO DEEP LEARNING

The historical progress of AI coupled with the increase of the data size used for model development and hardware improvement in drug discovery is summarized in Figure 3. The concept of AI was born in the 1950s (66) and was used in drug discovery after the first study of QSAR was presented in the 1960s (67). In the early stage of drug discovery (e.g., before the 1990s), the common computational approaches used for model developments were linear regressions (68). In these early studies, the chemical descriptors used for modeling were also limited to chemical structural features, such as atomic type and fragmental descriptors (69, 70). The advancement of AI in drug discovery was first facilitated by the development of novel chemical descriptors such as topological descriptors (71) and molecular fingerprints (72, 73), which greatly increased the size/categories of descriptors calculated from training sets. Instead of using all available descriptors, descriptor selection was integrated into the modeling procedure, e.g., the genetic algorithm (74, 75) and simulated annealing (76). Instead of using linear regression, new machine learning approaches, which were developed based on nonlinear modeling algorithms such as k-nearest neighbors (77), support vector machines (78), and random forest (79, 80), were used frequently in modeling studies from the 1990s to the 2000s. In the same period, model validation was emphasized and treated as a must-have component of modeling (81). Instead of only showing self-correlations, the developed models using these new machine learning approaches were always validated using cross-validations, external validations, and/or experimental validations (14, 63, 82, 83). In addition, the applicability domain became standard practice for model development (17, 84–86). In the early 2000s, QSAR modeling, together with relevant studies (e.g., docking), became a well-developed workflow based on the progress of AI discussed above (Figure 3). These milestones of AI in drug discovery are emphasized in other reviews (9, 87–91).

Figure 3.

The historical progress of artificial intelligence in drug discovery coupled with increasing data size and computer power (shown as processor improvement).

In addition to the development of AI, the computational power of hardware and the available data for modeling were also significantly improved to facilitate this progress (Figure 3). The early-stage computational modeling of small training sets by simple algorithms (e.g., linear regressions) did not require significant computational power. The advancement of computational power and the availability of biological data for drugs enabled the application of novel modeling techniques such as large-scale networks to address challenges in drug discovery. The first application of the neural network, which was designed as a computational tool in the 1980s (92), in drug discovery was reported in 1989 (93). Since then, various neural network approaches have been applied to drug discovery (90, 94). The first popular approach was the artificial neural network (ANN) (95, 96), which focuses on the variable selection procedure (97). This approach is a machine learning algorithm inspired by biological neural networks such as those in the human brain. With several variables as the input (e.g., chemical descriptors), ANN approaches form hundreds of artificial neurons, which are connected with relationships (quantified as weights) in the form of a network. A single neuron might have some effectiveness in predicting output, but the actual predictions are made by the network consisting of hundreds or even thousands of neurons. Since they learn from the input data, ANNs represent an excellent machine learning approach for constructing nonlinear relationships among the variables and the target biological activities (98). The advanced computational models using various machine learning approaches, such as ANNs, required powerful computers and benefited directly from the hardware developments in the 1990s (Figure 3).

The concept of deep learning was originally presented together with ANNs in the 1980s (4). However, neural networks did not show significant advantages over other machine learning approaches when data used for model development are limited (99, 100). From the 1990s to 2000s, computer hardware was still not adequate for training neural networks with many hidden layers and/or when the data sets for model development were large. In the 2010s, hardware development reached the milestone of using GPUs and cloud computing, which directly benefited neural network modeling studies (Figure 3). Advanced as one of the major interests of AI by various information technology companies, deep neural networks (DNNs), sometimes referred to as deep neural nets, with many hidden layers were developed to address challenging questions such as speech recognition (101). In the Google DeepMind project of 2015, an AI program based on a DNN with 13 hidden layers first mastered the game of Go, which has long been viewed as the most challenging of the classic games for AI (102). The milestone paper of deep learning was published at almost the same time (103), and the big data concept was proposed the next year (41, 104). Deep learning was immediately applied to the life sciences and demonstrated its capability to identify complex patterns in biological systems (4, 105). The first project in which deep learning approaches showed significantly better performance than other machine learning approaches for drug discovery was a QSAR machine learning challenge supported by Merck (38). Another similar effort organized by the National Center for Advancing Translational Sciences of the National Institutes of Health (NIH) was to model around 12,000 chemicals, including many drugs, for 12 different toxic effects (106). In this competition, DeepTox, a computational toxicity model based on DNNs outperformed other models based on machine learning approaches (107).

Besides the modeling challenges mentioned above, there have been various individual deep learning studies for drug discovery in the past three years. For example, Wen et al. (108) reported a deep learning model developed to predict interactions between drugs and their biological targets based on 15,524 drug-target pairs obtained from the DrugBank database. Another similar deep learning study was performed using transcriptome data obtained from the Library of Integrated Network-Based Cellular Signatures program (109). Furthermore, multitask learning based on DNNs is a modeling approach that allows multiple related tasks to be modeled simultaneously. Modeling several biologically related end points (i.e., bioactivities sharing similar mechanisms) for drug discovery purposes through multitask learning has shown superior performance to traditional QSAR models by reducing overfitting, solving issues of biased data, and identifying variables from related tasks (110–113). The high performance of these DNN models demonstrates the advantages of using deep learning approaches to model large data sets and select meaningful features. However, there were also recent reports that showed mixed results from the comparison between deep learning and machine learning modeling (114, 115). Since deep learning is a brand-new concept being applied to computer-aided drug discovery, there are no universal criteria for selecting relevant modeling parameters and/or constructing the modeling workflow (115).

OTHER AREAS OF COMPUTATIONAL MODELING UTILIZING ARTIFICIAL INTELLIGENCE FOR DRUG DISCOVERY

Rational Nanomaterials Design

Modern nanotechnology highly impacts drug discovery by offering biocompatible nanomaterials (e.g., nanomedicines with desirable therapeutic activities and low side effects) to the drug research and development process, especially as versatile yet reliable carriers for the delivery of drugs to treat systemic diseases such as cancers (116, 117). Early efforts of using AI in nanomodeling for drug discovery were based on molecular dynamic (MD) simulations. For example, several studies using MD simulations detected the insertion of nanoparticles in the plasma membranes of the recipient cells and an overall change in the cell membrane structure (118). Later, the same approach was used to estimate the affinity of carbon nanotubes to organic molecules (119). In another study, a set of nanoparticles was tested in vitro in four cell lines, and the potential membrane perturbation effects of these nanoparticles were studied (120). The reaction behaviors of individual nanoparticles were also investigated under certain conditions using MD simulations (e.g., interactions with or passing through membranes), along with the effects of the size, density, position, distribution, length, and type of surface ligands on the biological properties of the nanomaterials (121). The advantage of MD simulations is that they can precisely simulate molecular structures, but the clear disadvantages are that modeling procedures are computationally expensive and cannot provide rapid predictions for big databases due to the current limitations of computational resources. Another computational approach is to apply traditional QSAR modeling methods to nanomaterials. For example, the QSAR technique was used to create predictive models for nanoparticles with similar or different metal cores (122). Recently, membrane-nanoparticle interactions were modeled based on the atomization energy of the metal oxide, the period of the nanoparticle metal, and the primary size of the nanoparticle (123).

The current application of AI approaches in nanomodeling has been limited to designing new nanomaterials due to a lack of suitable chemical descriptors. Although descriptors calculated from only the surface ligands are useful in predicting specific bioactivities/properties of nanomaterials, as described above, the effects of the nanomaterial’s size/shape, density, position, distribution, length, and type of surface ligands were not considered in these studies. Some other nanomodeling studies have incorporated descriptors derived from experimental properties (e.g., nanoparticle size) (123, 124) or even biological data (e.g., proteomics data) (125, 126). Due to the diversity and complexity of nanomaterial structures, Puzyn et al. (127) argued that no universal nano-QSAR model can accurately predict the biological properties of variable nanomaterials. Figure 4 presents a recent methodology for nanostructure simulations in the modeling procedure (128). Briefly, the properties and bioactivities of nanomaterials were largely determined by their surface chemistry. To simulate the nano surface chemistry correctly, the surface ligand orientations and accessibility of functional groups needed to be considered in the calculations (Figure 4). For example, in an early modeling of nanohydrophobicity, the contributions of heavy atoms and functional groups to nanologP values were correlated with their accessibility by solvent molecules (128). In a recent study, an advanced method of integrating the solvent-accessible surface into calculations can be viewed as a universal nanologP calculator (129). A similar modeling strategy has been applied to model nanocellular uptake capacities (130) and several other nano bioactivities. The resulting models were utilized to design and synthesize several new nanoparticles with desired nano bioactivities (130).

Figure 4.

Nanomaterial surface simulations for computational modeling: surface ligand orientations and accessibility assessments.

Convolutional Neural Networks and Image Modeling

The CNN is a special network modeling approach inspired by neuroscience to imitate images within the visual cortex, where individual neurons respond to stimuli only in the receptive fields. Different neurons can partially overlap with each other to cover the entire receptive field. The CNN architecture is constructed in a way that hidden layers are particularly adept at screening multidimensional input such as the red, green, and blue saturation values obtained from thousands of pixels for an image. In the training process, the CNN approach uses kernels and grids of a predefined dimension to scan the image and learn to recognize certain critical features such as lines and contours for a human face. The concept of CNNs was proposed in the 1980s for image recognition purposes but did not draw great attention until the 2010s (4). This approach has become well known, as it has dominated all image recognition challenges since 2012, and it is now the base of image/speech recognition, video analysis, language understanding, and other relevant applications (131).

As one of the most popular deep learning approaches, CNNs have been used for image modeling in clinical diagnoses such as cancer (132), Alzheimer’s disease (133), and heart disease (134). In traditional drug discovery, CNNs were also applied to analyze image data obtained from experimental drug testing, such as HTS results (135). Due to its unique advantages in image recognition, CNNs were also used to recognize 3-D experimental and virtual images to predict ligand-protein interactions (136, 137). In some studies, CNNs were coupled with other computational approaches to realize specific goals. For example, CNNs were used as a new approach to recognize molecular features from drug molecular graphs (138). In this study, drug molecules were treated as 2-D graphs with atom features. The CNN was used to transform the input molecular graphs into new molecular features for training purposes. In another study, an advanced CNN approach called the survival convolutional neural network was used to predict the cancer outcomes of patients based on histological images and genomic biomarker data (139). Furthermore, CNNs were able to function as a text-mining technique to extract drug-drug interaction data from biomedical literature (140).

Personalized Medicine

A drug commonly interacts with multiple targets, including both on- and off-targets, and drug efficacy and side effects are greatly affected by this (141). The perturbation of an individual biological system (e.g., a patient) by a drug molecule is determined by various genetic, epigenetic, and environmental factors. To identify this hidden hierarchical information, personalized medicine was designed to respond to the individual characteristics of each patient (142). Personalized medicine strongly relies on a scientific understanding of how an individual patient’s unique characteristics, such as molecular and genetic profiles, make this patient vulnerable to a disease and sensitive to a therapeutic treatment. Driven by biomarker studies starting in the late 1990s, hundreds of genes have been identified for their contributions to human illness, and genetic variability in patients has been used to distinguish individual responses to dozens of treatments (142). Along with the huge amount of data generated by these studies, such as the Human Genome Project (143), computational modeling has become one of the most important tools for personalized medicine. Drug-target predictions (144), metabolic network modeling (145), and population genetics pattern identifications (146) are several recent advancements in this field that rely on computational modeling. Under the NIH Precision Medicine Initiative (147), many data generation and sharing initiatives and computational modeling efforts have arisen to support the expansion of precision medicine. For example, the Genomic Data Commons program of the National Cancer Institute aims to provide a data repository that enables data sharing across cancer genomic studies in support of precision medicine (148). So far, 33,549 case studies have been submitted and shared via this portal (https://gdc.cancer.gov/). Although it is not the focus of this review, genome sequencing analysis has been a widely applied approach involving AI techniques, and there are many reviews available on this popular bioinformatics topic (149–151).

CONCLUSIONS

AI is a promising method to greatly reduce the cost and time of drug discovery by providing evaluations of drug molecules in the early stages of development. In the current big data era, clinical and pharmaceutical data continue to grow at a rapid pace, and novel AI techniques to deal with big data sets are in high demand. The recent deep learning modeling studies have shown advantages compared to traditional machine learning approaches for this challenge. However, standard criteria and modeling workflows are still needed for deep learning models to be applicable. The applications of AI have been widely extended into all relevant areas beyond traditional drug discovery. Coupled with database curation, web portal development as data repository servers, and the improvement of computer hardware, AI and recent deep learning studies have paved the road to modern drug discovery.