Recent progress on cheminformatics approaches to epigenetic drug discovery

Abstract

The ability of epigenetic markers to affect genome function has enabled transformative changes in drug discovery, especially in cancer and other emerging therapeutic areas. Concordant with the introduction of the term ‘epi-informatics’, the size of the epigenetically relevant chemical space has grown substantially and so did the number of applications of cheminformatic methods to epigenetics. Recent progress in epi-informatics has improved our understanding of the structure–epigenetic activity relationships and boosted the development of models predicting novel epigenetic agents. Herein, we review the advances in computational approaches to drug discovery of small molecules with epigenetic modulation profiles, summarize the current chemogenomics data available for epigenetics targets, and provide a perspective on the greater utility of biomedical knowledge mining as a means to advance the epigenetic drug discovery.

Introduction

The term ‘epigenetics’ has its historical roots in the work of C. Waddington (1940s) and D.L. Nanney (1950s), where it was initially defined to denote a cellular memory, persistent homeostasis in the absence of an original perturbation, or an effect on cell fate not attributable to changes in DNA [1,2]. However, the term is now used with multiple meanings. For example, epigenetics has been used to describe the (i) heritable phenotype (cellular memory) without modification of DNA sequences [3]; (ii) ‘the structural adaptation of chromosomal regions to register, signal, or perpetuate altered activity states’ in the genome [4]; and (iii) the mechanism by which the environment conveys its influence to the cell, tissue, or organism [5]. Despite these ambiguous definitions, the growing importance of epigenetics in new therapeutics discovery and understanding disease etiology can be illustrated by the trends in the number of scientific publications on the subject. Figure 1a shows the number of papers published in the literature with the word ‘epigenetics’ and six co-occurring diseases per year; now, 25 years after the first six published reports on ‘epigenetics’ in 1994 [6], there are cumulatively >105 000 relevant publications. Although the most frequent co-occurring term is ‘cancer’ and most epidrugs are approved for cancer treatment, epigenetic modifications are not limited to malignant conditions. Instead, the data suggest that exceptionally deadly diseases were the first to be studied in the epigenetic field [6].

Figure 1.

Title. (a) The number of papers in PubMed with the keyword ‘Epigenetics’ alone or co-occurring with multiple diseases (cancer, diabetes, neurodegenerative, cardiovascular, central nervous system, and rare disease) by year. (b) The number of papers in PubMed with the keywords ‘epigenetics drug discovery’ and ‘epigenetics + computer-aided drug discovery / computational drug discovery’ by year.

As one can see from Figure 1, epigenetics has implications for many different areas of research, as indicated by many extensive reviews on different topics, such as the waves of epigenetic drugs [4], the role of epigenetics in cancer [7], autism diagnosis [8], multiple sclerosis [9], depressive disorders [10], and rare diseases [11], to name a few. Although the general trends in drug discovery show that kinases and G-protein-coupled receptors to be the most investigated protein families [12], epigenetic research has predominantly focused on different classes of target, primarily histone deacetylases (HDACs), histone methyltransferases (HMTs), DNA methyltransferase (DNMT) 1, and other chromodomain and/or Tudor domain-containing proteins (TUD). Interest in the bromodomain and external terminal protein (BET) family has shed light on 42 bromodomain (BRD) targets [13], and further investigation of histone acetyltransferases (HATs) led to a distinctive class of lysine acetyltransferases (KATs). Since the report of the first adenine-specific DNMT in Escherichia coli [14], the first discovered epigenetic target, 515 human epigenetic targets (136 mappable onto epigenetic phylogenetic groups) have been described, along with respective chemogenomics data (Table S1 in the supplemental information online). Despite the immense growth of the field, several epigenetic proteins associated with important diseases have been largely understudied, such as protein arginine methyltransferases (PRMTs), histone methyl readers (HMRs), and DNMT3; limited funding allocated to the studies of such ‘dark’ targets could be a contributing factor [15]. Many compounds have been tested in histone demethylase (HDM) assays, some of them reaching preclinical and clinical stages, but none of them have been approved by the US Food and Drug Administration (FDA) as yet. Considering their potential impact on the field, more work needs to be done on dark targets to better understand their mechanisms and expand the number of available targets.

Epigenetics has a major role in the understanding of inheritance, development, and progression of diseases. Consequently, the discovery and development of epidrugs (or epigenetic drugs) have become a major research focus for multiple biotechnology companies [4]. As of August 2020, there were eight epigenetic drugs approved for clinical use [4,16]: two DNMT inhibitors (azacytidine and decitabine), five HDAC inhibitors (vorinostat, belinostat, panobinostat, romidepsin, and tucidinostat), and one HMT inhibitor (tazemetostat). Figure 2a and 2b show the chemical structures of representative FDA-approved drugs and compounds in clinical development as epidrugs, respectively [4]. An ongoing clinical trial is examining the FDA-approved DNMT inhibitor, azacitidine, in combination with a novel BET inhibitor, INCB057643, and a novel lysine dependent demethylase (LSD1) inhibitor, INCB059872, for solid metastatic tumors [17]. An extensive description of ongoing clinical trials related to epigenetic drugs has been reviewed recently [4]. Currently, most drugs focus on DNMT1 and HDACs (Figure 2a). The increased interest in epigenetic applications outside of cancer has not yet translated to a rise in viable drugs or drug candidates, despite the expansive therapeutic benefits of epigenetic modulation.

Figure 2.

Title. (a) Chemical structures of epigenetic drugs approved for clinical use. (b) Chemical structures of selected compounds in clinical development as epidrugs. Epigenetic targets are grouped as writers, erasers, and readers. A comprehensive review of compounds in clinical development the reader was recently provided by Ganesan et al. [4]. For definitions of abbreviations, please see the main text.

In recent years, the substantial growth in epigenetics-related data has prompted the development of cheminformatics methods with application to this field. In 2015, the term ‘epi-informatics’ was introduced and conceptualized [6] to summarize advances in epigenetic drug and chemical probe discovery driven by computational methods. Since then, molecular modeling and cheminformatics approaches have made substantial contributions to the field (Figure 1b) [18–20]. Computational methods have helped to explore the mechanism of action of active compounds at the molecular level and guide lead optimization programs. Herein, we review recent advances in computational approaches to epigenetic drug discovery and summarize, to the best of our knowledge, all the publicly available chemogenomics data for epigenetics targets.

Chemogenomics data and databases

Over the past decade, several open-source databases have compiled targets and compounds with epigenetic profiles. However, most do not use the same criteria for target selection/classification or use a systematic drug target ontology [21]. Table 1 summarizes the chemogenomics databases related to epigenetics published thus far, including web links to each resource. Although all of them were accessible as of August 2020, none have been updated since their release. This fact highlights the need for tools that automate the analysis of the constantly increasing body of epigenetic data in the public domain.

Table 1.

Major public databases relevant to epigenetic drug discovery

Database	Last update	Number of targets	Number of compounds	URLa	Refs
Human epigenetic enzyme and modulator database (HEMD)	2012	269	4377	http://mdl.shsmu.edu.cn/HEMD/	[22]
EpiDBase	2015	Not available	5784	www.epidbase.org/	[23]
EpiFactors	2015	815	–	http://epifactors.autosome.ru	[24]
Database of Epigenetic Modifiers (dbEM)	2016	167	–	http://crdd.osdd.net/raghava/dbem	[25]
Epigenomics chemical database	2018	54	7820	www.difacquim.com/d-databases/	[18]

^aLast accessed: 25 August 2020.

One of the first web-accessible databases was the Human Epigenetic Enzyme and Modulator Database (HEMD). This resource, published in 2012, includes 4377 small-molecule modulators and 269 epigenetic targets that are annotated with information on epigenetic mechanisms, catalytic processes, and related diseases [22]. EpiDBase is also a web-accessible database that was released in 2015 [23]. It contains 5784 different ligands annotated with experimental activity (IC₅₀) against writers, erasers, and readers, as well as calculated properties.

EpiFactors [24], also published in 2015, is a manually curated database that provides information on epigenetic regulators, their complexes, targets, and products. The database contains 815 human epigenetic proteins and 69 protein complexes involved in epigenetic regulation. EpiFactors also provides the corresponding genes and their expression levels in 458 human primary cell samples, 255 different cancer cell lines, and 134 human post-mortem tissues.

The Database of Epigenetic Modifiers (dbEM) is a web-accessible database released in 2016 [25] that contains the genomic information on 167 epigenetic targets. This resource maintains the information of mutations, copy number variation, and gene expression in tumor samples, cancer cell lines, and healthy samples.

In 2018, a database of epigenetic small molecules inhibitors was released [18] integrated from other public access databases such as ChEMBL (www.ebi.ac.uk/chembl/) and PubChem (https://pubchem.ncbi.nlm.nih.gov/). The epigenomics database includes 7820 unique compounds, of which 3456 have information for more than one epigenetic target. The database has 16 102 compound–target associations, of which 15 887 have quantitative potency data associated with them. The database has associations with 60 epigenetic targets.

Current approaches, models, and best achievements

Epigenetic-relevant chemical space

According to the Chemical Space project, the total number of synthetically feasible organic molecules exceeds 166 billion compounds [26]. Through chemical clustering and visualization, cheminformatics has allowed the navigation of large databases to identify therapeutically relevant chemical spaces. In this context, the ‘epigenetic-relevant chemical space’ (ERCS) was the first attempt to comprise a list of HDACs, DNMTs, and BET inhibitors (2772 compounds in total) [27]. The design of selective inhibitors remains a challenge in epigenetic drug discovery. For instance, most HDAC inhibitors approved so far are non-isoform selective, even though it has been hypothesized that selective inhibitors might have fewer adverse effects [3,28]. A comprehensive characterization of the ERCS can uncover multitarget relationships as well as guide the design of selective inhibitors, which could significantly improve the effectiveness of drug therapies [29].

We surveyed the status of the compounds tested against the epigenetic targets and currently available in the public domain. Figure 3 visualizes the current ERCS using a Tree Manifold Approximation and Projection (TMAP) [30], generated with Statistical-Based Database Fingerprints (SB-DFPs) to represent target-associated compound datasets [31]. The SB-DFPs are novel condensed representations of compound data sets that have also been shown to capture these relationships between the compound data sets. This representation is based on statistical comparisons of molecular fingerprint bit proportions among any two data sets. These condensed representations are connected by a minimum spanning tree, as described in detail in the original publication [30]. The TMAP displays the relationships among compounds with biological data for a particular target as nodes that are connected to similar data sets of other targets through branches and subbranches. This map could also help to guide the design of multitarget epigenetic agents. In the map, several targets that are in the same or neighboring branches share the same role in histone modifications, such as kinases (KINs; writers), the BET family (readers), HDACs and HDMs (erasers), as well as chromatin remodelers (CHR). This observation is consistent with the known difficulty of designing a selective inhibitor within each class [4]. The map also shows a close relation between targets related to DNA methylation (i.e., DNMTs) to those involved in RNA modification (mRNAs) as well as different histone writers (HMTs, PRMTs, and HATs). Targets associated with other histone modifications tend to appear in the terminal regions of these branches, illustrating a weak relationship between different target classes. By contrast, Polycomb-group proteins (PcG) and other targets, such as the chromatin remodeler PARG2 and methyltransferases PRMT3 and PRMT5, do not appear to be closely related to targets with the same function, suggesting them as promising targets for the development of selective compounds.

Figure 3.

Visual representation of the current epigenetics-relevant chemical space. The visualization was generated with the recent Tree Manifold Approximation and Projection method and Statistical-Based Database Fingerprints as condensed representations of compounds associated to the epigenetic targets [31]. Individual nodes of the tree are epigenetic targets, colored by their class. The nodes are clustered according to the pairwise similarity of their condensed representations. The branches represent the connection between the chemical space of two different targets. Proximal data sets can contribute to the discovery of compounds acting on multiple targets. For definitions of abbreviations, please see the main text.

Structure-epigenetic activity relationship analyses

The increasing number of reports on compounds with experimental activity against one or more epigenetic targets has allowed the exploration of the structure–activity landscape for several epigenetic target datasets. Here, we review the respective models developed and published in recent years.

Quantitative structure-activity relationships (QSAR) modeling

QSAR modeling is a major computational approach to drug discovery [32] and, with the growth of chemogenomics datasets for epigenetic targets, its application to find bioactive compounds with epigenetics profile has grown. A recent review discussed 3D-QSAR studies performed with several different sets of HDAC inhibitors [33]. More recently, models were reported for 11 sets of inhibitors of HDAC using conformal prediction [34], showing that these models demonstrated high accuracy for both training and external test sets. The efficiencies for the predictions were >80% for most data sets and >90% for four data sets at different significance levels. Another paper expanded the discussion on HDAC-1 inhibitors, specifically, aminophenyl benzamide derivatives, and was able to produce an externally validated model with a high correlation (R²) and predictive (r²_{ext_ts}) values of 0.96 and 0.79, respectively [35]. García-Sánchez et al. [36] reviewed more recent examples of QSAR models developed, again, mostly for HDACs and DNMTs. In an interesting effort to obtain novel DNMT1 inhibitors, the authors modified the structures of known inhibitors and used QSAR models to predict the activity of the novel compounds. They also identified electronegativity and the bond information content index as the most influential descriptors [37]. Unfortunately, other epigenetic targets have been minimally explored. BET inhibitors have been studied using QSAR modeling approaches; one study used such models to predict six putative multitarget BET bromodomain inhibitors [38]. Six reversible LSD1 inhibitors were proposed based on 3D-QSAR, molecular docking, and dynamics studies, but these predictions have not been supported by the experimental binding affinity measurements [39]. Similar approaches have been applied to lysine methyltransferase DOTL1 inhibitors, resulting in the computer-assisted design of two compounds that demonstrated inhibition at micromolar levels in confirmatory assays [40].

Activity landscape modeling

Activity landscapes can be defined as representations that compare compound similarity and activity relationships [41]. One of the main goals of this approach is to identify activity cliffs that can help lead optimization efforts [42] as well as data sets with a smooth (continuous) SAR that would be more likely to yield predictive models. Similarly, one can identify epigenetic datasets with a ‘rough’ (i.e., discontinuous) SAR landscape, for which the development of predictive models is expected to be difficult. Based on this concept, the SEAR of various epigenetic target data sets has been analyzed, including DNMTs [43,44], BRDs [36], HDACS [45], and lysine methyltransferases (G9a or EHMT2) [46]. To characterize the epigenetic activity landscape, a SAR analysis of 52 compounds tested against different epigenetic targets was used [47], which showed DNMT3B and DOT1L to have the highest percentage of activity cliffs (i.e., the respective data sets had more discontinuous SAR). The study also revealed that HDACs were the targets with the most continuous SAR, making them most suitable to carry out hit-to-lead optimization. This result is also consistent with the generation of a large number of predictive QSAR models. Finally, this large-scale epigenetic activity landscape study highlighted SMARCA2 and HAT as the epigenetic targets more prone to scaffold hopping (i.e., the search for compounds with different scaffolds but similar activity).

Epigenetic Target Profiler

Recently, a free online service was developed to predict the potential activity of small molecules against epigenetic targets. Briefly, the Epigenetic Target Profiler is a user-friendly web application that uses binary classification models relying on machine-learning algorithms (support vector machines and artificial neural networks) to predict the most probable epigenetic targets for a small molecule. Classification models were built on the available bioactivity data for 35 epigenetic regulators from ChEMBL26; the development of a new version using the latest release of ChEMBL is in progress. Figure S1 in the supplemental information online shows the graphical user interface of the webserver.

Reshaping the discovery and development of epidrug candidates with current and emerging technologies in cheminformatics

Molecular modeling and cheminformatics have made notable contributions to drug discovery [48]. Several important drugs have been developed with computational methods, including imatinib (a kinase inhibitor used to treat certain types of cancer), dorzolamide (a carboanydrase II inhibitor used to treat high pressure inside the eye, including glaucoma), enfuvirtide (a first-in-class antiretroviral drug used in combination therapy for the treatment of HIV-1), oseltamivir (a neuraminidase inhibitor used to treat flu), and many others [49,50]. The increasing chemogenomic body of data in the field of epigenetics has facilitated the application of several well-established computational methodologies. One example is the development of the Epigenetic Target Profiler discussed in the previous section. However, the application of computational approaches in epigenetic drug discovery is still recent and, so far, no epi-drugs have been discovered with computational strategies. Therefore, the development and application of new epi-informatics methodologies should continue to be an actively growing area of research.

Over the past few years, several groups have used multiple computational approaches to epigenetics research [18–20]. A recent review [51] summarized the advantages and limitations of several structure-based approaches in multitarget drug design, including examples from epigenetics. For example, Kuang et al. [52] reported on a comprehensive modeling study of several BRD isoforms with molecular dynamics simulations; the authors exploited the non-negligible ligand-binding kinetics features of these proteins, enhancing the understanding of the binding site of the BRD family. Although these techniques are promising, some limitations associated with molecular simulations [53] need to be overcome to enable their broader application to epigenetic research. For instance, an improvement in parametrization protocols for both the simulation of drug-like molecules [54] and protein–protein interfaces [55] is necessary.

Computational approaches can be used to find hidden allosteric binding sites [56] and protein–protein interaction hotspots [57] for epigenetic targets. Computational approaches can also help improve both the pharmacodynamics and pharmacokinetics of compounds in the hit-to-lead stage. Recently, Letfus et al. [58] used quantum/molecular mechanics (QM/MM) approaches to improve toxoflavin-based inhibitors of the human histone lysine demethylase enzymes of the subfamily KDM4C (a molecular target related to prostate and breast cancer). These studies led to compounds with activity in the low nM range in biochemical assays and mM activity in cell-based assays that also had enhanced pharmacokinetics properties in vitro. QM/MM methods were used because conventional force fields used in docking programs cannot model most metalloenzymes and metal–ligand interactions properly.

Epigenetic drug discovery can mostly benefit from cheminformatics to identify and prioritize hits for experimental validation. Recently, Tao et al. [59] reported the identification of a novel small molecule with in vitro enzymatic inhibition at a submicromolar level for PRMT5, an anticancer therapeutic target, using molecular docking and pharmacophore-based virtual screening. In another study, Song et al. [60] used a combination of pharmacophore searches, molecular docking, and molecular dynamic simulations and found compounds with selective activity against the proliferation of cancer cells.

We expect that QSAR modeling will continue to be an important methodology to accelerate the discovery of epidrugs. This approach has been substantially enriched over the past few years with modern machine learning and artificial intelligence algorithms [32]. Although some studies have reported the development of QSAR models and the identification of virtual hits for epigenetic targets [37,38], these hits were not experimentally tested. However, compounds have been identified computationally that were later validated experimentally [61–66], confirming how this approach could help identifying promising bioactive compounds for epigenetic targets. We have summarized available libraries that have been designed for epigenetics targeted using computational approaches, and a comprehensive list of these libraries is available in Table S2 in the supplemental information online. Meanwhile, structural diversity and scaffold content of such libraries, as well as their in silico pharmacokinetic and toxicological profiles, still need to be estimated [67].

Molecular docking and molecular dynamics simulations as well as artificial intelligence methods have gained attention because of increased computational capabilities and efficient scaling with the advent of graphics processing units. For instance, Lyu et al. [68] recently ran molecular docking on a 170-million compound library leading to the experimental discovery of a new scaffold of phenolate inhibitors of AmpC. After optimization, one of the compounds showed inhibitory activity of 77 nM. The authors also identified 30 compounds with submicromolar activity for the D₄ dopamine receptor. Gentile et al. [69] developed Deep Docking, a hybrid platform built on deep learning QSAR models trained on docking scores, which afforded rapid accurate prediction of docking scores for billions of molecular structures. Several recent applications of deep learning and generative neural network models have been described in the literature. For instance, a deep reinforcement learning algorithm, termed ReLEASE (Reinforcement Learning for Structural Evolution) was developed that can help to design chemical libraries with a bias toward desired inhibitory activity [70]. By using a similar approach, Zhavoronkov et al. [71] reported the discovery of a potent candidate for DDR1, a kinase target implicated in fibrosis and other diseases, noting that the entire project was accomplished in 21 days. Although this work received criticism because the new molecule was similar to an approved drug [72], the new compound was de facto not available in any open database. The application of these advanced and accelerated computational approaches to epigenetic targets is pending.

The potential of polypharmacological profiles allowed by epigenetic targets is an important consideration of direct relevance to epigenetic drug discovery. Very often, drugs hitting undesired targets are the primary source of toxicity, also known as ‘off-target’ effects [73]. Indeed, toxicity and lack of efficacy are the major causes of drug attrition [74]. However, recent studies highlighted the advantages of multitarget design (MTD), taking advantage of the ‘selective synergism’ [75]. Structure-based and machine-learning methods have been successfully applied in MTDs [76]. Epigenetic modulation is suitable for MTD because several targets are involved in the same pathway, which could lead to synergic results [18,77]. For the purposes of epigenetics drug discovery, promising MTD strategies should focus on designing compounds that inhibit different aspects (reading, writing, and erasing) of the same gene modification, such as (i) reader/writer (e.g., BET/HDAC inhibition) [78]; (ii) writer/writer of distinct changes (e.g., HDAC/DNMT) [79]; and (iii) writer/writer of the same modification (e.g., G9a/DNMT) [46,80]. Therefore, it is evident that epigenetic polypharmacology, or even the combination therapy of different epidrugs, offer a promising avenue for future epigenetic drug discovery [16,18].

Concluding remarks

Over the past 25 years, the number of scientific publications related to epigenetics has increased impressively from six papers in 1994 to >100 000 overall. Although cancer remains the primary therapeutic area associated with epigenetic drug discovery, other conditions, such as cardiovascular, neurodegenerative, central nervous system-related diseases, diabetes, and rare diseases, have been rapidly gaining interest in relation to epigenetics. In addition, we have seen a slight uptick in computational approaches used in this field. So far, epi-informatics has allowed the creation and maintenance of target-compound databases, exploring the increasing ERCS and SEARs, which eventually led to the development of the Epigenetic Target Profiler, a webserver to generate a predicted profile of potential inhibition of small molecules across a panel of epigenetic targets of pharmaceutical relevance.

Herein, we have highlighted recent advances in the use of cheminformatics in epigenetic drug discovery as well as discussed recent cutting-edge computer-aided drug design technologies that could be used to enable breakthrough discoveries in the field. Although no epigenetic drug has been discovered by computational approaches yet, the diversity of therapeutic areas associated with epigenetic targets, the growth of epigenetic chemical space, and the proliferation of robust computational approaches to drug discovery promise a substantial increase in the number of publications and the emergence of novel epigenetic drug candidates discovered by cheminformatics methods.

Highlight.

• The interest in developing epigenetic drugs is substantially increasing.

• We reviewed the recent epigenetics structure-activity relationships.

• Epi-informatics has seen recent advances in databases and web servers.

• Polypharmacology profiling is a must in the future of epigenetic drug discovery.

• We briefly discuss the future directions of epi-informatics.