In 2014, the outbreak of the Ebola virus (EBOV) in West Africa highlighted the need for broad-spectrum antiviral drugs for this and other emerging viruses ¹ . Several groups had previously performed high throughput screens (HTS) and identified FDA approved drugs (amodiaquine, chloroquine, clomiphene and toremifene) with in vitro growth inhibitory activities against EBOV ^{2, 3} . It appears none of these molecules were tried during the epidemic in Africa ⁴ , likely due to the lack of efficacy data in higher order species. We have previously summarized the numerous small molecules described in the literature as possessing antiviral activity that could be further evaluated for their potential EBOV activity alongside the few new antivirals. We have found that there is considerable prior knowledge regarding these small molecules possessing activity against EBOV in vitro or in animal models ^{5– 8} , and this includes a number of accessible FDA-approved drugs ^{2, 3, 9} . Another recent study has shown three approved ion channel blockers (amiodarone, dronedarone, and verapamil) inhibited EBOV cellular entry ⁹ . The drugs were given at concentrations that would be achieved in human serum, and were effective against several of filoviruses ⁹ . None of the FDA approved drugs described in these various studies were designed to target the Ebola virus. For example amodiaquine and chloroquine are well known antimalarials, clomiphene and toremifene are selective estrogen receptor modulators, while amiodarone, dronedarone, and verapamil are anti-arrhythmics ⁴ . It may or may not be of importance but all of these compounds have a common tertiary amine feature ^{10, 11} . What is important is that they are all orally bioavailable and generally safe for humans at their approved doses. Some have suggested that G-protein-coupled receptors (GPCRs) may play a role in filoviral entry and receptor antagonists could be developed as anti-EBOV therapies ¹² . The compounds which are FDA-approved drugs for other diseases ^{2, 3, 9} but with activity against EBOV in vitro or in vivo may represent useful starting points with the advantage that much is known regarding their absorption, distribution, metabolism and excretion (ADME) and toxicity properties. Thus, these repurposed drugs may represent a more advanced starting point for therapeutic development and approval compared with new chemical entities for preventing the spread and mortality associated with EBOV.

Beyond these early stage drugs, there are a number of other compounds that have also been identified as active against EBOV (summarized in a review ¹³ ). A thorough literature search identified 55 molecules suggested to have activity against EBOV in vitro and/or in vivo which were evaluated from the perspective of an experienced medicinal chemist as well as using simple molecular properties and ultimately 16 were highlighted as desirable ¹⁴ . This dataset overlaps to some extent with another review that identified over 60 molecules ¹⁵ . Two recent repurposing screens identified 53 ¹⁶ and 80 ¹⁷ compounds with antiviral activity which also overlap the earlier screens. Additional studies have identified small number of inhibitors ^{18, 19} . In total there may now be close to several hundred compounds identified with activity against EBOV in vitro.

Approaches with more capacity to screen compounds include using computational methods as a filter before in vitro testing. Computational models for anti-EBOV activity include one which used the average quasi valence number (AQVN) and the electron-ion interaction potential (EIIP), parameters determining long-range interaction between biological molecules for virtual screening of DrugBank and suggested hundreds of compounds to test ²⁰ . A follow up to this study proposed ibuprofen for testing ²¹ . Others have also used computational docking studies to propose multi-target inhibitors of VP40, VP35, VP30 and VP24 ²² , inhibitors of VP40 ²³ or have suggested molecules to test in the absence of computational approaches ^{24, 25} . We are unaware of any validation of these compounds. A further computational approach used a pharmacophore ²⁶ that was generated from four FDA approved compounds resulting from the two earliest high throughput screens against EBOV ^{2, 3} . This pharmacophore closely matched the receptor-ligand pharmacophores for the EBOV protein 35 (VP35) ⁵ . Follow-up docking studies suggested that these compounds may also have favorable inhibitory interactions with this receptor. The pharmacophore was further used to screen several compound libraries ²⁷ . We proposed that if we could learn from the many compounds already screened for anti-EBOV activity, we could more efficiently find additional compounds and perhaps understand the key molecular features needed for antiviral activity ¹⁴ . We speculated then that Laplacian-corrected Naïve Bayesian classifier models might be useful as they have been for M. tuberculosis ^{28, 29} and more recently for T. cruzi ³⁰ . To our knowledge machine learning approaches to identify EBOV inhibitors have not been attempted elsewhere. The current study extends the machine learning approach to EBOV and uses both commercially available Bayesian, Support Vector Machines (SVM) and recursive partitioning methods and open source Bayesian software for model generation and compound scoring. We report the identification of three novel EBOV inhibitors with nanomolar EC ₅₀ values as validation of this approach.

Our recent work on neglected diseases has shown that we can learn from existing assay datasets. Specifically we have previously analyzed large datasets for Mycobacterium tuberculosis to build machine learning models that use single point data, dose-response data ^{43, 45} , combine bioactivity and cytotoxicity data (e.g. Vero, HepG2 or other model mammalian cells) ^{28, 29, 46} or combinations of these sets ^{47, 48} . These models in turn have been validated with additional non-overlapping datasets, demonstrating that it is possible to use publically accessible data to find novel in vitro active antituberculars. We have also applied the same approach recently to identify a molecule with in vitro and in vivo activity against T. cruzi ³⁰ . In the current study we found that different machine learning methods produced similar 5-fold cross validation data, although the Bayesian models had ROC values consistently above 0.80, which is preferable. One of the issues with computational models is that they are rarely accessible to others due to the commercial software licensing requirements. We have previously showed that models built with open source tools can produce validation statistics comparable to commercial modeling tools ⁴⁹ . We recently made “function class fingerprints of maximum diameter 6” (FCFP6) and “extended connectivity (ECFP6) fingerprints,” open source and have described their implementation with the Chemistry Development Kit (CDK) ⁵⁰ components ⁴¹ . In addition we described an open source Bayesian algorithm that can be used with these descriptors ^{39, 40} . One way to make such models more accessible is to use mobile devices for their delivery and we have developed cheminformatics mobile apps ^{41, 51– 55} . Several of these apps combine Bayesian models and open source fingerprint descriptors to enable models that can be used within a mobile app (TB Mobile, MMDS, Approved Drugs and MolPrime). This enables a scientist to select a molecule and score it with models. In the current study we used the same training sets for the anti-EBOV activity using replication and pseudotype screening data to build open source models that we can share with the community ( http://molsync.com/ebola/).

The Bayesian models allowed us to select three compounds from the MicroSource compound library that scored highly and were not in the model training sets. The Open Bayesian models also scored the three hits favorably, which bodes well for screening other compounds of interest. Two of these molecules had also been identified with our earlier pharmacophore model ²⁶ . When tested in vitro the three compounds possessed EC ₅₀ values 230–420 nM, much lower than the positive control chloroquine (EC ₅₀ 4.0 μM) used in this study and identified previously ³ . Tilorone is an investigational agent that has been known for over 40 years as an antiviral ⁵⁶ and is an inducer of interferon in mice ⁵⁷ . It has been shown to possess a broad array of biological activities including cell growth inhibition in PC3 CDK5dn prostate cancer cells (IC ₅₀ 8–12 μM) ⁵⁸ , inhibition of Primase DnaG from Bacillus anthracis (IC ₅₀ 7.1μM) ⁵⁹ , in a mouse model of pulmonary fibrosis it decreased lung hydroxyproline content and the expression of collagen genes ⁶⁰ , α7 nicotinic receptor (nAChR) agonist activity (K _i 56 nM) ⁶¹ , activated human alpha7 nAChR with an EC ₅₀ value of 2.5 μM ⁶² , radioprotective activity ⁶³ , potent modulation of HIF-mediated gene expression in neurons with neuroprotective properties ⁶⁴ and induction of the accumulation of glycosaminoglycans, delay infectious prion clearance, and prolong prion disease incubation time ⁶⁵ . Quinacrine is an old antimalarial drug now more widely used as an antiprotozoal for the treatment of giardiasis ⁶⁶ and as an anthelmintic. Pyronaridine is a potent antimalarial (IC ₅₀ 13.5 nM) ⁶⁷ , has activity against Babesia spp. ⁶⁸ , is active in vitro (EC ₅₀ 225 nM) and in vivo (85.2% efficacy 4 days treatment at 50 mg/kg) against T. cruzi ³⁰ and is a P-glycoprotein inhibitor ⁶⁹ . Pyronaridine is used in combination with artesunate in the European Medicines Agency approved Pyramax ⁷⁰ which has performed well in clinical trials for malaria ⁷¹ . As this molecule has already been approved this may have a more direct path to clinical testing if it is found to be active in standard animal models infected with the Ebola Virus.

As stated before in perspectives by us ⁷² and others ^{2, 16, 20, 73} , the fact that approved drugs may be repurposed for other diseases should not be viewed as a negative aspect of the small molecules, belying undesirable target promiscuity ⁷⁴ . Instead, we prefer to reference recently published crystallographic analyses ⁷⁵ demonstrating that small molecules may bind multiple proteins in different types of binding sites and with distinct conformations to ultimately facilitate molecular repurposing. While it would be most desirable to repurpose an approved drug and, thus, catapult a discovery effort into a Phase II trial, one should not ignore the significance of utilizing the discovery of a new use for an old drug to seed efforts in the lead optimization phase ⁷⁶ . Such an expedited program would be expected to have a high probability of producing novel small molecules, closely related to or inspired by the drug, with the opportunity to translate quickly to clinical trials.

In summary, this study has added to the previous work that identified several FDA approved compounds active against EBOV in vitro. We propose that these three molecules may warrant further evaluation in vivo as they are significantly more active than chloroquine. Larger scale virtual screening could be performed on the millions of commercially available molecules or more complete sets of approved and older no longer used drugs than have already been screened. These computational efforts can then prioritize molecules for testing. Such an approach may be a useful way to leverage the HTS data that has already been developed at great cost. In this study we have focused on just the data from a single group ^{3, 31} but it may also be possible to combine this with the data from the other high throughput screens ^{2, 16, 17} to provide a much larger training set. There is also the opportunity to apply many different computational approaches beyond those described here to identify whole cell active compounds against EBOV. Ultimately, we should be able to identify additional compounds that could be immediately useful to treat patients with the disease while we await the approval of a vaccine.

Supplemental data contains results from Bayesian models and SVM models as well as the output of predictions with Bayesian models and open Bayesian models.

The training sets used in the models are available as SDF files ( http://molsync.com/ebola/).