Recognition of Potential COVID-19 Drug Treatments through the Study of Existing Protein–Drug and Protein–Protein Structures: An Analysis of Kinetically Active Residues

We report the results of our in silico study of approved drugs as potential treatments for COVID-19. The study is based on the analysis of normal modes of proteins. The drugs studied include chloroquine, ivermectin, remdesivir, sofosbuvir, boceprevir, and α-difluoromethylornithine (DMFO). We applied the tools we developed and standard tools used in the structural biology community. Our results indicate that small molecules selectively bind to stable, kinetically active residues and residues adjoining them on the surface of proteins and inside protein pockets, and that some prefer hydrophobic sites over other active sites. Our approach is not restricted to viruses and can facilitate rational drug design, as well as improve our understanding of molecular interactions, in general.


Introduction
The Coronaviridae positive stranded RNA virus family includes a substantial number of members, many of whom are known to cause a broad range of illnesses from common cold to sever diseases like Severe Acute Respiratory Syndrome (SARS), Middle East Respiratory Syndrome (MERS), etc. [1,2].The latest worldwide rapidly spreading disease, COVID-19, is caused by a new member of this virus family, SARS-COV-2.The disease originally emerged in China in December 2019 with most common symptoms being fever and cough, as well as shortness of breath, sore throat, headache, muscles ache, nausea, and diarrhea [1].In some cases initial symptoms also involve a still unexplained loss of smell and taste [3].The SARS-COV-2 virus spreads through respiratory droplets, directly via physical contact, or through contact with contaminated objects, and can severely affect patients with immune systems weakened by preexisting conditions, such as hypertension, diabetes mellitus or cardiovascular diseases [4].The virus has lower fatality rate than SARS and MERS, but spreads more easily due to high binding affinity between the virus spike glycoprotein (S) and the host receptor [5][6][7][8], making it, potentially, more deadly.The virus is therefore a primary cause or initiator of a significant number of deaths all over the world.By restricting access to work and slowing down supply lines the virus also directly affects the global economy, which experiences a significant decline in gross national products worldwide and unemployment rates reaching levels not encountered since the Great depression.
As this manuscript is being prepared there are several efforts and clinical trials underway to develop a vaccine and evaluate potential drugs for COVID-19, but such investigations usually take months or even years to yield a successful treatment.Drug repurposing, on the other hand, may offer an immediate solution, because it considers already approved compounds as potential treatments for COVID-19.There are two paths toward a viral treatment.One path directly attacks the virus and interrupts its replication machinery or its ability to attack host cells [9].This path is often hard to implement due to rapid emergence of new viral strains with acquired resistance to implemented drugs.The second path should therefore aim to block the host-viral interactions on the host side due to difficulties viral single point mutations should have in recovering the loss of host factors [10].A recent study of human-virus protein-protein interactions (PPIs) detected 332 high-confidence SARS-CoV-2-human PPIs [11].The study showed that 40% of SARS-CoV-2 proteins interact with endomembrane compartments or vesicle trafficking pathways, and that viral proteins also interact with multiple innate immune pathways, the host translation machinery, bromodomain proteins, enzymes involved in ubiquitination regulation, and Cullin ubiquitin ligase complex.Importantly, it showed that the SARS-COV-2 human PPI map is very similar to the interaction maps of West Nile Virus (WNV) and Mycobacterium tuberculosis (Mtb).Among the human proteins involved in interactions with viral proteins, the study detected 66 druggable human (host) proteins targeted by 69 compounds (29 FDA-approved drugs, 12 drugs in clinical trials, and 28 preclinical compounds).It identified two groups of compounds with noticeable antiviral activity: inhibitors of mRNA translation/protein biogenesis (zotatifin, ternatin-4, PS3061, and plitidepsin), and predicted regulators of the Sigma1 and Sigma2 receptors (Haloperidol, PB28, PD-144418, Hydroxychloroquine, Clemastine, Cloperastine, Progesterone, and the clinical molecule Siramesine).The first group of compounds directly affects the viral cap-dependent mRNA translation because coronaviruses use the host translation machinery for their own mRNA translation.The compounds affecting the second group of proteins are approved and long established human therapeutics [11].As much as they are informative, such screening associative studies rarely offer detailed insights into mechanisms of molecular interactions, whereas structural studies [5][6][7][8]12] give snapshots into residue and atom level physical interactions between molecules, but cannot offer general principles of molecular interactions.
To facilitate the drug screening we undertook a comparative study of binding modes of four antiviral candidate drugs.We analyzed compounds that bind to parasitic and to human proteins.We use those results to anticipate their binging affinities.The drugs we have studied so far include (hydroxyl)chloroquine, Ivermectin, Remdesivir (and sofosbuvir), and α-difluoromethylornithine (DMFO) (see Table 1).
Chloroquine [13] and its less toxic derivative hydroxychloroquine [14] are drugs used to prevent and treat acute attacks of malaria.They are also used to treat discoid or systemic lupus erythematosus and rheumatoid arthritis in patients whose symptoms have not improved with other treatments.These drugs are subject of a number of clinical trials worldwide as potential treatment for Covid- 19 [15, 16].Interestingly, the study mentioned above [11] showed that PB28 was ~20 times more potent viral inhibitor than hydroxychloroquine.
Ivermectin is a medication used to treat many various types of parasite infestations [17].They include, but are not limited to, head lice, scabies, river blindness (onchocerciasis), strongyloidiasis, trichuriasis, ascariasis, and lymphatic filariasis.Depending on the kind of treatment, the drug is taken by mouth or applied to the skin for external infestations.Ivermectin molecular structure is rather complex and made of a set of macrocyclic lactone isomers.Ivermectin binds to glutamate-gated chloride channels and increases the permeability of chloride ions.The drug was shown to inhibit the replication of SARS-COV-2 in vitro [18] and is currently the subject of clinical trials as a potential COVID-19 treatment [19].
Remdesivir is a nucleoside analog RNA-dependent RNA Polymerase (RdRp) inhibitor initially developed to treat Ebola and Marburg virus diseases [9,20].The drug decreases the viral RNA production by affecting the function of RdRp and proofreading by viral exoribonuclease (ExoN).Remdesivir is a subject of clinical trials as a potential COVID-19 treatment [21], as it was shown to reduce the lung viral load and improve pulmonary function with SARS infection [9].α-difluoromethylornithine (DMFO), is a medication primary used to treat African trypanosomiasis (sleeping sickness) and excessive facial hair in women [22].Specifically, it is used for the second stage of sleeping sickness caused by T. b. gambiense and may be used with nifurtimox [23].It is used by injection or applied to the skin.The drug prevents binding of the natural product ornithine to the active site of ornithine decarboxylase.We did not find any record of this drug ever being tried, so far, for COVID-19.However, since it is a halogenated organic molecule with somehow similar active sites as Chloroquine we decided to study it towards treatment of COVID-19.
Mass media, internet and even professional publications are currently immersed in controversial debates over the effectiveness, or lack of it, of some of those drug for the treatment of COVID-19.We avoid getting involved in such discussions and only report our research findings based on the existing bioinformatics predictive methods and tools [24,25] and the method which we have implemented to recognize binding patches through the analysis of normal modes in proteins [26,27], Self Adjustable Gaussian Network Model (SAGNM) [28].The method predicts binding areas without any information on the binding partner's properties, position or orientation.

Results
The analysis of chloroquine binding patterns to human lysosomal protein saposin B (pdb id 4v2o) and plasmodium falciparum lactate dehydrogenase (pdb id 1cet) reveals that chloroquine binds to kinetically active sites recognized by the SAGNM algorithm [28] which are mostly hydrophobic (Figure 1).Other kinetically active sites although exposed to solvent are not binding targets.The analysis of Chloroquine's nondiscriminatory binding to human and parasitic proteins may offer an explanation of its efficiency against parasitic infections as well offer a glimpse into its toxicity.Our results suggest that Chloroquine's binding to Covid-19 proteins should follow the same patterns of attachment to residues which are both hydrophobic and kinetically active (or close to kinetically active sites).
The drug Ivermectin binds glutamate-gated chloride channels and thus increases their permeability to chloride ions.We analyzed the Ivermectin's binding to the human glycine receptor alpha-3 (pdb id 5vdh).This structure, besides Ivermectin, also has glycine and the potentiator AM-3607 (7c6) bound to the glycine receptor.The receptor is a pentamer, so we only analyzed the binding to its chain A. The analysis shown in Figure 2 reveals that all three compounds bind to kinetically active and adjoining residues [28], some of which are highly hydrophobic, but Ivermectin binds to almost exclusively hydrophobic residues.That means that this drug well seek similar sites on the surface of the Covid-19 proteins.We used the recently cryo-EM determined structure of SARS-COV-2 RdRp with double-stranded template-primer RNA and Remdesivir (pdb id 7bv2, [12]) to analyze the RNA and drug binding to residues in RdRp.The structure reveals that the double stranded RNA is inserted into RdRp's central channel and that the active triphosphate form of Remdesivir is covalently bound to the primer strand at the first replicated base, which effectively terminated the chain elongation.It shouldbe noted that the prodrug form of Remdesivir does not have any inhibitory effect on the polymerization activity of the purified enzyme [12].Our analysis reveals that the residues recognized via the fastest two normal modes (corresponding to kinetically active residues [28]) delineate the central channel (Figure 3).The enzymatically important residues K500, S501, K545 and R555 are all recognized by the SAGNM algorithm using just the fastest normal mode, while the residue D761 of the catalytic center (out of residues 759-761 that form the catalytic center) is also emphasized with the two fastest modes.Residues K545 and R555 are important because they stabilize the incoming nucleotide in the correct position for catalysis.The crystal structure shows that the catalytic center of RdRp, NSP12 protein (Non Structural Protein 12), does not have any contacts with base pairs of RNA emphasizing RdRp's sequence-agnostic polymerization ability [12].This is in concordance with our coarse grained analysis, based on the positions of C-α atoms only, that shows that stable, kinetically active residues outline the enzyme's central channel.
We also performed a comparative analysis of the Hepatitis C virus (HCV) RdRp (chain A in pdb id 4wtg; the structure is give with the drug sofosbuvir bound to it) and the Covid-19 RdRp (chain A in pdb id 6m71).We followed the steps of Y. Gao and collaborators [29] and attempted to compare predictions of the binding residues in HCV RdRp to sofosbuvir, to binding residues predictions in Covid-19 RdRp.The binding residues in HCV are buried deep inside the polymerase catalytic core.Our analysis shows that they are generally delineated by the kinetically active residues and are thus stable and characterized by the two fastest normal modes (Figure S1a), but they are not explicitly hydrophobic (Fig. S1b-c).The structural alignment of HCV and Covid-19 RdRp (Figure S1d) shows that they share the structure of the binding pocket, and also reveals that the catalytic cores in both are bounded by kinetically active residues, but the overall distribution of residues is partially different between the two proteins (Figure S1e).The similarities suggest that the interior of the RdRps in coronaviruses are attractive binding spot for small compounds in general.
We also analyzed the binding patterns of the three proteins that form the SARS-COV-2 RdRp (pdb id 6m71).The main enzymatic unit is NSP12, which mostly keeps its conformation between RNA free and RNA bound structures [12].Figure S2 shows that cofactors NSP7 and NSP8 seek patches with kinetically active residues on the surface of NSP12.However, they are also in contact with kinetically less active areas.This should be analyzed in the light of fact that SARS-COV-2 RdRp (NSP12) cannot perform its function without NSP7 and NSP8 [12].The distribution of kinetically very active and kinetically dormant residues may be important for the overall stability of NSP12, and also act as stochastic transformer that translates random fluctuations of solvent and proteins into a regular vibrations that produce a regular rhythm of translation (i.e.act as a regular clock/oscillator).
The analysis of the contact patterns between the ACE2 receptor and the spike glycoprotein receptor binding domains (RBD) in SARS (pdb id 6cs2) and SARS-COV-2 (pdb id 6m0j) reveals a difference in the distribution of kinetically active residues important for binding between RBD and ACE2 (Figure 4).SARS-RBD has a smaller number of kinetically active and adjoining residues in direct contact with ACE2 (Figure 4a-c), while kinetically active residues in Covid-19 RBD are directly oriented and are in contact with the active residues in ACE2 (Figure 4d-f).In SARS active residues are mostly perpendicular to the interfacial plane (compare the distributions of C-alpha atoms in Figs.4a and 4d).That should make the binding affinity between the Covid-19-RBD and ACE2 receptor stronger than between the SARS-RBD and ACE2 receptor.In both cases the predicted residues are recognized via the fastest vibrational mode (see [28]).
The analysis of kinetically active and adjoining residues in the SARS monomer (pdb id 6nb6) reveals that they are attractive binding spots for glycans (Figure S3).Glycans form the glycan shield, which was already suggested to assist in immune evasion similarly to the HIV-1 envelope trimer [30].The kinetically active residues recognized by the SAGNM algorithm [28] can be used as target areas for drugs aimed at removing/disrupting the viral glycan shield.Those residues are not particularly hydrophobic and should be targeted by drugs that bind to hydrophilic patches, and electrostatically complementary.
We also analyzed the kinetically active residues in the structure formed by the SARS Spike Glycoprotein RBD and human neutralizing S230 antibody FAB fragment (pdb id 6nb6).The analysis reveals that S230 antibody binds to kinetically active residues in SRAS RBD, while heavy and light chains in S230 communicate via kinetically active residues (see Figure S4).The binding residues are mostly neutral to hydrophilic, thus any potential drug should be able to bind to similar surfaces (neutral/hydrophilic and stable).
The drug α-difluoromethylornithine (Eflornithine) prohibits binding of the natural non-coded amino acid ornithine to the active site on the surface of Trypanosoma brucei ornithine decarboxylase (ODC, pdb id 1njj).The binding of this drug should follow the binding patterns of ornithine.Figure S5.shows that SAGNM algorithm accurately detects binding sites for both ornithine and G418 (Geneticin), an aminoglycoside antibiotic.In contrary to chloroquine, both compounds bind preferably to the hydrophilic sites on the surface of ODC.If applied to treat Covid-19, the drug Eflornithine should bind to similar sites on the surface of Covid-19 proteins (hydrophobic and kinetically active, i.e. stable).

Conclusion
Covid-19 is the first severe global pandemic caused by a coronavirus, and there are no guarantees that it will be the last.We therefore need not only to develop an effective and efficient treatment for the current pandemic, but also have to design a set of protocols to address all future, similar pandemics.In this manuscript we presented our strategy to recognize potential drug binding residues in human and viral proteins.We analyzed four currently approved drugs (Chloroquine, Ivermectin, Remdesivir, and Eflornithine).Our results indicate that small, drug like compounds preferentially bind to kinetically active and adjoining residues, thus seeking stable residues characterized by fast normal modes with small amplitude of fluctuations [28].Some drugs preferentially seek active patches that are hydrophobic (Chloroquine, Ivermectin), while others prefer hydrophilic surfaces (Remdesivir, Sofosbuvir, Eflornithine).We can postulate that in water environment drugs binding to hydrophilic patches will be more stable, as their removal will lead toward the reduction in structural entropy, but a full account of this proposition will require calculations of binding free energy differences based on full atom molecular dynamics, using, for instance, steered molecular dynamics simulations (SMD) [31].We can also propose that the drugs/small molecules that bind to deep pockets will be more stable, and thus more effective.Our algorithm accurately recognizes such pockets as binding spots for drugs (Figure 3), and small peptides (see, in particular, Figure 6a in [28]).
Multidrug cocktails are frequently used to treat viral diseases [32].Our analysis shows that in designing antiviral drug cocktails, binding affinity to and between kinetically active (stable) sites should be combined with the information on their hydrophobic and hydrophilic properties to avoid conflicts, increase drug cocktail efficiency, and reduce toxicity and other unwanted side effects.
In our analysis we used both viral-parasitic, as well as human proteins.The analysis shows that kinetically active residues exist in both human and non-human proteins/enzymes and that drugs bind indiscriminately to them regardless of their origin.The compounds that bind to human proteins potentially offer longer lasting treatments as host cells and tissues have less chance of developing drug resistance through single point mutations.
The protocol depicted here may also help in accessing drug toxicity.Binding spots in various proteins can be very effectively predicted with our SAGNM approach and accessed with other bioinformatics tools for charge and shape complementarity, binding affinity, atomic mass and other properties as well.
Recent advances in machine learning helped advance our ability to predict and design protein structures [33], but the full theoretical foundation is still lacking.The quality of the machine learning protocol directly depends on the quality and size of training data sets and thus in many ways follows classical methods based on statistical potentials and homology modeling [34].Our results can also help in that respect as they offer interpretation on how residue packing inside protein segments guides their assemblage.
The procedure we described here is fast and effective, and can analyze protein structure much faster than computationally demanding molecular dynamics simulations.The only requirement is the protein structure.Its advantage is not in its efficiency, but also in its ability to suggest general binding patterns between proteins and drugs or small peptides.It can be used to filter binding areas on protein surfaces and thus facilitate preclinical stages in drug design.Recognition of potential Covid-19 drug treatments through the study of existing protein-drug and protein-protein structures: an analysis of kinetically active residues Ognjen Perišić 1 (1) Big Blue Genomics, Vojvode Brane 32, 11000 Belgrade, Serbia, ognjen.perisic@gmail.com May 13, 2020

Methods and Materials
Our aim was to analyze presently available structures existing drugs bound to parasitic and human proteins and predict their binding patterns, as well as the binding patterns of SARS and SARS-COV-2 binding patterns to the human ACE2 receptor.To predict binding residues in protein we applied our Self Adjustable interpretation of Gaussian Network Model (SAGNM) [28].The structure alignment, visualization and analyses were performed with programs Chimera [24] and VMD [25].
We focused our study on pdb structures with the listed drugs present as ligands.For chloroquine we analyzed 2 structures (pdb ids 1cet and 4v2o).For Ivermectin we analyzed the binding pattern of the drug to the human glycine receptor alpha-3 (pdb id 5vdh).For Remdesivir we analyzed its binding patterns in the recently released structure [12] (pdb id 7bv2).We also performed the analysis of binding pattern of drug Sofosbuvir to the hepatitis C virus (HCV) RdRp (pdb id 4wtg) and compared them to the Covid-19 RdRp predictions (pdb id 6m71).Sofosbuvir was already analyzed in light of similarities between HCV and SARS-COV-2 RdRp and similarities between Remdesivir's and Sofosbuvir's [29].For α-Difluoromethylornithine we analyzed a structure of Trypanosoma brucei ornithine decarboxylase (ODC) with D-ornithine bound to it (pdb id 1njj).α-Difluoromethylornithine binds to the active site of ODC and inhibits ornithine binding to it.We performed the comparative analysis of the binding patterns between the ACE2 human receptor and the spike glycoproteins from SARS (pdb id 6cs2) and SARS-COV-2 (pdb id 6moj).We also analyzed the binding patterns between the SARS RBD with S230 human neutralizing antibody, and between SARS RBD and glycan shield (pdb id 6nb6).

Figure 4 .
Figure 4. SARS spike glycoprotein chain B RBD and ACE2 receptor (pdb id 6cs2) in comparison to Covid-19 spike glycoprotein chain A RBD and ACE2 receptor (pdb id 6m0j).In a) and d) spike glycoprotein's predictions are shows through C-alpha atoms only.In b, c, e and f predictions are shown through all atoms.

Figure S3 .
Figure S3.SARS-Cov spike glycoprotein (Chain B, pdb id 6nb6) and glycans (NAG, BMA, MAN) bound to it.The algorithm predictions are yellow.The predictions accurately predict binding spots for glycans.They can be used as target areas for drugs aimed at removing/disrupting the virus glycan shield.Cyan bars (Figure S4a) represent missing segments.

Figure S4 .
Figure S4.Receptor binding domain (RBD) of SARS-COV spike glycoprotein (Chain A, pdb id 6nb6) with human neutralizing S230 antibody FAB fragment.a) SRAS-COV RBD (blue, chain A) with heavy (green, chains H and I) and light chains (red, chains L and M).Predictions are cyan (SARS), yellow (S230 light) and light green (S230 heavy).b) Hydrophobic surface of SARS RBD and S230 (chains H and L).c) Transparent hydrophobic surface of SARS RBD and S230 (chains H and L) with predictions.

Table 1 .
Comparison of existing drugs currently being tested for the antiviral treatment and prevention of Covid-19 through drug repurposing.