Deep Reinforcement Learning and Structure-Based Approaches in the de novo Design of a New Potential Inhibitor of F13 Protein from Monkeypox Virus

Abstract

Monkeypox (MPOX) is a zoonotic infectious disease caused by the monkeypox virus (MPXV) and has recently emerged as a significant concern for public health organizations globally. In 2022, the World Health Organization (WHO) reported thousands of laboratory confirmed cases, mobilizing the scientific community to control this phenomenon due to its emergency nature. Tecovirimat (TPOXX), a drug primarily recognized for the treatment of smallpox, has also been recommended for managing MPOX. It works by inhibiting the viral F13 protein (VP37), a critical component in the replication cycle of the virus. Some issues related to the possibility of drug resistance by the virus, the intrinsic chemical complexity of this molecule and the limited availability of therapeutic alternatives highlight the urgent need to explore and identify new effective compounds. In this paper, we propose the combination of modern machine learning techniques (deep reinforcement learning) with structure-based drug design (SBDD) approaches (molecular docking and dynamics) in the de novo design of molecular scaffolds with affinity for the F13 protein, lower structural complexity than TPOXX and easy synthetic accessibility, contributing to efforts in the search for therapeutic alternatives for MPOX.

Keywords:
monkeypox virus; tecovirimat; deep reinforcement learning; de novo design; molecular dynamics; molecular docking

Introduction

Monkeypox (MPOX) is an infectious zoonosis described since 1958, caused by the monkeypox virus (MPXV), belonging to Orthopoxvirus genus, which also includes smallpox, bovine smallpox and vaccinia viruses.¹ The transmission occurs mainly by direct contact with lesions, crusts, body fluids or respiratory perpetrators of infected people or animals, such as primates and rodents.² Clinically, symptoms are similar to those of smallpox, but less severe, manifesting with high fever, headache, lymphadenopathy and skin lesions that evolve from pustules, with lethality rate between 1 and 10%.³

In the first half of 2022, a global outbreak resulted in a significant increase in the number of cases, including individuals with no history of travel to endemic areas. The World Health Organization (WHO)⁴ reported 3,413 laboratory confirmed cases and one death in 50 countries/territories, covering five of the six WHO regions, mobilizing the scientific community to control this phenomenon due to its emergency nature.

Three vaccines are available for orthopoxvirus, the “ARM2000”, “Jynneos” and “LC16”, which may also prevent MPXV infections, although they have not been developed specifically to combat this virus.⁵ The US FDA (United States Food and Drug Administration) approved the first two for the prevention of monkeypox.⁶ In addition, the “Jynneos” licensed vaccine, which uses a non-replicating living virus, was recommended for people at risk of occupational exposure.⁷ However, the effectiveness of these vaccines in recent outbreaks, observed in several countries, is not yet fully proven, showing the need to develop a new generation of specific vaccines for MPXV, which are safer and more effective.⁵ An important fact is that, with the eradication of smallpox and, therefore, the interruption of vaccination for more than forty years ago, the MPXV has found conditions to resurface, now with different characteristics.⁸

Currently, three antivirals stand out in the fight against orthopoxvirus infections, including monkey smallpox virus: tecovirimat (TPOXX), brincidofovir and cidofovir.⁹ Among them, tecovirimat and brincidofovir were approved by FDA specifically for the treatment of smallpox,¹⁰ while cidofovir is approved only for the treatment of cytomegalovirus retinitis (CMV) in patients with acquired immunodeficiency syndrome (AIDS).¹¹ Nevertheless, cidofovir has been used off-label in patients with severe immunosuppression that have advanced MPOX. The justification for this use stems from extrapolated data from brincidofovir, which acts as a DNA (deoxyribonucleic acid) polymerase inhibitor, thus blocking viral replication. Although randomized clinical trials in humans are not yet available, studies on animal models indicate that brincidofovir may have effectiveness against MPXV.¹²

TPOXX acts by inhibiting F13 (VP37) protein activity, which is essential for the dissemination of orthopoxvirus. Furthermore, it has more consistent evidence regarding its safety, pharmacokinetics and pharmacodynamics, not only from studies in animal models, but also some data in humans.¹³,¹⁴ This makes TPOXX a promising option, with more detailed information about its efficacy and security profile compared to brincidofovir and cidofovir, which still lack more comprehensive clinical investigations in humans. However, the possibility of viral resistance to TPOXX is a potential limitation to be considered.¹⁵,¹⁶ Therefore, restricting therapeutic options highlights the importance of continuing to seek new molecules and approaches to face MPOX.

The de novo design of drugs is an approach that creates unpublished molecules from atomic blocks and can depend on the inspiration of pre-existing structures, using advanced computational methods. It can be used to adjust ligands to specific biological targets, such as disease-related proteins, maximizing effectiveness and/or reducing side effects. Traditionally, drug design is based on evaluation of binding sites (structure-based design) or known ligands (ligand-based design). Currently, the use of artificial intelligence (AI), especially machine learning (ML) and deep learning (DL) techniques, has revolutionized the development of new drug structures accurately, acting as a promising strategy, particularly to treat difficult targets, playing a central role in the modern pharmaceutical industry.¹⁷

Machine learning is a field of AI focused on the study of computer programs that can improve performance through experience.¹⁸ The performance of traditional machine learning algorithms heavily depends on the representation of the data they receive as input.¹⁹ This representation is manually crafted and requires domain-specific knowledge. A solution to reduce this dependency is to use machine learning to discover not only the mapping from input representation to output but also the input representation itself.²⁰ This approach is known as representation learning. Deep learning methods are representation learning techniques with multiple levels of representation, achieved by composing simple but non-linear modules, where each module transforms the representation at one level (starting from the raw input) into a higher-level, slightly more abstract representation.²¹

Several machine learning approaches have been explored in the de novo development of new drugs.²² Among these, reinforcement learning (RL) stands out, aiming to learn which actions should be taken within an environment to maximize a reward signal. In RL, the agent interacts with the environment without prior knowledge of ideal actions, discovering, through trial and error, those that yield the highest reward.²³

In RL, the environment is described as a discrete-time Markov decision process (MDP), consisting of four main elements: S represents the set of possible states of the environment; A is the set of actions the agent can perform; P(s’ |s, a) is the transition function, defining the probability of reaching a new state (s’) upon taking action (a) in the current state (s); R is the reward function, assigning a reward value to the agent.²³

The agent learns through successive interactions, guided by the reward function R, which provides feedback for each action executed. This iterative process models the relationship between the agent and the environment. At each time step t, the agent observes the current state (s), selects an action (a), and, as a result, transitions to a new state (s’). The transition function P(s’|s, a) determines this transition, while the set of action-selection rules is called the Policy (π), which defines the behavior of the agent.

The goal of learning is to build an optimal policy that maximizes the cumulative reward over time. The total reward in a state (s) following a policy n is the expected value for an agent, given by the value function:

(1) ${\text{V}}^{\text{π}}\left(\text{s}\right)=\text{π}={\displaystyle \sum _{\text{k}=0}^{\text{∞}}{\text{γ}}^{\text{k}}\times {\text{R}}_{t+\text{k}+1}}$

where γ is a discount factor, making rewards in the near term more desirable than those in the distant future, and R_t is the reward at time instant t. The ultimate goal of the agent is to learn the optimal policy π* that maximizes this reward over time. In the context of drug discovery, the agent is a computational model that seeks to optimize the structural attributes of a molecule. The environment is the chemical space, where each molecule represents a state. The actions of the agent involve structural modifications, such as adding or removing functional groups. Rewards are assigned based on desirable properties, such as physico-chemical descriptors, chemical similarity or binding energy values. Figure 1 illustrates this process. RL enables the agent to explore efficiently the chemical space, learning from its experiences, which structural modifications lead to molecules that are more desirable. This approach is particularly useful for complex problems, where the relationship between molecular structure and biological activity is challenging to model explicitly.

Figure 1
Reinforcement learning process for drug discovery (adapted from reference 24).

A popular class of algorithms in RL is the policy gradient method.²⁵ In this approach, the policy is often represented by a parameterized neural network that maps states to a probability distribution over possible actions. Optimization occurs by updating the parameters of the network, commonly denoted by θ in the literature.²² The agent’s learning is thus guided by the gradient of an objective function J(θ), which depends on the policy parameters. To maximize performance, network parameters are updated in the direction of the ascending gradient of J(θ), as described in equation 2:²⁶

(2) $\text{θ}\left(\text{t}+1\right)=\text{θ}\left(\text{t}\right)+\text{α∇J}\left(\text{θ}\right)$

The term ΔJ(θ) is known as the “policy gradient.” The symbol a represents the learning rate, controlling the update step size to ensure the algorithm converges efficiently to an optimal policy.

Also addressing the computational design of drugs, molecular docking can be a useful tool in the de novo design process, allowing to predict the interaction between candidate molecules and their biological targets, such as proteins and nucleic acids. By simulating how a potential drug interacts into a binding site, the docking provides valuable information about affinity and specificity of these connections. This technique not only helps identify the best molecules to continue in vitro experimental studies, but also optimizes interactions by modeling different structures, conformations and optimizing possible interactions through human interpretation and chemical sense.²⁷

In fact, molecular docking provides a quick and relatively accurate method for affinities assessment and comparison of interactions performed by different ligands in a macromolecule, considering its three-dimensional aspect. However, it usually has important limitations such as total or partial rigidity of the macromolecule and absence of solvation by water molecules. In this sense, molecular dynamics (MD) calculations have consolidated as an important step in rationalizations before or after dockings. MD allows observing the temporal behavior of the target-ligand complexes, subjected to specific force fields, bringing the system closer to physiological conditions. It offers a detailed view of the stability and flexibility of the formed complexes. Thus, MD can reveal critical information related to geometry, energy and temporal dynamic of solvated systems. Together, molecular docking and dynamics form a powerful synergy, that optimizes the prediction of interactions and supports the development of new drugs.²⁸

Methodology

De novo design by deep fragment-based multi-parameter optimization

In this study, we applied the deep fragment-based multi-parameter optimization (DeepFMPO) method,²⁴ which is based on an actor-critic model for reinforcement learning.²⁹ The actor-critic model advances traditional policy gradient methods by employing a neural network that simultaneously learns the policy and the reward estimation. This method consists of two outputs (or can be seen as two separate networks): “actor” that is responsible for recommending the action to be taken by generating a probability distribution over possible actions for a given state, similar to the policy gradient model. In the context of computational chemistry, the actor can decide which atoms should be added or removed from a molecule or which bonds to modify; “critic” evaluates the actions suggested by the actor, estimating potential future rewards. This evaluation provides feedback to the actor, indicating whether the molecule modifications bring it closer to or further from the desired properties.

The actor-critic method integrates reinforcement learning with policy gradients, allowing the agent not only to propose actions but also to receive evaluative feedback based on estimated future rewards. This dual-network structure enables more stable training, as the critic helps to reduce variance in gradient updates by providing an estimate of future rewards, an essential factor in tasks with complex reward structures, such as molecular optimization.

DeepFMPO²⁴ operates with two molecular datasets. The first dataset represents molecules that the algorithm will modify to generate new compounds with desirable properties. The second dataset comprises molecules used to construct a fragment library, which forms the basis for generating new molecules from existing ones. This library is built by fragmenting a set of molecules that typically show activity against the biological targets of interest in a drug discovery project. Broadly, the method learns to generate new compounds with desired properties by starting with an initial set of lead molecules and improving them by replacing some of their fragments.

Fragments are derived by breaking single bonds extending from ring atoms in molecules, typically those with known activity against specific biological targets. Molecules can come from various datasets, including FDA-approved drugs, chemical reagents, and screening collections. This fragmentation method does not sort fragments into classes, treating all fragments equally. Attachment points from fragmentation are recorded for later assembly. The process involves exchanging fragments with differing attachment points to maintain consistent overall structure. The RDKit, version 2024.09.3, Python library³⁰ is used for this purpose. Although ring bonds are not broken during fragmentation, the assembly step allows for the replacement of rings with open chains and vice versa. Fragments exceeding 25 heavy atoms or having four or more attachment points are excluded to simplify the process while generating a diverse set of candidates. Figure 2 illustrates the molecule fragmentation process.²⁴

Figure 2
Illustration of the molecular fragmentation process (adapted from reference 24).

In this paper, the DeepFMPO²⁴ configuration includes both the actor and critic models, each designed with bidirectional long short-term memory (LSTM) networks.³¹ The actor model consists of a time-distributed dense layer with 128 neurons, followed by two LSTM layers (one forward and one backward) with 32 neurons each. The outputs are flattened, concatenated with an auxiliary input, and passed through a dense layer with 32 neurons, followed by a final softmax layer for action prediction. The critic model shares a similar structure, starting with a time-distributed dense layer of 128 neurons, followed by a bidirectional LSTM layer with 64 neurons. The output is concatenated with an auxiliary input, passed through a dense layer with 32 neurons, and ends with a single-neuron output layer with linear activation to estimate the value function.

Two in silico experiments were performed using the DeepFMPO²⁴ algorithm to design molecules with potential inhibitory effects on F13. The standard TPOXX molecule (Figure 3) was used as a base for constructing the lead molecule “dataset”. As your turn, another dataset was used to build the fragment library. This dataset consists of 6,000 molecules with higher values of the MCE-18 drug likeness parameter,³² which provides good molecular inspiration for potentially promising drug candidates. This is a molecular descriptor given by the following equation 3:

Figure 3
Structural representation of tecovirimat (TPOXX) (obtained from PubChem database).

(3) $\text{MCE}18=\left(\text{AR}+\text{NAR}+\text{CHIRAL}+\text{SPIRO}+\frac{{\text{sp}}^3+\text{Cyc}+\text{Acyc}}{1+{\text{sp}}^3}\right)\times {\text{Q}}^1$

where AR is the presence of an aromatic or (hetero)aromatic ring (0 or 1); NAR is the presence of an aliphatic or a (hetero)aliphatic ring (0 or 1); CHIRAL is the presence of a chiral center (0 or 1); SPIRO is the presence of a spiro point (0 or 1); sp³ is the portion of sp³-hybridized carbon atoms (from 0 to 1); Cyc is the portion of cyclic carbons that are sp³ hybridized (from 0 to 1); Acyc is a portion of acyclic carbon atoms that are sp³ hybridized (from 0 to 1); and Q¹ is the normalized quadratic index.³² This descriptor is used in drug development campaigns, bringing together a set of characteristics of promising new chemical entities. The DeepFMPO²⁴ process was also optimized to ensure that generated structures possessed certain values of the following properties: lipophilicity (cLog P), polar surface area (PSA), and molecular weight (MW). Their reference values were first considered in a range that encompasses TPOXX itself and, in a second in silico experiment, using the values from the study of Olivecrona et al.²⁵ (Table 1).

Table 1
Reference values for molecular properties used in DeepFMPO²⁴ in silico experiments

Property	Experiment 1	Experiment 2
MW / (g mol-1)	320-420	320-420
cLog P	2.2-3.2	2.0-3.0
PSA / Å2	50-80	40-60

• DeepFMPO: deep fragment-based multi-parameter optimization; MW: molecular weight; cLog P: lipophilicity; PSA: polar surface area.

TPOXX-F13 interactions via molecular dynamics simulations

Initially, TPOXX molecule was downloaded from PubChem database,³³ in a molecular mechanics pre-optimized three-dimensional (3D) format. The 3D structure of the F13 protein was modelled by Alphafold2 platform,³⁴ using a FASTA sequence obtained from National Center for Biotechnology Information (NCBI, GenBank:³⁵ WBN90803.1 “envelope phospholipase F13 [Monkeypox virus]”), that corresponds to an isolated strain in Brazil.³⁶ To obtain the interaction model of TPOXX with the protein, a molecular dynamics simulation was performed using the GROMACS 2019.2 package³⁷ and the GROMOS 54A7 force field. The specific interaction site and initial ligand geometry was determined based on previous literature,³⁸ which demonstrated the main amino acids and complementary interactions. Based on the reference and using the Chimera software,³⁹ residues Asn55, Ser58, and Arg89 were spatially oriented toward the (trifluoromethyl) benzene moiety of TPOXX and Phe52 and Leu118 were oriented toward the tricyclic part of the ligand.³⁸ This orientation can be further justified in terms of hydropathy scale, where the more hydrophilic amino acids tend to interact with the (trifluoromethyl)benzene moiety, while the hydrophobic amino acids tend to interact with the tricyclic moiety of TPOXX.⁴⁰ Once this positioning was guaranteed, the system was subjected to energy minimization by steepest descent algorithm, before the equilibration and production stages. In this process, a cubic box enclosing the entire complex was drawn and filled with water molecules based on the GROMACS SPC (simple point charge)-216 model, which consists of a pre-equilibrated model system with 216 SPC water molecules that replicates to fill the entire dimension of the box, following periodic boundary conditions. Na⁺ and C1^– were used replacing solvent molecules to neutralize the total charges, considering a final physiological concentration of 0.1 M. The long-range electrostatic interactions were modelled using the particle Mesh-Ewald method, with a cutoff of 1.2 nm.⁴¹ The same cutoff was used for the calculation of the van der Waals interactions. Bond lengths involving hydrogens and water molecules were initially restrained using the P-FINCS⁴² and SETTFE algorithms,⁴³ respectively. The leap-frog algorithm⁴⁴ was applied to integrate the motion equation. The systems were submitted to an NVT (constant number of atoms, volume and temperature) ensemble, performed at 310 K, with a time step of 2 fs for 100 ps, using the modified Berendsen thermostat. Subsequently, was performed an NPT (constant number of atoms, pressure and temperature) ensemble under the same conditions, with a temperature of 310 K and a pressure of 1.0 bar, using a Parrinello-Rahman barostat⁴⁵ and the modified Berendsen thermostat. Finally, a full MD simulation of 100 ns was performed at 310 K and 1.0 bar of pressure, with an integration time step of 2 fs, collecting data every 10 ps. Analysis of the RMSD (root mean square deviation) plot for the system allowed the selection of a balanced production range, from which 10 protein conformations were extracted for subsequent docking calculations.

Molecular docking analysis

The compounds obtained via de novo design were subjected to molecular docking calculations, considering as targets ten conformations selected from well-balanced TPOXX-F13 dynamic complex (70-80 ns). At this stage, we aim to verify which of the compounds obtained would be most promising in affinity criteria for future in vitro/in vivo considerations. The calculations were performed by AutoDock Tools⁴⁶ and AutoDock Vina⁴⁷ software. The proteins and ligands were prepared using the Chimera software,³⁹ where the water molecules were removed. The grid box was defined for each of the ten conformations of the complexes and centered at respective TPOXX location. The internal grid points spacing was 1.0 Å, a default value commonly used in AutoDock Vina.⁴⁷ A redocking was then performed on these 10 complexes (frames), checking which ones would be suitable for subsequent comparisons with the molecules originating from the deep reinforcement learning experiments.

Results and Discussion

From the two types of DeepFPMO²⁴ in silico experiments, two molecular structures were generated from the first condition and three from the second. The two sets are clearly differentiated, in the first, by the presence of a spiro system with saturated heterocycles and, in the second, a tetrahydro-thioindolizinone nucleus (Figure 4).

Figure 4
Two molecular sets generated by DeepFPMO²⁴ algorithm; (a) employing conditions of the first in silico experiment; (b) results for the second experiment.

The computational cost of training the deep actor-critic model in this study is influenced by the size of the fragment library, the number of lead molecules, and the hardware used. For the dataset, which consisted of a limited number of lead compounds (only one - TPOXX), the training process required approximately 1 h. This was achieved using the free graphics processing unit (GPU) infrastructure available on Google Colab,⁴⁸ which, despite its limitations compared to a specialized hardware, proved sufficient for handling this task. Considering the above results, a molecular docking analysis was proposed to verify the affinity of the designed compounds in the F13 protein, compared with the standard ligand TPOXX. In the absence of crystallographic data of the TPOXX-F13 complex, the protein was modeled through the Alphafold 2 platform³⁴ and then the complex was subjected to molecular dynamics, as described in detail in the Methodology section. Figure 5 presents some temporal parameters evaluated after MD calculations. It can be observed that the protein and ligand (Figures 5a and 5b) acquire temporal stability after 20 ns of simulation, maintaining a vertical fluctuation in RMSD of at most 0.5 Å (0.05 nm), where the section between 70-100 ns can be highlighted as a feasible production range. Additionally, the minimum ligand-protein distance (Figure 5c) remains stable in this production range (1.8-2.8 Å), where the establishment of two hydrogen bonds complemented by a third (Figure 5d), more spaced one predominates. These data together indicate the formation of a stable complex, similar to the data found by Ali et al.³⁸

Figure 5
Geometric parameters obtained from 100 ns molecular dynamics calculations of the TPOXX-F13 complex. (a) RMSD plot for alpha-carbons; (b) RMSD plot for TPOXX; (c) minimal distances between TPOXX and F13 along the trajectory; (d) number of ligand-protein hydrogen bonds along the trajectory.

From the production range between 70-80 ns, 10 frames were selected, corresponding to a conformational sampling of the ligand-protein interactions, taken every 1 ns of this interval. This range was chosen because the system was well balanced, being the number of frames and spacing arbitrary, with the aim of improving the representation for docking, instead of using a single conformation of the complex. These ten complexes were then used as a model for redocking calculations and subsequent evaluation of the designed de novo generated structures.

The redocking results can be observed in the Table 2, which denotes that only five of the ten selected complexes had a good validation by AutoDock Vina.⁴⁷ In this sense, these 5 complexes with satisfactory ligand RMSD (< 2.0 Å) were used for the affinity evaluations of the designed compounds. In molecules with RMSD greater than 2.0, a “flip” was observed, with a 180-degree rotation of the structure in relation to what was expected. This can be explained by the hydrophobic nature of the two terminal regions of TPOXX (Figure 3), which may provide this possibility within the analysis via docking.

Table 2
RMSD and binding energies for the ten TPOXX-F13 redocked complexes, extracted from 70-80 ns range (1 ns spacing) of molecular dynamics calculations

Frame (F13)	RMSD	Binding energy / (kcal mol-1)
1	9.227	–9.2
2	8.708	–9.4
3	8.871	–9.7
4	1.867	–9.2
5	1	–9.7
6	1.512	–9.7
7	1.501	–9.8
8	3.588	–8.2
9	9.059	–10.0
10	1.211	–9.9

• TPOXX: tecovirimat; RMSD: root mean square deviation.

Table 3 presents the docking energies for the five de novo designed compounds (1a, 1b, 2a-2c) in the five conformations (frames) of the F13 protein, used as targets. In this case, in each complex the respective TPOXX molecule existing in the frame was removed so that the test ligand could be allocated. The values in Table 3 corresponds to the averages on energies for the ten best poses of AutoDock Vina.⁴⁷ We can observe that compound 1b presents the highest success on average binding energy.

Table 3
Docking energies for de novo designed compounds in the selected conformations of F13

F13-Frame	Docking energy / (kcal mol-1)
	1a	1b	2a	2b	2c
F13-4	–7.4	–7.7	–7.5	–7.1	–7.1
F13-5	–7.7	–8.1	–7.7	–7.5	–7.3
F13-6	–7.6	–7.8	–7.6	–7.4	–7.7
F13-7	–7.1	–7.8	–7.4	–7	–7.4
F13-10	–8	–8.1	–7.7	–7.4	–7.5

Two observations are important at this stage. First, the stabilization energies of the de novo-designed molecules are less negative than those of the tecovirimat (TPOXX) prototype. In this context, less negative values indicate weaker binding affinities. Second, the generation of anew drug is often a nonlinear process, relying, at some point, on the skill and experience of the Medicinal Chemist. Other important step in the new drugs investigations is to reduce costs and facilitate the synthetic development, to enable the use of a new bioactive substance. In this sense, the spiro system obtained by the AI design may still represent a certain complexity in the search for viable bioactive alternatives for TPOXX. In this way, the structural inspiration of 1b led to the proposal of four relatively simpler systems than TPOXX, which could satisfy the requirements of the F13 binding site. Thus, the compounds 3a-3d (Figure 6) were designed and submitted to a similar docking analysis, under the same previous conditions.

Figure 6
Structures of the analogous compounds proposed based on the structures generated by the de novo design

After affinity docking analyses, compound 3b proved to be the most promising (Table 4) regarding the pharmacodynamics affinity criteria, with better binding energy value than its analogues and de novo precursor molecules. Considering the best frame for protein in redocking (F13-10), the energy of the best pose of 3b was –9.9 kcal mol^-1, exactly comparable to TPOXX redocking. Observing the three-dimensional geometry of both at the binding site (Figure 7), we verified the physical meaning of such a result, indicating 3b as the prototype compound to be considered from the present investigation.

Table 4
Average docking energies for compounds 3a-3d in the selected conformations of F13

F13-Frame	Average docking energy / (kcal mol-1)
	3a	3b	3c	3d
F13-4	–8.3	–8.9	–7.9	–7.8
F13-5	–8.9	–8.8	–8.4	–8.6
F13-6	–8.7	–9	–8.3	–8.4
F13-7	–8.3	–8.2	–7.7	–7.8
F13-10	–9	–9.3	–8.4	–8.2

Figure 7
Redocked geometry for TPOXX (orange) and docking best pose of 3b (green) in the binding site of F13 protein (frame 10), binding energy = –9.9 kcal mol^-1.

Finally considering the synthetic accessibility of 3b, we rationalize a possible retro synthetic analysis of this new proposed compound, inspired by the paper of Ganwir et al.,⁴⁹ which is demonstrated in Figure 8.

Figure 8
Proposal for retrosynthetic analysis of the novel compound 3b from commercial sources.

Conclusions

In this paper, we present the use of a modern machine learning tool in the de novo design of candidate molecules for inhibitors of the F13 protein (VP37) of the monkeypox virus. The “deep fragment-based multi-parameter optimization” (DeepFMPO),²⁴ based on actor-critic model for reinforcement learning, was successfully used to generate molecular inspirations from a prototype drug, tecovirimat (TPOXX), and a useful fragment library, also having as filter some properties related to pharmacokinetics (cLog P, MW, PSA). These strategies, together with structure-based approaches (molecular docking and dynamics), led to a molecular analogue (3b) with binding energy and docking geometry comparable to TPOXX, a more simplified chemical structure, and a viable synthetic profile, considering a retrosynthetic analysis also proposed. In this sense, this work contributes in the search for useful and potentially active molecular scaffolds, for subsequent in vitro/in vivo evaluations against monkeypox virus.

Supplementary Information

Supplementary data (link to the source code of the algorithm hosted on GitHub and fragment library used, in SMILES notation) are available free of charge at http://jbcs.sbq.org.br as PDF file.

Acknowledgments

The authors are grateful to National High-Performance Processing Centre at the Universidade Federal do Ceará (CENAPAD-UFC) and the High-Performance Computing Centre at the Universidade Federal do Rio Grande do Norte (NPAD/UFRN) for the computer cluster facility. Molecular graphics and analyses were performed with UCSF Chimera. We would like to thank FACEPE (Fundação de Amparo à Ciência e Tecnologia do Estado de Pernambuco), CNPq (Conselho Nacional de Desenvolvimento Científico e Tecnológico), CAPES (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior) and Univasf (Fundação Universidade Federal do Vale do São Francisco) for their financial support.

References

• ¹ Mitjà, O.; Ogoina, D.; Titanji, B. K.; Galvan, C.; Muyembe, J. J.; Marks, M.; Orkin, C. M.; Lancet 2023, 401, 60. [Link] accessed in January 2025
  » Link
• ² Kaler, J.; Hussain, A.; Flores, G.; Kheiri, S.; Desrosiers, D.; Cureus 2022, 14, e26531. [Crossref]
  » Crossref
• ³ Singhal, T.; Kabra, S. K.; Lodha, R.; Indian J. Pediatr. 2022, 89, 955. [Crossref]
  » Crossref
• ⁴ World Health Organization (WHO), Multi-Country Monkeypox Outbreak: Situation Update, https://www.who.int/emergencies/disease-outbreak-news/item/2022-DON396, accessed in January 2025.
  » https://www.who.int/emergencies/disease-outbreak-news/item/2022-DON396
• ⁵ Abdelaal, A.; Reda, A.; Lashin, B. I.; Katamesh, B. E.; Brakat, A. M.; Al-Manaseer, B. M.; Kaur, S.; Asija, A.; Patel, N. K.; Basnyat, S.; Rabaan, A. A.; Alhumaid, S.; Albayat, H.; Aljeldah, M.; Al Shammari, B. R.; Al-Najjar, A. H.; Al-Jassem, A. K.; Al Shurbaji, S. T.; Al Shahrani, F. S.; Alynbiawi, A.; Alfaraj, Z. H.; Alfaraj, D. H.; Aldawood, A. H.; Sedhai, Y. R.; Mumbo, V.; Rodriguez-Morales, A. J.; Sah, R.; Vaccines 2022, 10, 1419. [Crossref]
  » Crossref
• ⁶ United States Food and Drug Administration (US FDA), Key Facts About Vaccines to Prevent Mpox Disease, https://www.fda.gov/vaccines-blood-biologics/vaccines/key-facts-about-vaccines-prevent-mpox-disease#:~:text=There%20are%20two%20FDA%2Dapproved,the%20prevention%20of%20mpox%20disease, accessed in January 2025.
  » https://www.fda.gov/vaccines-blood-biologics/vaccines/key-facts-about-vaccines-prevent-mpox-disease#:~:text=There%20are%20two%20FDA%2Dapproved,the%20prevention%20of%20mpox%20disease
• ⁷ Rao, A. K.; Petersen, B. W.; Whitehill, F.; Razeq, J. H.; Isaacs, S. N.; Merchlinsky, M. J.; Outcalt, D. C.; Morgan, R. L.; Damon, I.; Sánchez, P. J.; Bell, B. P.; Morb. Mortal. Wkly. Rep. 2022, 71, 734. [Crossref]
  » Crossref
• ⁸ Bunge, E. M.; Hoet, B.; Chen, L.; Lienert, F.; Weidenthaler, H.; Baer, L. R.; Steffen, R.; PLoS Neglected Trop. Dis. 2022, 16, e0010141. [Crossref]
  » Crossref
• ⁹ Siegrist, E. A.; Sassine J.; Clin. Infect. Dis. 2023, 76, 155. [Crossref]
  » Crossref
• ¹⁰ United States Food and Drug Administration (US FDA), Mpox: FDA ‘s Role in Mpox Preparedness and Response, and Information about Mpox (Formerly Referred to as Monkeypox), https://www.fda.gov/emergency-preparedness-and-response/mcm-issues/mpox, accessed in January 2025.
  » https://www.fda.gov/emergency-preparedness-and-response/mcm-issues/mpox
• ¹¹ United States Food and Drug Administration (US FDA), Cytomegalovirus in Transplantation: Developing Drugs to Treat or Prevent Disease, https://www.fda.gov/regulatory-information/search-fda-guidance-documents/cytomegalovirus-transplantation-developing-drugs-treat-or-prevent-disease, accessed in January 2025.
  » https://www.fda.gov/regulatory-information/search-fda-guidance-documents/cytomegalovirus-transplantation-developing-drugs-treat-or-prevent-disease
• ¹² Rao, A. K.; Schrodt, A. C.; Minhaj, F. S.; Waltenburg, M. A.; Goldwasser, S. C.; Yu, Y.; Petersen, B. W.; Hutson, C.; Damon, I. K; Morb. Mortal. Wkly. Rep. 2023, 72, 232. [Crossref]
  » Crossref
• ¹³ Russo, A. T.; Grosenbach, D. W.; Chinsangaram, J.; Honeychurch, K. M.; Long, P. G.; Lovejoy, C.; Maiti, B.; Meara, I.; Hruby, D. E.; Expert Rev. Anti-Infect. Ther. 2021, 19, 331. [Crossref]
  » Crossref
• ¹⁴ Nguyen, B. T.; Marc, A.; Suñer, C.; Marks, M.; Ubals, M.; Hernández-Rodríguez, A.; Melendez, M. A.; Hruby, E. D.; Russo, A. T.; Mentré, F.; Mitjà, O.; Grosenbach, D. W.; Guedj, J.; PLoS Biol. 2023, 21, e3002249. [Crossref]
  » Crossref
• ¹⁵ Smith, T. G.; Gigante, C. M.; Wynn, N. T.; Matheny, A.; Davidson, W.; Yang, Y.; Condori, R. E.; O’Connell, K.; Kovar, L.; Williams, T. L.; Yu, Y. C.; Petersen, B. W.; Baird, N.; Lowe, D.; Li, Y.; Satheshkumar, P. S.; Hutson, C. L.; Emerging Infect. Dis. 2023, 29, 2426. [Crossref]
  » Crossref
• ¹⁶ Van Dijckl, C.; Crozier, I.; Vercauteren, K.; Brosius, I.; Mbala-Kingebeni, P.; Dodd, L.; Bottieau, E.; Mbaya, O. T.; Nussenblatt, V.; Neyts, J.; Liesenborghs, L.; Clin. Microbiol. Infect. 2024, 30, 276. [Link] accessed in January 2025
  » Link
• ¹⁷ Mouchlis, V. D.; Afantitis, A.; Serra, A.; Fratello, M.; Papadiamantis, A. G.; Aidinis, V.; Lynch, I.; Greco, D.; Melagraki, G.; Int. J. Mol. Sci. 2021, 22, 1676. [Crossref]
  » Crossref
• ¹⁸ Janiesch, C.; Zschech, P.; Heinrich, K.; Electron. Mark. 2021, 3, 685. [Crossref]
  » Crossref
• ¹⁹ Krizhevsky, A.; Sutskever, I.; Hinton. G. E.; Commun. ACM 2017, 60, 84. [Crossref]
  » Crossref
• ²⁰ Goodfellow, I.; Bengio, Y.; Courville, A.; Deep Learning; The MIT Press: Cambridge, MA, 2016.
• ²¹ Lecun, Y.; Bottou, L.; Bengio, Y.; Haffner, P.; Proc. IEEE 1998, 86, 2278. [Crossref]
  » Crossref
• ²² Kang, S.; Cho, K.; J. Chem. Inf. Model. 2018, 59, 43.[Crossref]
  » Crossref
• ²³ Sutton, R. S.; Barto, A. G.; Reinforcement Learning: An Introduction, 2^nd ed.; The MIT Press: Cambridge, MA, 2018.
• ²⁴ Ståhl, N.; Falkman, G.; Karlsson, A.; Mathiason, G.; Bostrom, J.; J. Chem. Inf. Model. 2019, 59, 3166. [Crossref]
  » Crossref
• ²⁵ Olivecrona, M.; Blaschke, T.; Engkvist, O.; Chen, H.; J. Cheminf 2017, 9, 48. [Crossref]
  » Crossref
• ²⁶ Sutton, R. S.; McAllester, D.; Singh, S.; Mansour, Y.; Neural Inf. Process. Syst. 1999, 12, 1057. [Link] accessed in January 2025
  » Link
• ²⁷ Bhagat, R. T.; Butle, S. R.; Khobragade, D. S.; Wankhede, S. B.; Prasad, C. C.; Mahure, D. S.; Armarkar, A. V.; J. Pharm. Res. Int. 2021, 33, 46. [Crossref]
  » Crossref
• ²⁸ Morris, C. J.; Corte, D. D.; Mod. Phys. Lett. B 2021, 35, 2130002. [Crossref]
  » Crossref
• ²⁹ Konda, V.; Tsitsiklis, J.; Adv. Neural Inf. Process. Syst. 2000, 12, 1008. [Link] accessed in January 2025
  » Link
• ³⁰ RDKit: Open-Source Cheminformatics, https://www.rdkit.org/, accessed in January 2025.
  » https://www.rdkit.org/
• ³¹ Hochreiter, S.; Schmidhuber, J.; Neural Comput. 1997, 9, 1735. [Crossref]
  » Crossref
• ³² Ivanenkov, Y. A.; Zagribelnyy, B. A; Aladinskiy, V. A.; J. Med. Chem. 2019, 62, 10026. [Crossref]
  » Crossref
• ³³ Kim, S.; Thiessen, P. A.; Bolton, E. E.; Chen, J.; Fu, G.; Gindulyte, A.; Han, L.; He, J.; He, S.; Shoemaker, B. A.; Wang, J.; Yu, B.; Zhang, J.; Bryant, S. H.; Nucleic Acids Res. 2016, 44, D1202. [Crossref]
  » Crossref
• ³⁴ Yang, Z.; Zeng, X.; Zhao, Y.; Chen, R.; Signal Transduction Targeted Ther. 2023, 8, 115. [Crossref]
  » Crossref
• ³⁵ Sayers, E. W.; Cavanaugh, M.; Clark, K.; Pruitt, K. D.; Schoch, C. L.; Sherry, S. T.; Karsch-Mizrachi, I.; Nucleic Acids Res. 2022, 50, D161. [Crossref]
  » Crossref
• ³⁶ Nunes, D. D. S.; Higa, L. M.; Oliveira, R. L.; da Costa, L. C.; Bomfim, L. M.; Gonçalves, C. C.A.; Mariani, D.; Hruby, D. E.; Voloch, C. M.; Castiñeiras, T. M. P. P.; Tanuri, A.; Damaso, C. R.; Mem. Inst. Oswaldo Cruz 2023, 118, e230056. [Crossref]
  » Crossref
• ³⁷ Abraham, M. J.; Murtola, T.; Schulz, R.; Páll, S.; Smith, J. C.; Hess, B.; Lindahl, E.; GROMACS, version 5.0.1; Stockholm University, Stockholm, Sweden, 2024.
• ³⁸ Ali, Y.; Imtiaz, H.; Tahir, M. M.; Gul, F.; Saddozai, U. A. K.; Ur Rehman, A.; Ren, Z. G.; Khattak, S.; Ji, X. Y.; Viruses 2023, 15, 570. [Crossref]
  » Crossref
• ³⁹ Pettersen, E. F.; Goddard, T. D.; Huang, C. C.; Couch, G. S.; Greenblatt, D. M.; Meng, E. C.; Ferrin, T. E.; J. Comput. Chem. 2004, 25, 1605 [Crossref]; UCSF Chimera, version 1.18; UCSF Resource for Biocomputing, Visualization, and Informatics/University of California/National Institutes of Health, CA, USA, 2004. [Link] accessed in January 2025
  » Crossref» Link
• ⁴⁰ Kyte, J.; Doolittle, R. F.; J. Mol. Biol. 1982, 157, 105. [Crossref]
  » Crossref
• ⁴¹ Darden, T.; York, D.; Pedersen, L.; J. Chem. Phys. 1993, 98, 10089. [Crossref]
  » Crossref
• ⁴² Hess, B.; Kutzner, C.; Van Der Spoel, D.; Lindahl, E.; J. Chem. Theory Comput. 2008, 4, 435. [Crossref]
  » Crossref
• ⁴³ Miyamoto, S.; Kollman, P. A.; J. Comput. Chem. 1992, 13, 952. [Crossref]
  » Crossref
• ⁴⁴ Van Gunsteren, W. F.; Berendsen, H. J. C.; Mol. Simul. 1988, 1, 173. [Crossref]
  » Crossref
• ⁴⁵ Parrinello, M.; Rahman, A.; J. Appl. Phys. 1981, 52, 7182. [Crossref]
  » Crossref
• ⁴⁶ Morris, G. M.; Huey, R.; Lindstrom, W.; Sanner, M. F.; Belew, R. K.; Goodsell, D. S.; Olson, A. J.; J. Comput. Chem. 2009, 16, 2785 [Crossref]; AutoDockTools4, version 4.2.6; The Scripps Research Institute, USA, 2009. [Link] accessed in January 2025
  » Crossref» Link
• ⁴⁷ Trott, O.; Olson, A. J.; J. Comput. Chem. 2010, 31, 455 [Crossref]; AutoDock Vina, version 1.2; The Scripps Research Institute, USA, 2011. [Link] accessed in January 2025
  » Crossref» Link
• ⁴⁸ Google Colaboratory, version 24.11.11; Google, USA, 2017. [Link] accessed in January 2025
  » Link
• ⁴⁹ Ganwir, P.; Jaydeokar, S.; Chaturbhuj, G.; Results Chem. 2022, 4, 100293. [Crossref]
  » Crossref

Edited by

• Editor handled this article:
  Paula Homem-de-Mello (Associate)

Publication Dates

• Publication in this collection
  03 Mar 2025
• Date of issue
  2025

History

• Received
  08 Nov 2024
• Accepted
  06 Feb 2025