There are over 150 crystal structures of the SARS-CoV-2 main protease (Mpro) available at the Protein Data Bank,and few of these structures have been already used as input for molecular docking and virtual screening efforts, aiming at identifying drug inhibitors that can be used to treat COVID-19. Unfortunately, these efforts have not yet produced effective drug inhibitors for this protein. One of the main limitations of computational efforts to identify inhibitors for Mpro and other SARS-CoV-2 proteins is not to account for receptor flexibility. Receptor flexibility is a known challenge in molecular docking (Antunes et al., 2015) and there is now evidence it might be particularly important for SARS-CoV-2 proteins (Jin et al., 2020).

To address this challenge, we are running molecular dynamics simulations of SARS-CoV-2 proteins, and creating ensembles of representative conformations of these proteins in solution. In the case of Mpro, in addition to ensembles produced by simulations using two different force fields (e.g., charmm and gromos), we have also created an ensemble of experimentally-determined structures which captures most of the subtle variation observed across available crystal structures of Mpro. These different ensembles of receptor conformations were then made available for ensemble docking with DINC (Devaurs et al., 2019), through the DINC-COVID webserver (Fig. 1). In summary, users can easily upload their own ligands of interest, select one of the available ensembles, and obtain the best predicted binding modes, ranked by three different scoring functions (e.g., Vina, Vinardo and AutoDock4).

Overview of the DINC-COVID webserver workflow. The top left figure is a schematic representation of a given SARS-CoV-2 dimeric receptor, with one protomer depicted in blue, and another in red. Note that in this case there are two identical binding sites (i.e., BS1 and BS2), formed by the combination of both protomers. 1. Ensemble generation: Three different ensembles of the dimeric receptor are pre-computed (e.g., from available crystal structures or molecular dynamics simulations) and made available through the DINC-COVID webserver. Different shades represent different receptor conformations within each ensemble. 2. Input selection: The user can select one of the available ensembles, and upload the ligand of interest (e.g., drug or peptide). 3.. Ensemble docking: The parallelized meta-docking approach DINC is used to sample alternative binding modes using each of the receptor’s conformations included in the ensemble. 4. Scoring and Ranking: All generated binding modes are rescored and ranked according to three different scoring functions (e.g., Vina, Vinardo and AutoDock4). 5. Output complexes: A number of top scoring complexes are returned to the user, reflecting the flexibility of both the ligand and the receptor.

The first SARS-CoV-2 protein added to the webserver was MPro, but ensembles of other valuable targets will be soon made available to users. For each SARS-CoV-2 protein of interest, a total of 10 independent 200 ns molecular dynamics simulation is executed with the GROMACS 19 package. One half of the simulations are performed with the CHARMM36 force field, and the other half with the GROMOS53a6 force field. This is an effort to avoid limitations on conformational sampling due to force field bias. CHARMM and GROMOS force fields use distinct representations for hydrogen atoms (e.g., all-atom and united-atom, respectively), which can have an impact on the results of MD simulations. Therefore, we will continue to provide ensembles produced with both of these popular force fields, to address different user needs and preferences.

Conformations derived from simulations using the same force field are taken together and used to create a representative ensemble, in a process that relies on a combination of dimensionality reduction and clustering. When crystal structures are available, as in the case of MPro, they are used to create a separate (experimentally-based) ensemble of conformations. Note that each of these ensembles also accounts for different levels of receptor flexibility. For instance, the crystal-based ensemble accounts for the least amount of flexibility, and might be better suited to identify binding modes that are similar to those observed in available crystal structures. On the other hand, the gromos-based ensemble accounts for the largest range of receptor flexibility sampled in our simulations, being best suited to identify potentially new binding modes for ligands of interest.


DINC is a meta-docking algorithm, in the sense that it relies on a standard docking tool, currently AutoDock Vina, to perform the sampling and scoring at each docking round (Devaurs et al., 2019). Here, the DINC algorithm was adapted to performed parallelized ensemble docking for SARS-CoV-2 proteins (Fig. 2).
The DINC-ensemble algorithm. DINC uses parallelization to speed up both the sampling and the scoring of protein-ligand binding modes. Once a ligand strcuture and an ensemble of receptor conformations are selected (input), the algorithm will trigger multiple parallel docking jobs with the fast docking method Vina. Each independent job starts with an alternative (randomized) conformation of the ligand and one of the receptor conformations from the selected ensemble. This first batch of docking jobs produces a diversity of binding modes for the ligand against a single receptor conformation. While there are other receptor conformations available in the ensenble, the process is repeated. Once all dockings are completed, the rescoring phase starts. In this phase, all binding modes predicted for each receptor conformation are rescored with an alternative scoring function. Currently, the final results include rankings provided by three popular scoring functions: Vina, Vinardo and AD4. A number of top scoring conformations, defined by the user, is provided as output at the end of the process. These results include both alternative conformations of the ligand, as well as alternative conformations of the receptor (Fig. 1).