The DockThor-VS platform

The DockThor-VS web server is based on a suite of programs mainly developed by our research group GMMSB. The docking engine is the DockThor program^25,26, a non-covalent docking program that utilizes a topology file for the ligand and cofactors (.top) and a specific input file for the protein (.in) containing the atom types and partial charges from the MMFF94S49 force field. The .top file of the ligand is generated by the MMFFLigand program, which utilises the facilities of the OpenBabel chemical toolbox¹¹⁶ for deriving partial charges and atom types with the MMFF94S force field¹¹⁷, identification of the rotatable bonds and the terminal hydroxyl groups, and calculating the properties necessary for computing the intramolecular interactions. In the MMFFLigand, all hydrogen atoms are considered explicitly. The PdbThorBox program is used to set the protein atom types, the partial charges from the MMFF94S force field considering the nonpolar atoms implicitly, and reconstruct missing residue side-chain atoms. Thus, in the DockThor-VS platform, both protein, cofactor and ligands are treated with the same force field in the docking experiment. All the molecular force field parameterisations are performed automatically by the programs cited without the users’ need for intervention. However, some preparation steps of the input molecules can be done interactively in the web server, such as changing the protonation state of some amino acid residues, adding hydrogen atoms and freezing rotatable bonds of the ligand (only available for the upload of a single ligand).

The search space is represented as a grid box where the potentials are stored at the grid points, significantly reducing the computational cost. The user can interactively set the grid box’s configuration in the web server through the parameters: center of coordinates, size of the grid and discretization (i.e., the spacing between the grid points). The initial population is randomly generated within the grid box using random values for the rotational, translational, and conformational degrees of freedom of the ligand. For each SARS-CoV-2 therapeutic target, DockThor-VS provides a recommended set of parameters for the grid box (i.e., center and grid sizes) that the user can use or modify according to the objectives of his docking experiment (Table ^S5).

The DockThor docking program was specially developed to deal with highly flexible ligands, and it is very suitable to dock highly flexible peptides²⁶. The program uses a phenotypic crowding-based multiple solution steady-state genetic algorithm as the search method²⁵. In this strategy, the parental replacement method follows the Dynamic Modified Restricted Tournament Selection (DMRTS), which provides a better exploration of the energy hypersurface and allows identifying multiple minima solutions in a single run, preserving the population diversity of the generated structures. The default parameters of this algorithm (named Standard) are set in the web server as follows: (i) 24 docking runs, (ii) 1.000.000 evaluations per docking run, (iii) population of 750 individuals, (iv) maximum of 20 cluster leaders on each docking run. For virtual screening experiments, we also provide an alternative set of parameters to accelerate the docking experiment without significantly losing accuracy (named Virtual Screening): (i) 12 docking runs, (ii) 500.000 evaluations per docking run, (iii) population of 750 individuals, (iv) maximum of 20 cluster leaders on each docking run. The docking experiments are performed on CPU nodes of the SDumont supercomputer, each one containing two processors Intel Xeon E5-2695v2 Ivy Bridge (12c @2,4 GHz) and 64 Gb of RAM memory. We validated the docking experiments through the redocking of the non-covalent ligands present in the complexes 6W63 (Mpro) and 6WXC (Nendo-U) using the standard configuration, successfully predicting the co-crystallized conformation of each complex (Table ^S5). The experimentally observed interactions between the uracil ring from tipiracil and Nendo-U binding site were correctly predicted for the top-energy solution of tipiracil. In contrast, the iminopyrrolidin was predicted on an inverted conformation to optimize the interactions with the protein. In the crystallographic structure, this moiety is exposed to the solvent and has insufficient electronic density data.

The scoring function used to score the docked poses of the same ligand is based on the sum of the following terms from the MMFF94S force field and is named “Total Energy (Etotal)”: (i) intermolecular interaction energy calculated as the sum of the van der Waals (buffering constant δ = 0.35) and electrostatic potentials between the protein–ligand atom pairs, (ii) intramolecular interaction energy calculated as the sum of the van der Waals and electrostatic potentials between the 1–4 atom pairs, and (iii) torsional term of the ligand. All docking poses generated during the docking step are then clustered by the in-house tool DTStatistics. The top energy-poses of each cluster are selected as representatives and made available to the users. In the results analysis section, users can choose the clustering criteria and the maximum number of cluster representatives available on the web server. The affinity prediction and ranking of distinct ligands are performed with the linear model and untailored for specific protein classes, DockTScore_GenLin scoring function. The DockTScore is a set of empirical scoring functions recently developed by our research group¹¹⁸. These scoring functions take into account the following important terms for protein–ligand binding: (i) intermolecular interactions terms; (ii) a torsional entropy term that penalizes the "frozen" rotatable bonds due to binding, (iii) a protein–ligand lipophilic interaction term, (iv) a polar solvation term which accounts for the loss of polar interactions of the charged groups of both protein and ligand after binding and (v) a favorable nonpolar solvation term that is proportional only to the solvent-accessible surface. The scoring function was trained and tested in a large set (> 2900) of high-quality 3D structures associated with diverse physicochemical profiles (including an extensive range of binding affinities, MWs and number of rotatable bonds) and relevant therapeutic targets for drug design. In addition to the current scoring function, shortly we will also provide the affinity prediction using models developed for specific target classes such as proteases and protein–protein interactions (PPIs) and trained with sophisticated machine-learning algorithms.

The DockThor program is freely available as a web server at www.dockthor.lncc.br, which provides to the user the main steps for protein and ligand preparation with PdbThorBox and MMFFLigand, and the analyses of the results using DTStatistics. The visualization of protein, cofactors and compounds, the grid location superposed with the protein and the docking results are generated with NGL, a WebGL-based library for molecular visualization¹¹⁹. Guest users are allowed to submit VS experiments with up to 200 compounds, whereas registered users with approved projects can submit up to 5,000 compounds per job. All users have access to the COVID-19 repurposing and Macrocycle datasets, but only registered users can access the complete e-Drug3D repurposing dataset. The web server utilises the computational facilities of the Brazilian high-performance platform (SINAPAD, https://www.lncc.br/sinapad/) and the supercomputer SDumont (https://sdumont.lncc.br/).