2  Materials and Methods

2.1  Pdb files
The SARS-CoV-2 spike protein trimer in the prefusion conformation was obtained from pdb file # 6VSB [20]. Hydroxychloroquine (CLQ-OH) is (RS)-2-[{4-[(7-chloroquinolin-4-yl)amino]pentyl}(ethyl)amino]ethanol. CLQ-OH was generated by hydroxylation of chloroquine (CLQ) and validated as previously described [10]. CLQ was retrieved from pdb file # 4V2O (CLQ co-crystallized with saposin B) [21].
Azithromycin (ATM) is (2R,3S,4R,5R,8R,10R,11R,12S,13S,14R)-11-[(2S,3R,4S,6R)-4-(dimethylamino)-3-hydroxy-6-methyloxan-2-yl]oxy-2-ethyl-3,4,10-trihydroxy-13-[(2R,4R,5S,6S)-5-hydroxy-4-methoxy-4,6-dimethyloxan-2-yl]oxy-3,5,6,8,10,12,14-heptamethyl-1-oxa-6-azacyclopentadecan-15-one. The 3D structure of ATM was obtained from pdb file # 5UXD (ATM co-crystallized with macrolide 2′-phosphotransferase) [22].

2.2  Molecular modelling and visualization tools
Molecular modelling studies were performed using Hyperchem (http://www.hyper.com), Deep View/Swiss-Pdb viewer (https://spdbv.vital-it.ch) and Molegro Molecular viewer (http://molexus.io/molegro-molecular-viewer) as described previously [[16], [17], [18], [19],23]. Lennard-Jones parameters and atomic charges of ATM obtained from pdb file # 5UXD were checked with Hyperchem. The molecular modelling protocol consisted of docking, equilibration, and subsequent 50-ns molecular dynamics (MD) simulations with CHARMM force field [16,24] in Hyperchem. The starting point for ATM docking was done by manual positioning on the NTD and receptor-binding domain (RBD) surfaces of SARS-CoV-2 spike protein (chain A) with full consideration of the properties of the drug. In addition, a series of 20 randomly positioned ATM molecules covering the whole NTD surface were analysed. Energy minimization of each system was then performed with the Polak-Ribière conjugate gradient algorithm, with CHARMM force field in Hyperchem, using a maximum of 3 × 105 steps, and a root-mean-square (RMS) gradient of 0.01 kcal. Å−1.mol−1 as the convergence condition. According to the validation ligand-binding protocol [25], the optimized docked structures were used as the initial structures for MD simulations with the following parameters: heat time 5 ps, run time 10 ps, step size 1 fs, starting temperature 0 K, simulation temperature 300 K, temperature step 30 K, bath relaxation time step size 0.1 ps. Each complex was submitted to 50-ns long MD trajectories, replicated three times, with convergent trajectories obtained. Frames were saved at 10 ns intervals for subsequent analysis. The final complex obtained with chain A was reintroduced in the trimeric spike structure (pdb file # 6VSB) with Molegro Molecular viewer and with Deep View Swiss-Pdb viewer. Similar results were obtained with both programs. The energies of interaction were extracted from the Ligand Energy Inspector function of Molegro Molecular viewer. Molecular volumes were calculated by the QSAR function of Hyperchem.

2.3  Ganglioside GM1
The initial coordinates of ganglioside GM1 were obtained from CHARMM-GUI Glycolipid Modeler [24] (http://www.charmmgui.org/?doc=input/glycolipid), which uses the internal coordinate information of common glycosidic torsion angle values, orientates the ganglioside perpendicular to the membrane, and performs Langevin dynamics with a cylindrical restraint potential to keep the whole GM1 molecule cylindrical, particularly the membrane-embedded ceramide part. In the next step, the saccharide part of the ganglioside was included in a periodic box solvated with 1128 water molecules (dimensions: 34848 Å3 with x = 33 Å, y = 32 Å, z = 33 Å). The system was energy-minimized 6 times switching alternatively between runs using steepest descent gradients or Polak-Ribière conjugate gradients until convergence to machine precision [16]. The dimer of ganglioside GM1 interacting with 4 CLQ-OH molecules was obtained by MD simulations of a previously described model [10]. To mimic GM1 gangliosides in a typical lipid raft membrane domain, two GM1 molecules were merged with eight cholesterol and two sphingomyelin lipids. The whole system was optimized, merged with SARS-CoV-2 spike protein (chain A) and submitted to MD simulations with the same conditions as those used for the ATM-spike protein complex (50-ns run in triplicate). For comparison, similar MD simulations were performed on an isolated GM1 dimer in the CHARMM-GUI membrane-compatible topology but without surrounding lipids.