QSAR screening
The quality of the developed machine learning and deep learning based QSAR classification models was assessed using the receiver operating characteristic (ROC) curve, which plots the true positive rate against the false positive rate [14]. Notably, all the classification models achieved an area under the curve (AUC) values higher than 0.80 as shown in Figure 1. The performance of the QSAR models was further evaluated using various performance metrics, as presented in Table 1. The convolutional neural network (CNN) model performed exceptionally well on the testing and validation sets and was selected to predict the activity of Ambinter natural compounds library. At this stage, 1,193 compounds out of 6,151 were predicted as active against A. baumannii and were selected for the structure-based virtual screening study.
Docking screens of natural products
The quality assessment of the AlphaFold model of OmpW, according to Ramachandran plot, shows 92.2% of residues of are in most favourable regions, 7.2% in allowed regions, 0.6% in generously disallowed regions and 0.0% in disallowed regions. Validation of the OmpW structure using PROSA-web shows Z-score value of −4.95 which is within the range of scores typically found for native proteins of similar size (Figure 2A and 2B). The predicted active compounds were subjected to molecular docking screens, and their binding affinities were ranked accordingly. Specifically, we observed that the highest-ranking compounds exhibit binding scores ranging from -7.0 to -7.8 kcal/mol and belong to curcuminoids as shown in Figure 2C. The amino acids involved in the ligand binding are presented in Table 2.
Docking poses of the highest-ranking compounds are displayed in Figure 3. In brief, the structural analysis of the docked compounds reveals consistent hydrogen bond formation between the hydroxyl (-OH) group of the phenyl ring in curcuminoids and the amino acid residue GLN-23. Furthermore, we detected additional hydrogen bond interactions implicating key residues, namely ASN-104, THR-109, and LYS-195, situated within the periplasmic site of OmpW. Additionally, our analysis reveals multiple instances of hydrophobic interactions, with notable involvement of amino acid residues PHE-59, HIS-101, ASN-144, and GLN-146.
ADME evaluation
A significant proportion, approximately 40%, of drug candidates fail during clinical trials primarily due to inadequate ADME properties [15]. In silico ADME prediction offers a rapid method to assess the drug-likeness of a compound by calculating its physicochemical properties. This approach substantially reduces the time and resources required during the overall drug development process. In this study, SwissADME (http://www.swissadme.ch/) was employed to compute various pharmacokinetic properties of the highest-scoring compounds to evaluate their drug-likeness and suitability for further experimental studies [16]. ADME properties for the selected compounds are shown in Table 3. The results reveal that all the compounds possess a good lipophilicity in accordance with Lipinski’s rule of five, moreover water solubility values were found to be in the recommended range for most drugs. Intestinal absorption was found to be high in all the compounds. Out of the top ten compounds tested for blood-brain barrier (BBB) permeability, only five were found to be unable to penetrate the BBB. This is a crucial finding, as antibacterial compounds should not exert their effects on the central nervous system (CNS). None of the compounds were found to act as a P-gp substrate, thus their bioavailability is not impacted by this protein. Finally, PAINS test has revealed four compounds presenting one alert in their structure due to the presence of the catechol group which can result in non-specific binding with various target proteins.
Molecular dynamics simulations and binding free energy
In the molecular docking study, the protein structure was treated as rigid. To gain deeper insights into the protein-ligand interactions, molecular dynamics simulations were performed on the docked complexes in a water environment for 100 ns. The root-mean square deviation (RMSD) was measured relative to the OmpW structure bound to the selected candidates. Figure 4A illustrates the protein RMSD values for the top four complexes, showing a consistently stable RMSD of 0.3 nm during most of the simulation, except for Amb22174074, which displayed higher fluctuations exceeding 0.3 nm in the last 20 ns. The analysis of the ligand RMSD showed values between 0.1 and 0.25 nm for most ligands, suggesting minor conformational changes during the simulation. However, the ligand Amb8399162 deviated from this trend, with an RMSD of 0.35 nm, suggesting a more significant conformational change (Figure 4B). In Figure 4C, the graph illustrates the variations observed in each amino acid. Notably, the N-terminal region exhibited the highest fluctuations, which is a common characteristic. For all other residues, minor fluctuations of approximately 0.1 nm were observed, except for Amb8399162, which displayed fluctuations higher than 0.2 nm in certain regions of the periplasm. Finally, hydrogen bonds within a proximity of 0.35 nm were documented. Figure 4D depicts the hydrogen bonds observed at 100 ns, with Amb2698241 forming four hydrogen bonds, highlighting its stable and consistent binding to the protein. The average free binding energy of the selected complexes was determined using the g_mmpbsa package [17].
The binding energy was computed by combining the scores of Van der Waals energy, electrostatic energy, polar solvation, and SASA energy as presented in Table 4. The highest binding energy was observed in Amb2698241 (-45.23 kJ/mol) suggesting a strong binding to the target protein.
Antibacterial activity
The best compound exhibiting the lowest docking score as well as favourable ADME properties was demethoxycurcumin (Amb2698241). The MIC was then assessed using microdilution assays against different reference A. baumannii ATCC 17978 strain, its isogenic mutant deficient in OmpW, and colistin-resistant A. baumannii clinical isolates. Demethoxycurcumin inhibited bacterial growth at a concentration of 64 µg/mL for all the studied strains (Table 4).
Colistin potentiation is critical for safeguarding this last resort antibiotic as it is often our only treatment option against highly resistant Gram-negative pathogens. We examined whether demethoxycurcumin can sensitize colistin-resistant clinical strain CR17. Chequerboard assay showed that demethoxycurcumin at >=1 mg/L demonstrated synergy with colistin against CR17 strain. Demethoxycurcumin >= 8 mg/L in combination with colistin increased the activity of colistin against CR17 strain, with a fractional inhibitory concentration index (FICI) of <0.2 (Figure 5A). In addition, the combination between 16 mg/L demethoxycurcumin and 1 mg/L colistin exhibited a synergistic effect during 2 and 4 h, reducing significantly the bacterial growth compared with colistin demethoxycurcumin alone (Figure 5B).
Using bacterial growth assays, we examined the antibacterial activity of demethoxycurcumin against ATCC 17978 and ΔOmpW strains. Figure 5C reveals that A. baumannii ATCC 17978exhibits rapid growth, reaching 0.5 OD within the first 4 h. However, a noticeable disparity in growth is observed between the control sample and the samples treated with demethoxycurcumin, particularly at higher compound concentrations (2xMIC and 4xMIC). A similar trend of growth inhibition is observed in the ΔOmpW strain, although it demonstrates a higher OD value compared to A. baumannii ATCC 17978 in presence of demethoxycurcumin treatment.This disparity in growth can be attributed to the resistance of the mutant strain to the compound, as the absence of OmpW may hinder the compound's ability to exert its effect, as indicated by the findings of the molecular docking study.
In addition, and to evaluate the effect of demethoxycurcuminon A. baumannii interaction with host cells, we studied the adherence of ATCC 17978 and ΔOmpW strains to HeLa cells for 2 h in the presence of demethoxycurcumin. Treatment with demethoxycurcumin at 1xMIC reduced the adherence of ATCC 17978 and ΔOmpW strains to HeLa cells by 36% and 16%, respectively (Figure 5D).