2.1. Design and Application of the Three Filters Used for the MNP Database Screening
The first filter used for the identification of FABP4 ligands was a statistical (based on 2D and 3D descriptors) one, as already used successfully by us [
31]. We selected 2922 molecules among the MNP database by a statistical/2D descriptors filter using DataWarrior software [
32]. The appropriate range of values to be considered for each chosen descriptor was obtained by analyzing the most potent and selective compounds present in a recently published dataset of FABP4 ligands, giving a total of 120 entities [
28,
29]. Therefore, the ranges for molecular weight (MW, 224/501), cLogP (−0.84/6.1), cLogS (−8.77/−2.33), H-bond-acceptors (HBA, 2/7), H-bond-donors (HBD,0/2), total surface area (TSA, 170/400), polar surface area (PSA, 40/125), and relative polar surface area (RPSA, 0.07/0.4) (
Figure 2), belonging to the 120 potent and selective FABP4 inhibitors, were associated with each descriptor and applied to the dataset of 14,492 MNP molecules to give 2922 marine filtered compounds.
These skimmed molecules were then subjected to a second filtration using a mixed ligand- and structure-based approach. Firstly, the 3D molecular structures of the 2922 marine compounds were aligned to our previous published 3D-QSAR model for the FABP4 protein, and the compounds were then evaluated, as previously reported, employing Forge software (v10.4.2, Cresset, New Cambridge House, United Kingdom) [
33]. Over the whole dataset of the first filtered marine natural products, 1854 molecules resulted in an excellent or good description by the model. This means that most of the features in the evaluated molecules were well described by the training set of the 3D-QSAR model, and the predicted activity could be considered reliable. Among these compounds, 198 molecules resulted in a predicted pIC
50 activity between 6.0 and 7.6. The 3D molecular structures of the 2922 marine compounds were then passed to the structure-based approach, adapting the docking procedure already reported for the identification of FABP4 inhibitors [
34,
35]. The AutoDock software (v. 4.2.6, Molecular Graphics Lab at The Scripps Research Institute, La Jolla, CA, USA) [
36] was used for all docking studies. The validation of the adopted docking procedure was assessed by using linear regression analysis upon a benchmark data set of 34 known FABP4 inhibitors (
Table S1, Figure S1).
All of the generated binding poses were manually inspected in order to ensure correct positioning within the binding pocket with respect to the interactions of ligand moieties with the amino acid residues relevant for the catalytic activity. The residues Phe19, Met20, Ala33, Pro38, Lys58, Phe57, Ala75, Glu72, Arg106, and Arg126 play an important role in the interactions of FABP4 with inhibitors [
37,
38]. Initially, these residues were used as a filter to discard the incorrect poses derived from the docking. In addition, molecular dynamics (MD) simulation studies of three of the most promising compounds (
Table 1, 5339, 14123, and 13575) were conducted to verify the effectiveness of the poses selected. In particular, for each selected ligand, we performed three 20 ns MD simulations using three different poses, named P1–P3, where the P1 ones are those that visually satisfy the aforementioned reported key interactions, whereas the others two (P2 and P3) lack some of them. The results reported in
Figure 3 as root-mean-square deviation (RMSD) fluctuations of the ligand coordinates clearly highlight the two unfavorable poses, i.e., those with higher RMSD fluctuations [
39]. These present even less persistence of the hydrogen bond interactions with the key residues. On the contrary, the best poses (P1) show very low RMSD fluctuation.
In
Figure S2, it is possible to note that the compound 5339 forms a hydrogen bond with the residue Arg106,
π-
π stack interactions with the residue Pro38, and hydrophobic interactions with the residues Phe16, Ala 33, Ala36, Phe57, Ile62, and Ala75 with aromatic regions of the ligand. Compound 13575 establishes hydrogen bonds with the residues Tyr19 and Ala75, and the hydrophobic region of the molecule establishes hydrophobic interactions with the residues Phe16, Ala33, Pro38, Phe57, and Ile 62 (
Figure S3). Finally, compound 14123 shows a hydrogen bond with the residue Arg106, while other hydrophobic interactions with Phe16, Met20, Ala 33, Pro38, Ala40, Ala, Ile 62, and Ala75 reinforce the bond with the hydrophobic region of FABP4 (
Figure S4).
Re-docking experiments conducted after 20 ns of MD simulation of P1 poses gave a calculated pKi value of 7.92, 7.96, and 7.96 for the ligands 5339, 13575, and 14123, respectively, which are slightly better than those calculated on the crystallographic laying of the FABP4 receptor, according to a correct arrangement of the ligand in the catalytic pocket.
The 3D-QSAR and docking evaluation results are reported in
Table S2.
2.2. Merged Ligand- and Structure-Based Filters
The results derived from the ligand-based calculation (i.e., 3D-QSAR evaluation) and the structure-based calculation (i.e., docking calculations) were then merged with the aim to create a final filter. For this purpose, the best (lowest calculated IC
50 or
Ki) 2% and 5% of the molecules obtained from each of the two approaches were retrieved, and those simultaneously present in both filters were selected (
Table 1). The 2% filter resulted in only one molecule (5339), whereas the 5% one returned six other molecules (14123, 13575, 7846, 3164, 2076 and 1534).
All the binding poses retrieved from the molecular docking calculations of the seven compounds reported in
Table 1, which showed the classical interactions of the most common FABP4 ligands, are reported in the
Supplementary Material (Figures S2–S8). In particular, the best-docked pose of compound 5339, chosen as representative, superposed with the co-crystallized structure of the BMS309403 FABP4 inhibitor in the binding pocket of the enzyme shows that the two compounds are partially overlapped and occupy almost the totality of the catalytic orthosteric site (
Figure 4).
Among the seven filtered best potential inhibitors, molecule 5339 (indole alkaloid) has been reported as an inhibitor of the Ca
2+-ATPase of the sarcoendoplasmic reticulum (SERCA) [
40]. Compound 14123 (steroid) was identified as cytotoxic and an anti-tumor agent [
41]. Compound 13575 (diterpene) was tested for its cytotoxicity against several tumor cells, but it lacked any activity [
42]. Compound 7846 is a cembrane diterpenoid; similar compounds were reported to exert growth-inhibition effects toward tumor cells [
43]. Compound 3164 (pentacyclic hydroquinone) was reported as cytotoxic by acting on DNA topoisomerase I [
44]. Compound 2076 (alkaloid) was reported as cytotoxic against several tumor cell lines [
45]. Compound 1534 (sesquiterpene) has anti-inflammatory activity by acting as a phospholipase A2 (PLA2) inhibitor [
46]. Interestingly, PLA2 catalyzes the hydrolysis of phospholipids to produce free fatty acids. The fatty acid is the substrate for the biosynthesis of eicosanoids, which are known to mediate inflammation. Based on this mode of action, compounds that inhibit PLA2 activity have been targeted as potential therapeutic agents in the treatment of inflammation. The association between PLA2 and FABP4 in the regulation of inflammatory responses has already been proven [
47], and dual inhibition of such proteins would be advantageous in inflammation treatment. Moreover, compound 1534 is a sesquiterpene, and this class of natural products, together with steroids, diterpenes, diterpenoids, quinones, and alkaloids, has already been identified as a candidate for the inhibition of FABP4 [
48].
2.3. ADMET Properties
As the interaction of an inhibitor with an enzyme cannot guarantee its suitability as a drug, to further strengthen the results of 3D-QSAR and docking studies, we also performed in silico ADMET studies on the seven molecules reported in
Table 1. The ability to reach targets in bioactive form was assessed using the SwissADME (
http://swissadme.ch) and pkCSM (
http://biosig.unimelb.edu.au/pkcsm/) web platforms. Importantly, the technologies implemented in these platforms are able to predict, with a fair degree of certainty, the false-positive results commonly observed in biochemical assays of small molecules [
49].
The oral availability of our proposed bioactive compounds is shown in the bioavailability radar plots (
Figure 5), which provide a graphical snapshot of the drug-likeness parameters of the investigated molecule. Notably, five compounds (5339, 13575, 7846, 3164, and 1534) have been predicted as orally bioavailable, whereas compounds 14123 and 2076 present only one off-shoot relative to the lipophilicity (LIPO) and unsaturation (INSATU) vertexes, respectively, leading to suboptimal physicochemical properties for their oral bioavailability.
In addition to the Lipinski rule of five [
51], another four drug-likeness rules named Ghose [
52], Egan [
53], Veber [
54], and Muegee [
55], were contemporarily satisfied by six compounds with the exception of molecule 14123 (
Table 2). Instead, the stringent lead-like criteria of Teague [
56] were passed by compounds 5339 and 2076. As lead-likeness tests are intended to provide leads with high affinity in high-throughput screens that allow for the discovery and exploitation of additional interactions in the lead-optimization phase, molecules 5339 and 2076 are excellent candidates for investigation based on the scaffold hopping approach.
Finally, the outcome of the pan assay interference structures (PAINS) model [
57], conceived to exclude small molecules that are likely to show false positives in biological assays, post only one alert for compound 1534, concerning the presence of a quinone moiety.
Human gastrointestinal absorption (HIA) and blood–brain barrier penetration (BBB), relative to the absorption and distribution parameters, respectively, have been graphically represented by the extended and renewed version of the Edan–Egg model, named the Brain or IntestinaL EstimateD (BOILED) permeation predictive model (BOILED-Egg). The visual analysis of
Figure 6 highlights that all investigated molecules, with the exception of 3164, were predicted to be passively absorbed by the gastrointestinal tract, and three of them, 5339, 14123, and 2076, passively permeate through the BBB, the first with the aid of the P-glycoprotein and the other two without it. These data are reflected in the values shown in
Table 3.
Regarding the absorption parameters, compounds 5339, 14123, 13575, and 2076 present a promising oral availability, due to the optimal Caco-2 cell permeability and HIA (>0.9 and >90%, respectively,
Table 3), and skin permeability (log
Kp < −2.5,
Table 3).
The volume of distribution (VDss) and unbound fraction are two of the most important pharmacokinetic drug parameters. Values of the VDss > 0.45 indicate that the drug will be distributed in tissue, whereas values < −0.15 indicate that the drug will be distributed in plasma. So, VDss describes the extent of drug distribution, and the unbound fraction describes the portion of free drug in plasma that may extravasate. Except for compounds 3164, the other ones showed intermediate values of VDss, and should have an adequate plasma distribution profile, with a fraction of the unbound drug between 0 and 0.157. These values indicate that the molecules can be well distributed and present a significant unbound fraction in the plasma, thus becoming available to interact with the pharmacological target. Only compound 3164 is entirely unable to penetrate the central nervous system (CNS).
The predicted values of the total clearance (
Table 3), which measure the efficiency of the body in eliminating a drug, indicate that all compounds have a good renal elimination (1.5–8.4 mL/min/kg) and are not substrates of the renal organic cation transporter 2 (OCT2), with the exception of compound 13575. Finally, compounds 2076 and 14123 did not pass the AMES and Minnow toxicity tests, respectively, whereas all others did not present any particular toxicity problems.
The overall lecture of
Table 3 highlights that compounds 5339, 13575, and 1534 could be excellent candidates as drugs, or could lead to further studies and manipulations.