Definition of Fatty Acid Transport Protein-2 (FATP2) Structure Facilitates Identification of Small Molecule Inhibitors for the Treatment of Diabetic Complications

Abstract

Diabetes is a major public health problem due to morbidity and mortality associated with end organ complications. Uptake of fatty acids by Fatty Acid Transport Protein-2 (FATP2) contributes to hyperglycemia, diabetic kidney and liver disease pathogenesis. Because FATP2 structure is unknown, a homology model was constructed, validated by AlphaFold2 prediction and site-directed mutagenesis, and then used to conduct a virtual drug discovery screen. In silico similarity searches to two low-micromolar IC₅₀ FATP2 inhibitors, followed by docking and pharmacokinetics predictions, narrowed a diverse 800,000 compound library to 23 hits. These candidates were further evaluated for inhibition of FATP2-dependent fatty acid uptake and apoptosis in cells. Two compounds demonstrated nanomolar IC₅₀, and were further characterized by molecular dynamic simulations. The results highlight the feasibility of combining a homology model with in silico and in vitro screening, to economically identify high affinity inhibitors of FATP2, as potential treatment for diabetes and its complications.

1. Introduction

The US and worldwide prevalence of diabetes was 37.3 million and nearly 500 million, respectively, in 2019 (https://diabetes.org/about-us/statistics/about-diabetes). Diabetes is a major public health problem due to morbidity and mortality associated with end organ complications. Diabetic kidney disease (DKD) affects as many as one in three people with diabetes. DKD is by far the most common cause of kidney failure [1], which is only treatable with dialysis or kidney transplantation. Despite effective antihypertensive and glucose-lowering regimens for DKD, there is still residual DKD progression [2–4], emphasizing the need for new therapies. Cardinal features of DKD pathology include glomerular lesions, but downstream interstitial fibrosis and tubular atrophy are superior predictors of DKD progression [5, 6]. Multiple mechanisms of tubular atrophy pathogenesis have been proposed, but no specific therapies have been developed.

Long chain (C14-C20) fatty acids (FA) circulate in a non-esterified form bound to albumin, or as glycerol esters (triglycerides). Both complexes are poorly filtered by the kidney glomerulus due to their size. In DKD, injured glomeruli permit filtration of albumin-bound long chain FA, which are then reabsorbed by apical Fatty Acid Transport Protein-2 (FATP2), causing lipotoxicity [7–10]. Global FATP2 deletion in mouse models of progressive DKD normalized kidney histology and function, and markedly reduced plasma glucose [11]. In vivo administration of FATP2 shRNAs also reduced hepatosteatosis [12], which is a common complication of diabetes that can lead to cirrhosis. These data substantiate FATP2 as a therapeutic target for diabetic end organ complications. FATP2 inhibitors have been identified by functional screens [13], but lead compounds Lipofermata and Grassofermata, are not yet sufficient for clinical trials due to prohibitively high IC₅₀ values and lack of pharmacokinetic data.

FATPs form a family of six proteins that are conserved across species [14–16]. FATP2 is a 70 kDa integral membrane protein with extracellular N-terminal and cytosolic C-terminal domains [15, 17], which is most prominently expressed in kidney and liver [11, 14]. FATP2 mediates both long-chain FA transport and activation via intrinsic acyl coenzyme-A synthetase (ACS) activity through vectorial acylation [14, 18–20]. The yeast homologue Fat1p is most closely related to FATP1 and has 39% sequence identity to FATP2. In mutagenesis studies, 16 essential Fat1p amino acids correlated with FATP2 [15], and several of these residues have been implicated in long-chain FA transport and ACS activity [15]. Discovery of more potent FATP2 inhibitors would be facilitated by a three-dimensional target structure, which has not been experimentally determined for any FATP family members. Our goal was to use in silico techniques to define the FATP2 structure, which could then be employed as a guide for FATP2 inhibitor drug discovery, using in silico small molecule library screening and functional assays.

2. Materials and methods

2.1. Generation of a three-dimensional hFATP2 model.

Homology modeling, which employs known structures with similar sequence identity to the target protein, was used to construct the three-dimensional structure of the human FATP2 model. A search of PDB databases revealed that the closest homologue to hFATP2 with known structure is a Mycobacterium tuberculosis protein (PDB id-3R44) with intrinsic ACS activity, similar to the FATP2 cytosolic domain.

TMHMM (Expasy-TMHMM Server v. 2.0) is a hydrophobicity and membrane protein topology prediction method based on a hidden Markov model. It predicts transmembrane helices and discriminates between soluble and membrane proteins with >99% specificity and sensitivity (http://www.cbs.dtu.dk/services/TMHMM-2.0/). The program predicted that the 52 N-terminal FATP2 residues, which contains putative transmembrane helices that facilitate ACS binding and vectorial acylation [13], are missing in the M.tb 3R44 protein. As there are no homologous structures containing this membrane region, we excluded these residues in our model. We reasoned that the cytoplasmic domain, containing the critical catalytic site residues, would be sufficient for the docking studies, and ligands binding to the catalytic site would inhibit transport. Human FATP2 3D models were constructed using the programs Swiss modeller, I-TASSER (Iterative Threading ASSEmbly Refinement), Swiss modeller (https://swissmodel.expasy.org/) and Modeller 4.0 (https://salilab.org/modeller/). Unlike I-TASSER, Swiss Modeller and Modeller 4.0 do not employ threading-based routines. Although all three programs provided very similar three-dimensional homology models of hFATP2, further examination with preliminary docking studies using the ligands Lipofermata and Grassofermata [13] revealed that a larger number of the essential yeast Fat1p residues required for fatty acid transport [15], interacted with the ligand using I-TASSER. The I-TASSER-generated FATP2 model was superimposed upon the M.tb 3R44 and AlphaFold2 [21] structures, and RMSD for Cα main chain backbone and all-atom structures, respectively, were calculated. The model was energy minimized and a 20 ns MD simulation was conducted to optimize the stereochemistry of the model and reach a global energy minimum prior to in silico docking.

2.2. Model validation.

Further validation of the structure was tested using the PROCHECK and ERRAT program (http://servicesn.mbi.ucla.edu/ERRAT/). The PROCHECK server checks the main chain dihedral Φ and Ψ angles in a Ramachandran plot, to ensure that they lie within the allowed regions. ERRAT analyzes the statistics of non-bonded interactions between different atom types and plots the value of the error function versus position of a 9-residue sliding window, calculated by comparison with statistics from highly refined structures [22].

2.3. In silico similarity search.

The quickest way to screen millions of compounds involves ligand-based similarity algorithms that search for similar structures to a lead compound. We employed two lead compounds, Lipofermata and Grassofermata, which were previously identified as the best candidates from a 100,000 compound high throughput screen (HTS) [13]. Schrödinger software (Canvas, version 1.5, New York, NY) was used for high throughput screening of small molecules with structural similarity to the two moderate affinity FATP2 inhibitor lead compounds. Because Lipofermata and Grassofermata are heterocyclic structures, and FATP2 contains an ATP-binding motif, we used Canvas to select the top 5% most similar compounds from combined 800,000 compound ChemBridge diversity, kinase, fragment-based, macrocyclic and spirocyclic libraries.

2.4. In silico docking.

In silico docking of the compounds was performed using the Glide module of the Schrödinger 2017 [23–25]. A prerequisite for docking experiments is to generate a grid that defines a cube where the ligand is proposed to bind. In yeast Fat1p mutagenesis studies 16 crucial contact residues were identified [15], and Arg480 was identified as an essential residue for long chain FA transport by vectorial acylation. We therefore drew multiple grids with radii between 5–15 Å centered on this conserved Arg in our model. For each of the docking runs, molecular interactions were analyzed to maximize the number of the 16 residues that are predicted to interact with the candidate ligand using the program PyMOL (Schrödinger Inc.). The first step was standard precision (SP) docking, which uses a series of hierarchical filters to exhaustively search for all possible locations of the ligand in the binding pocket of hFATP2. The 40,000 compounds filtered from the Canvas module for Lipofermata and Grassofermata similarity were docked to the active site using Glide SP. The top 5% hits from SP were subject to Xtra precision docking (XP). The top 10% of hits based on XP docking scores and examination of poses were subject to QikProp [26] analysis, PAINS filter and optimization of structural diversity. QikProp predicts membrane permeability and absorption, distribution, metabolism and excretion (ADME) properties. Pains filters are PAIN interference assays that identify false positives from the docking screens. Structurally redundant compounds were then eliminated after visual inspection, resulting in 26 final compounds (23 hits, 2 positive and 1 negative control) for functional analyses.

2.5. Theoretical pharmacokinetic, membrane permeability and pan interference assays.

Pharmacokinetics evaluation was conducted using QikProp [26]. The program derives parameters such as the rule of five [27]. Docking hits were also subjected to pan interference assays that identify non-drug-like features (http://cbligand.org/PAINS/). Membrane permeability estimates were derived for Madin-Darby canine kidney (MDCK) and colorectal adenocarcinoma (Caco-2) cells with QikProp. Hits with MDCK values >500 and Caco-2 values >100 were considered membrane permeable.

2.6. Molecular dynamics (MD) simulations.

MD simulations were conducted to evaluate the dynamic stability of FATP2 models over 100 ns using DESMOND [28]. The NPT ensemble with the temperature 300k and a pressure with 1 bar was applied on all runs. The relaxation time for each complex was set as 1 ps. The OPLS2005 force field parameter was used in all simulations [29, 30]. The long-range electrostatic interactions were calculated using the particle mesh EWLD method [29–32]. The cut off radius of coulomb interactions was 10 Å.

The system was prepared prior to simulation using Protein Preparation Wizard (Schrodinger). The boundary condition was defined by default setting of a 10 Å × 10 Å × 10 Å orthorhombic box around the protein that only defines the binding site (the whole protein is not required because only residues within 8Å are able to make ionic, van der Waals or hydrogen bonding interactions with the ligand). The system was solvated using TIP3P solvent model [33], within a 0.15 M NaCl solution to mimic physiological conditions. The energy was recorded at regular interval of 1.2 ps. The trajectories were saved at 4.8 ps intervals for analysis. The behavior and interactions between the ligands and protein were analyzed using the simulation interaction diagram (SID) tool in the Desmond MD package. The stability of MD simulations was monitored by examining the RMSD of the ligand and protein atom positions in time.

MD simulation yielded the RMSD of protein atoms with reference to the starting structure. All protein frames were first aligned on the reference frame backbone, and then RMSD was calculated based on the atom selection. Protein RMSD ≤3 Å was considered acceptable during MD simulations. The protein regions with maximum fluctuations during MD simulations were identified using protein root mean square fluctuations (RMSF), where RMSF is defined as a measure of protein residue fluctuations from an initial reference position [34]. The protein-ligand interaction plot (Ligplot) is based on the coordinates of the end time point of the simulation using the program protein–ligand profiler [22].

2.7. Estimation of binding free energy.

The binding free energy of the post-MD protein-ligand complex was calculated using Prime MM/GBSA (Molecular Mechanics, Generalized Born Surface Area) module in Maestro [35]. A total of 50 conformations from the stable trajectories between 70 to 100 ns interval of the MD run were extracted for the MM/GBSA calculation. The ΔG binding free energy was computed using the thermal_mmgbsa.py script and per residue interaction energy was calculated using the script breakdown_MMGBSA_by_residue.py [35].

2.8. Site–directed mutagenesis.

Of the residues that were most important for yeast Fat1p-dependent FA transport [15], four were predicted to be most critical in the hFATP2 structure (N339, T365, D465, R480). To create substitution mutations at these FATP2 ligand-binding residues, site-directed mutagenesis was conducted according to manufacturer’s instructions [GENEART Site-Directed Mutagenesis Kit, Thermo Fisher, Waltham, MA) and In-Fusion HD Cloning Kit (Takara Bio USA, Mountain View, CA)]. The 30 nucleotide PCR primers with the mutations centrally located within the sequence (mutated nucleotides in lower case below), and without overhangs were purchased from Eurofins Genomics (Louisville, KY). The T365K/D465K/R480E plasmid was synthesized by VectorBuilder (Chicago, IL).

N339A-forward: GACTGGCACTGGGAgcTGGCTTACGAGGAGATG

N339A-reverse: CATCTCCTCGTAAGCCgcTCCCAGTGCCAGTC

T365A-forward: GAGTTCTATGCTGCCgCTGAAGGCAATATTG

T365A-reverse: CAATATTGCCTTCAGcGGCAGCATAGAACTC

D465A-forward: TTTCAACAGTGGAGcTCTCTTAATGGTTGA

D465A-reverse: TCAACCATTAAGAGAgCTCCACTGTTGAAA

R480A-forward: CCACGACgcAGTTGGAGATACATTCCGGTGG

R480A-reverse: CCAACTgcGTCGTGGAAATAGATGAAATTTTCA

2.9. Kidney cell line.

HRPT cells are an immortalized cell line derived from human kidney proximal tubule, which retain proximal tubule characteristics [6, 36]. HRPT cells maintain polarity, with basolateral (blood) side adhering to plastic, and apical (luminal) side exposed to media. Early passage cells, which express FATP2 in small amounts, were stably transfected to express a bi-cistronic human FATP2 cDNA and Tomato red fluorescence tag. FATP1, FATP3 and FATP4 were detectable by qPCR, but at much lower expression levels than FATP2 (not shown). Of the FATPs, only FATP2 [17], and to a lesser extent FATP3 (Figure S1), were expressed in the proximal tubule plasma membrane in vivo.

2.10. FA uptake assays.

Experiments were conducted according to previously described methods [17]. Briefly, freshly thawed HRPT cells were seeded in 96-well, black-walled, clear bottom plates, and cultured to confluence over 24 hr. Wells were washed with serum-free, phenol-free media for two hours at 37°C, and candidate molecules were robotically incubated for the final hour at 37°C prior to assays. All compounds were purchased from ChemBridge (Hit2Lead.com, San Diego, CA) and met the high quality standards of 100% identity and >95% or >85% purity by ¹H-NMR and/or LC-MS. BODIPY-conjugated C16 FA (Molecular Devices QBT assay, San Jose, CA; 2.5 μM complexed with 0.2% FA-free albumin in PBS + drug) were robotically added at time = 0. Excitation λ = 490 nm pulses were delivered, and emission λ = 510 nm was recorded at 15-sec intervals for 10 min. BODIPY-labeled FA uptake velocity was determined from the 45–90 sec net (total minus background) fluorescence values. Plates were imaged on the Synergy Neo2 HTX Multi-Mode Microplate reader (BioTek) and averaged from six fields captured from each well using Gen5 software. Results were normalized to Tomato red fluorescence, as an index of cell density. IC₅₀ values were calculated using GraphPad Prism software. The uptake assay signal:noise ratio (BODIPY fluorescence with and without cells) = 3.51. The Z-factor = 1− [3(σ_p + σ)/absolute value of (μ_p − μ_n)], as defined by four parameters: the means (μ) and standard deviations (σ) of the positive (p, no drug) and negative (n, 50 μM Lipofermata) controls = 0.52.

2.11. Apoptosis assays.

Lipoapoptosis in response to palmitate incubation was measured by TUNEL as described [9, 17]. HRPT cells were grown to near confluence on glass coverslips. Experimental compounds were incubated for two hours prior to addition of albumin complexed with palmitate (100 μM, 16 hrs, 37°C). Apoptosis was measured with a Nikon fluorescence microscope by operators blinded to experimental conditions. In experiments where compounds were incubated at multiple doses, IC₅₀ values were calculated using GraphPad Prism software.

3. Results and discussion

3.1. Model to define FATP2 structure.

Using the nearest homologue to FATP2 with known structure (Mycobacterial tuberculosis ACS protein PDB id-3R44) as a template, we developed a homology- and threading-based three-dimensional model for human FATP2 containing amino acids 54–583 (Figures 1A and 1B). The model was used for in silico screening of FATP2 inhibitors from a subset of the ChemBridge library. 3R44 has functional similarity to FATP2, since both proteins have intrinsic ACS activity associated with the cytosolic domain. Though the overall sequence identity is 24%, it is 29% at the putative catalytic region. Homology models based on 20–25% sequence identity have yielded three-dimensional structures that have been successfully used for in silico screening [37]. As expected, superimposition of the hFATP2 homology model on the 3R44 structure is similar (Figure 1A), with RMSD for Cα main chain backbone = 2.45 Å. The hFATP2 model is also quite similar to the FATP2 structure predicted by AlphaFold2 [21] (Figure 1B), with all-atom RMSD for residues 54–583 = 3.8 Å.

Figure 1. hFATP2 model structures.

(A) Superimposed hFATP2 model structure (cyan) with Mycobacterium 3R44 structure (magenta). (B) Superimposed hFATP2 model structure (cyan) with full-length AlphaFold2 (tan) structure.

FATP2 is an integral membrane protein. The soluble domain contains the catalytic site, which we hypothesized is a drug target through inhibition of FA transport by vectorial acylation. The I-TASSER model predicts that the two structurally different FATP2 inhibitors, Lipofermata and Grassofermata [13], dock in the same active site and make contact with 14 FATP2 residues (Figure S2), including six of 16 critical residues predicted from yeast Fat1p studies [15]. A valid model should contain <4% of residues in the disallowed region of the Ramachandran plot and ERRAT score >80. The I-TASSER model meets these criteria; 1.5% of the residues are in the disallowed region of the Ramachandran plot and the ERRAT score is 81.5.

To validate the structural model for FATP2 we created site-directed alanine substitution mutations corresponding to four FATP residues that are highly conserved across species (N339, T365, D465 and R480), and predicted to be binding site contacts, which regulate enzymatic activity in yeast [15]. Unlike FA uptake studies in yeast [15], none of the mutant FATP2 proteins affected baseline FA uptake in kidney proximal tubule epithelial cells, and thereby minimized confounding of subsequent small molecule inhibitor results. Inhibition of FA uptake by Lipofermata was abrogated in cells expressing T365A, D465A or R480A, whereas no effect was observed with N339A (Table 1). The magnitude of IC₅₀ shifts was consistent with the effects of single amino acid substitutions on inhibitors of other transporters or ion channels [38–40]. FA uptake in cells expressing non-conserved mutations of all three FATP2 residues (T365K/D465K/R480E) revealed a greater effect on IC₅₀ compared to the individual alanine mutants (Table 1). The data suggest that T365, D465 and R480 are contacts within the binding pocket. We conclude from the in silico and functional studies that a valid FATP2 structural model was created, and it is sufficient to proceed with small molecule docking experiments.

Table 1.

BODIPY-FA uptake velocity was determined in the human HRPT proximal tubule cell line with transient overexpression of each human FATP2 variant. Cells were preincubated with Lipofermata (2 hrs, 37° C, 0–25 μM). Maximum FA velocity was normalized to values in wild-type cells in the absence of Lipofermata. Samples were assayed in duplicate, from 3–6 experiments.

FATP2 variant	IC50 ± SE (μM)
Wild-type	1.7 ± 0.2
N339A	1.6 ± 0.3
T365A	3.4 ± 0.7
D465A	5.4 ± 1.1
R480A	3.8 ± 0.7
T365K/D465K/R480E	13.0 ± 3.9

3.2. In silico screening and hit characterization.

In silico screening of a library of 800,000 small molecules with known structures was conducted according to the strategy shown in Figure 2 [23]. The library was carefully selected for diversity, and then initially subjected to a fingerprint-based similarity search, for which there is precedent to narrow candidates to a more manageable number, and ultimately identify high affinity inhibitors [41]. The 40,000 compounds most similar to Lipofermata and Grassofermata were tested with docking studies as described in Methods. The docking was accompanied by physicochemical, pan-assay interference, and membrane permeability analyses, as well as adherence to the rule of five (molecular weight <500Da, hydrogen bond donors <5, hydrogen bond acceptors <10, cLogP <5, and <10 rotatable bonds) in the 200 candidates with the best docking scores.

Figure 2.

Flow chart shows in silico screening strategies and experimental validation for the discovery of hFATP2 inhibitors.

Table 2 shows the docking scores, molecular characteristics and QikProp program estimates for cell membrane permeability for the 23 hits with superior docking scores compared to positive controls Lipofermata and Grassofermata, as well as negative control FA transport inhibitor PBI-4050 [42, 43]. No drugs in Table 2 at 40 μM concentration exhibited auto-fluorescence or direct BODIPY quenching. Despite screening by similarity, the top 23 hits demonstrated chemical diversity (Table 3). In silico studies revealed that the two best ranked compounds (1 and 2) dock to the same FATP2 binding site as Lipofermata and Grassofermata (Figures 3A and 3B).

Table 2. Molecular characterization of the top hits and controls.

Docking scores approximate binding affinity, with the greatest negative integer value being the most significant. IC₅₀ values were determined by cell-based FA uptake assays; all other values were determined in silico using Schrodinger software. All hits were verified as non-violators of the pan interference assay. GF, Grassofermata; LF, Lipofermata; ND, not done.

Compound	MW	Docking Score	IC50 (μM) mean ± SEM	Rotatable Bonds	H Bond Accept or	H Bond Don or	Membr Perm : Caco-2 (nm/s)	Membr Perm : MDCK (nm/s)	Log P(O/W)	Oral Absorption
GF	457	−7.30	4.1 ± 2.4	5	3	1	223	241	5.18	94
LF	360	−6.19	2.0 ± 1.1	0	4.5	2	738	1300	2.91	94
1	295	−7.62	0.4 ± 0.3	0	4.5	2	735	490	2.65	94
2	352	−7.57	1.0 ± 0.3	6	5	2	592	830	3.14	94
3	362	−7.43	8.2 ± 7.4	5	4	1	220	255	2.86	86
4	374	−7.38	15.4 ± 26.8	3	6	1	1361	690	4.22	100
5	308	−6.76	308.1 ± 652.	6	3.9	1	4397	2452	4.72	100
6	281	−6.60	479.3 ± 383	4	4.25	1	516	242	2.68	95
PBI-4050	206	−4.76	>500.0	ND	ND	ND	306	175	3.51	ND
7	331	−7.00	8.6 ± 0.9	2	2	4	420	193	3.07	86
8	352	−6.80	22.0 ± 5.9	2	4.5	2	856	686	2.87	94
9	354	−8.70	85.6 ± 30.7	0	4	2	909	1071	1.85	94
10	329	−7.20	268.8 ± 71.1	1	3	1	352	250	1.76	98
11	353	−7.31	72.4 ±	2	4	2	430	593	2.99	94
12	323	−8.02	74.8 ± 27.8	0	4	3	338	277	2.97	95
13	320	−7.50	10.2 ± 2.8	3	4	3	1490	1245	3.02	94
14	380	−6.74	38.8 ±	2	4	4	400	282	2.90	94
15	349	−7.90	9.0 ± 4.1	4	3	2	594	449	3.22	91
16	324	−7.30	165.2 ± 48.3	4	3	2	364	246	1.74	94
17	367	−7.70	142.6 ±	2	3	3	258	206	2.18	85
18	316	−6.70	206.5 ±	2	4	4	1459	1225	2.54	94
19	371	−7.43	108.3 ± 27.7	2	3	3	183	129	2.06	89
20	351	−7.44	76.8 ±	2	3	2	263	196	2.5	94
21	322	−7.53	167.0 ±	4	2	4.5	1206	605	3.44	87
22	369	−7.42	104.4 ±	2	1	3	1191	598	2.66	94
23	352	−7.00	109.3 ±	1	3	2	732	529	3.51	94

Table 3.

Chemical structures of the top 23 hits (and 3 controls) from virtual screens

Figure 3. FATP2 models with inhibitors docked in the binding pocket.

(A) Compound 1 (elemental grey) docked in hFATP2 active site. The contact residues and ligands are shown as ball and stick. Carbon, oxygen, nitrogen and sulfur atoms are shown in cyan, red, blue and yellow, respectively. Hydrogen bonds are shown as solid red lines, pi-pi interactions as green dashed lines and hydrophobic interactions as yellow solid lines. (B) Compound 2 (elemental orange) docked in hFATP2 active site. The contact residues and ligands are shown as ball and stick. Carbon, oxygen, nitrogen and chlorine atoms are shown in cyan, red, blue and green, respectively. Hydrogen bonds are shown as red solid lines and hydrophobic interactions as yellow solid lines. (C) Effect of the bromine (red) to methyl group (yellow) substitution in Lipofermata (elemental magenta) versus compound 1 (elemental yellow) after MD simulation. The protein residues (elemental cyan) and ligands are shown as ball and stick.

A comparison of the 23 hit structures revealed that three compounds (1, 3 and 6) were structurally similar to Lipofermata, containing the oxindole-based skeleton (Table 3). Compound 1 was previously described as an analogue of Lipofermata [44], and shares the same 3’H-spiro[indoline-3,2’-[1,3,4]thiadiazol]-2-one core structure. Compounds 3 and 6 do not have the spiro scaffold, but are ring-opening analogues of Lipofermata, with an intramolecular H-bond between hydroxyl and carbonyl groups (Table 3). On the other hand, a “U” shape conformation is observed in the cluster of compounds 9, 13, 16 and 23, which possess pyrrolidine or piperidine at the U-turn position. Although the scaffolds of these compounds differ from Lipofermata and Grassofermata, the common heterocycles such as fused pyrazole and benzimidazole are favored, which should contribute to FATP2 activation site binding.

3.3. Model validation of the docked complexes by MD simulation.

The five compounds with the best docking scores were refined by MD simulation using the DESMOND package [28]. The RMSD fluctuations for these complexes are below 3 Å (Figures S3A–E), which indicates that they are stable and valid. The RMSF shows residue fluctuations during the simulation with respect to time (Figures S3F–J). The RMSD fluctuations of the atomic positions of both protein and bound ligand were 1–3 Å after equilibration, which is within the thermal fluctuation limits. More specifically, FATP2 complexes with either compound 1 or 2 have a fluctuation of 2.0–2.2 Å square. The deviation is defined as the difference of RMSD after a pre-equilibration period that occurs between 5–7 ns and the 100 ns end point (Figures S3A–E). Moreover, RMSD trajectory behavior was observed during simulation, indicating the stability of these complexes. The RMSD (Figures S3A–E) and RMSF (Figures S3F–J) fluctuations for compounds 1 and 2 showed the least fluctuation and enhanced stability compared to other hits.

Simulation analysis of compound 1 revealed interactions of the methyl group with residues Ala363, Ala364, Thr365, and Arg480, whereas Lipofermata interacted with Asn339, Tyr362, and Ala363. The compound 1 complex simulation shows that the ligand is stabilized by five hydrogen bonds with the FATP2 structure (Figures S4A–C and Table S1), including between the N1 atom of compound 1 and side chain hydroxyl of Tyr267 (2.6 Å). The O atom of Tyr267 forms a H bond with the N2 and O1 atoms of the ligand (3.4 Å). N1 of the ligand interacts with the main chain nitrogen of Ser269 (2.8 Å), as well as the N atom of Thr365 (3.4 Å) during 86% and 54% of the total trajectory time (Figures S4A–C). Tyr224, Thr225, Thr365 and Asn491 make significant water-assisted hydrogen bonds. Compound 1 is also stabilized by 41 van der Waals interactions (Table S1) over 30% of the simulation time. The catalytically important residue Arg480 makes pi cation interactions with phenyl ring of the ligand and water-mediated polar interactions with ligand oxygen atoms. The summary of interactions is shown in the protein-ligand interaction plot (Figure S4A). As shown in Figure 2C the bromine to methyl substitution appears to cause an imperfect superposition of the heterocyclic structure. This conformational change could result in compound 1 forming several interactions, including three additional hydrogen bonds compared to Lipofermata. Compound 1 has predicted favorable ADME properties (Table 2), similar to Lipofermata.

Unlike compound 1, compound 2 has no spiro skeleton and consists of a linear structure with an amide linkage and substituted pyrazole. Compound 2 is stabilized by four hydrogen bonds and 56 van der Waals contacts. Specifically, Tyr362, Gln366, Asn462 and Asp465 form hydrogen bonds with O and N atoms of the ligand (Figures S4D–F). Based on the simulation studies the Glu366 O atom interacts with the N2 (2.7 Å) and O (3.1 Å) atom of the ligand, while the Asn462 O atom interacts with the N3 (3.4 Å) and N4 (2.9 Å) atoms of the compound 2 ligand (Table S2). The critical Asp465 residue makes hydrogen and ionic bonds, as well as water-mediated hydrogen bonds during 90% of total trajectory time. The hydrogen bonds were present for over 50% of the time (Figures S4D–E) and the van der Waals interactions occur 30% of the time. Although compounds 1 and 2 make different interactions, signifying distinct modes of binding due to the diversity of chemical structure, binding to residues Thr225, Thr365, Gln366, Arg480 and Asn491 is critical for both compounds.

The free energy (ΔG) estimates, derived using MMGBSA calculations, demonstrate that compound 1 had a higher binding affinity for human FATP2, compared to all hits (including Lipofermata and Grassofermata). Compound 2 possessed slightly lower ΔG binding free energy, but ranked third among all hits (Table S3). Per residue free energy calculations show the residues most involved in interactions (Figure S4G–H). In general, there is good agreement with the docking models, and only minor variations were observed after MD simulation.

3.4. FA transport inhibition.

The top 23 candidate small molecules were tested for inhibition of long-chain FA transport in HRPT cells stably expressing human FATP2. IC₅₀ values were calculated for each compound and compared to results with positive controls Lipofermata and Grassofermata [13]. Lipofermata and Grassofermata demonstrated modest IC₅₀ values (2.0 μM and 4.1 μM, respectively) (Table 2), consistent with results for these compounds in other epithelial cell lines [13]. Competition curves for Lipofermata, Grassofermata and compounds 1 (IC₅₀ = 0.44 μM) and 2 (IC₅₀ = 0.99 μM) are shown in Figure 4A. Compound 1 displayed a five-fold lower IC₅₀ compared to Lipofermata in kidney proximal tubule cells (Figure 4A and Table 2), which contrasts with slightly higher IC₅₀ for FA uptake in HepG2 liver epithelial cells (IC₅₀ = 5.9 μM) compared to Lipofermata [44].

Figure 4. Characteristics of the two best candidate FATP2 inhibitors.

(A) Human HRPT proximal tubule cells overexpressing wild-type FATP2 were preincubated with compounds (2 hrs, 37° C, 0–50 μM) in triplicate. BODIPY-FA uptake velocity was then determined. Results are from three to six experiments. Data are mean values and error bars were omitted for clarity. (B) To establish potency for lipoapoptosis inhibition HRPT cells pre-incubated with indicated drug (5 μM, 2 hr, 37° C), and then palmitate (100 μM, 16 hr, 37° C) was added. Fixed cells were assayed for apoptosis by TUNEL. Each data point represents a mean value from ten random fields at 400X magnification from one experiment. Data are expressed as mean ± SD. * P <0.05 compared to other groups by ANOVA. (C) HRPT cells were incubated with compounds 1 and 2 at indicated concentrations (2 hr, 37° C), then palmitate (100 μM, 16 hr, 37° C), and TUNEL as in (B). Data are from three experiments and expressed as mean ± SD.

3.5. Inhibition of lipoapoptosis.

Major diabetic complications are largely due to apoptosis of parenchymal cells in target organs, such as kidney, liver and eye. FATP2 has been implicated as a mediator of lipoapoptosis [17]. Compounds 1 and 2, the top candidates from docking, cell-based FA uptake, and MD simulation analyses, as well as Lipofermata and Grassofermata controls, were tested for inhibition of palmitate-induced lipoapoptosis. Figure 4B shows that Lipofermata and compounds 1 and 2 were protective, whereas Grassofermata increased apoptosis, suggesting that it is cytotoxic to kidney tubule cells. Compounds 1 and 2 demonstrated IC₅₀ of 0.6 μM and 0.2 μM, respectively, for palmitate-induced apoptosis (Figure 4C).

3.6. Study limitations.

Although we identified 23 hits among 800,000 compounds through virtual screening, including two lead compounds that are an improvement over existing FATP2 inhibitors, there are limitations to the study approach. It is possible that in silico screening could have missed additional hits due to utilization of a rigid receptor model and inability of the docking program to completely exclude false negatives. Despite good agreement between the homology model and AlphaFold2 FATP2 structures, a crystal or cryo-EM structure would confirm hits and guide chemical modifications to identify even higher affinity candidates. The in vitro experimental screening models may not completely recapitulate in vivo physiology, though we previously demonstrated good correlation between in vivo, ex vivo and cell culture models for kidney proximal tubule FA uptake and apoptosis [11, 17].

4. Conclusions

Contemporary drug discovery often employs high throughput laboratory screening, which is cumbersome and expensive. We used an efficient and inexpensive alternative, which utilized a homology model, similarity search, in silico docking, pharmacokinetic predictions, and molecular dynamics simulations to screen a chemically diverse 800,000 compound library. This strategy yielded 23 hits, which were then evaluated by more labor-intensive functional assays, with the ultimate identification of two FATP2 inhibitors with IC₅₀ <1 μM and favorable ADME properties. Rational next steps to achieve lead compounds with low nM IC₅₀ could include additional virtual screening of even larger libraries [45, 46] and/or medicinal chemistry modifications of the current best candidates. The next generation of compounds might then be amenable for testing in animal models for therapeutic utility against lipotoxic diabetic complications.