Introduction

Fatty acid-binding proteins (FABPs) belong to the calycin superfamily of proteins, sharing a similar overall structure with a ten-stranded β-barrel that encloses a large interior cavity for fatty-acid binding. Access of ligands to the binding site is regulated by a small α-helical subdomain (lid) that can be fixed by interaction with an opposing latch (Fig. 1[link]). The different FABP isoforms regulate the uptake, metabolism and intracellular trafficking of fatty acids (reviewed in Storch & McDermott, 2009[Storch, J. & McDermott, L. (2009). J. Lipid Res. 50, S126-S131.]; Supplementary Table S1).

Based on epidemiological studies and animal knockout models, the FABP4 and FABP5 isoforms were identified as potential diabetes and atherosclerosis targets (reviewed in Furuhashi & Hotamisligil, 2008[Furuhashi, M. & Hotamisligil, G. S. (2008). Nat. Rev. Drug Discov. 7, 489-503.]). Both isoforms are produced, among others, in adipocytes and macrophages and share a sequence identity and similarity of 54% and 72%, respectively. Deletion of the FAB4 gene in mice reduces hepatic steatosis, improves glucose tolerance and increases insulin sensitivity. At the same time, in several genetic models, inflammation and atherosclerotic lesion size is reduced. In humans, FABP4 haplo-insufficiency is associated with a decreased risk of type 2 diabetes and cardiovascular disease. Increased concentrations of FABP4 have been detected in diabetic patients, leading to obesity and atherosclerotic lesions (reviewed in Furuhashi & Hotamisligil, 2008[Furuhashi, M. & Hotamisligil, G. S. (2008). Nat. Rev. Drug Discov. 7, 489-503.]). However, it has also been found that a lack of FABP4 can be functionally complemented by FABP5, which is upregulated in the adipose tissue of FABP4−/− mice (Shaughnessy et al., 2000[Shaughnessy, S., Smith, E. R., Kodukula, S., Storch, J. & Fried, S. K. (2000). Diabetes, 49, 904-911.]; Hertzel et al., 2006[Hertzel, A. V., Smith, L. A., Berg, A. H., Cline, G. W., Shulman, G. I., Scherer, P. E. & Bernlohr, D. A. (2006). Am. J. Physiol. Endocrinol. Metab. 290, E814-E823.]). A double knockout of the FABP4 and FABP5 genes in mice displayed a stronger phenotype than the individual knockouts, including protection from diet-induced obesity, insulin resistance, type 2 diabetes and fatty liver disease (Maeda et al., 2005[Maeda, K., Cao, H., Kono, K., Gorgun, C. Z., Furuhashi, M., Uysal, K. T., Cao, Q., Atsumi, G., Malone, H., Krishnan, B., Minokoshi, Y., Kahn, B. B., Parker, R. A. & Hotamisligil, G. S. (2005). Cell Metab. 1, 107-119.]). Thus, an efficient inhibitor for these indications would have to bind to both FABP isoforms with high affinity. FABP3, on the other hand, is prominent in the heart, and FABP3 knockout mice display improved cardiac function and decreased cardiac myocyte apoptosis after myocardial infarction (Zhuang et al., 2019[Zhuang, L., Li, C., Chen, Q., Jin, Q., Wu, L., Lu, L., Yan, X. & Chen, K. (2019). Am. J. Physiol. Heart Circ. Physiol. 316, H971-H984.]). To prevent unwanted effects on cardiac energy metabolism, an optimal FABP inhibitor would display dual activity against FABP4 and FABP5 while not binding to FABP3. As a further requirement, the testes-specific isoform FABP9 might be included, as male FABP9 knockout mice display abnormal sperm morphology (Selvaraj et al., 2010[Selvaraj, V., Asano, A., Page, J. L., Nelson, J. L., Kothapalli, K. S., Foster, J. A., Brenna, J. T., Weiss, R. S. & Travis, A. J. (2010). Dev. Biol. 348, 177-189.]). Taken together, while partial inhibition of FABP9 may be tolerated, a high degree of selectivity against FABP3 appears paramount to avoid any potential adverse cardiac effects.

Human FABP isoforms hFABP4 and hFABP5 crystallize reproducibly, and apo or fatty acid-bound crystals can be soaked with a variety of ligands. The hFABP4 construct proved to be highly reproducible in crystallization and very tolerant to high ligand and DMSO concentrations during soaking, while retaining diffraction to high resolution. As an alternative to individual crystallization, structural information on FABP isoforms may be gained by mimicking the binding site of the isoform in the context of hFABP4. Octuple variants of hFABP4 changed the binding sites to those of isoforms 3 and 5, termed hFABP4_3 and hFABP4_5, while retaining the favorable surface properties of hFABP4 for crystallization.

These efforts led to a set of 229 crystal structures, 216 of which have a ligand bound. Several of the structures are associated with additional affinity data. In total, 75 structures contain a ligand for which IC50 values were measured for hFABP isoforms 4 and 5, while for 50 structures IC50 values are available for hFABP isoforms 3, 4 and 5. The IC50 values of all disclosed FABP inhibitors have been determined using the same biochemical assay, thus the data are intrinsically consistent and should serve as a rich resource for the training of machine-learning or other algorithms for the prediction of ligand-binding poses and relative binding affinities. Affinity estimation in particular has proven to be notoriously difficult to achieve and is always target-centered (Shortridge et al., 2008[Shortridge, M. D., Hage, D. S., Harbison, G. S. & Powers, R. (2008). J. Comb. Chem. 10, 948-958.]; Leidner et al., 2019[Leidner, F., Kurt Yilmaz, N. & Schiffer, C. A. (2019). J. Chem. Inf. Model. 59, 3679-3691.]; Parks et al., 2020[Parks, C. D., Gaieb, Z., Chiu, M., Yang, H., Shao, C., Walters, W. P., Jansen, J. M., McGaughey, G., Lewis, R. A., Bembenek, S. D., Ameriks, M. K., Mirzadegan, T., Burley, S. K., Amaro, R. E. & Gilson, M. K. (2020). J. Comput. Aided Mol. Des. 34, 99-119.]). We recently developed a docking workflow based on the piecewise linear potential scoring function to generate physically plausible ligand poses and compared prediction models in different scenarios encountered in drug discovery for the target PDE10 (Tosstorff et al., 2022[Tosstorff, A., Rudolph, M. G., Cole, J. C., Reutlinger, M., Kramer, C., Schaffhauser, H., Nilly, A., Flohr, A. & Kuhn, B. (2022). J. Comput. Aided Mol. Des. 36, 753-765.]). The prediction models derived from that data set are relevant for that target only, which is a standard situation in lead optimization, but unsatisfactory in general. To generalize the predictions to other targets, there is a need to publish diverse sets of high-quality protein–ligand co-crystal structures plus accompanying affinity data, to which this FABP data set shall be a contribution.

By Andreas Ehler,a Joerg Benza and Markus G. Rudolph