Computational Engineering of siderocalin to modulate its binding affinity to the antihypertension drug candesartan

Abstract

Lipocalin family proteins have been shown to bind a vast array of small molecules and have subsequently been adapted to selectively bind specific ligands. In this study, candesartan, an antihypertension drug, was identified to bind mouse and human siderocalin in biomolecular NMR experiments, allowing for structural insights into the candesartan-siderocalin interaction. The ligand binding site was determined through an integrative structural biology approach using in silico ligand docking guided by NMR experiments. Building on this structurally informed binding model, we used rational protein design to modulate the binding pocket for increased or decreased ligand binding affinity. The predicted mutations were evaluated in vitro using isothermal titration calorimetry. This resulted in a mutant with a 50-fold increase in binding affinity in addition to a second mutant with a five-fold decrease in binding affinity. Thus, siderocalins have potential as a scaffold for creation of various ligand binding-based tools, including drug scavengers.

1. Introduction

Lipocalins are a group of proteins defined by their characteristic fold. Their prominent motif consists of an eight-stranded, anti-parallel β-barrel that forms the central cavity, also known as calyx, complemented with a conserved α-helix outside of the central β-barrel. The strands comprising the β-barrel are continuously hydrogen-bonded, with flexible, mostly β-hairpin loops (Flower et al., 1993). Though the structural conservation within the lipocalin family is high, the sequence is not conserved: lipocalins with different biological functions can have less than 10 % sequence identity, but still retain the typical fold (Cowan et al., 1990).

The lipocalin calyx allows for the binding of a diverse array of small molecules (Flower, 1996). The lipocalin fold tolerates large variability in amino acid sequences while keeping a rigid ligand binding site. This makes lipocalins scaffolds that can be designed to bind ligands if interest for various applications. Among these are medicinal or biotechnological applications like drug targeting, drug scavenging in the case of intoxication and the purification or stabilization of small molecules (Skerra, 2000). To achieve this, a robust and fast design pipeline capable of modifying lipocalins for specific binding of selected small molecule ligands would be needed.

The human proteome contains several members of the lipocalin family including retinol-binding proteins, α1-acid glycoprotein, apolipoproteins, α1-microglobulin, and siderocalin (Scn) (Schiefner and Skerra, 2015). Among those, α1-acid glycoprotein is known to promiscuously bind pharmaceutical compounds. However, it is highly glycosylated and its ligand selectivity is profoundly influenced by the specific glycosylation pattern (Israili and Dayton, 2001), which makes it a non-optimal scaffold for the design of robust and selective small molecule binders. Scn, on the other hand, has no O-linked glycans and only one N-glycosylation site that is occupied in 7–9 % of recombinantly produced Scns (Zhao et al., 2010). Scn has successfully been used before as a scaffold for the introduction of novel binding sites, proving that it tolerates sequence modifications well (Rothe and Skerra, 2018; Schönfeld et al., 2009; Barkovskiy et al., 2019). On top of that, Scn has been modified to alter its plasma half-life, the renal reuptake, the degree of tissue penetrance, and immunogenicity (Gebauer and Skerra, 2012; Barasch et al., 2016). This makes Scn a great scaffold for the introduction of small molecule binding sites, especially for therapeutic applications. The Scn orthologs in rat and mouse share ~ 60 % sequence homology with their human counterpart (Åkerström, 2006), allowing for the potential transfer of promising therapeutic small molecule binders based on the Scn scaffold into these important model organisms for further analysis.

The ligand-binding properties of Scns differ from most members of the lipocalin family. While the majority of lipocalins preferably binds small, lipophilic compounds like retinoids, fatty acids and porphyrins (Schiefner and Skerra, 2015; Bratt, 2000; Bozhanova et al., 2021), Scn only weakly binds some of these compounds (Bao et al., 2015; Chu et al., 1998). The native ligands of Scn are catechol-type siderophores, which bind at the open end of the typical lipocalin calyx (Goetz et al., 2002). This difference in ligand preference is due to the shallower calyx that is lined with hydrophilic amino acids (the most important of which are K125, K134, and R81) as opposed to the narrower and mostly hydrophobic calyxes of other lipocalins (Bao et al., 2015). A comparison of Scn to other common members of the lipocalin family with respect to their calyx properties and preferred ligands as well as details on different Scn ligands and their respective binding sites are shown in Fig. 1. Details on different Scn ligands and their respective binding regions can be accessed in the SI Fig. 1.

Fig. 1. Lipocalins and their ligands.

Comparison of calyx shapes and surfaces of common lipocalin family proteins binding hydrophobic ligands (α1-microglobulin, retinol binding protein 4, and apolipoprotein D) with hScn (PDB IDs 3QKG, 5NU7, 2HZQ, and 1DFV, resp.). The protein surfaces are colored according to their hydrophobicity (red – hydrophobic, blue – hydrophilic). Ligand binding sites are indicated with dashed circles (Eichinger et al., 2007; Zanotti et al., 1993; Rutardottir et al., 2016; Clifton et al., 2019). Ligand structures are shown to the right of the protein.

Lipocalins have been previously designed for high affinities to various new binding partners. A group of these engineered lipocalins were called anticalins (Rothe and Skerra, 2018; Morales-Kastresana et al., 2022) because four loops at the open end of their calyx resemble antibody complementarity-determining regions (CDRs) and allow lipocalins to bind proteins in a similar manner as antibodies (Richter et al., 2014). One anticalin example is human siderocalin (hScn) optimized to bind cytotoxic T-lymphocyte-associated antigen 4 (CTLA-4) (Schönfeld et al., 2009). However, anticalins have also been designed to bind therapeutically relevant small molecules like colchicine and digoxigenin (Barkovskiy et al., 2019; Schlehuber et al., 2000). Anticalins for various therapeutic applications have reached the clinical stage (Rothe and Skerra, 2018).

In previous design efforts, sites optimized for ligand binding were predominantly the loops at the open end of the calyx, sometimes the loops at the opposing end, the outside of the β-barrel, and the α-helix. These regions are less conserved than the protein core and tolerate the introduction of mutations well. Still, the core of the β-barrel shows sequence diversity across lipocalins (Skerra, 2000). Since the native binding site for small molecule ligands in lipocalins resides inside the calyx, it is interesting to see how binding specificity can be introduced and modulated there. While the lipocalin Blc redesign (Bozhanova et al., 2017) provides a rare example of such modulation, further studies are needed to expand and generalize our understanding.

In this study, we report a novel ligand for human and mouse Scns, the antihypertension drug candesartan, determine its binding site in hScn, and support it by NMR studies and mutagenesis analysis. The ligand was identified with NMR-based screening, and its binding mode was elucidated using computational ligand docking with experimental restraints. The identified candesartan binding pocket differs from the siderophore binding site and is located at the narrow end of the calyx, where other lipophilic ligands have been shown to bind. Using the Rosetta protein design protocol, we predicted hScn mutants with increased and decreased binding affinity to candesartan that were further tested in vitro with isothermal titration calorimetry (ITC) and nanoscale differential scanning fluorimetry (nanoDSF). Two of the proposed mutants showed a marked decrease and increase in binding affinity, respectively. These results show the ability of Scn to bind small molecule drugs and that its binding modalities can be modified using the structural and computational methods applied herein, hence opening a possibility for the use of Scn in medical and biotechnological applications.

2. Results

2.1. Preparation of mouse Scn (mScn) and hScn samples

Uniformly ¹⁵N labeled mScn and hScn proteins were produced in E. coli. mScn was expressed as described before (Moeller et al., 2024) resulting in a protein with two additional amino acids (GP) on the N terminus – a “scar” from the cleavage with HRV3C protease. Due to expression in poor M9 minimal medium, mScn might have been co-purified with iron-chelating enterobactin. hScn was expressed with C-terminal 6His-tag and a point mutation. C87, which is lacking in the mouse ortholog, is known to form a disulfide bond to produce intermolecular homodimers. To ensure a homogeneous monomeric hScn sample, the C87S mutant (hScn_C87S) was used (Clifton et al., 2019). To reduce variability in the sample due to inconsistent co-purified enterobactin level, hScn was grown in the presence of iron in the culture medium, thus inhibiting bacterial enterobactin production (Barasch et al., 2016). No absorption at 335 nm confirms that the protein samples did not contain significant amount of enterobactin-Fe complex (data given in SI Fig. 2) (Goetz et al., 2002).

2.2. Identification of candesartan as a novel ligand for mScn in biomolecular NMR titrations

¹H-¹⁵N Trosy HSQC NMR screening experiments were performed using an FDA-approved small molecule drugs library (1184 compounds) and 1140 compounds from the Vanderbilt discovery library to probe for novel mScn binders. For compounds with multiple significant peak shifts, we performed confirmatory NMR titration experiments (mScn: ligand ratios of 1:0; 1:0.2; 1:0.5; 1:1; 1:2; 1:4, 1:8, 1:10). Eight ligands were found to interact with mScn (resulting spectra except those for candesartan are given in SI Fig. 3). Among those, the marketed antihypertension drug candesartan showed the tightest binding. Interestingly, azilsartan, valsartan, and Vanderbilt discovery library compound VU0363925 were also among the identified hits, allowing to analyze which substructures of the molecules might play the crucial role in the observed interaction.

Based on previously reported backbone assignment of mScn (BMRB ID 52144) (Moeller et al., 2024), the list of residues potentially involved in interaction of mScn with candesartan (either directly interacting or placed in the proximity of the directly interacting residues) included W31, Y32, V33, V34, G35, L36, A37, A40, K43, T45, S53, V66, L70, R72, F94, L96, G97, Q106, S107, A113, T114, A121, M122, V123, F124, T128, Y134, F135, I137, T138, L139, G141, T143. Their peaks – labeled in Fig. 2A – lose intensity in the original position and reappear in a different location with increasing ligand concentration. This behavior suggests binding in the intermediate or slow exchange regime. Theoretical K_D of 10–100 μM would be expected for such interaction, assuming k_on equals not the diffusion limit of 10⁹ M⁻¹s⁻¹ but rather 2*10⁶ M⁻¹s⁻¹ as previously reported for Fe-enterobactin and NGAL in the calyx (Hohlbaum et al., 2018). Hence, candesartan was identified as a novel mScn ligand with medium to low affinity.

Fig. 2. NMR titration experiments.

Overlay of 2D ¹H-¹⁵N Trosy HSQC spectra of (A) mScn or (B) hScn in the presence of different concentrations of candesartan. Labeled residues (sequence number and amino acid one letter code) are highly influenced and thus potentially involved in binding candesartan. Residues found to be influenced upon binding in both, human and mouse siderocalin, are underlined. Peaks enclosed in a dashed circle in (B) represent aliased Arg NHε signals.

2.3. Binding of candesartan to hScn

Candesartan is an antihypertension drug of the angiotensin II receptor blocker type, one of the most prescribed classes of antihypertension drugs worldwide (Abdelkader et al., 2023). To evaluate whether candesartan not only binds mScn but also its human ortholog, NMR titration experiments were performed analogous to the ones already described for mScn.

¹H-¹⁵N Trosy HSQC spectra were acquired for uniformly ¹⁵N labeled, monomeric hScn with increasing concentrations of candesartan (hScn: ligand ratios of 1:0; 1:0.25; 1:0.5; 1:0.75; 1:1; 1:2; 1:4; 1:10). Spectra were well-dispersed and showed similar perturbation patterns confirming that candesartan not only binds mScn but also hScn (Fig. 2B). Amino acids most influenced were F22, Q23, Q28, G29, W_ε31, Y32, V33, G35, L36, A40, R43, A53, T54, L58, L70, R72, K73, K74, C76, D77, Wε79, R81, L94, G95, T104, F122, K125, V126, Y132, F133, I135, T136, G139, K142, E143, T145, L161, F168 and V170. As in mScn:candesartan spectra, we observed an intermediate to slow exchange behavior in the NMR spectra of hScn:candesartan, suggesting a similar K_D. However, when higher than equimolar amounts of the ligand were added to hScn, not one but often two new distinct resonances were observed for affected amino acids. Peaks where this behavior could be unambiguously assigned to the specific signals are highlighted in SI Fig. 4 in grey.

Spectra with two visible distinct binding states are rather unusual. To get insights into possible binding modes which can explain this behavior, the spectra of the highly influenced peaks identified in all states (namely Q23, Q28, Y32, G139, and E143; G95 was excluded due to low SNR for one bound state) were fitted using TITAN (Waudby et al., 2016) and assessed for induced fit, conformational selection, and three state model (sequential and parallel). While the induced fit and three state parallel models were able to reproduce the spectral behavior visually, the induced fit model produced parameters with more reasonable uncertainties. This slightly better performance of the induced fit model can also be seen in SI Fig. 5 for peak Y32, where, for example, the final state is too much populated at low candesartan to Scn ratios (green contour level for state C visible, middle panel). The resulting parameters including their standard error and a jackknife analysis are shown in the SI Table 1. The three states model assumes the presence of two different binding sites that can’t be occupied at the same time. However, there is also a possibility that there are two binding sites for candesartan that can be occupied simultaneously. This wasn’t simulated as fitting the spectra to this even more complex four states model as well as determining the K_D from this small subset of residues is not reasonable.

2.4. Comparison of the candesartan binding sites of mScn and hScn

The NMR titration experiments for mScn and hScn showed distinct chemical shift perturbations upon candesartan addition, clearly proving candesartan binding. Though there are differences in the regions predicted to be affected by binding, there are many overlaps between the orthologues, with substantial sequence conservation (Fig. 3). Based on the perturbed residues’ distribution, in both proteins the candesartan binding pocket seems to be located at the protein center, inside the typical lipocalin calyx, but at a binding site distinct from that of enterobactin and other siderophores. However, a few chemical shift changes can also be observed in both Scn orthologues for amino acids in the open end of the calyx, where enterobactin would typically bind.

Fig. 3. Putative candesartan binding sites based on NMR data.

Residues highly influenced by candesartan presence are highlighted blue (mScn, A) or red (hScn, B) on the corresponding Scn structures (PDB IDs 3S26 (A) and 1DFV (B)). C – hScn structure with mScn-specific (blue), hScn-specific (red), and common for both (yellow) affected residues shown. mScn is depicted with bound enterobactin to indicate the binding site of the natural, high affinity ligand (bound in the Scn calyx, stick representation with colored heteroatoms). Residues found to be influenced upon candesartan binding to both mouse and human Scn are given in D.

2.5. In silico ligand docking and protein design for modulated ligand binding affinity

To further confirm the proposed location of the candesartan binding pocket, docking and mutagenesis studies were performed in silico. First, Rosetta ligand docking application (Meiler and Baker, 2006) was used to generate an atomic model of the candesartan-hScn complex. To prevent successive narrowing of the hScn calyx during ligand docking as well as to account for inability of candesartan to outcompete enterobactin that binds with (0.41 ± 0.11) nM affinity (Goetz et al., 2002; Hohlbaum et al., 2018), we used hScn:enterobactin complex as an input structure.

The residues identified as potentially interacting in the NMR titration experiments were used as restraints to favor ligand placement near them. After five rounds of docking, there were ten distinct ligand clusters (based on root-mean-square deviation (RMSD) to the best scoring model). Upon closer inspection, eight of the ten clusters represented varying ligand conformations occupying the same hScn binding pocket (Fig. 4A) in the narrow end of the calyx and towards the flexible loops that are situated there. The remaining two clusters represented binding positions outside of the binding pocket, failing to fit the experimental restraints to the same extent that the eight other clusters did. Thus, we presumed that candesartan binds inside the hScn calyx but away from the siderophore binding site. Though the remaining eight clusters were distinct in regards to candesartan conformation and binding site, the ligand positions at the protein core still largely overlapped. This is why these clusters were combined to define one putative binding site, even though the specific ligand conformation could not be elucidated. The best-scoring model from each cluster was used as an input structure for further analysis. Additionally, the best-scoring model from each of the two clusters outside the hScn calyx was included to make sure to not discard potential binding sites in our downstream analysis. To further confirm the supposed candesartan binding site, we predicted a set of interaction-modulating mutations and experimentally tested them.

Fig. 4. In silico ligand docking and hScn redesign using Rosetta.

A – Computational docking of candesartan into enterobactin-bound hScn. A.1 – The hScn-candesartan interaction energy estimated in Rosetta energy units [REU] in relation to the position of the ligand, calculated in respect to the best scoring docking model. A.2 – Candesartan positions in the representative docking models. Protein is shown as grey cartoon, ligands are shown as sticks. Enterobactin carbon atoms are shown in grey, candesartan carbon atoms are shown in purple (for ligands with binding positions inside the binding pocket) or in pink (for ligands with binding positions outside of the binding pocket), as indicated in A.1. Oxygen atoms are colored red, nitrogen in blue. B – Loss of binding mutations predicted by an in silico alanine scan in Rosetta. B.1 – Predicted changes in interface energy in the mutants selected for testing. Shown are the top ten computationally predicted loss of binding mutants chosen for in vitro verification and position A119 as a control. Highlighted in orange is the mutant that showed the most marked decrease in binding, Y138A. The plot shows the median as a horizontal line with the box indicating the 25^th to 75^th percentile. All individual values are given as dots. B.2 – Amino acids mutated to alanine are shown as orange sticks on the hScn-candesartan-enterobactin complex model. Positions where mutations potentially disrupt candesartan binding to the protein core are indicated in dark orange. Positions where mutations potentially lessen candesartan binding to the outside of the protein (less probable binding sites) are indicated in yellow. Protein is shown as grey cartoon, candesartan and enterobactin are shown as pink and grey sticks, correspondingly. C – Gain of binding mutations. C.1 – Sequence logo showing mutations seen in the top 20 (ranked by interface score) designed structures. Amino acids are given in one letter code and size correlates with observed frequency. The wild type sequence is shown below. The position of the best in vitro performing mutant, T54R, is highlighted in green. C.2 – All mutations present in the top 20 designed structures were evaluated in silico one by one for their ligand interface score in relation to the wild type protein. Depicted are the interface scores of the ten mutations chosen for in vitro evaluation in relation control (As a control, no mutation was inserted at the corresponding position. Results are depicted for T54). The plot shows the median (horizontal line) in the 25^th to 75^th percentile box as well as the individual values of the 30 models with the lowest ligand interface score per mutation as dots. The best in vitro performing mutation, T54R, is highlighted in green. C.3 – The eight residues chosen for in vitro evaluation are shown as green sticks on the hScn-candesartan-enterobactin complex model. Protein is shown as grey cartoon, candesartan and enterobactin are shown as pink and grey sticks, correspondingly. PDB ID hScn in A, B and C: 1NGL.

First, we performed an in silico alanine scanning to identify potential loss of binding mutations. If mutating residues, predicted to significantly contribute to the protein–ligand interaction, to alanine would result in weaker binding when tested, that would suggest that the binding site was indeed determined correctly. After mutating each of the possibly interacting residues (except for alanine and glycine) to alanine, interface scores representing the interaction energy between hScn and candesartan were compared to the wild-type (detailed results see SI Table 2). The resulting models were also visually inspected in PyMOL. Mutants representing the energetically most unfavorable ligand–protein complexes and that seemed to result in the most distinct disruption of key interactions with the ligand when inspecting the model in PyMOL were selected for expression and further characterization: Q26A, F27A, W31A, Y56A, Y78A, F92A, T113A, Y115A, M120A, Y138A (Fig. 4B).

We also used RosettaDesign (Kuhlman, 2019) with simultaneous docking to search for mutations that would improve the calculated binding energy between hScn and candesartan (Davis and Baker, 2009). The integrated docking steps added extra flexibility and allowed adjustment of the ligand position in response to changing side chain topographies of the interface. Of the mutations found in the top 20 designed structures (as determined by interface score) we omitted mutations to alanine that showed reduced binding affinity in the alanine scan, mutations to cysteine to prevent the formation of novel disulfide bonds, and mutations of proline because of their potential disruptive effect on the protein fold. The remaining mutations were introduced one by one into the wild type hScn to evaluate the effect they had on estimated candesartan binding. The remaining mutations were ranked based on their frequency of occurrence in the first design round and their respective interface scores in relation to control (SI Fig. 6 and 7). Point mutations chosen for testing were F27M, W31M, W31N, T54R, F83Y, A119T, V121S, V121T, F123Y, R140Q (positions indicated in Fig. 4C).

2.6. In vitro evaluation of predicted loss and gain of binding mutations

The interaction of candesartan with hScn variants was investigated in two steps. NanoDSF was used for the initial screening. The melting point of hScn_C87S increased with increasing candesartan concentrations, indicating a stabilizing effect of the ligand (ΔT_m = 5.07°C). Stability of other Scn:candesartan complexes could be deduced from the changes in melting temperature in comparison to wild-type protein. Mutations that showed distinct changes in the thermal unfolding were further characterized by ITC (detailed results in the SI).

Five out of the ten selected loss of binding mutants were expressed and tested for their binding affinity to candesartan. Expression yields and binding data are given in the SI. The loss of binding mutation with the strongest decrease in stabilizing effect of the ligand according to nanoDSF measurements was Y138A. In ITC experiments, it showed an increased K_D and reduced reaction enthalpy compared to hScn_C87S (Fig. 5A). The K_D is increased fivefold, but the mutant is still able to bind candesartan (Fig. 5B).

Fig. 5. In vitro characterization of designs with ITC measurements.

Binding curves for candesartan and hScn obtained with ITC. A – wild type Scn_C87S, B – loss of binding mutant Y138A and C – gain of binding mutant T54R. The thermodynamic parameters and estimated number of binding sites calculated from the triplicate measurements for the wild type and loss of binding mutant and the singular measurement for the gain of binding mutant are indicated in the graphs.

Similarly, five out of the ten selected for testing gain of binding mutants were successfully expressed. Surprisingly, none of these mutants showed any change in melting temperature in nanoDSF experiments, but for some mutants changes in binding affinities were observed when evaluated with ITC. For example, T54R mutation led to a 50-fold improvement of the K_D (115 nM in relation to 5 μM for wild type). However, two different binding events could be deduced with ITC. The second binding site exhibited a K_D of 3.5 μM (Fig. 5C).

3. Discussion

Here, we describe a novel ligand for hScn and mScn, the antihypertension drug candesartan. We estimated its binding affinity to the wild-type proteins with NMR titration experiments, and for hScn supplemented these with ITC measurements. The putative binding site of the novel ligand was elucidated using NMR titration experiments results, in silico docking, and the introduction of loss and gain of binding mutations into hScn and their subsequent in vitro testing. We found the binding site to be situated to the narrow end of the typical lipocalin calyx, where predominantly lipophilic ligands have been known to bind. In silico modeling suggests that candesartan should be able to bind in the presence of enterobactin, one of the high affinity native ligands of siderocalin.

Previously the binding of lipophilic compounds to Scn has been described (Chu et al., 1998; Bratt et al., 1999) and involved residues F27, W31, V33, V66, F83, F92, L94, V108, V110, V121 and F123 (hScn) (Bao et al., 2015). Of these previously published residues, chemical shifts of W31, V33, V66, F94, L96 and V123 were changed in our NMR titration experiments with candesartan and mScn (note that due to differences in the amino acid sequence between human and mouse Scn, F92, L94 and V121 in human correspond to F94, L96 and V123 in mouse). In case of hScn, W31 Hε, V33, and L94 disappeared during the titration experiments. Apart from the residues located in the narrow end of the calyx, perturbations were also observed in the loops and at the enterobactin binding site (residues found to be influenced in mScn and hScn and have been reported to be involved in enterobactin binding in hScn: Y132 and F133 (Clifton et al., 2019)). While the perturbations in the narrow region of the calyx were expected, considering that candesartan is more lipophilic than enterobactin, the perturbations in the enterobactin binding site were surprising. It is possible that the binding to candesartan leads to conformational changes propagating into the open end of the calyx, where siderophores are bound, thus resulting in the observations in the NMR spectra. In enterobactin-free hScn, where the ITC data suggests multiple binding sites, candesartan might even occupy a binding pocket towards this broader end of the Scn calyx. However, this binding event would be of no relevance in the presence of the native, higher affinity ligand enterobactin. Further insight into the candesartan binding site in presence and absence of the native ligand enterobactin could be gained by obtaining crystal structures.

Our attempt to fit the hScn:candesartan NMR titration data revealed that a determination of the binding model is not possible with this method alone. The induced fit model and the three-state parallel model both fit the data very well but result in totally different presumptions. The three-state parallel model assumes two binding sites, both of which cannot be occupied simultaniously. The induced fit model supposes a conformational change of the protein after ligand binding. Taken together, this data points towards a binding mode with more than one distinct binding site or bound protein state. Comparing these results with the ITC data, which was evaluated with the one set of sites model (assuming N binding sites with the same K_D values), shows similar results. The number of binding sites (1.5) also indicates several binding sites that have the same K_D values. Hence, as a dimerization of hScn upon binding of candesartan could be excluded, there are probably several protein states in which either one or two molecules of candesartan are bound to hScn.

To gain a more detailed understanding of the possible candesartan binding site in hScn, we performed computational ligand docking in Rosetta using NMR-derived constraints. As expected, candesartan was predicted to be bound inside the narrow end of the calyx. This is in concordance with the more lipophilic nature of candesartan compared to the phenolic or catecholic ligands bound to Scn at the open end of the calyx (Bao et al., 2015). Because of the difference in binding sites between the ligands (enterobactin and the novel ligand, candesartan), it is possible that both ligands can bind simultaneously. Enterobactin, however, will probably block access of candesartan to its predicted binding site through the open end of the calyx, and candesartan might need to enter its binding pocket through the opposite end of the barrel.

In addition to determining the binding site, a dissociation constant of the binding of hScn_C87S to candesartan was estimated as (5.0 ± 0.8) μM by ITC. While this interaction might be classified as specific, such affinity is of near to no relevance therapeutically. With maximal candesartan plasma concentrations of around 0.227 μM and plasma protein binding exceeding 99 % (Gleiter and Mörike, 2002; Buter et al., 1999), the effect of the observed candesartan-hScn interaction is probably not influencing candesartan pharmacokinetics.

To further confirm the candesartan binding pocket location, we assessed 20 hScn point mutants. Not all selected mutants could be expressed successfully. Mutants that expressed often showed decreased yields compared to wild-type hScn, which was not necessarily surprising, as our prime selection criterion was ligand binding energy, not protein stability. Still, many mutations at the protein core were tolerated well enough to produce substantial amounts of stable protein (expression yields accessible in the SI). This means that hScn is not only easily allowing mutations in the loops and α-helix, but also in its core.

All expressed mutants were first assessed by nanoDSF. In case of the loss of binding mutants, only the variants that showed decrease in the stabilizing effect of the candesartan were further assessed by ITC. Interestingly, none of the gain of binding mutants showed any detectable change in the melting temperature measured by nanoDSF. Hence, four out of the five mutants that expressed were evaluated by ITC.

Our best performing loss of binding mutant, Y138A, showed fivefold increase of K_D value compared to the wild type (Fig. 5). Upon closer examination of the docked ligand pose, it seems that the tyrosine at position 138 forms a polar interaction with the tetrazole ring that is lost when mutation to the apolar alanine is introduced (Fig. 6A). A complete loss of binding behavior would probably require mutation of more than one amino acid or different mutation approaches like changing polarity or reversing charge in interacting residues.

Fig. 6. Structural insights on the effects of best-performing loss and gain of binding mutations.

The two best-performing mutations evaluated in vitro are indicated in the depicted hScn structure in orange for loss of binding mutant Y138A (A) and green for gain of binding mutant T54R (B). The wild type amino acids are colored grey. hScn (PDB ID 1NGL) is shown as grey cartoon, ligands and side chains of the mutated amino acids are shown as sticks. For the loss of binding mutant, it can be presumed, that a polar interaction between the tyrosine and the tetrazole ring of candesartan is disrupted upon mutation to alanine. The best gain of binding mutation might lead to a salt bridge between the tetrazole ring and the introduced arginine, where before the threonine could not provide this interaction.

The best-performing gain of binding mutant, T54R, not only demonstrated improved K_D value, but also made the complex nature of hScn_C87:candesartan interaction, previously only observable in the NMR titration results, visible in the ITC curves manifesting as two quantifiable binding events with estimated K_D values of (115 ± 18) nM and (3.5 ± 0.1) μM (Fig. 5C). Interestingly, T54 is located close to the enterobactin binding site and T54 has even been reported to be involved in enterobactin binding (Clifton et al., 2019). However, since the binding data for this mutant was obtained only using enterobactin-free sample, the effect of this mutation on enterobactin binding could not be evaluated.

The T54R mutation replaces a small, polar amino acid with a larger, positively charged one. Based on the predicted candesartan binding mode it can be speculated, that the newly introduced arginine can form a salt bridge with the acidic tetrazole ring of candesartan (Lieber et al., 1951), thus strengthening the interaction in comparison to the native amino acid, threonine, that can only form polar interactions. Details on the structural interpretation of loss and gain of binding mutations Y138A and T54R are given in Fig. 6. However, until a crystal structure of the candesartan-Scn complex is determined, these explanations remain hypothetical.

4. Conclusion

Using the combination of NMR titration experiments, computational ligand docking, protein design, and affinity measurements with ITC and nanoDSF, a new ligand for siderocalin, candesartan, was identified and its binding mode was suggested. The binding affinity between hScn and its novel ligand was modulated in a rational manner yielding a mutant with a 50-fold tighter binding. Though lipocalins have been designed before for modulated binding affinities to given ligands (Rothe and Skerra, 2018; Schönfeld et al., 2009; Vopel et al., 2005; Müller and Skerra, 1994; Muslinkina et al., 2020; Bozhanova et al., 2021), the targeted sites for mutagenesis were predominantly located in the flexible loops of the proteins and only seldom reached the protein core. Working with the binding sites that are located closer to the protein core offers the possibility of burying large parts of the ligand, giving sufficient shape complementarity, and thus resulting in many ligand–protein interactions and higher affinities. The possible ligands could be small, drug-like compounds like candesartan, making siderocalin a potential scaffold for the design of drug scavengers deployed in the case of intoxications or in other biotechnological applications.

5. Methods

5.1. Experimental methods

Expression and purification of mScn.

The mScn gene, codon-optimized for E. coli expression (GenScript), was fused with the hexa-His-tagged B1 domain of protein G (GB1, cloned from GEV 2 plasmid that was a gift from John Louis (Addgene plasmid #12616; https://n2t.net/addgene:12616; RRID:Addgene_12616) (Huth et al., 1997) using the LEVLFQGP linker containing an HRV3C protease cleavage site. The resulting 6His-GB1-mScn construct was inserted into a pET21a vector and verified using Sanger sequencing. The verified plasmid was transformed into E.coli BL21(DE3) cells for protein expression. One colony was used to inoculate the 55 mL Lysogeny Broth (LB) culture supplemented with 100 μg/mL ampicillin. This starting culture was grown at 37 °C in a shaking incubator at 220 rpm until the optical density (OD) exceeded 0.6. 50 mL of the culture was spun down, resuspended in 10 mL mQ H₂O, and transferred into a 2.8 L baffled flask containing 1 L of minimal media (M9) supplemented with 0.4 % glucose, 1x MEM vitamins (New England Biolabs), 100 μg/mL ampicillin, 2 mM MgSO₄, 1g of ¹⁵N-(NH₄)₂SO₄, 1x trace metals (Teknova), 0.1 mM CaCl₂, and 5 μM ZnCl₂. The M9 cultures were grown at 37 °C in an incubated shaker at 220 rpm until an OD of 0.8–1.0 (usually took 4–5 h to reach). After reaching the desired OD, isopropyl β-d-thiogalactopyranoside (IPTG) was added to a final concentration of 0.2 mM alongside with additional 100 μL of 100 mM ZnCl₂, and the temperature was lowered to 22°C. The bacteria were harvested the following morning (12–14 h after induction) by centrifugation at ~ 9,000 × g for 20 min using a JLA-8.1000 rotor in a Beckman J-20 XP centrifuge. The supernatant was discarded, and cell pellets were collected and either used immediately for protein purification or stored at − 80°C until further use.

Cell pellets were resuspended (after thawing, if frozen) in an equilibration buffer (50 mM Tris, 150 mM NaCl, pH 8.2) supplemented with cOmplete, EDTA-free Protease Inhibitor cocktail (Roche) according to the manufacturer’s instruction using 5 mL of the buffer per gram of wet cells. The cells were destroyed by sonication on ice at 50 % Amp, 5 s on/10 s off, for a total of 10 min. The lysate was clarified by centrifugation at ~ 50,000 × g for 20 min using a JA25.50 rotor in a Beckman J-20 XP centrifuge, and the supernatant was filtered using 0.45 μm syringe filters. Ni-NTA resin (1 mL for 10 mL of lysate) was placed in a column and equilibrated with 5 column volumes (CV) of equilibration buffer. The filtered supernatant was slowly run through the column twice, then the column was first washed with 5 CV of equilibration buffer followed by 5 CV of washing buffer (50 mM Tris, 150 mM NaCl, 10 mM Imidazole, pH 8.2). The protein was eluted with 7 CV of elution buffer (50 mM Tris, 150 mM NaCl, 200 mM Imidazole, pH 8.2), and the eluant was collected into 7 separate 1 CV fractions. The elution fractions were assessed by SDS-PAGE, and fractions containing 6His-GB1-mScn fusion protein were pulled together, concentrated, and buffer exchanged into equilibration buffer using an Amicon 10 kDa MWCO centrifugal filter unit. To remove the solubility-enhancing GB1 tag, the protein was mixed with an in-house produced HRV3C protease at a 20:1 6His-GB1-mScn:HRV3C molar ratio and incubated for 2 h at room temperature on a nutator. A fresh batch of Ni-NTA resin was placed in a column and equilibrated with 5 column volumes (CV) of equilibration buffer. The cleavage reaction mix was run through the column twice to remove His-tagged GB1, His-tagged HRV3C protease, and uncleaved 6His-GB1-mScn fusion protein. The purity of the mScn protein sample (the flow-through) was assessed by SDS-PAGE. If required, the protein was further purified by size-exclusion chromatography on a HiLoad 16/600 Superdex 75 pg column (GE Healthcare) pre-equilibrated with 100 mM sodium phosphate buffer, pH 6.0. Otherwise, the flow-through was concentrated to ~ 10 mg/mL and buffer exchanged into 100 mM sodium phosphate, pH 6.0 using an Amicon 10 kDa MWCO centrifugal filter unit. The protein concentration was determined by UV280 absorbance using the theoretically calculated Abs0.1 % (1g/L) = 1.238 (ProtParam). The protein that was not planned to be used within the next week was aliquoted and stored at − 80°C

5.2. Expression and purification of hScn_C87S and further mutants

For the expression of monomeric hScn, the C87 responsible for dimerization was replaced with S in all used constructs. The gene sequence of hScn_C87S was codon-optimized for E. coli expression and inserted into a pET21a(+) vector, containing a C-terminal His6-tag for purification and an ampicillin resistance. The plasmid was transformed into the E.coli BL21(DE3) cells. The mutants were generated with site-directed mutagenesis using the Platinum SuperFi II polymerase from Thermo Fisher Scientific (Waltham, MA, USA) according to the manufacturer protocol and checked with Sanger sequencing. The cells were cultivated in a preculture (50 mL LB medium and 0.27 mM ampicillin) overnight at 37°C and 220 rpm. Cells were harvested by centrifugation for 10 min at 4347 × g, 20°C, and transferred into the minimal medium (50 mL 20x M9 stock solution (0.85 M Na₂HPO4, 0.44 M KH₂PO4, 0.17 M NaCl), 900 mL water, 1 g NH₄Cl (¹⁵N labeled for NMR experiments), 4 g glucose, 0.1 mM CaCl₂, 0.1 mM ZnCl₂, 1 mM MgSO₄, 10 mL MEM vitamins, 0.1 mM FeSO₄ and 0.28 mM ampicillin). The cells were grown at 37°C, 220 rpm to an optical density of 1–2. Then, the protein expression was induced with 0.5 mM IPTG. 4 h after induction the cells were harvest by centrifugation (4°C, 10628 × g, 20 min). The supernatant was discarded and the pellet was frozen at −80°C. The cell pellet was resuspended in a buffer containing 25 mM Tris, 240 mM NaCl and 20 mM imidazole, pH 8. Ultra sonication was used for cell disruption (Fisherbrand: Fisher Scientific GmbH, Vienna, Austria; 10 min, Amplitude 60 %). The cell suspension was then centrifuged two times (20 min, 33745 × g, 4°C) to separate the cell components. Afterwards, the resulting supernatant was filtrated and purified using Ni–NTA agarose (Qiagen GmbH, Hilden, North Rhine-Westphalia, Germany). Protein was eluted with 25 mM Tris, 240 mM NaCl and 400 mM Imidazole, pH 8. Elution fractions were combined, concentrated and purified further with size exclusion chromatography (HiLoad 16/600 Superdex 75 pg) on an A^¨ KTA instrument (Cytiva, Chicago, Il, USA) in 100 mM sodium phosphate buffer pH 6. The obtained protein fractions were analyzed by SDS PAGE and the pure protein fractions were pooled and used for further experiments.

5.3. Check for bound enterobactin

The purified hScn was tested for bound enterobactin at a concentration of 240–300 μM with an absorption scan from 280–700 nm (Goetz et al., 2002). The spectrum was recorded with NanoDrop One^C (Thermo Fisher Scientific). Purified mScn (470 μM) was scanned from 220–600 nm with a UV-1800 UV/Vis Spectrophotometer (Shimadzu Corporation).

5.4. NMR titration experiments mScn

Screening of the libraries of compounds was performed similar to described before by Castro et al. (Castro et al., 2022). Briefly, first, the mixtures of 12 compounds (1 μl of 10 mM stock solution in DMSO of each) were dispensed in 96-well plates and mixed with ¹⁵N labeled mScn to achieve the final volume of 550 μl of 20 μM mScn in 100 mM sodium phosphate, pH 6.0, with 5 vol% D₂O, 2 vol% DMSO. The samples were transferred into 5 mm NMR tubes, and spectra were collected using Bruker Avance III 600 MHz NMR spectrometer equipped with a cryoprobe (CPTCI) and automated sample changer (SampleJet). Data were collected at 298 K using the ¹⁵N-¹H BEST-TROSY (b_trosy3gpph.2) experiment with 64 scans for each of 96 increments and processed using Bruker Topspin 3.6.1. Compounds from the wells where significant peak shifts were observed compared to the control sample with no ligands added were re-dispensed 1 per well (1 μl of 10 mM stock solution in DMSO supplemented with 11 μl of DMSO), mixed with 20 μM mScn, and the NMR spectra were measured again. Identified hits were further confirmed by NMR titration experiments using 50 μM mScn in 100 mM sodium phosphate, pH 6.0, with 5 vol% D₂O, 5 vol% DMSO and mScn: ligand ratios of 1:0; 1:0.2; 1:0.5; 1:1; 1:2; 1:4, 1:8, 1:10.

5.5. NMR hScn_C87S

For the NMR titration experiments of candesartan to human siderocalin, ¹⁵N labeled hScn_C87S was expressed and purified as described above. Each NMR sample was prepared independently and consisted of 200 μM hScn_C78S in 100 mM sodium phosphate buffer, pH 6.0, with 5 vol% D₂O, 2 vol% DMSO and 10 μM TSP-d₄ added. The varying candesartan:protein ratios were achieved by adding the protein mix to 10 μl of DMSO with varying candesartan concentrations (up to 100 mM) to achieve a final volume of 500 μl. This rapid dilution prevented undissolved aggregates of candesartan. Measurements were performed on a Bruker NEO 700 MHz spectrometer equipped with a 5 mm HCN triple resonance probe with z-gradient at 25°C. For each hScn:candesartan ratio a ¹H-¹⁵N Trosy HSQC with water suppression using the WATERGATE scheme was acquired with a spectral width of 16.2 ppm for ¹H, 41 ppm for ¹⁵N and 128 data points in the indirect and 8 k points in the direct dimension (Pervushin et al., 1997; Piotto et al., 1992). STATES-TPPI acquisition scheme was used to alias the Arg NHε in the circled region in Fig. 4 SI. Depending on the candesartan concentration, 64 to 400 scans per increment were collected, to gain a sufficient spectral quality. Processing and referencing to TSP-d₄ (Hoffman, 2006; Cavanagh et al., 2007) was done in Topsin 4.0.6, while for further evaluation POKY (Lee et al., 2021) was used. The resonance assignment for the hScn was transferred from BMRB-Id 4267 (Coles et al., 2011).

5.6. Isothermal titration calorimetry (ITC) for candesartan-hScn affinity measurements

The MicroCal PEAQ-ITC from Malvern Panalytical (Malvern, Malvern Hills, UK) was used for the measurements. Two solutions had to be prepared: 1) protein solution for the cell, 2) the ligand solution for the syringe. The buffer of ligand and protein solution were matched and contained 100 mM sodium phosphate, pH 6.0 and 2 % DMSO to avoid buffer mismatches. The following setup was chosen for the binding measurements: 25°C, 41.9 μW reference power, High Feedback, 750 rpm stirrer speed, initial delay of 120 s, a minimum of 19 injections of 2 μL in 4 s and a spacing time of 150 s (first injection: 0.4 μL with a duration of 0.8 s and a spacing time of 150 s). Reference measurement with titration of the ligand to the buffer was performed. For the data analysis the MicroCal PEAQ-ITC Analysis Software was used.

5.7. NanoDSF for candesartan hScn affinity measurements

Thermal unfolding measurements were performed with Prometheus Panta from Nanotemper (NanoTemper Technologies GmbH, Munich, Bavaria, Germany). The samples containing 50 μM of the protein were prepared in 100 mM sodium phosphate buffer, pH 6.0 with 2 % DMSO and were filtered through a 0.2 μm filter. The excitation power was set by the instrument, a temperature ramp from 25 – 90°C with a heating ramp of 0.5°C/min. Every condition was measured at least as triplicates. The tryptophan to tyrosine ratio in the protein is 2:10 which leads to an insensitivity for the ratio of 350:330 nm. This results in the ratio signal being very small and prone to error. Therefore, only the 350 nm signal was taken into account.

5.8. Computational methods

For a detailed protocol capture including scripts and command lines, please refer to the SI.

Ligand docking and protein design were performed and analyzed in Rosetta 3.10, a software suite for biomolecular modeling that is maintained and continuously refined by the RosettaCommons (Bender et al., 2016; Koehler Leman et al., 2020).

5.9. In silico ligand docking of candesartan to hScn

Ligand docking of candesartan to hScn was performed with RosettaLigand (Lemmon and Meiler, 2012; Combs et al., 2013). Protein structures were collected from the PDB (https://www.rcsb.org/, PDB IDs 1DFV and 1NGL (Goetz et al., 2000; Coles et al., 1999). These template structures are not mutated at position C87, as opposed to the C87S mutant used in in vitro studies. The structures were relaxed in Rosetta and the best-scoring structures were used for docking. Ligand structures were obtained from PubChem (candesartan, https://pubchem.ncbi.nlm.nih.gov) and the PDB (enterobactin, PDB ID 3CMP (Clifton et al., 2019). Confoccsion 4.2.0 (Brown et al., 2022). For both candesartan and enterobactin Rosetta-readable params files (containing conformers in the case of candesartan) as well as a pdb files were produced. To prevent an artificial narrowing of the lipocalin calyx during the following docking rounds, enterobactin was placed in its known Scn binding site as observed in crystal structure 3K3L (Clifton et al., 2019). Candesartan was placed at differing positions around Scn in PyMOL to generate a diverse set of 20 input structures in total. Docking was restrained with experimental data obtained from NMR using Rosetta’s restraint framework (Fischer et al., 2021). Candesartan was docked to Scn in five rounds. In each round 20,000 models were generated, sorted for the best interface score between ligand and Scn (measured in Rosetta Energy Units (REU)), and the best 1,000 structures were selected and used as input for the next round of docking. Clustering was monitored by plotting interface energy to ligand root-mean-square deviation (RMSD) of the resultant structures. Clusters were evaluated concerning restraints met, size, energy and plausibility of ligand conformations.

5.10. In silico protein design of hScn for modulated candesartan binding properties

Protein design was performed in Rosetta with the top 100 docked structures from ligand docking as input. Scn residues predicted to be involved in binding candesartan after docking were identified using the residue energy breakdown in Rosetta.

To verify the predicted binding site, each identified interface residue was mutated in silico to alanine. The resultant change in interface energy was evaluated in Rosetta and mutations were inspected in PyMOL (Schrodinger, Version 2.5 (Schrödinger et al.)). Mutations were chosen for in vitro examination based on their interface scores (energetically most unfavorable) and the visual inspection.

Potential gain of binding mutations were investigated using a Rosetta protocol that allows for interface residue design and flexibility as well as adjustment of the ligand position and conformation in the binding pocket (Moretti et al., 2016). Specifically, residues involved in ligand binding were defined in a resfile and then allowed to be altered to any amino acid. For each of the 100 input structures, 100 designs were generated. Parallel to the design, control structures were generated using the same protocol but without changes to the native amino acids. Designs were scored in Rosetta for the ligand–protein interface score. Mutations that were present in the 20 best-scoring designs were reintroduced into the hScn:candesartan complex one by one. Again, control structures were prepared using the same design protocol but without the insertion of a mutation. Designs were scored separately for their interface score. Excluded from this procedure were mutations to alanine that have been shown to be unfavorable in the alanine scan, mutations to cysteine and mutations of proline. For each tested mutation the interface scores of the 30 best-scoring designs were compared to the corresponding controls. Mutations for in vitro evaluation were selected based on the interface score difference to control and the frequency among the 20 best-scoring structures after the first round of design.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.jsb.2025.108180.

Data availability

Data will be made available on request.