Introduction

In early drug discovery it is often the case that the only information available about a biological target is its activity against a select few ligands. Due to the absence of a 3D model (or any other structural information) the ligands’ active (bound) conformation is unknown. The lack of such models can be due to a number of reasons, such as the protein structure being too unstable to be successfully crystallized. In any case, ligand-based computational methods can be utilized in lieu of structural methods. Approaches include developing models to relate physical properties to activity, such as 2D/3D QSAR or pharmacophore model building. FieldTemplater™ in Flare™, Cresset’s structure-based and ligand-based drug design platform, is a version of the latter which first performs a conformational hunt on a handful of active ligand structures (starting in 2D or 3D), and then attempts to align the subsequent 3D conformations to each other in a figurative web of links.¹ The result is a series of templates of aligned ligands in a predicted binding pose despite not having a structure of the target. These templates may then be used as a reference against which to align new ligand designs or a virtual library.² Of course, the assumption for this type of experiment is that the ligands bind to the same binding site: ligands which bind to different sites will not lead to a reliable binding site hypothesis, as this basic assumption is not correct. It is also important that the reference ligands be of similar size, as large differences in molecular weight may lead to less accurate hypotheses of bioactive conformations.

Dipeptidyl peptidase IV – an illustrative example

To illustrate the utility of FieldTemplater, we will consider dipeptidyl peptidase IV (DPPIV), a serine protease with notable implications in diabetes and cancer. One of the primary functions of DPPIV in normal cells is the regulation of metabolism and glucose homeostasis, such as by cleaving glucagon-like peptide-1 (GLP-1), a hormone involved in appetite suppression and insulin secretion.³ The FDA-approved DPPIV inhibitors, also known as gliptins, were designed for glycemic control in type 2 diabetes mellitus.⁴ In the context of cancer, DPPIV has been implicated in tumor progression and metastasis, as it modulates immune responses and influences cell adhesion.⁵ The multifaced role of DPPIV highlights its importance as a therapeutic target. Four known potent DPPIV inhibitors are displayed in 2D and 3D in Figure 1.

As shown in Figure 1, the four inhibitors have very different chemical structures but are approximately the same size. (Their Morgan 2D Similarity calculated with respect to compound 1 range from 0.39 – 0.57). Based on just the 2D chemical structures, all four compounds have nitrogen heterocycles on a hinge-like structure that are likely to present negative electrostatics to complement a protein pocket with positive electrostatics. The aromatic groups likely participate in aromatic interactions with aromatic residues in the active site. The one functional group they all have in common, however, is the free amine/ammonium group. It so happens to be that in the protein, this ammonium group forms hydrogen bonds and salt bridges to two glutamic acid residues in the active site, as illustrated with compound 1 (Figure 2 below). For the FieldTemplater experiment, we will proceed along with the assumption that they all form very similar sets of interactions in the protein, particularly the hydrogen bonding and salt bridges with the ammonium, as it is much simpler in this case to focus on one single functional group constraint as opposed to a larger electrostatic constraint. The goal will be to start from the ligand 2D structures and compare the output template to the aligned crystal structures.

Methods

Obtaining a template from the four chemically diverse inhibitors

To begin, the ligands were loaded into Flare as SMILES strings in the physiological protonation state, i.e., the amines protonated to ammoniums, and a minimization was done using the Cresset XED force field,6 though this step is not necessary. FieldTemplater was then run with the default settings, including a maximum number of conformations of 100, a Sim score of a 50% weight from shape similarity, a maximum number of comparisons per pair of 100, and a maximum delta score of 0.10 per pair (Figure 3). The top four templates, sorted by average Sim between the contributing compounds, are highlighted in the results table and displayed in the 3D view.

It’s immediately clear that the top four templates do not represent the crystal structures; the ammonium groups are rarely aligned to one another. Furthermore, there doesn’t seem to be much consensus in where the protein binding site might be more electrostatically positive or negative; the field point clusters are all mixed without defined regions of positive and negative electrostatics. These two observations may be key indicators that these templates are not representative of the protein binding site. In order to obtain a more realistic template, we can set pairwise constraints (Figure 4). Pairwise constraints tell the algorithm to prioritize alignments that place the selected atoms within a set distance to each other. In this case, we set the nitrogen of the amines to be within 1 Å of each other (default settings).

With these constraints set, a new experiment was run with the default conformation hunt, alignment, and templating settings. The top results are shown in Figure 5.

In the experimental results, the output contained only one template with all four compounds, and the rest of the templates used three out of four compounds. If we look at the top two results, both templates focus the alignment on the pairwise constraint of the free ammonium. However, the similarity scores are different. The top template which uses all four ligands has a similarity score of 0.572, while the second template which uses only three of four ligands has a higher similarity score of 0.637. Without knowing the true bioactive conformations of the ligand, FieldTemplater predicts that the more likely alignment is given by template 2. Furthermore, template 1 contains all four ligands but it requires a very high energy conformation, the 99th of 100, of compound 3 (A S14 1001). When excluding compound 4 (A 1MD 801) from the template, the program can use lower energy conformations of the remaining three ligands to build a stronger consensus alignment. Taking both of these templates and rotating and translating them to align them to the crystal structure overlay, we can see that template 2 is, in fact, much closer to the crystal structure. The results are shown in Figure 6 and the RMSD values in Table 1.

Despite having the structural dissimilarity, these templates are acceptable results and can be used as reference conformation. Additional parameters can be altered in the FieldTemplater experiment to try and optimize the results, including increasing the maximum score delta per pair or the maximum number of alignments to use for each pair. This option specifies the maximum difference in similarity score between the best-scoring alignment and any alignment to be used. The minimum link density may also be increased, which would increase the minimum fraction of possible pairwise links that must be used in constructing this template. However, it was found that increasing the delta score and the link density didn’t much improve the results of this experiment.

Table 1. Distances of the templates from the crystal structure orientations.

As predicted by Sim scores and confirmed with the RMSD values to the crystal structure, template 2 is much closer to the crystal structure conformations and orientations than template 1. We can also see that the electrostatics are much closer to the crystal structure in template 2. The lower region of template 2 and the crystal structure is more electrostatically positive, while the upper region is more electrostatically negative with areas of positivity. Template 1, with its lower Sim score and higher energy conformations, has strong positive electrostatics throughout, while the negative field points seem to be spread apart. In this experiment, using only three of four reference ligands generates a more accurate template. (To adjust the minimum number of compounds in a template, the Templating settings can be altered).

Conclusions

In this study, FieldTemplater was used to generate ligand-based predictions of bioactive conformations of DPPIV binders. The structurally diverse input ligands gave acceptable results when applying a constraint based upon a simple assumption: the only common functional group is key to the activity of these ligands. Under these restrictions, FieldTemplater was able to predict their bioactive conformations within approximately 1.3 – 2.8 Å. Moreover, the clusters of field points closely mirrored those of the crystal structure. These results highlight the utility of FieldTemplater in cases where very little structural information may be available on the biological targets. FieldTemplater may be an especially useful tool for compounds with a common anchoring point, including zinc binding moieties, covalent warheads, or a key common functionality such as with this example. The top templates generated here can now be taken forward as bioactive conformation references for ligand alignment experiments or ligand-based virtual screenings.

References

1. https://cresset-group.com/method/fieldtemplater/
2. Chisholm, T. S., Mackey, M., Hunter, C. A. J. Am. Chem. Soc. 2023, 145, 29, 15936–15950. https://pubs.acs.org/doi/10.1021/jacs.3c03749
3. Deacon, C. F. Front. Endocrinol. 2019, 10, 80. doi.org/10.3389/fendo.2019.00080
4. Ikuma, Y., et al. Bioorg. Med. Chem. 2012, 20, 5864–5883. https://doi.org/10.1016/j.bmc.2012.07.046
5. Zhang, T., Tong, X., et al. Front. Pharmacol. 2021, 12, 731453. https://doi.org/10.3389/fphar.2021.731453
6. https://cresset-group.com/science/