Hands-On with ApherisFold: Reproducing a SIK3–AMPK Selectivity Study Using OpenFold3

Predicting how small molecules interact with proteins is one of the biggest challenges in drug discovery. Co-folding models, such as OpenFold3, can help address this challenge by predicting the structure of protein-ligand complexes. This allows users to model binding poses and understand molecular interactions. To explore how well OpenFold3 performs on a real-world example, we integrated the latest version of OpenFold3 into ApherisFold; an enterprise software product that enables pharmaceutical organizations to securely run, benchmark, and fine-tune co-folding models directly within their own IT environments. In this post, we focus on a recently published structure of salt-inducible kinase 3 (SIK3) bound to a potent and selective pan-SIK inhibitor.

We selected this study because it was published in December 2023, after the model’s training cutoff of September 2021, ensuring the structure was unknown to OpenFold3 during training. This makes it an ideal test case to evaluate the model’s ability to predict novel structures accurately.

Our goal is to explore some of the experimental findings reported by Temal-Laib et al. through ApherisFold. We set out to determine whether OpenFold3 can be used to rationalize the selectivity of compounds for SIK3 over Adenosine Monophosphate-activated Protein Kinase (AMPK), in a manner consistent with the hypotheses proposed by the authors.

Results and Discussion

As described in their 2023 article in the Journal of Medicinal Chemistry, Temal-Laib et al. aimed to achieve selective pan-SIK inhibition. As part of that, they investigated some of the factors driving selective inhibition of SIK3 over AMPK. They found that fluorination of the methoxy group on compound 15, producing the closely related analogue 20, led to an increase in both potency and selectivity toward SIK3. The chemical structures of compounds 15 and 20 are shown below, showing the modification resulting in improved SIK3 selectivity.

To rationalize these findings, we must first assess whether these targets fall within the domain of applicability of the OpenFold3 model. To validate OpenFold3 could replicate the experimental binding pose, we replicated the reported crystal structure of SIK3 in complex with a close analogue of compound 20 from the same chemical series (compound 22) using the query builder in ApherisFold. By providing the protein sequence, ligand SMILES and reference structure, readers can replicate the co-folding results in their own private instance of ApherisFold (see figure below). The details are supplied in the Appendix.

Shortly after submitting the job, the results were available on the Results Page in ApherisFold. Here, the predicted structure could be superimposed with the reference structure. As shown in the figure below, the ligand pose was predicted accurately relative to the ground truth structure 8OKU, with a calculated ligand RMSD of 1.01 Å. While performance could likely be improved by fine-tuning the OpenFold3 model on available data for this target, a service offered by Apheris, the accuracy achieved with the public model is sufficient to proceed with our analysis.

With the crystal structure successfully reproduced, we next explored whether OpenFold3 could rationalize the increased selectivity of compound 20 for SIK3 over AMPK compared to compound 15. We set up four co-folding jobs: SIK3 with 20, SIK3 with 15, AMPK with 20, and AMPK with 15. The results, shown in the figure below, demonstrate that the increased selectivity toward SIK3 can be explained by the steric clash between the added OCHF₂ group in compound 20 and Met95 in AMPK, a much bulkier residue than the corresponding Thr142 in SIK3.

Temal-Laib et al. reached the same conclusion through a docking study using Glide. However, a key advantage of co-folding is that it can reach this insight even in the absence of a crystal structure, which was of course required for their docking study.

To further support these results, we scored the co-folded poses using the open-source docking tool Smina, including protonation and minimization of the predicted structures. The results are clear: the scores of the minimized co-folded poses closely mirror the measured activity trends. Compound 20 has measured pIC₅₀ values of 9.0 and 5.4 for SIK3 and AMPK, respectively, while compound 15 shows pIC₅₀ values of 8.5 and 6.2.

Conclusion

In this blog post, we have shown how ApherisFold can accurately reproduce the experimentally determined binding pose of compound 22 in complex with SIK3. We further demonstrated how an experimentally observed increase in selectivity for SIK3 over AMPK, reported by Temal-Laib et al., can be rationalized using co-folding through a steric clash with a bulkier residue in AMPK.

Knowing that these targets fall within the applicability domain of OpenFold3 provides confidence that researchers can use the model in ApherisFold to investigate structure-based selectivity and guide rational drug design, even in the absence of crystal structures. Likely, the results could be improved even further by fine-tuning the OpenFold3 model within ApherisFold using proprietary crystallographic data in a secure, locally hosted instance of the application.

Appendix

Please use the details in the table below to replicate the results in your own ApherisFold instance.

¹The reference structure was used only for visualization in the Results Page and was not used to template or otherwise bias the co-folding job.

Authors: Alwin Otto Bucher, Jonathan Cremers