Our project aimed to leverage the BioSolveIT suite of ultra-large chemical space navigation programs to identify make-on-demand agonist candidates for a therapeutically relevant, and structurally unresolved GPCR with few experimentally validated agonists. For this case study, we targeted bitter taste receptor 38 (TAS2R38), a promising drug target for type 2 diabetes given its role as an upstream stimulator of endogenous GLP-1 release. We constructed a homology-based model of TAS2R38 using the closely related TAS2R14 as a template. The homology-based model was refined in its native membrane environment via 20 µs apo molecular dynamics (MD) simulations, and 4 µs of MD simulation with a validated agonist bound. In parallel, we obtained a set of conformationally diverse AF2-based models (AFsample2T). The 240,000 MD simulation frames and 1,000 AF2-based models were clustered using CPPTRAJ to balance conformational diversity and efficiency. Ensemble docking of 15 validated agonists and a set of property-matched decoys, followed by ROC-curve-logAUC analysis, identified the binding pocket most tuned to early enrichment of agonists. The best-performing model stemmed from the homology-based model refined via holo MD simulation. This binding pocket was then screened against the Enamine Hit Locator Library (460,160 compounds) for virtual hits. Filtering and clustering of top-scoring hits yielded a set of 1,000 seed compounds for optimization. We then used FTrees, SpaceLight, and SpaceMACS to identify virtual hit analogs from the Enamine Real Space (83 billion compounds, September 2025 release). The resulting TAS2R38-optimized ligand library of 300,000 make-on-demand compounds was docked against the previously validated binding pocket. Docking of this optimized library shifted the score distribution toward stronger predicted binders. Hits filtering, clustering, and pose inspection prioritized chemically diverse compounds, with Enamine successfully synthesizing all 30 purchased candidates. Compound potencies are now being validated in vitro by our collaborators at The University of North Carolina at Chapel Hill using a TRUPATH BRET2 G-protein dissociation assay. Overall, this project enabled the design of a tandem virtual screening workflow that used both structural and ligand-based approaches to efficiently navigate chemical space. This approach can now be applied prospectively to other unresolved targets with few known actives. Patients who are currently underserved by market-available drugs may experience life-changing outcomes from this alternative approach to treating type 2 diabetes through TAS2R38 agonism.
After 1 year, Dylan has achieved the following goals:
- Construction and validation of an active-state model of our type 2 diabetes target, TAS2R38, was the first hurdle to in silico agonist identification for this structurally unresolved target. Through our competing structure-prediction approaches, we sampled a large subset of the physiologically relevant conformational space of TA2R38. Ensemble docking validated a single structure model capable of consistently ranking validated agonists above property-matched decoy molecules in a structure-based virtual screening procedure. This active-state pocket was utilized in our tandem virtual screening workflow, which identified hits with extremely favorable scores. A set of 30 promising ligands was purchased and synthesized by Enamine. These hits are currently being validated by collaborators in vitro. 100-ns MD simulations with the top 4 novel hits bound evidenced the stability of the predicted docked poses in silico.
- To maximize campaign hit rates, we utilized ligand-based virtual screening of an ultra-large chemical space to generate a TAS2R38-specific ligand library for structure-based screening. We used SpaceLight, SpaceMACS, and FTrees to build this target-optimized library of 300,000 compounds from our initial 1,000 seeds identified via virtual screening of the Enamine Hit Locator Library. Docking of this target-optimized library improved dock scores and yielded a top analog with a 13% better dock score than the top hit from the original commercial library screen. We retrospectively analyzed the per-seed fraction of analogs with improved dock scores and the best Δ dock score to interrogate the performance of each tool in optimizing binder structure. All tools yielded high-scoring hits ultimately purchased for validation, and each exhibited a negative average Δ dock score. Thus, we concluded that each tool serves a unique role in maximizing chemotype diversity and dock scores.
- We also aimed to interrogate the capabilities of traditional, physics-based protein structure prediction methods versus deep learning approaches. Our use of homology-based modeling with physics-based refinement yielded significantly improved enrichment capabilities for known agonists relative to the AF2-based models (AFsample2T). Homology-based modeling of TAS2R38, followed by holo MD-simulation structure refinement, generated a conformation best tuned for agonist identification. The logAUC of the top-performing MD frame was roughly 30.5% higher than that of the top-performing AF2-based model, and 93% higher than the AF3 model. While the AF-based methods were able to rapidly generate a large number of predicted structures, the more computationally-intensive, physics-based approach proved superior in this case for agonist identification and will remain the primary approach to structure prediction for future campaigns.