Before the advent of AlphaFold, homology modeling was the most common method for predicting G protein-coupled receptor (GPCR) structures, especially when experimental data were limited. Homology modeling relies on using the known structure of a homologous protein as a template to model the target protein (Carlsson, J. et al., 2011). However, this method has limitations: its accuracy is heavily dependent on the availability and quality of the template structure. For proteins with low sequence similarity to available templates, homology models tend to be less accurate, a notable challenge for GPCRs. GPCRs are dynamic, switching between multiple conformations, and many have only distant homologs with resolved structures, making accurate predictions difficult.
The introduction of AlphaFold2 marked a major advancement in the field of GPCR structure prediction. AlphaFold2 uses an AI-driven approach that eliminates the need for structural templates, allowing it to predict structures for GPCRs that were previously difficult to model due to their low sequence homology to known structures. In a recent study on TAAR1, a GPCR linked to central nervous system disorders, researchers compared the performance of AlphaFold and homology models in virtual screening efforts (Diaz-Holguin, A. et al., 2024). The results were striking: AlphaFold2 outperformed homology models, achieving a 60% hit rate for compounds targeting TAAR1, more than double that of the homology models. The AlphaFold-predicted models revealed distinct ligand-binding site shapes, enabling the prioritization of different chemotypes during docking. One of the identified TAAR1 agonists demonstrated promising antipsychotic-like effects in rodent models, reinforcing the value of AlphaFold in drug discovery, especially when experimental structures are unavailable.
While AlphaFold2’s ability to predict static structures was groundbreaking, however, GPCRs are highly dynamic proteins and continuously adopt different conformations based on their interactions with ligands and the cellular environment which are crucial for understanding how drugs can selectively target different receptor conformations to achieve therapeutic effects. AlphaFold3, the latest version, has improved on this by modeling interactions with various compounds, including natural ligands, ions, DNA, and RNA. However, AlphaFold3 faces challenges with synthetic ligands, which are central to pharmaceutical development. It excels in modeling interactions with natural ligands but struggles to generalize this accuracy to synthetic compounds, limiting its utility in drug discovery involving novel drug-like molecules.
While AlphaFold provides an excellent foundation for generating hypotheses in drug discovery, it is not a standalone solution. To fully understand GPCR behavior and optimize drug targeting across multiple receptor states, experimental validation such as X-ray crystallography, cryo-EM or NMR are still essential for confirming predictions, refining models, and exploring the functional states of GPCRs. Additional kinetic validations must also be verified to comprehend the functions of the proteins. Future iterations of AlphaFold, or its integration with other computational techniques, could help overcome these limitations, leading to more comprehensive models that bridge the gap between computational predictions and real-world data.
References:
Carlsson, J. et al. Ligand discovery from a dopamine D3 receptor homology model and crystal structure. Nat Chem Biol 7, 769-778 (2011). https://doi.org/10.1038/nchembio.662
Diaz-Holguin, A. et al. AlphaFold accelerated discovery of psychotropic agonists targeting the trace amine-associated receptor 1. Sci Adv 10, eadn1524 (2024). https://doi.org/10.1126/sciadv.adn1524
Comments