Comparing models and experimental structures of the GPR101 receptor: Artificial intelligence yields highly accurate models

Abstract

Experimental structures solved through cryo-electron microscopy have recently been published for GPR101, a G protein-coupled receptor (GPCR) implicated in the genetic condition X-linked acrogigantism (X-LAG). Here, we compared these experimental structures with computational models that we previously published, including our internally developed homology models and third-party models generated through the AlphaFold2 and AlphaFold-Multistate artificial intelligence (AI) methods. Our analysis revealed considerable accuracy for both homology models and AI-generated models. However, it also revealed the general superiority of AI methods. Particularly noteworthy is the model generated by AlphaFold2, which captured with high fidelity various structural aspects, including the challenging second extracellular loop. Our previously published homology model of the GPR101-G_s protein complex, based on the β₂-adrenergic receptor, accurately predicted the binding mode of the G protein to the receptor. Moreover, this model predicted the structure of the sixth transmembrane domain (TM6) significantly more accurately than the others, including those built through AI methods, suggesting that homology modeling based on templates solved in complex with the G protein of interest might be the most reliable way of modeling this transmembrane domain. Lastly, our analysis revealed that our molecular dynamics simulations did not have a significant and consistent effect on the accuracy of the models, increasing the accuracy for some domains while decreasing it for others. This work provides insights into the relative strengths of different modeling approaches for our case study on GPR101. More broadly, when considered alongside other assessment studies, it contributes to the growing body of knowledge that can guide the modeling of GPCRs for which experimental structures are not yet available.