Modeling Active-State Conformations of G-Protein-Coupled Receptors Using AlphaFold2 via Template Bias and Explicit...

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Protein structure prediction tools, including deep learning models like AlphaFold2, are advancing rapidly.
• However, accurately modeling diverse protein conformations, crucial for function, remains a challenge.
• G-protein-coupled receptors (GPCRs) exhibit distinct active and inactive states essential for signaling, involving significant conformational changes.

Purpose of the Study:

• To benchmark the capability of deep learning tools in predicting GPCR conformational states, particularly the active state.
• To evaluate the accuracy of these models in capturing allosteric effects within GPCRs.
• To assess the potential impact of these limitations on structure-based drug design.

Main Methods:

• Benchmarking AlphaFold2 and similar deep learning tools against known GPCR structures.
• Analyzing predictions with and without conformational bias or ligand binding information.
• Evaluating prediction accuracy at both intracellular and extracellular sites of GPCRs.

Main Results:

• Deep learning tools can successfully model the active state of GPCRs when complexed with G-proteins.
• Predictions show reduced accuracy for the extracellular ligand-binding site, indicating limitations in modeling allosteric effects.
• The inactive state is often the preferred conformation predicted by AlphaFold2 for isolated receptors.

Conclusions:

• Deep learning offers significant potential for modeling active GPCR-G-protein complexes.
• Current models exhibit limitations in accurately predicting allosteric effects and extracellular structural rearrangements.
• These limitations may affect the utility of deep learning predictions in structure-based drug design for GPCRs.