AlphaFold2 structures template ligand discovery.

Area of Science:

• Computational Biology
• Drug Discovery
• Structural Biology

Background:

• AI-driven protein structure prediction tools like AlphaFold2 (AF2) have increased available structures for drug discovery.
• Previous retrospective studies questioned the direct utility of AF2 models for structure-based ligand discovery.
• The study aimed to prospectively evaluate unrefined AF2 models for large-scale ligand screening.

Approach:

• Compared prospective large library docking hit rates and affinities using unrefined AF2 models versus experimental structures for σ2 and 5-HT2A receptors.
• Performed retrospective docking screens against AF2 models and experimental structures to assess recapitulation of known ligands.
• Determined a cryo-EM structure of a potent ligand discovered via AF2 model docking against the 5-HT2A receptor.

Key Points:

• Retrospective docking against AF2 models struggled to identify previously found ligands, aligning with prior research.
• Prospective docking against AF2 models achieved hit rates similar to experimental structures for both receptors.
• Successful prospective screening occurred despite conformational differences in orthosteric pockets between AF2 models and experimental structures.
• The 5-HT2A receptor yielded its most potent, subtype-selective agonists from docking against the AF2 model, outperforming the experimental structure.

Conclusions:

• Unrefined AlphaFold2 models can be effectively used for prospective structure-based ligand discovery.
• AF2 models may sample biologically relevant conformations, expanding the scope of structure-based drug design.
• These findings suggest AI-predicted structures hold significant potential for identifying novel therapeutic agents.