From data to scent: validating an ensemble AI model that predicts novel insect odourant interactions

Area of Science:

• Computational biology
• Structural biology
• Chemo-sensation research

Background:

• Insect navigation relies on odorant receptor (OR)-mediated chemo-sensation.
• Experimental structures of OR-ligand interactions are scarce, limiting understanding.
• Deep learning offers potential for predicting OR structures and interactions.

Purpose of the Study:

• To evaluate deep learning models for insect OR structure and binding prediction.
• To identify limitations in current models for predicting OR-ligand interactions.
• To develop an improved computational approach for accurate OR-ligand binding prediction.

Main Methods:

• Assessed AlphaFold2 and Boltz2 for insect OR structure and pose prediction.
• Evaluated existing deep learning models for ligand binding affinity prediction.
• Developed and tested a weighted meta-model combining multiple prediction approaches.

Main Results:

• AlphaFold2 and Boltz2 accurately predict insect OR structures and poses.
• Current deep learning models exhibit high variability in predicting ligand binding affinity.
• The developed meta-model significantly improves accuracy in predicting OR-ligand binding affinities.
• Empirical validation confirmed the meta-model's predictions for naphthalene binding to ORs.

Conclusions:

• Meta-models enhance the accuracy of insect OR structure and binding predictions.
• This approach provides insights into the generalized prediction of membrane protein complexes and ligands.
• Improved computational tools are crucial for advancing insect chemo-sensation research.