Decrypting orphan GPCR drug discovery via multitask learning.

Area of Science:

• Computational chemistry
• Pharmacology
• Structural biology

Background:

• Drug discovery for G protein-coupled receptors (GPCRs) is hindered by limited 3D structures and bioactivity data, especially for orphan GPCRs.
• Existing computational models face challenges in predicting ligand-GPCR interactions due to data scarcity.

Purpose of the Study:

• To develop and validate multitask models for predicting the half maximal effective concentration (EC50) of chemical-GPCR pairs.
• To enable drug discovery for human orphan GPCRs by utilizing protein sequence and chemical property information.

Main Methods:

• Utilized protein multiple sequence alignment features and chemical physicochemical properties/fingerprints for encoding.
• Employed multitask learning to transfer data from known GPCRs to orphan receptors based on feature similarity.
• Trained models on agonist and antagonist data from 200 GPCRs.

Main Results:

• Achieved an excellent mean squared error (MSE) of 0.24 on the validation dataset.
• Demonstrated a reasonably good MSE of 1.51 on an independent orphan dataset, improvable to 0.53 with feature-based transferability.
• Identified informative features and mapped them to 3D structures to understand GPCR-ligand interactions.

Conclusions:

• The proposed multitask models offer a novel approach to learning ligand bioactivity within the GPCR superfamily.
• This method can significantly accelerate the discovery of therapeutic agents for orphan GPCRs.
• Insights into GPCR-ligand interactions were gained by analyzing informative features and their structural mapping.