Combining graph neural networks and transformers for few-shot nuclear receptor binding activity prediction

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Molecular biology and pharmacology

Background:

• Nuclear receptors (NRs) are critical targets in drug discovery, but identifying endocrine disruptors is challenging.
• Current computational methods for NR-binding prediction often focus on single receptors, limiting their effectiveness.
• Knowledge transfer among multiple NRs can enhance molecular predictor performance and accelerate drug development.

Purpose of the Study:

• To develop a computational model that predicts the binding activity of compounds to nuclear receptors (NRs).
• To identify potential NR-modulators using limited data by leveraging knowledge transfer across multiple NRs.
• To improve the efficiency and accuracy of drug discovery targeting NRs.

Main Methods:

• Integration of Graph Neural Networks (GNNs) and Transformers into a few-shot model named Meta-GTNRP.
• Meta-GTNRP captures local graph structures and global semantic information of molecular graph embeddings.
• A few-shot meta-learning approach optimizes model parameters across different NR-binding tasks, utilizing complementarity among tasks.

Main Results:

• Meta-GTNRP effectively predicts NR-binding activity by combining information from multiple NRs.
• The model demonstrates superior performance compared to other graph-based approaches on a compound database with 11 NRs.
• The few-shot meta-learning framework enables accurate predictions even with limited labeled molecules in imbalanced datasets.

Conclusions:

• Meta-GTNRP is a data-efficient approach that combines GNNs and Transformers for robust NR-binding prediction.
• The model's ability to leverage knowledge transfer across NRs makes it valuable for identifying potential NR-based drug candidates.
• This approach facilitates the discovery of effective drugs by improving the prediction of compound interactions with NRs.