EnGCI: enhancing GPCR-compound interaction prediction via large molecular models and KAN network

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• G protein-coupled receptor (GPCR) research

Background:

• GPCR-compound interactions (GCI) are crucial for drug discovery and chemogenomics.
• Deep learning models, especially large molecular models, show promise in enhancing GCI prediction accuracy.
• Evaluating the integration of large molecular models with other deep learning architectures is a key research area.

Purpose of the Study:

• To investigate the effectiveness of large molecular models in GCI prediction.
• To develop and evaluate a novel deep learning framework for GCI prediction.
• To explore the fusion of multimodal information for enhanced GCI prediction.

Main Methods:

• Introduction of the EnGCI model with two distinct modules: MSBM and LMMBM.
• MSBM utilizes Graph Isomorphism Network (GIN) and Convolutional Neural Network (CNN) with Kolmogorov-Arnold Network (KAN) for feature extraction and decision-making.
• LMMBM employs two large-scale pre-trained models for feature extraction, also using KAN for decision-making.

Main Results:

• The EnGCI model achieved an Area Under the Curve (AUC) of approximately 0.89 on a curated GCI dataset.
• The proposed model significantly outperformed existing state-of-the-art benchmark models.
• Fusion of multimodal information from both modules enhanced overall GCI prediction accuracy.

Conclusions:

• EnGCI integrates feature learning from scratch with pre-trained large molecular models for GCI prediction.
• The model enables a deeper understanding of complex GPCR-compound interaction relationships.
• EnGCI provides a robust framework to advance GCI predictive capabilities and aid GPCR drug discovery.