Combining GCN Structural Learning with LLM Chemical Knowledge for Enhanced Virtual Screening

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• Traditional machine learning methods for virtual screening often lose information due to predefined molecular representations.
• Deep learning, especially graph convolutional networks (GCNs), offers a more expressive approach by directly processing molecular graphs.
• Large language models (LLMs) excel at capturing complex chemical patterns from large datasets.

Purpose of the Study:

• To develop a novel hybrid deep learning architecture for improved virtual screening in drug discovery.
• To integrate GCNs with LLM-derived embeddings for more effective molecular representation.
• To demonstrate the model's broad applicability across different target classes.

Main Methods:

• Proposed a hybrid architecture combining GCNs with LLM-derived embeddings.
• Implemented a layer-wise concatenation strategy for progressive enrichment of molecular representations.
• Evaluated the model on kinase and non-kinase datasets, comparing against GCN, Molformer, SVM, and XGBoost.

Main Results:

• The hybrid model achieved a superior accuracy of 88.7%, outperforming standalone GCN (86.8%), Molformer (85.1%), XGBoost (85.0%), and SVM (84.7%).
• The layer-wise injection of LLM embeddings progressively enriched molecular representations with global chemical context.
• LLM embeddings can be precomputed, ensuring computational efficiency during training and inference.

Conclusions:

• The proposed hybrid GCN-LLM architecture significantly enhances virtual screening performance.
• This approach offers a more accurate and efficient method for identifying promising drug candidates.
• The model's effectiveness across diverse datasets highlights its potential for broad application in drug discovery.