KnoMol: A Knowledge-Enhanced Graph Transformer for Molecular Property Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in drug discovery

Background:

• Molecular property prediction (MPP) is crucial for efficient drug development, but current deep learning models struggle with representational capacity.
• Existing methods often require large datasets, posing challenges in data-scarce scenarios.

Purpose of the Study:

• To develop a novel knowledge-based Transformer framework, KnoMol, to enhance molecular structure understanding and improve MPP accuracy.
• To address the challenge of data scarcity in deep learning models for MPP by integrating expert chemical knowledge.

Main Methods:

• Developed KnoMol, a Transformer framework incorporating expert chemical knowledge and a multiperspective attention mechanism.
• Evaluated KnoMol on benchmark datasets (MoleculeNet) and small-scale data, comparing its performance against existing models.

Main Results:

• KnoMol achieved state-of-the-art performance, outperforming existing models in accuracy and generalization on MPP tasks.
• The integration of knowledge significantly reduced KnoMol's reliance on data volume, mitigating data scarcity issues.
• KnoMol successfully identified novel HER2 inhibitors in a case study, showcasing its practical applicability.

Conclusions:

• KnoMol offers a powerful and data-efficient tool for molecular property prediction, advancing computer-aided drug discovery.
• This research establishes a successful precedent for embedding domain knowledge into Transformer models, benefiting the broader field of MPP algorithm development.