Geometry-based BERT: An experimentally validated deep learning model for molecular property prediction in drug...

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular representation learning

Background:

• Deep learning significantly impacts drug discovery, but identifying novel active compounds remains challenging.
• Existing methods struggle with characterizing complex molecular structures and their 3D conformations.
• There is an urgent need for advanced deep learning techniques to accelerate early-stage drug discovery.

Purpose of the Study:

• To introduce Geometry-based Bidirectional Encoder Representations from Transformers (GEO-BERT), a self-supervised framework for molecular property prediction.
• To enhance molecular structure characterization by integrating 3D conformational information and novel positional relationships.
• To demonstrate the practical utility of GEO-BERT in identifying novel drug candidates.

Main Methods:

• Developed GEO-BERT, a self-supervised representation learning framework using Transformers.
• Input includes atomic and chemical bond information, alongside 3D molecular conformation.
• Integrated atom-atom, bond-bond, and atom-bond positional relationships to improve structural characterization.

Main Results:

• GEO-BERT achieved optimal performance across multiple benchmark datasets for molecular property prediction.
• A prospective study screened for DYRK1A inhibitors using the GEO-BERT model.
• Identified two novel and potent DYRK1A inhibitors with IC50 values below 1 μM.

Conclusions:

• GEO-BERT is an effective open-source tool for molecular property prediction, advancing early-stage drug discovery.
• The framework's ability to characterize molecular structures and predict properties is validated.
• GEO-BERT demonstrates significant potential for accelerating the discovery of novel therapeutic agents.