Employing Molecular Conformations for Ligand-Based Virtual Screening with Equivariant Graph Neural Network and Deep...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Ligand-based virtual screening (LBVS) is crucial for early-stage drug discovery.
• Current deep learning LBVS methods often overlook 3D molecular conformations, which significantly impact bioactivity.
• A lack of comprehensive bioactivity benchmark datasets hinders deep learning model development.

Purpose of the Study:

• To develop a novel deep learning framework for LBVS that utilizes 3D molecular conformers.
• To create a large-scale bioactivity benchmark dataset for training and validation.
• To enhance the accuracy and efficiency of virtual screening in drug discovery.

Main Methods:

• Extraction of molecular conformers from public bioactivity data to build a benchmark dataset (millions of endpoints, 954 targets).
• Development of EquiVS, a deep learning architecture using Graph Convolutional Networks (GCN) and Equivariant Graph Neural Networks (EGNN).
• Application of attention-based deep multiple-instance learning (MIL) for bioactivity prediction from aggregated molecular representations.

Main Results:

• The constructed benchmark dataset demonstrated high quality and utility.
• EquiVS outperformed 10 traditional and deep learning-based LBVS methods in performance.
• Ablation studies confirmed the importance of molecular conformation and the effectiveness of the EquiVS architecture.

Conclusions:

• The developed benchmark dataset and EquiVS method show significant promise for virtual screening applications.
• Incorporating 3D molecular conformation is vital for accurate bioactivity prediction.
• EquiVS offers a powerful tool for identifying potential drug candidates and discovering optimal conformers.