GeminiMol: Conformational Space Profiling for Molecular Representation

Abstract

GeminiMol enhances generic molecular representation for AI-powered ligand-based drug discovery by profiling the conformational space of molecules. Instead of relying on single static molecular structures, GeminiMol systematically explores and encodes the three-dimensional conformational diversity of molecules, capturing the dynamic nature of molecular flexibility that is critical for biological activity.

Method

GeminiMol introduces a novel approach to molecular representation:

• Conformational Space Profiling: Systematically generates and encodes multiple 3D conformations of each molecule
• Multi-Conformer Encoding: Uses deep learning to aggregate information from diverse conformations into a single comprehensive representation
• Similarity Learning: Trains the model to capture molecular similarities that are relevant to biological activity through contrastive learning
• Generic Representation: The learned representation transfers effectively across different downstream tasks

Key Results

• State-of-the-Art Performance: Achieves top results on multiple molecular property prediction benchmarks
• Ligand-Based Drug Discovery: Significantly improves virtual screening performance for various drug targets
• Conformational Sensitivity: Captures activity-relevant conformational information that static representations miss
• Broad Applicability: Effective across diverse chemical spaces and drug discovery tasks

BibTeX

@article{wang2024conformational,
  title={Conformational Space Profiling Enhances Generic Molecular Representation for AI-Powered Ligand-Based Drug Discovery},
  author={Wang, Lin and Wang, Shihang and Yang, Hao and Li, Shiwei and Wang, Xinyu and Zhou, Yongqi and Tian, Siyuan and Liu, Lu and Bai, Fang},
  journal={Advanced Science},
  volume={11},
  number={40},
  pages={2403998},
  year={2024},
  publisher={Wiley Online Library}
}