MolPLA: a molecular pretraining framework for learning cores, R-groups and their linker joints

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Molecular core structures and R-groups are crucial in drug development.
• Integrating these concepts into graph pre-training can enhance molecular understanding.

Purpose of the Study:

• To introduce MolPLA, a novel pre-training framework for understanding decomposable molecular parts.
• To enable MolPLA to assist chemists in identifying replaceable R-groups during lead optimization.

Main Methods:

• MolPLA employs masked graph contrastive learning.
• A secondary framework is formulated for R-group replacement suggestions.

Main Results:

• MolPLA achieves predictability comparable to state-of-the-art models in molecular property prediction.
• Qualitative analysis confirms MolPLA's ability to distinguish core and R-group substructures.
• MolPLA aids lead optimization by suggesting R-group replacements for given core templates.

Conclusions:

• MolPLA effectively learns molecular core structures and R-groups.
• The framework shows promise in accelerating drug discovery and lead optimization processes.