Molecular Property Prediction Based on Motif-Centric Multi-Grain Pretaining and Finetuning Strategy.

Area of Science:

• Computational chemistry
• Machine learning for drug discovery

Background:

• Accurate molecular property prediction is vital for drug discovery and retrosynthesis.
• Existing Graph Neural Network (GNN) methods struggle with detailed chemical structures and multi-level graph relationships.

Purpose of the Study:

• To develop an advanced framework, MMGSF, for improved molecular property prediction.
• To enhance the capture of intricate chemical information within molecular graphs.

Main Methods:

• Introduced the Multi-level Molecular Graph Self-supervised Learning (MMGS) component with a hierarchical GNN encoder.
• Implemented Multi-grain Finetuning (MGF) using a novel mol-adapter with cross-attention for adaptive feature fusion.
• Designed pretraining tasks focusing on inter-node relationships across multiple graph levels.

Main Results:

• MMGSF demonstrated superior performance in molecular property classification tasks.
• Visualization confirmed that learned representations effectively capture chemical semantics and properties.
• The framework successfully integrates structural and semantic information from different grains.

Conclusions:

• MMGSF offers a significant advancement in self-supervised learning for molecular property prediction.
• The proposed methods enhance the understanding of complex molecular features and relationships.
• This work provides new directions for designing effective GNN-based predictive models.