Molecular property prediction based on graph contrastive learning with partial feature masking.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Molecular representation learning is crucial for tasks like molecular property prediction (MPP) and drug design.
• Self-supervised learning (SSL), particularly contrastive learning (CL), shows promise in addressing data scarcity in MPP by learning generalizable molecular knowledge.
• Generating semantically preserved augmented samples is a key challenge in CL for molecular data.

Purpose of the Study:

• To propose a novel contrastive learning framework, FMGCL, for enhanced molecular representation learning.
• To address the challenge of generating effective augmented molecular samples that retain core chemical semantics.
• To improve the performance of downstream tasks like MPP through better pre-trained molecular encoders.

Main Methods:

• Developed FMGCL, a partial feature masking-based graph contrastive learning model.
• Constructed masked molecular graphs by masking partial atom and bond features, preserving molecular structure and semantics.
• Incorporated relative sample distance within batches to enhance regression task performance.

Main Results:

• FMGCL demonstrated superior performance on 12 benchmark datasets from MoleculeNet and ChEMBL.
• The proposed partial feature masking strategy effectively preserves molecular semantics during augmentation.
• The method successfully captured valuable prior molecular knowledge during pre-training.

Conclusions:

• FMGCL offers a robust approach for molecular representation learning, outperforming existing methods.
• Preserving chemical semantics in augmented samples is vital for effective contrastive learning in cheminformatics.
• The FMGCL framework provides a promising direction for advancing machine learning applications in drug discovery and molecular property prediction.