MMPCS: multi-view molecular pretraining based on consistency information and specific information.

Area of Science:

• Cheminformatics
• Drug Discovery
• Machine Learning

Background:

• Molecular representation learning automates feature extraction for cheminformatics and drug discovery.
• Integrating multiple molecular data views (SMILES, 2D graphs, 3D conformations) remains a challenge for superior representations.

Purpose of the Study:

• To propose a novel multi-view molecular pretraining method, MMPCS, to address the challenge of integrating diverse molecular data views.
• To explicitly factorize molecular representations into consistency and view-specific information for enhanced downstream task performance.

Main Methods:

• Utilized Graph Isomorphism Network for 2D molecular graphs and RoBERTa for SMILES sequences.
• Decomposed molecular embeddings into shared consistency and view-specific components, aligned via an autoencoder.
• Combined consistency and view-specific representations for task-aware predictions.

Main Results:

• MMPCS achieved superior average performance across 16 state-of-the-art methods in molecular property prediction (classification and regression).
• Demonstrated outstanding results in predicting drug-target binding affinity and cancer drug response.
• Case study on SARS-CoV-2 Omicron variant highlighted potential for drug repurposing.

Conclusions:

• MMPCS effectively integrates multi-view molecular data, yielding improved representations.
• The method shows robustness and broad applicability in drug discovery and development.
• MMPCS facilitates precise and task-aware predictions for various cheminformatics applications.