MSMPP: Molecular Property Prediction by Integrating Multi-scale Multi-view information with pretrained 3D molecular...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Current deep learning models for molecular property prediction often use a single molecular view, neglecting multi-dimensional features and inter-molecular information.
• Limited labeled data hinders the generalization of existing computational methods to novel molecules.
• Conventional drug discovery is characterized by high costs, long development cycles, and low success rates.

Purpose of the Study:

• To propose MSMPP, a multi-scale, multi-view fusion framework for molecular property prediction.
• To address limitations of existing methods by integrating intra- and inter-scale molecular features.
• To enhance the generalization capability of models for novel molecules and accelerate drug discovery.

Main Methods:

• MSMPP integrates 1D sequence, 2D topological graph, and 3D conformational features using TxGemma, Graph Transformer, and Uni-Mol.
• It learns intra-scale features by combining representations from pretrained models and Graph Transformer.
• Inter-scale feature learning involves constructing an inter-molecular graph (IMG) to model pairwise interactions and extract cross-task features.

Main Results:

• MSMPP significantly outperforms state-of-the-art models on eight MoleculeNet datasets.
• The framework effectively integrates multi-view intra-molecular, inter-molecular, and cross-task information.
• Demonstrated improved generalization to novel molecules through task-agnostic prior knowledge.

Conclusions:

• MSMPP offers a competitive tool for molecular property prediction, enhancing computational drug discovery.
• The multi-scale, multi-view approach effectively captures complex molecular information.
• This framework supports the acceleration of drug discovery workflows by improving prediction accuracy and generalization.