SAMPN

SAMPN predicts molecular properties, specifically lipophilicity and aqueous solubility, using a self-attention based message passing neural network operating on chemical graphs.

Key Features:

• Graph-Neural-Network Framework: Uses a graph neural network that takes chemical graphs as input to capture molecular structure for property prediction.
• Self-Attention Mechanism: Implements self-attention during message passing to quantify each atom's contribution to target properties.
• Interpretability and Visualization: Produces chemically interpretable atom-level importance scores via the attention mechanism to visualize structure–property relationships.
• Performance Superiority: Demonstrates superior predictive performance compared to random forests and the MPN implementation from DeepChem.
• Multi-SAMPN Formulation: Provides a Multi-SAMPN variant for simultaneous prediction of multiple chemical properties.

Scientific Applications:

• Drug discovery and development: Predicts lipophilicity and aqueous solubility to inform rational compound design and selection.
• Compound analysis and optimization: Enables multi-property assessments to support compound analysis and optimization workflows.
• Materials science: Supports materials research by predicting solubility and lipophilicity relevant to material function.

Methodology:

SAMPN applies a self-attention based message passing neural network on chemical graphs, and Multi-SAMPN extends this architecture to simultaneous multi-property prediction.