TrimNet

TrimNet predicts molecular properties and identifies compound–protein interactions (CPIs) using a graph-based message passing neural network to improve molecular representation learning for computational drug discovery and biomedicine.

Key Features:

• Triplet Message Mechanism: Implements a triplet message mechanism within the message passing neural network to capture higher-order bond information.
• Lightweight Message Passing Neural Network: Uses a lightweight MPNN architecture that reduces model parameters relative to traditional approaches.
• Bond Information Extraction: Efficiently captures essential bond information from molecular graph data.
• Property and CPI Prediction: Learns representations to predict molecular properties and identify compound–protein interactions (CPIs).
• Atom-level Interpretability: Focuses on atoms critical to target properties to provide interpretable predictions.
• Empirical Performance: Demonstrates superior performance over previous state-of-the-art methods across various datasets.
• Dataset Evaluation: Evaluated on quantum and drug datasets from MoleculeNet.

Scientific Applications:

• Quantum Property Prediction: Predicts quantum characteristics of molecules as represented in MoleculeNet quantum datasets.
• Bioactivity Prediction: Predicts compound bioactivity for target identification and prioritization.
• Physiological Effect Prediction: Predicts physiological effects relevant to pharmacology and toxicology.
• Compound–Protein Interaction Identification: Identifies CPIs to support interaction mapping studies.
• Computational Drug Discovery and Biomedicine: Provides accurate, interpretable molecular representations to support drug discovery and biomedical research.
• Interpretability for Mechanistic Insight: Offers atom-focused explanations that aid mechanistic interpretation of predictions.

Methodology:

Learns molecular graph representations via a lightweight message passing neural network that implements a triplet message mechanism to capture bond information while reducing model parameters.