AutoMSR: Auto Molecular Structure Representation Learning for Multi-label Metabolic Pathway Prediction

Area of Science:

• Bioinformatics
• Computational Chemistry
• Drug Discovery

Background:

• Understanding metabolic and molecular pathways is crucial for drug development and optimizing drug metabolism.
• Multi-label prediction of metabolic pathways using bioinformatics is a key approach to decipher these relationships.
• Graph Neural Networks (GNNs) show promise in extracting molecular structural features for pathway prediction.

Purpose of the Study:

• To develop an automated framework, AutoMSR, for designing optimal GNN models for molecular structure representation learning.
• To eliminate the need for manual GNN architecture design and hyperparameter tuning in metabolic pathway prediction.
• To enhance the efficiency and accuracy of predicting metabolic pathways from molecular data.

Main Methods:

• Designed an end-to-end automatic molecular structure representation learning framework named AutoMSR.
• Proposed a multi-seed age evolution (MSAE) algorithm to search for optimal GNN architectures.
• Utilized a tree-structured Parzen estimator for efficient hyperparameter optimization.

Main Results:

• AutoMSR successfully automates the selection of GNN architectures and hyperparameters for metabolic pathway prediction.
• The framework was evaluated on the KEGG dataset, demonstrating superior performance compared to baseline methods.
• Achieved improved results across various multi-label classification metrics.

Conclusions:

• AutoMSR provides an effective, automated solution for molecular structure representation learning in bioinformatics.
• The framework significantly advances the prediction of metabolic pathways, aiding in drug discovery and optimization.
• Automated GNN design accelerates research and reduces reliance on expert experience in the field.