metabolic pathway prediction

metabolic pathway prediction predicts metabolic pathway classes for biochemical compounds by extracting molecular shape features from SMILES using graph convolutional networks (GCNs) and classifying pathways with random forest (RF) classifiers.

Key Features:

• Graph Convolutional Networks (GCNs): Extracts molecular shape features directly from SMILES representations of chemical structures.
• Random Forest Classifier: Uses GCN-derived features as inputs to an RF classifier to assign metabolic pathway classes.
• High Predictive Accuracy: Reports single-class prediction accuracy of 95.16% and multi-label prediction accuracy of 97.61%, exceeding competing methods at 84.92% or less.
• Multi-Label Classification: Supports multi-label predictions where compounds may belong to multiple pathway classes.
• Physicochemical Feature Analysis: Demonstrates that global physicochemical features can be predicted from GCN-extracted shape features using linear or logistic regression models.

Scientific Applications:

• Reaction Predictor Design: Informs design of reaction predictors by mapping structural features of molecules to metabolic pathway assignments.
• Metabolic Research: Enables exploration of metabolic processes and identification of pathway involvement for novel compounds within cellular biochemistry studies.

Methodology:

Input SMILES are processed through a GCN to capture molecular shape features, these features are used by an RF classifier to predict metabolic pathways, the framework supports single-class and multi-label predictions, and linear/logistic regression is used to predict global physicochemical features from shape features.