Metabolic-Pathway-Prediction

The understanding of processes and structural mappings between molecules and route classes is required for designing reaction predictors for the synthesis of new compounds. This project effort addresses the topic of determining types of metabolic pathways in which a particular biological molecule participates. To predict the metabolic pathway classes of the compounds, we use a hybrid graph-based deep learning model called Graph Convolutional networks (GCN). Using RDKit, our model pulls crucial shape information directly from input SMILES representations, which are atom-bond specifications of the chemical structures that make up the molecules, unlike previously utilised machine learning techniques for this topic. Unlike earlier route prediction approaches that did not naturally extend to multi-class classification, our design allows for multi-classification of substances into several pathway classes. To allow multi-class categorization, we made suitable adjustments to the GCN architecture. We focus on the features that are most essential in determining route groups in this study. We also discovered that the shape features generated by the GCN architecture can predict the values of the top-ranked features for a given molecule, indicating that the GCN approach can be used in chemical classification applications as a data-driven alternative to time-consuming expert-driven feature engineering. Finally, we create a web application that allows users to submit a query compound and see the results of our trained GCN model's class prediction.

Built With

• jupyter-notebook
• procfile
• python