This repo contains a Python implementation for training a neural network for predicting drug-protein interactions using Keras and Tensorflow.
The objective is to create a model that is able to predict which xenobiotics (drugs) can be catalyzed by Cytochrome P450 enzymes in the human body. Using only the string representations of the drugs and the protein sequences, I am able to extract quantitative features from both types of molecules, which I later represent as a column vector that will be used as the input to a Deep Neural Network. The network achieves accuracies close to 80% using a batch size of 10 and 150 epochs.
This repo contains 5 Notebooks that implement a Neural Network trained on drug-protein interactions from a subset of enzymes (Cytochrome P450) and drugs. The networks is trained to recognize substrates that are transformed by CYP450 enzymes.
Each notebooks is numbered according to the order in which they must be executed and contains relevant documentation for each step. The first Notebooks (1-4) deal with downloading, cleaning, and pre-processing the data using the KEGG API and Mole DB.
The last Notebook shows the implementation of the Neural Network and how to visualize the results