Predicting Small Molecule Potency to Inhibit Estrogen Receptors using Machine Learning and Deep Learning Approaches
Hetal Rahul Rajpura, Alioune Ngom · 2018
Uncovering new therapeutic potentials of existing approved drugs is an accelerated process of drug discovery as compared to designing a new drug from scratch. Launching a new drug into the market is challenging in terms of laborious efforts, time, cost and risks attached. Identifying the potency of the drug in terms of their binding affinity offers a new avenue of research concerning faster and cheaper health-care solution. In this regards, predicting the binding affinity of various drugs to a specific target receptor-like Estrogen receptor can lead to deeper insights. Estrogen receptor plays a significant role in diseases like breast cancer, ovarian cancer and endometrial cancer. Computational techniques like advanced deep learning models have shown effective results with complex data. In the proposed model, we construct a deep neural network with open-source Tensorflow python package to predict the binding affinities of small molecules with respect to Estrogen receptors. Small molecules are represented as feature vectors comprising of binding affinities to the other targets within the dataset. Based on the behaviour of a compound to inhibit the rest of the targets, we predict its potential binding affinity to bind to the estrogen receptors. The linear regression model is trained on the binding data from BindingDB database consisting of 7962 small molecules and a unique set of 995 target receptors. The performance of simple linear regression technique is compared to the deep neural network based linear regression estimator function in terms of mean squared error estimates. Better performance obtained with deep learning approach indicates that these advanced techniques in the domain of artificial intelligence should be further investigated for drug-target binding affinity prediction. Sources to reproduce the analysis are available at https://github.com/hetalraj/Bindingaffinity.