A machine learning approach towards the prediction of protein–ligand binding affinity based on fundamental molecular properties

There is an exigency of transformation of the enormous amount of biological data available in various forms into some significant knowledge. We have tried to implement Machine Learning (ML) algorithm models on the protein–ligand binding affinity data already available to predict the binding affinity of the unknown. ML methods are appreciably faster and cheaper as compared to traditional experimental methods or computational scoring approaches. The prerequisites of this prediction are sufficient and unbiased features of training data and a prediction model which can fit the data well. In our study, we have applied Random forest and Gaussian process regression algorithms from the Weka package on protein–ligand binding affinity, which encompasses protein and ligand binding information from PdbBind database. The models are trained on the basis of selective fundamental information of both proteins and ligand, which can be effortlessly fetched from online databases or can be calculated with the availability of structure. The assessment of the models was made on the basis of correlation coefficient (R2) and root mean square error (RMSE). The Random forest model gave R2 and RMSE of 0.76 and 1.31 respectively. We have also used our features and prediction models on the dataset used by others and found that our model with our features outperformed the existing ones.