Machine Learning for drug discovery in dopamine receptors

About

This project focuses on the field of biochemistry and drug discovery. Using machine learning technique of QSAR (Quantitative Structure-Activity Relationship) modeling we explored the relationships between biological activities of the drugs to dopamine receptors, in relation with their chemical structural properties. This helps us discover new paths of drugs to test on human experiments with better accuracy, precision, and recall. 

Contribution

The assay data is collected, analyzed, and classified using machine learning and neural network methods. Different ways of exploring the chemical structures like Morgan Fingerprints and Atom tokenization were attempted. High accuracy, precision and recall were achieved for classifying the drugs’ bioactivity class using their chemical structures and fingerprints.

Key Findings

Using machine learning and neural network models, were were able to predict the bioactivity classes of chemical compounds and their chemical structures useful in inhibiting dopamine receptors. Among machine learning models, ensemble models worked well in predicting, and in neural networks, RNN models performed the best. In predicting chemical fingerprints, RNNs with Bi-LSTM achieved about 72% accuracy. This can be used as baseline for future exploration in predicting chemical fingerprints.