Exploration of Chemical Space with Partial Labeled Noisy Student Self-Training for Improving Deep Learning Performance: Application to Drug Metabolism

Motivation Drug discovery is time-consuming and costly. Machine learning, especially deep learning, shows a great potential in accelerating the drug discovery process and reducing its cost. A big challenge in developing robust and generalizable deep learning models for drug design is the lack of a large amount of data with high quality and balanced labels. To address this challenge, we developed a self-training method PLANS that exploits millions of unlabeled chemical compounds as well as partially labeled pharmacological data to improve the performance of neural network models. Result We evaluated the self-training with PLANS for Cytochrome P450 binding activity prediction task, and proved that our method could significantly improve the performance of the neural network model with a large margin. Compared with the baseline deep neural network model, the PLANS-trained neural network model improved accuracy, precision, recall, and F1 score by 13.4%, 12.5%, 8.3%, and 10.3%, respectively. The self-training with PLANS is model agnostic, and can be applied to any deep learning architectures. Thus, PLANS provides a general solution to utilize unlabeled and partially labeled data to improve the predictive modeling for drug discovery. Availability The code that implements PLANS is available at https://github.com/XieResearchGroup/PLANS