Deep Semi-supervised Learning for Virtual Screening Based on Big Data Analytics

Abstract

Nowadays, scientists and researchers, are facing the problem of massive data processing, which consumes relatively too much time and cost. That is why researchers have turned to Deep Learning (DL) techniques based on Big Data Analytics. On the other hand, the ever-increasing size of unlabelled data combined with the difficulty of obtaining class labels has made semi-supervised learning an interesting alternative of significant practical importance in modern data analysis. In the same context, drug discovery has reached a state and complexity that we can no longer avoid using Deep Semi-Supervised Learning and Big Data Processing Systems. Virtual Screening (VS) is a computationally intensive process which plays a major role in the early phase of drug discovery process. The VS has to be made as fast as possible to efficiently dock the ligands from huge databases to a selected protein receptor. For these reasons, we propose a deep semi-supervised learning-based algorithmic framework named DeepSSL-VS for pre-filtering the huge set of ligands to effectively do virtual screening for the breast cancer protein receptor. The latter combines stacked autoencoders and deep neural network and is implemented using the Spark-H2O platform. The proposed technique has been compared to twenty-four different machine learning algorithms applied all on the same reference datasets, and preliminary performance assessment results have shown that our approach outperforms these techniques with an overall accuracy performance more than 99%.