Molecular similarity searching based on deep belief networks with different molecular descriptors

Abstract

Molecular 2D similarity searching is one of the most widely used techniques for ligand-based virtual screening (LBVS). This study has used the concepts of deep learning by adapted deep belief networks (DBN) and data fusion concept with DBN to enhance the molecular similarity searching of chemical compounds in LBVS. The MDDR Datasets represented by different descriptors to convert the molecule shape to numerical values and each descriptor has different important features rather than the others. The DBN with data fusion is adapted to obtain a lower detection error probability and a higher reliability by using data from multiple distributed descriptors and analyzing the performance of combination and individual descriptors target by target and showed that the combination descriptor did better than both original descriptors. The overall results of this research showed that the use of DBN with data fusion in similarity-based is found to significantly outperform the conventional, industry-standard Tanimoto-based similarity search systems and some others benchmarks witch have been adapted by others researchers, with 1 % and 5% performance improvement in the average recall rates.