Portal de Periódicos da CAPES

Neural network and deep-learning algorithms used in QSAR studies: merits and drawbacks

2018; Elsevier BV; Volume: 23; Issue: 10 Linguagem: Inglês

10.1016/j.drudis.2018.06.016

1878-5832

Fahimeh Ghasemi, Alireza Mehridehnavi, Alfonso Pérez, Horacio Pérez‐Sánchez,

Spectroscopy and Chemometric Analyses

• In this article, historical challenges of neural networks is reviewed in the QSAR studies. • An overview of the related deep learning works in chemoinformatics are presented. • After that, drawbacks of neural networks in QSAR studies are considered. • Future perspective of deep learning algorithms is investigated. • Finally, it is tried to provide comparison between neural network and deep learning in drug discovery. The past two decades are regarded as the golden age of using neural networks (NNs) in chemoinformatics. However, two major issues have arisen concerning their use: redundancy problems when dealing with small data sets, and the large number of compounds with thousands of descriptors, which gives rise to serious overfitting problems. Various NN algorithms, based on feature selection methods and learning algorithms, were devised to avoid these predicaments in drug discovery. Pruning the overfitting problem has emerged as another challenge in recent years, leading to the advent of deep-learning (DL) networks using innovative techniques. Here, we discuss the advantages and disadvantages of the proposed NN algorithms, especially the innovative DL techniques used in ligand-based virtual screening (VS).