An overview of neural networks for drug discovery and the inputs used

ABSTRACT

Introduction: Artificial intelligence systems based on neural networks (NNs) find rules for drug discovery according to training molecules, but first, the molecules need to be represented in certain ways. Molecular descriptors and fingerprints have been used as inputs for artificial neural networks (ANNs) for a long time, while other ways for describing molecules are used only for storing and presenting molecules. With the development of deep learning, variants of ANNs are now able to use different kinds of inputs, which provide researchers with more choices for drug discovery.

Areas covered: The authors provide a brief overview of the applications of NNs in drug discovery. Combined with the characteristics of different ways for describing molecules, corresponding methods based on NNs provide new choices for drug discovery, including de novo drug design, ligand-based drug design, and receptor-based drug design.

Expert opinion: Various ways for describing molecules can be inputs of NN-based models, and these models achieve satisfactory results in metrics. Although most of the models have not been widely applied and tested in practice, they can be the basis for automatic drug discovery in the future.

KEYWORDS:

• Artificial neural networks
• deep learning
• de novo drug design
• drug discovery
• ligand-based drug design
• receptor-based drug design

Article highlights

• With descriptors and fingerprints as inputs, ANNs have been widely applied and they can be further improved. DNNs improve models with additional functions, and DL can be used to automatically produce fingerprints and more effective features.

• Applications of molecular graphs are not limited to LBDD, as de novo drug design and RBDD can also be realized with the application.

• With sequences of molecules as inputs, DL, RL, and TL can be used for de novo drug design. The sequences of molecules can also be encoded into new representations by NNs.

• With 3-D grids as inputs, NNs can help analyze docking results and predict receptor–ligand interactions. Even models based on 2-D grids of molecules achieve satisfactory results, and intuitive information of molecules can also be appropriate inputs.

• Most models have been tested on existing datasets in metrics, but their performance on real problems is still unknown.

Declaration of interest

The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Additional information

Funding

This work was supported by funds from the National Key R&D Program of China [2018YFC0311004] and the "Double First-Class" University Project, China [CPU2018GY15].