Application of Computational Biology and Artificial Intelligence in Drug Design

Abstract

Traditional drug design requires a great amount of research time and developmental expense. Booming computational approaches, including computational biology, computer-aided drug design, and artificial intelligence, have the potential to expedite the efficiency of drug discovery by minimizing the time and financial cost. In recent years, computational approaches are being widely used to improve the efficacy and effectiveness of drug discovery and pipeline, leading to the approval of plenty of new drugs for marketing. The present review emphasizes on the applications of these indispensable computational approaches in aiding target identification, lead discovery, and lead optimization. Some challenges of using these approaches for drug design are also discussed. Moreover, we propose a methodology for integrating various computational techniques into new drug discovery and design.

Keywords: artificial intelligence-aided drug design (AIDD); computational biology; computer-aided drug design (CADD); deep learning.

Figure 1

The process of drug research and development. The details in drug development have been improved in the past forty years. Nowadays, the complete process in drug research includes drug discovery, clinical testing, and approval for production. The process of drug discovery includes target identification, lead discovery, lead optimization, and preclinical testing. This process usually takes 7–10 years and $600 M–$800 M. Then approximately 200 compounds enter preclinical testing step while about 5 compounds enter clinical testing process. This process includes three steps, phase I, II and III clinical trials, respectively. It is a long and expensive process that costs 6–12 years and billions of dollars. The compounds that have passed clinical testing enter the process of approval for production. Approved compounds by FDA/EMA can commercialize on the market. This process takes 1–2 years and about $50 M.

Figure 2

The workflow of structure-based drug design (SBDD) and ligand-based drug design (LBDD). For SBDD, it starts with target identification. Then, binding site of the target requires identifying and compound library needs to be prepared. Next, dock each compound from the library into the identified binding site evaluate the score. In molecular docking, MD simulations can be utilized to obtain more flexible target and rescore for the docking process. Additionally, MD simulations can be applied to lead optimization through ligand-target interactions. Through these steps, leads are obtained primarily. For LBDD, it starts with known ligands with bioactivity. Then, extracting the chemical features of these ligands and build pharmacophore or QSAR model. Next, according to the information of known ligands (e.g., ligand similarity), ligand-based virtual screening is performed in the compound library and leads are screened. These leads are further optimized in wet and dry lab.

Figure 3

Overview of the machine learning-based de novo drug design procedure, from left-top to right-bottom: appropriate data selection; data filtering and classification; molecular data and de-sired properties storage; feature representation for molecules and properties; molecule generation by machine learning methods; generation model optimization by reinforcement learning strategy and property prediction models; de novo molecules generation.

Figure 4

Example of a message passing neural network. The left side is an example of a graphical molecule. Ten atoms are regarded as nodes, with each node connecting with one or more nodes. The right side shows the message passing procedure for the target node I through a multilayer neural network (or one layer). The information from the directly connected nodes J, C, and D is passed to the target node I. As nodes J, C, and D have their own directly connected nodes (G, I, A, E), message is also passed by each of their neural networks. Message passing between nodes in a graph is a circular iteration process.

Figure 5

Variational autoencoder architecture. It consists of an encoder and a decoder, and they are deep neural networks. In general, the encoder maps input molecular data x into latent codes z by parameterizing a posterior distribution q_Ø(z|x), and the decoder reconstructs molecular data from the learned distribution p_θ(x|z).

Figure 6

Flowchart of drug design using GANs. It contains a generator and a discriminator, and they are all deep neural networks. The generator transforms latent vectors that are sampled from a prior distribution such as Gaussian into novel molecular data samples, and the discriminator distinguishes fake molecular data generated by the generator from the actual points sampled from the distribution of training data and gives feed-back.

Figure 7

Architecture of normalizing flows. It contains a series of invertible functions for transforming molecular data into simple distribution and converting the distribution into high-dimensional molecular data for de novo drug design. Each of the functions optimizes the data distribution.