Artificial Intelligence in Drug Design

Abstract

Artificial Intelligence (AI) plays a pivotal role in drug discovery. In particular artificial neural networks such as deep neural networks or recurrent networks drive this area. Numerous applications in property or activity predictions like physicochemical and ADMET properties have recently appeared and underpin the strength of this technology in quantitative structure-property relationships (QSPR) or quantitative structure-activity relationships (QSAR). Artificial intelligence in de novo design drives the generation of meaningful new biologically active molecules towards desired properties. Several examples establish the strength of artificial intelligence in this field. Combination with synthesis planning and ease of synthesis is feasible and more and more automated drug discovery by computers is expected in the near future.

Keywords: artificial intelligence; de novo design; deep learning; neural networks; property prediction; quantitative structure-activity relationship (QSAR); quantitative structure-property prediction (QSPR).

Figure 1

Neurons are connected to each other. Incoming signals are multiplied by a weight. The output signal z_j is given by the sum of this products plus a bias transformed by an activation function. Examples of activation functions are graphically shown, like the rectified linear unit (ReLU) or the Gaussian function. For each neuron in the neural net, weights and biases need to be trained. Deep neural networks have several hidden layers with many neurons. The number of neurons typically varies between different layers.

Figure 2

Toxicophoric features identified from the Tox21 dataset by the neural network [8].

Figure 3

(A) Recurrent neural networks (RNNs) use sequential data. The output for the next element depends on the previous element. Thus, RNNs have a memory. h_i represent the hidden state at each neuron. They are updated based on the input x and the hidden state from the previous neuron. (B) In the UG-RNN approach, molecules are described as undirected graphs and fed into a RNN. Each vertex of a molecular graph is selected as a root node and becomes the endpoint of a directed graph. Output for all nodes is traversed along the graph until the root node is reached. All signals are summed to give the final output of the RNN, which enters into the NN for property training. (C) Graph convolutional models use the molecular graph. For each atom a feature vector is defined and used to pass on information for the neighboring atoms. In analogy to circular fingerprints different layers of neighboring atoms are passed through convolutional networks. Summation of the different atomic layers for all atoms results in the final vector entering the neural network for training.

Figure 4

(A) A variational autoencoder consists of two neural networks. The encoder network transforms the molecular description into a description vector, the latent space, while the decoder network is trained to translate a latent space vector into a molecule. (B) The adversarial autoencoder comprises a standard autoencoder, which learns to generate chemical structures. The discriminator network compares descriptions from a defined distribution to structures generated from the autoencoder.