De novo generation of dual-target compounds using artificial intelligence

Abstract

Drugs that interact with multiple therapeutic targets are potential high-value products in polypharmacology-based drug discovery, but the rational design remains a formidable challenge. Here, we present artificial intelligence (AI)-based methods to design the chemical structures of compounds that interact with multiple therapeutic target proteins. The molecular structure generation is performed by a fragment-based approach using a genetic algorithm with chemical substructures and a deep learning approach using reinforcement learning with stochastic policy gradients in the framework of generative adversarial networks. Using the proposed methods, we designed the chemical structures of compounds that would interact with two therapeutic targets of bronchial asthma, i.e., adenosine A2a receptor (ADORA2A) and phosphodiesterase 4D (PDE4D). We then synthesized 10 compounds and evaluated their bioactivities via the binding assays of 39 target human proteins, including ADORA2A and PDE4D. Three of the 10 synthesized compounds successfully interacted with ADORA2A and PDE4D with high specificity.

Keywords: Bioinformatics; Biological sciences; Natural sciences; Pharmacoinformatics.

Graphical abstract

Figure 1

Overview of the proposed method for generating dual-target compounds (A) Construction of quantitative structure-activity relationship (QSAR) models for predicting the compound bioactivities on different therapeutic targets and generating dual-target compound candidates. Incorporation of bioactivity-value prediction models into (B) the fragment-based structure generator DualFASMIFRA and (C) the deep learning-based structure generator DualTransORGAN.

Figure 2

Molecular generation by the proposed structure generators (A and B) Scatterplots of the predicted versus observed pIC₅₀ values using the QSAR models for (A) ADORA2A and (B) PDE4D. (C) Ratios of elite compounds with high scores in the QSAR models, diverse (randomly selected) compounds, and mutant compounds (positional scan analogs of the merged set of elite and random molecular population) after one iteration of the genetic algorithm in DualFASMIFRA. (D) Optimization trajectories of the objective functions in DualFASMIFRA. The horizontal axis indicates the generation number of the genetic algorithm, and the vertical axis indicates the predicted molecular weight invariant, pIC₅₀ (averaged between the results of ADORA2A and PDE4D). (E) Distributions of molecular weights of the generated compounds (green) and compounds in the training dataset (gray) applied to DualTransORGAN. The horizontal and vertical axes indicate the molecular weights of the compounds and the densities of the compounds with the corresponding molecular weights, respectively. (F) Optimization trajectories of the objective functions in DualTransORGAN. The maximum (max), average (avg), and minimum (min) values are shown.

Figure 3

Synthesis of the AI-generated compounds (A) Computationally generated compounds 5, 9, 13, and 16 are predicted to interact with both ADORA2A and PDE4D. Schemes A, B, C, and D show the synthesis routes of compounds 5, 9, 13, and 16, respectively. Compound 6 is a derivative of 5. (B) The chemical structures of the synthesized 10 compounds and the corresponding compound numbers in (A) are shown.

Figure 4

List of target proteins in the binding assays and hit distributions of synthesized compounds for the target human proteins 39 human target proteins including 24 GPCRs, 3 transporters, 3 ion channels, 1 kinase, 2 nuclear receptors, and 6 nonkinase enzymes from the viewpoints of drug safety and pharmacological actions are selected (A). The binding activities of the synthesized compounds 5, 6, 7, 8, 9, 11, 12, 13, 15, and 16 are evaluated on this list including ADORA2A and PDE4D (B). The rows and columns in the heatmap indicate human target proteins and synthesized compounds, respectively. The cells are indicated by the percent inhibition colors of each compound for the 39 proteins. Cells with >100% inhibition scores are highlighted in the darkest red, those with “90% ≤ × < 100%” inhibition in orange, “80% ≤ × < 90%” inhibition in yellow, “70% ≤ × < 80%” inhibition in yellow-green, “60% ≤ × < 70%” in green, “50% ≤ × < 60%” inhibition in light blue, and “< 50%” inhibition in white.

Figure 5

Experimental validation of synthesized compounds using binding assays of 39 target human proteins The bioactivities of the 10 synthesized compounds were evaluated on 39 human target proteins, including ADORA2A and PDE4D. (A–J) are the binding assay results of synthesized compounds 5, 6, 7, 8, 9, 11, 12, 13, 15, and 16, respectively. The bar plots show the percentage inhibition scores of each compound for the 39 proteins. The red bars highlight the proteins with significant responses (>50% inhibition in the binding assay). The associated inhibition scores are also shown.