Combining Generative Artificial Intelligence and On-Chip Synthesis for De Novo Drug Design

Abstract

<p>Automation of the molecular design-make-test-analyze cycle speeds up the identification of hit and lead compounds for drug discovery. Using deep learning for computational molecular design and a customized microfluidics platform for on-chip compound synthesis, liver X receptor (LXR) agonists were generated from scratch. The computational pipeline was tuned to explore the chemical space defined by known LXRα agonists, and to suggest structural analogs of known ligands and novel molecular cores. To further the design of lead-like molecules and ensure compatibility with automated on-chip synthesis, this chemical space was confined to the set of virtual products obtainable from 17 different one-step reactions. Overall, 25 <i>de novo</i> generated compounds were successfully synthesized in flow via formation of sulfonamide, amide bond, and ester bond. First-pass <i>in vitro</i> activity screening of the crude reaction products in hybrid Gal4 reporter gene assays revealed 17 (68%) hits, with up to 60-fold LXR activation. The batch re-synthesis, purification, and re-testing of 14 of these compounds confirmed that 12 of them were potent LXRα or LXRβ agonists. These results support the utilization of the proposed design-make-test-analyze framework as a blueprint for automated drug design with artificial intelligence and miniaturized bench-top synthesis.<b></b></p>

Artificial intelligence in drug discovery: recent advances and future perspectives

“…Other popular deeplearning-based generative learning models, such as variational autoencoders [139], or generative adversarial networks [140,141] have also been commonly reported, as well as others based on graph convolutions [142,143]. Recently, instances of conditional generative approaches have been suggested, which leverage additional information guiding the design, such as three-dimensional shape [144], drug-likeness [142], synthesizability [142,145], molecular descriptors values [146], and gene expression signatures [147]. A major upcoming challenge in this context will be the definition of balanced objective functions that enable complex and constrained multi-parameter optimizations, similarly to those used in Pareto [148] or in desirability-based approaches [149][150][151][152], that are typically required in drug discovery.…”

“…Mixed approaches combining rule-free and rule-based methods might represent a promising middle ground for the design of novel bioactive and synthesizable molecular entities. Recently, a mixed strategy showed promise in designing bioactive molecules in a rule-free manner, while at the same time retaining synthesizability within a microfluidics system, thanks to a set of predefined virtual reactions [145].…”

Leveraging molecular structure and bioactivity with chemical language models for drug design

“…CLMs have already been successfully employed to generate focused virtual chemical libraries. Examples of de novo designed bioactive molecules include inhibitors of vascular endothelial growth factor receptor 2 kinase and the unfolded protein response pathway 7 , and nuclear hormone receptor modulators [20][21][22][23] .…”

Perplexity-based molecule ranking and bias estimation of chemical language models

“…Generative deep learning has become a promising method for chemistry and drug discovery [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17]19,21 . Generative models learn the pattern distribution of the input data and generate new data instances based on learned probabilities 12 .…”

“…Generative models learn the pattern distribution of the input data and generate new data instances based on learned probabilities 12 . Among the proposed generative frameworks that have been applied to de novo molecular design [1][2][3][4][6][7][8][9][10][11][12][13][14][15][16][17]19,21 , chemical language models (CLMs) have gained particular attention because of their ability to generate focused virtual chemical libraries and bioactive compounds 18,19,21,22 . CLMs are trained on string representations of molecules, e.g., simplified molecular input line entry systems (SMILES) strings (Fig.…”