Artificial intelligence driven de novo molecular design for nuclear receptor ligand discovery

Instances of artificial intelligence (AI) coupled with the availability of large chemical and biological datasets enable data-driven drug discovery1 . We have applied a generative machine learning strategy based on a deep recurrent neural network (RNN) for de novo molecular design2 . The computational model was first trained to capture the grammar of SMILES representations of drug-like small molecules, and then used to automatically generate SMILES strings of new chemical entities (NCEs). Transfer learning on small collections of bioactive templates enables finetuning of the model to generate target-focused sets of molecules. In a pioneering prospective study3 the generative RNN was trained on small molecules (540'000) from a public compound database (ChEMBL22) and fine-tuned on a set of 25 fatty acid mimetics with known activity on two nuclear receptors (RXR, PPAR). The computationally designed samples from this model resembled drug-like molecules and comprised favourable synthetic accessibility. Five top-ranked examples were selected for synthesis and biological characterization. Four designs were active on the studied nuclear receptors in specific hybrid reporter gene assays with up to nanomolar potencies (EC50 0.13 - 14 µM), resembling the activities of the fine-tuning set (EC50 0.024 - 31 µM). These results confirm the potential of AI-driven de novo design for the discovery of synthetically accessible and bioactive NCEs as lead compounds for medicinal chemistry. After this successful proof-of-concept application3, we studied the potential of generative AI models for the de novo design of bioactive natural product mimetics4 . The AI-algorithm was fine-tuned on small collections of RXR activating natural products. It generated synthetically accessible NCEs populating an unexplored chemical space at the interface between drug-like molecules (used for training) and natural products4 (used for fine-tuning). Again, top-ranked computational samples were synthesized and two out of four were active in vitro with similar potencies (EC50 16 - 27 µM) as the fine-tuning set (EC50 2.1 - 43 µM) confirming that the designs inherited the bioactivity profile of the natural product templates. Our results highlight generative AI as valuable data-driven tool for medicinal chemistry to obtain synthetically accessible and innovative NCEs5 that inherit properties and bioactivity of a template collection without the need of explicitly including molecule design rules. 1. Gawehn, E. et al.: Mol. Inf. 2016, 35, 3–14 2. Gupta, A. et al.: Mol. Inf. 2018, 37, 1700111. 3. Merk, D. et al.: Mol. Inf. 2018, 37, 1700153. 4. Merk, D. et al.: J. Med. Chem. 2018, 10.1021/acs.jmedchem.8b00494. 5. Schneider, G.: Nat. Rev. Drug Discov. 2018, 17, 97–113