Research

Generative deep learning

AI-driven de novo molecule design

Generating molecules from scratch with bespoke properties is arguably one of the most challenging tasks in chemistry. In the past few years, generative deep learning has remarkably enhanced the field of de novo design, by allowing to generate novel bioactive molecules from scratch, without the need of human-engineered features and assembly rules. By learning from the underlying data distribution, methods like Recurrent Neural Networks (RNNs) can be used to generate novel molecules from scratch that possess the desired bioactivity profile. The Molecular ML group is currently active in boosting the potential of generative deep learning to support medicinal chemistry and drug discovery.

Selected Publications

Grisoni F, Huisman BJ, Button AL, Moret M, Atz K, Merk D, Schneider G. (2021). Combining generative artificial intelligence and on-chip synthesis for de novo drug design. Science advances 7 , e3338. DOI: 10.1126/sciadv.abg3338t

Moret M, Helmstädter M, Grisoni F, Schneider G, Merk D (2021). Beam search for automated design and scoring of novel ROR ligands with machine intelligence. Angewandte Chemie International Edition 60, 19477. DOI: 10.1002/anie.202104405

Moret M, Grisoni F, Katzberger P, Schneider G (2021). Perplexity-based molecule ranking and bias estimation of chemical language models. ChemRxiv, under review.

Moret M, Grisoni F, Brunner C, Schneider G (2021). Leveraging molecular structure and bioactivity with chemical language models for drug design. ChemRxiv, under review.

Molecular property prediction

Molecule prioritization with AI

Machine learning approaches can be used to predict various molecular properties simultaneously, and used to screen millions of commercially available compounds. In the Molecular Machine Learning team, we are interested in boosting the potential of such approaches in challenging real-life scenarios, such as low-data regime training, out of distribution performance, and prediction of activity cliffs. Our ultimate goal is to augment the potential of molecular property prediction algorithms in the context of drug discovery.

Prospective applications

Bringing AI to the wetlab

Optimizing metrics and boosting model performance is something we have a lot of fun with. Ultimately, however, what we want to achieve is to bring the models to the wet-lab. This allows us to verify whether our computational hypotheses are correct, and if the AI approaches are applicable in real-world scenarios. Also, we are really into collaborating with experts in chemical biology and drug discovery, to ultimately tackle unsolved challenges with AI.