Multiresistant bacteria pose a major threat to human health. With numerous potent drugs discovered in the “golden age” of antibiotics in the mid of the 20th century, new developments have dramatically declined while strains resistant to common antibiotics are on the rise. Since many of the antibiotics currently in use focus on a rather narrow scope of cellular targets, multiple resistance strategies have already evolved. Given the vast number of essential proteins in bacteria there is a huge potential to decipher unprecedented antibiotic targets yet lacking resistance strategies.

Our goal is to identify unprecedented antibacterial targets beyond the scope of current antibiotics and to exploit these for chemical manipulation.

For this we apply a multi-disciplinary strategy comprising synthetic chemistry, functional proteomics, microbiology, protein biochemistry techniques as well as machine learning tools.

Discovering new bioactive small molecules is a lengthy and complex endeavor, especially in the context of antibiotic research. Leveraging the wealth of publicly available chemical and biochemical databases via data-driven modeling holds the promise of accelerating the drug discovery pipeline.

We aim to leverage Artificial Intelligence (AI) algorithms to develop new in silico models to identify novel small-molecule antibiotics. Specifically, we seek to create an end-to-end pipeline combining Generative Deep Learning (DL) and Quantitative Structure-Activity Relationship (QSAR) models to suggest new chemical moieties and optimize them in terms of antibiotic activity and pharmacological properties.

Our projects focus on two distinct areas. On one hand, we conduct theoretical cheminformatics research focused on developing new algorithms for molecular property prediction, while on the other we employ these methods to model public and in-house datasets to discover new antibiotics. Current projects focus on the following areas:

• Data valuation algorithms to explain the predictions of QSAR models and identify false positives in the training data.
• Multi-modal self supervised learning to develop performant QSAR models on endpoints where little or no experimental data is available for training.
• Using alternative sources of information for QSAR modeling to overcome the limits of chemical representations.

Publications: Follow our Google Scholar, linked in the Bio.