Improved tuberculosis drugs thanks to machine learning

New model can predict activity and efficacy of benzothiazinones

artwork of a screen, molecules and microorganisms
Machine learning models can predict activity of benzothiazinone compounds. Source: Freddy Bernal & Luo Yu/Leibniz-HKI

A research team at Leibniz-HKI has developed a computer model to predict the activity and potency of new chemical compounds. It is designed to help test new molecules from the substance class of benzothiazinones as antibiotics against the tuberculosis bacterium. The results were published in the Journal of Medicinal Chemistry.

Benzothiazinones (BTZs) are a very effective class of antibiotics against the causative agent of tuberculosis, Mycobacterium tuberculosis. Since their discovery, they have been developed at Leibniz-HKI in a partnership with the LMU University Hospital in Munich. One of the substances, BTZ-043, is already in clinical trials (phase II). In the "Next-Gen-BTZ" project, a team of scientists from Leibniz HKI also researched improved generations of active substances.

"New active substances often show unexpected challenges over the course of the long drug development process up to the finished drug. This can only be solved in a targeted manner by developing subsequent generations," explains Florian Kloß, head of the Antiinfectives Transfer Group, which is advancing the development of tuberculosis antibiotics. So far, several hundred different BTZs have been synthesised.

Under which conditions these substances are particularly effective, however, could hardly be predicted. "In our work, we were therefore interested in establishing relationships between the chemical structure and the antibiotic activity so that in future we can concentrate our work in the laboratory only on the production of promising substances," says Kloß. This is challenging in the case of BTZs because several processes have to work together in the bacterial cell in order to inhibit the growth of the bacteria in the best possible way.

The team has therefore developed a computer model that has been trained with data on 96 different BTZ. By combining different algorithms, it can identify highly effective substances with over 70 per cent reliability. "This enables us to achieve a higher hit rate in the development of new substances and avoids the unnecessary production of ineffective BTZs in the laboratory," says Kloß.

The Next-Gen BTZ research group was funded by the Free State of Thuringia with money from the European Social Fund.

Original publication

Schieferdecker S, Bernal FA, Wojtas KP, Keiff F, Li Y, Dahse H-M, Kloß F (2022). Development of Predictive Classification Models for Whole Cell Antimycobacterial Activity of Benzothiazinones. Journal of Medicinal Chemistry, doi:10.1021/acs.jmedchem.2c00098

Staff

Freddy Alexander Bernal
Hans-Martin Dahse
Francois Keiff
Florian Kloß
Yan Li
Sebastian Schieferdecker
Philip Wojtas