The researchers from the University of Warsaw developed an innovative artificial intelligence model that generates antibiotics and published their findings in “Nature Communications” journal.

According to the World Health Organization (WHO), treatment-resistant infections pose an extremely serious global health and economic threat. Not only is the number of antibiotic-resistant infections rising, but also the number of bacterial strains that cause these infections. At the same time, there has been a decline in the effectiveness of antibiotics used. Moreover, no new class of antibiotics has been developed for more than thirty years. There is a risk that we will soon lose any effective method of combating antibiotic-resistant infections.

For a century, compounds known as antimicrobial peptides (AMPs) have been studied. They are produced by both bacteria and mammals, including humans. AMPs are designed to protect the organism from the harmful effects of pathogens. AMPs have a key feature – bacteria are extremely slow to acquire resistance to them. However, among the thousands of peptides that have been discovered, it has not yet been possible to find one that is superior to conventional antibiotics in treating bacterial infections.

The untapped potential of AMPs caught the interest of Prof. Ewa Szczurek’s team. The researchers have developed HydrAMP, a new artificial intelligence model based on variational autoencoders. In a recent publication in “Nature Communications”, the researchers showed that HydrAMP can generate new AMP sequences with high antimicrobial activity. The publication is the result of a collaboration between researchers from the UW and the Medical University of Gdańsk.

How to teach biology to a computer program?

How does HydrAMP learn about existing AMPs and know how to generate a new peptide? Each peptide can be represented as a sequence of letters corresponding to the basic molecules of which it is composed. For example, the word “WARS” represents a peptide composed of four amino acids. Such a sequence contains information about the chemical structure of the peptide, namely the atoms and bonds that build it.

“Since we have a large enough number of such sequences, we can teach our model to distinguish random sequences from peptide coding sequences. We also have data on the activity of known peptides. The model can learn this too,” Paulina Szymczak, a doctoral candidate at the Interdisciplinary Doctoral School and first author of the paper published in “Nature Communications” says.

By combining knowledge of known sequences and their activity, HydrAMP can generate hundreds of new sequences, significantly increasing the chances of finding exceptionally effective peptides. HydrAMP can also improve existing peptides by making changes to their sequence. For example, we can increase the antimicrobial activity of a peptide that was not showing enough effectiveness.

Modified peptides offer a chance for effective treatment of bacterial infections

Prof. Wojciech Kamysz’s team from the Medical University of Gdansk synthesized and studied how HydrAMP-generated peptides act on bacterial strains and red blood cells.

“We discovered fifteen completely new active peptides. Among them is a peptide called Varsavian, which shows very promising activity against dangerous bacteria such as methicillin-resistant Staphylococcus aureus,” Prof. Ewa Szczurek of the UW’s Faculty of Mathematics, Mechanics and Computer Science, a co-author of the publication says.

The experiments confirmed that HydrAMP converted a completely inactive peptide into a peptide with a very promising activity profile without making it toxic. This is the first such model that can change the sequence of a peptide leading to increased antimicrobial activity.

In addition, to further the knowledge of the discovered peptides, the team of Dr. Piotr Setny of the UW’s Center for New Technologies studied how the peptides behave when they come into contact with the membrane of a bacterial cell.

The peptides that were discovered by Prof. Szczurek’s group are described in a patent application and could be tested in clinical trials and used as drugs in the future. The model itself is equipped with several innovative artificial intelligence solutions and is a powerful tool that could serve to accelerate research in the field of combating antibiotic resistance.