"AI will be a gamechanger in antibiotics research"

I am convinced that AI will become a gamechanger in antibiotics research. It can speed up the search for effective antibiotics among natural substances enormously and thus make development cheaper and faster. HIPS and Saarland University have launched the NextAID project, which is short for "Neuro-explicit AI for Drug Discovery". In the meantime, we are also cooperating with the Helmholtz Centre for Information Security, CISPA for short, and just yesterday my esteemed colleague Jilles Vreeken from CISPA and I met again to discuss this.

How exactly does artificial intelligence help in the search for suitable antibiotics?

At the moment, it mainly helps with prioritization, and that is already a decisive advantage. In practice, it works like this: We feed the AI with data about the germs, their genetic properties, molecular biological effects, the surface structure of the cells and how they react to various known antibiotics. The AI also "knows" the properties of the antibiotics. It then uses them to calculate the probabilities of which defense mechanisms lead to resistance and which natural substances are most likely to be able to circumvent or break through them. Then the computer gives us, for example, a top-20 list of those natural substances which, according to AI calculations, are most likely to crack certain multi-resistant germs.

In order for the AI to actually include promising substances in the list, it must first be fed with some data. Doesn't this data also have to be collected first, or at least put into a form that the AI can do something with?

It's true that the more data we give the AI - of the right quality, of course - the better its prioritization will be. This is where a development in international research comes in handy: The trend is moving strongly towards more data exchange according to the so-called FAIR principle, which stands for Findable, Accessible, Interoperable and Reusable. In addition, federated learning is being used: This involves different institutions training AI models with their data to help create a more robust model. Of course, collecting the data and training takes one to two years. But even if you factor in that time, AI can still save a huge amount of time.

Which pathogens are you currently working on at NextAID and HIPS?

One of our focal points is the treatment of tuberculosis. Globally, this is a huge problem: According to WHO estimates, more than ten million people fall ill every year and 1.6 million of them die from it. This might seem a long way off in Germany, but in view of global travel, migration flows and climate change, tuberculosis will also spread like that here in Germany. And resistance to the tried and tested combination therapies of different antibiotics is unfortunately becoming more and more common.

How do you proceed once the AI has produced a list of the most promising substances for treating tuberculosis, for example?

First, we have the active substances produced. In most cases, this is done biotechnologically with bacteria or unicellular fungi, such as yeasts. Then we test them in the laboratory to see how well they work against tuberculosis pathogens.

That actually sounds quite simple.

In principle it is, but in practice it is more complicated. First of all, it is not always the case that there is actually something effective directly among the 20 natural substances. And then, of course, you have to see whether the substance can be administered to people at all. To achieve optimal results, it is important that we feed the AI with as much data as possible on new natural substances, as well as search for new substances ourselves. To this end, HIPS has joined forces with the Naturlandstiftung Saar to launch a community science project called MICROBELIX: Every citizen can send us soil samples, for example, from their own garden, which are then analyzed in the laboratory for the natural substances they contain. Each of us carries another source of new substances in our intestines: The microbiome.

Does this mean that antibiotics are also produced in our intestines?

In principle, yes. The bacteria in the digestive system produce a number of substances, some of which are still unknown, and some of which act against bacteria that do not belong to the microbiome. Accordingly, we can search more specifically than usual for promising active substances. We ask ourselves, for example: Why is such and such a patient less susceptible to certain diseases? We look for answers in the microbiome, i.e., for certain bacterial strains in the intestine that are more prevalent in the "robust" people, and then we look at which substances these strains produce. We also let such data and information flow into the AI.

And the prioritization lists of the AI continue to improve.

Yes, we hope so, of course. It is also possible that such prioritization lists are only the beginning of the possibilities that AI offers us in the search for new antibiotics. Perhaps AI will soon be so far advanced that it will suggest certain molecular structures as candidates for effective antibiotics, and we will then try to design these structures. We would then speak of generative AI, i.e., AI that suggests something new itself. Maybe that will happen in one or two years, it's hard to predict, but it's possible! But even today, AI is already making a significant contribution to ensuring that humanity does not lose out in the fight against multi-resistant germs.