3 Questions for

Dagmar Krefting

Professor of Medical Informatics at the University of Göttingen and member of Plattform Lernende Systeme

Developing medicines with AI: "There is a lack of reliable legal requirements"

Research into new drugs is an expensive and lengthy undertaking. It takes around twelve years for a drug to be approved, at an average total cost of around 2.8 billion US dollars. In many cases, such as with antibiotics, the development of new active ingredients is no longer profitable - to the detriment of healthcare. Artificial intelligence (AI) can speed up the development of drugs. In this interview, Dagmar Krefting explains exactly how the use of AI can improve drug research and what needs to be done so that the population can benefit from cost-effective medicines. She is Professor of Medical Informatics at the University of Göttingen and a member of the Health Care, Medical Technology and Care working group of Plattform Lernende Systeme.


How does the use of AI improve drug research?

Dagmar Krefting: Artificial Intelligence is particularly useful wherever there are complex relationships between many different influencing factors. This is the case in the development of pharmaceuticals: a drug must not only have a desired effect, but should not cause any undesirable side effects, have as few interactions with other drugs as possible, be well tolerated, cost-efficient to produce, storable and easy to administer. Complex biochemical combinations of substances meet highly complex human physiological processes. Finding the proverbial needle in the haystack here, i.e. a new drug among the plethora of potential candidates, is only possible at great expense through laboratory experiments and classic clinical trials. As a result, drug development is a very expensive and lengthy process - on average, it takes 12 years from patent application to approval. Artificial Intelligence can help to identify potential drug candidates from the existing data on active ingredients, biochemical structures and processes as well as the effects and side effects of drugs in studies and in healthcare and to predict their effect on different patient groups. There is particularly high potential in the optimization of clinical trials through new data-supported study designs and virtual study groups.


How widespread is AI in the development of new drugs?

Dagmar Krefting: There are only around 700 companies worldwide that offer AI-supported solutions in drug development, particularly in the preclinical area. One area of application is the prediction of suitable ways of influencing disease development, known as targets. Here, for example, genetic data is linked with scientific literature in order to predict drug targets. Based on a target, there are AI approaches that predict active ingredient structures and, in some cases, already estimate important properties such as toxicity and manufacturability. The active ingredient structure can then be further developed with the help of AI, which means that both the manufacturing process and the interaction with the human metabolism can be tested and optimized in virtual experiments. So far, AI has been used very little in the clinical trials leading up to the approval of a drug. These account for between half and two thirds of the total development costs and are also particularly critical phases, as the drugs are used in humans. The above-mentioned potential of AI in evidence generation and process optimization in clinical trials is currently not being exploited due to a lack of legal requirements for the use of AI in this area.


What still needs to be done so that the healthcare sector can benefit from AI-supported drug research?

Dagmar Krefting: In general, AI in areas where its use entails high risks must be trained on a high-quality database that is representative of the respective use case and must be trustworthy. This means that the predictions generated with AI must be comprehensible and reproducible and, in particular, the limits within which AI can deliver reliable results must be clearly defined. The availability and quality of representative data on the development of an AI plays a key role in its performance. If we want AI models to be suitable for the local population and the German healthcare system, this data must be available in digital form and researchers must be able to use it. The Health Data Utilization Act has significantly improved research opportunities here, but this does not help if the database is not available. However, even a scientifically sound and trustworthy AI-supported system can only be used in clinical trials if the regulatory authorities accept it as a suitable procedure. Reliable guidelines are necessary for this. The extent to which EU-wide regulations such as the European Health Data Space, the Medical Device Regulation or the AI Act can provide planning security or represent further hurdles in international comparison remains to be seen in practice.


Further information on this topic can be found in the white paper (in German): Developing drugs with AI: From idea to approval

The interview is released for editorial use (provided the source is cited © Plattform Lernende Systeme).

Go back