“Research is undergoing fundamental change because of AI.”

What new possibilities do artificial intelligence methods open up in structural biology? This question is being addressed by biophysicist Dr. Andrea Thorn, who has been setting up the “AI and Biomolecular Structures” department at Helmholtz-Center Berlin (HZB) since July 2025. Her work combines AI-supported methods with experimental data from BESSY II, one of the world's leading sources of X-ray light, and builds on her international research experience in Göttingen, Cambridge, Oxford, and Hamburg.

In this interview, Thorn talks about her goal of bringing together molecular information in a way similar to a language model, explains the importance of explainable AI, and describes the role of collaboration in Berlin's science and innovation environment. She also reports on her experiences with the Coronavirus Structural Task Force and gives an outlook on upcoming developments.

Dr. Thorn, you have conducted research on the interactions of large molecules in Göttingen, Cambridge, Oxford, and Hamburg, among other places, developed software for analyzing diffraction data—i.e., the diffraction patterns of molecular crystals—and led your own research group. Against this background, what particularly appeals to you about your new role at HZB?

The interdisciplinary nature of the Helmholtz Association and the Helmholtz-Center Berlin opens up many new opportunities for me and my colleagues. Here, we can work as scientists without being bound by the traditional division between biology, physics, and chemistry. In addition, Helmholtz welcomes the development of software that helps other researchers make discoveries, for example, whereas at other institutions this is often dismissed as “auxiliary science” or “pure technology development.” Since my group and I firmly believe that the greatest discoveries are made collaboratively, this “make it possible” mentality suits us very well. And then, of course, with BESSY, we have a superb X-ray source with three beamlines for macromolecular crystallography right next to our office!

You are working on developing an AI-based pipeline that combines molecular information in a similar way to a language model – a “ChatGPT for molecules.” What exactly does that mean?

First of all, we are working on improving the automatic evaluation of three-dimensional molecular structures and the experimental data on which they are based. We started working on this pipeline back in 2020 during the coronavirus pandemic and were able to make a big difference in drug development. Even back then, we used AI in the pipeline, for example, to identify certain measurement errors. We now plan to increase this use, and in the long term, we hope to develop an expert system for biological molecular structures.

One focus of your work is on evaluating experimental data from BESSY II—an electron storage ring that generates extremely bright X-ray light for research—using machine learning. What new opportunities does this open up?

In crystallography, we need experts if we want to take really good measurements. However, automation is on the rise, and users are becoming less and less expert. For this automation, we have to make decisions automatically – for example, the measurement distance or crystal rotation during the measurement. In order to make these decisions as well as an expert would, we need machine learning and good statistical indicators. We are developing these.

You emphasize how important it is for AI analyses in molecular biology to remain traceable. Why is this so crucial – and how do you implement it in concrete terms?

This is called “explainable AI” or “XAI.” These methods are incredibly important because we often need to understand exactly what a neural network has learned during training. Unfortunately, the technology is still in its infancy – but we use algorithms, for example, that tell us which area of a data set the network has paid attention to when it gives us a specific answer – and that sometimes helps us understand what the AI has based its decision on. In general, scientists don't like “black boxes,” which is why XAI is really important for research—it's the only way we can learn together with AI.

With Charité, the Max Delbrück Center, and other partners, Berlin offers a strong infrastructure in the fields of AI and life sciences. What role does cooperation with other players play in your work, and what makes the location attractive for your research?

On the one hand, we already have a good international network of collaborations, but on the other hand, we started networking with the MDC and Berlin's universities over a year ago. Fortunately, Berlin's structural biologists are already in close contact with each other and meet once a year for the “Joint MX Day.” That naturally makes Berlin attractive. On the other hand, the location also has a lot to offer in the field of AI, and of course, it's a great city!

With Berlin-based deep tech companies such as PRAMOMOLECULAR (RNA-based gene silencing therapies) and Lucid Genomics (AI-supported genome analysis for diagnostics and biomarkers), new approaches are emerging to better diagnose diseases and develop more targeted therapies. Do you see similar opportunities for concrete applications or even future spin-offs in your research?

Yes, definitely! With the Coronavirus Structural Task Force, which investigated the molecular structures of the virus and combined information under my leadership from 2020 to 2022, we have delivered a very concrete application for vaccine and drug development. Similarly, our software for cryo-EM and crystallography is distributed to academics and industry, and we also have a web server for analyzing X-ray data. So our group is very application-oriented.

What scientific contributions and personal experiences have you made with the Coronavirus Structural Task Force – and how do these influence your current research?

During the pandemic, we systematically analyzed over 3000 molecular structures from SARS-CoV-1 and SARS-CoV-2. We produced and published the most accurate 3D model of the virus and also animated how a host cell is infected at the molecular level. Overall, we were able to help many researchers and developers worldwide. That makes me very proud – but it took us more than two years with 27 experts. Today, I wonder how much of this could be automated, especially with regard to so-called foundation models – perhaps known to you as large language models (LLMs).

Looking five years into the future, what would you like to achieve with your research at HZB?

I would like to see us being able to contextualize, combine, and evaluate structural biology information worldwide with just a few clicks – regardless of which laboratory the data was generated in. This would enable us to better understand the molecular basis of life and be better prepared for the next pandemic. I would also like to introduce the topic of “chemical processes in fungi” at HZB, but that is still up in the air...

How might the interaction between research and artificial intelligence develop in the future – in Germany, but also globally?

Research is undergoing fundamental change because of AI. This can already be seen, for example, in the field of scientific publishing. This is problematic because science itself is always cutting-edge—that's in the nature of things—but our systems, professorships, literature, Nobel Prizes, etc., follow very old patterns and change only slowly. Machine learning will help and support us, but it also forces us to rethink our own work and the education of the next generation. I'm very excited about it!

Thank you very much for this conversation.