It all started in my kitchen. After eating a homemade carrot salad one evening, I noticed that the inside of my mouth had become swollen. Later, I learned that pollen-related food allergies can sometimes trigger such reactions. The experience left me wondering: could we know in advance whether a food might trigger an allergic reaction?

Around the same time, Merlin and I joined Prof. Kjell Jorner’s Digital Chemistry course at ETH out of curiosity about AI in the life sciences. Over lunch, we began exploring whether AI could help predict allergenic proteins beyond those already catalogued. Encouraged by early conversations with allergy researcher Katja Bärenfaller at the Swiss Institute of Allergy and Asthma Research (SIAF), we decided to pursue the idea.

This grew into XAllergen, a model designed to predict protein allergenicity accurately while providing interpretable explanations. Yet one question remained: what exactly do these AI explanations capture, and how closely do they align with established immunological knowledge?

To investigate this further, we continued the project through an ETH Studios internship at SIAF. Arriving in Davos, we found ourselves surrounded not only by snow-covered mountains, but also by a welcoming scientific community.

Through weekly institute meetings and conversations over lunch, we met researchers working across different areas of immunology. These interactions helped us better understand the immunological context of our project and ask more meaningful biological questions.

We were particularly grateful to our supervisors, Katja Bärenfaller and Damir Zhakparov, who encouraged us to challenge assumptions and explore uncertain directions. This open and supportive environment gave us the confidence to continue working on a challenging problem at the intersection of AI and immunology.

We found that the regions highlighted by current AI models often captured signals that were useful for prediction, but did not consistently match the parts of proteins that immunologists believe are responsible for triggering allergic responses. This matters because such models are increasingly used for safety screening of new food proteins and could one day help develop foods like carrots that no longer cause allergy. Our results suggest that strong predictive performance does not necessarily imply alignment with established biological mechanisms.

Looking back, this experience taught us that curiosity-driven research often follows an unpredictable path. A simple observation in a student kitchen led to a course project at ETH, an internship in Davos, and ultimately a scientific finding. The project will continue beyond Davos through presentations and discussions with the international machine learning and computational biology community, but it has already reinforced a lesson we will carry forward: even when the answer is uncertain, follow a simple question and see where it leads. You may discover something you never expected to find.

Further reading

• Yao J., Song A., Baerenfaller K., Zhakparov D. Residue-Level Attributions in Protein Language Models Do Not Recover Allergen Epitopes. arXiv:2606.22181 (2026). https://doi.org/10.48550/arXiv.2606.22181