3D-Vac

Every day, our immune system protects us from diseases such as infections and cancer. It does so using specialized molecules called MHC (Major Histocompatibility Complex), which display small fragments of proteins—called peptides—on the surface of our cells. This allows immune cells to scan for and recognize potential threats. Understanding which peptides bind to which MHC molecules is therefore essential for designing effective vaccines and personalized cancer immunotherapies. However, predicting these interactions is a major challenge: humans carry thousands of different MHC variants (known as HLA alleles), and it is impossible to test all possible combinations experimentally.

The goal of this project was to develop more accurate and generalizable computational tools for predicting peptide–MHC binding. To achieve this, we used a modern class of artificial intelligence methods called Geometric Deep Learning, which can analyze the three-dimensional structure of molecules directly. Unlike traditional approaches that rely only on protein sequences, our methods learn from the physical shape and interactions of molecules, enabling better predictions—especially for MHC variants that have not been studied before.

A key outcome of the project is DeepRank2, a unified and user-friendly software platform that brings together several earlier tools into a single framework. Using this platform, we built new structure-based AI models that significantly improve prediction performance, particularly for underrepresented HLA alleles. During the project, we realize that the originally proposed two-step pipeline (physics-based 3D modelling first and then prediction) is slow. Thus we also developed SwiftMHC, an ultra-fast model capable of predicting both molecular structure and binding strength in milliseconds. By optimizing its training process—achieving a speed-up of more than 100 times—we made it feasible to scale to very large datasets.

This project matters because improving peptide–MHC prediction directly supports the development of personalized vaccines and immunotherapies, particularly for populations that are currently underrepresented in biomedical data. We successfully achieved our main objectives, and the project also opened up new directions we had not originally anticipated. For example, we found that incorporating generative AI models can further enhance structural predictions, creating exciting opportunities for future research.

Our results are most relevant to researchers in cancer vaccine, TCR therapy, immunology, and bioinformatics, as well as biotech and pharmaceutical companies developing vaccines and immunotherapies. Moving forward, we plan to expand our models to cover a broader range of HLA alleles and to collaborate with experimental partners to validate our predictions in the laboratory.

We invite researchers and organizations to explore and use our openly available tools—DeepRank2, 3D-Vac, and SwiftMHC—and to collaborate with us in translating these advances into real-world medical applications. All tools and documentation are publicly available on GitHub.