From Speed to Structure: SwiftMHC Advances AI in Immunology 

We are thrilled to announce that our SwiftMHC work, led by Dr. Li Xue’s Structural Bioinformatics Group in collaboration with the Netherlands eScience Center, has been accepted for publication in scientific publication journal, Cell Reports Methods. The article will appear in the April 20 issue. Read the full paper online.

What makes this announcement exciting is due to the best of our knowledge, it is the first demonstration that physics-based modeling can be used to build lightweight AI in this field, offering the potential to drastically reduce the carbon footprint associated with large AI models. At this speed, the tool enables rapid structural screening of candidate neoantigens, and aids the detection of potential cross-reactivity, and toxicity assessment.

Our very own Research Software Engineer (RSE), Cunliang Geng, with the help fellow RSE, Bart Schilperoort, contributed to the project by drastically speeding up the training of SwiftMHC by more than 100 times. The speed-up training has enabled Dr. Li Xue’s team to extend SwiftMHC to millions of data for other alleles (work ongoing). Cunliang will present his part of this work at the upcoming BioSB conference at Baarlo on 12 and 13 May.  

Why it matters 

Peptide binding to major histocompatibility complex (MHC) molecules is central to immunology, with direct impact on vaccine design, cancer immunotherapy, and T cell receptor (TCR)–based therapies. Accurate prediction of peptide–MHC structures has traditionally been slow and computationally intensive, limiting large-scale applications. 

What SwiftMHC does 

SwiftMHC is a transformer that predicts peptide–MHC 3D structures and binding affinities with high accuracy—delivering results in milliseconds in batch mode. It currently focuses on HLA-A*0201 9-mer peptides, one of the most common MHC alleles and peptide lengths, with plans to expand to other alleles and peptide types. 

Test out the tool in the SwiftMHC Colab: https://colab.research.google.com/github/X-lab-3D/swiftmhc-inference/blob/main/colab/SwiftMHC_colab.ipynb. 

Graphic abstract from Cell Reports Methods 

What SwiftMHC does 

  • Unmatched speed: Thousands of times faster than an AlphaFold2-based approach without sacrificing accuracy. 
  • Sequence meets structure: As fast as sequence-based predictors, yet provides full-atom 3D models for downstream immunogenicity studies. 
  • New AI paradigm: Demonstrates that physics-based modeling can be combined with lightweight, task-specific AI in immunology. 
  • Sustainable computing: Dramatically reduces computational cost, offering a greener alternative to large general-purpose AI models. 

Our team and vision 

Our lab pioneers geometric deep learning—a branch of AI specialized in 3D objects—to model 3D structures (see newsletter) for cancer immunotherapy, with a focus on improving generalizability and data efficiency. We develop next-generation, physics-aware, and task-specific AI systems that are both computationally efficient and data-efficient. The team works closely with Jolanda de Vries’ lab at Radboudumc, designing personalized dendritic cell vaccines for patients with CMMRD and Lynch syndrome. 

What’s next? 

We are in discussions with a leading company to conduct wet-lab validations.

This milestone represents a major step toward faster, more scalable, and sustainable AI for biomedical research.

Meet the team 

  • Li Xue, Principal Investigator, Radboud University Medical Center 
  • Coos Baakman, first author, Radboud Universitair Medisch Centrum 
  • Cunliang Geng, Research Software Engineer, Netherlands eScience Center 
  • Special thanks to Giulia Crocioni, Research Software Engineer, Netherlands eScience Center, for her support on paper editing 
  • Special thanks to Bart Schilperoort, Research Software Engineer, Netherlands eScience Center, for his support with speeding up training for SwiftMHC