Use SwiftMHC Online
Commercially Available SwiftMHC No-Code Web Server
SwiftMHC: A High-Speed Attention Network for MHC-Bound Peptide Identification and 3D Modeling
SwiftMHC is a structure-based deep learning framework that predicts peptide-MHC class I (pMHC-I) binding affinity and all-atom 3D structure in a single step. It was developed at Radboudumc with the Netherlands eScience Center and published in Cell Reports Methods in 2026.
Tumor-derived peptides that bind a patient's MHC molecules are central to cancer immunotherapy, but progress has been limited to a handful of well-studied alleles. Sequence-based tools are fast, but they generalize less well to rare alleles and give no 3D model. Structure-based tools can generalize better, but they have been too slow for large-scale screening. SwiftMHC is designed to remove that trade-off.
Metric | SwiftMHC result |
|---|---|
Binding affinity speed | 0.009 s per case (batch mode, single A100 GPU) |
3D structure speed | 0.9 s per case, or 2.2 s with energy minimization |
Binding prediction accuracy | Median AUC 0.91 |
Structure accuracy | Median Cα-RMSD 1.32 Å (1.19 Å with OpenMM) against X-ray structures |
Why It Matters
Only about 1,000 pMHC structures are in the Protein Data Bank, while more than 40,000 HLA variants have been identified.
Experimental structure determination can't keep up with this diversity, so computational modeling has to fill the gap.
Physics-based modeling tools such as PANDORA and APE-Gen 2.0 take seconds to minutes per case.
Structure-based affinity predictors have typically needed two steps: model the 3D structure, then estimate affinity.
Key Capabilities
Joint prediction: Binding affinity (numerical and binary) plus an all-atom 3D structure from one network.
Speed at scale: Faster than leading sequence-based tools in batch mode.
Variable peptide length handling: A residue-wise affinity predictor lets the model accommodate different peptide lengths.
Optional refinement: Short OpenMM energy minimization produces higher-quality models.
Open resources: Code, trained models, and data are publicly deposited.
How SwiftMHC Works
SwiftMHC takes two inputs: the structure of the MHC G-domain and a peptide sequence. It uses four modules built around residue-to-residue attention.
MHC self invariant point attention. Encodes each MHC residue from its amino acid type and structural neighbors, inspired by AlphaFold2's invariant point attention.
Peptide self attention. Encodes each peptide residue from its type and position in the sequence.
Cross-attention structure module. Updates peptide residues based on their interactions with MHC residues, and refines the peptide structure from a starting point at the center of the MHC groove. It predicts backbone frames, then side-chain torsion angles, then places atoms in 3D space.
Binding affinity predictor. A multilayer perceptron scores each peptide residue's contribution, and the contributions are summed. The training target is derived from experimental Kd or IC50 values.
Why It Is So Fast
No MSA module. MHC structures are highly conserved and short peptides lack evolutionary signal, so SwiftMHC drops the computationally heavy multiple-sequence-alignment attention used by AlphaFold.
Precomputed MHC data. Because the MHC structure stays fixed, its backbone frames and distance matrix are calculated once and reused.
Masking. MHC residues far from the peptide are masked out of the cross-attention step.
Quaternions. Rotations use quaternions instead of 3 × 3 matrices, reducing computation.
Training on Physics-Derived Data
SwiftMHC was trained on 7,726 binding affinity data points from the Immune Epitope Database (IEDB). Because the network needs structures, the authors generated a 3D model for each pMHC with the physics-based tool PANDORA. X-ray structures were held out of training, which tests how well a model trained purely on physics-derived models predicts experimental structures. The authors describe this as showing that small, task-specific models trained on physics-derived synthetic data can work well where experimental structures are scarce.
How SwiftMHC Compares
Binding Affinity Prediction
Tested with leave-one-cluster-out cross-validation across 10 peptide clusters for HLA-A*02:01 9-mers:
SwiftMHC reached a median AUC of 0.91, ahead of AlphaFold2-FineTune, MHCfold and a retrained MHCflurry 2.0.
Against NetMHCpan 4.1, accuracy was nearly comparable. NetMHCpan reached AUC 0.92, Pearson correlation 0.81 and AUPR 0.89. SwiftMHC reached 0.91, 0.80 and 0.88.
On speed, SwiftMHC took 0.009 s per case, against 0.081 s for NetMHCpan 4.1 and 0.020 s for MHCflurry 2.0.
The authors caution that AlphaFold2-FineTune and MHCfold were not retrained, so some of the test data may have been in their training sets. NetMHCpan 4.1 and MHCflurry 2.0 performance may also be slightly inflated by overlap.
3D Structure Prediction
Tested on 202 HLA-A*02:01 9-mer X-ray structures:
SwiftMHC with OpenMM reached a median Cα-RMSD of 1.19 Å, comparable to three other leading methods (0.74 to 1.21 Å) and well ahead of MHCfold (3.65 Å).
AlphaFold2-FineTune scored lower (0.87 Å), but the authors note this likely reflects overlap with its training data. They found 105 of the 118 unique test pMHCs were probably in its training set.
SwiftMHC produced far fewer outliers than AlphaFold2-FineTune, which placed one end of some peptides outside the MHC groove.
Robustness to Mutations
On 2,838 single-amino-acid peptide variants, SwiftMHC's predicted change in binding free energy (ΔΔG) correlated only modestly with the true change (r = 0.33). However, it rarely got binding status wrong, misclassifying only 11% of cases where a mutation did not flip binder status. The authors conclude that it reliably captures whether a mutation matters for binding, while fine-grained ΔΔG prediction remains challenging.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with SwiftMHC on Tamarind Bio
The authors describe several applications for fast pMHC structure and affinity prediction:
Large-scale peptide screening: Screen candidate peptides for vaccine development and immunotherapy research.
Neoantigen discovery: Identify novel neoantigens from the diverse pool of tumor-derived peptide fragments.
Safer TCR therapies: Fast 3D modeling could support building a comprehensive 3D library of self-peptide-HLA complexes. That would help identify therapy targets whose T cell-exposed surfaces differ from self-peptides.
TCR design: The 3D structures are intended to support downstream TCR design and cross-reactivity assessment.
Mutation screening: Rapidly assess whether patient-derived mutations change binding.
Tamarind already lists related immunology tools, including MHC-Fine, DeepImmuno, TLimmuno2 and TCRmodel2, so SwiftMHC output can feed a wider workflow.
How to Use SwiftMHC on Tamarind Bio
Access the platform. Log in to tamarind.bio.
Choose your allele and peptides. SwiftMHC currently supports HLA-A*02:01 and 9-mer peptides. Prepare a table linking each peptide sequence to its MHC allele. A measured binding value (affinity or binding/non-binding label) is optional.
Prepare the MHC structure. The model needs an MHC structure, and all MHC molecules must be aligned to a reference structure during preprocessing. The paper used PDB 3MRD as the reference.
Preprocess the data. The open-source code converts structures and peptide data into HDF5 files.
Run predictions. Use the open-source code. Results are written to a results table giving binding affinity and binder classification for each MHC and peptide pair, with optional 3D structures.
Choose refinement. By default, SwiftMHC skips energy minimization for speed. Add the energy-minimization flag to refine structures with OpenMM, which takes about 2.2 s per case instead of 0.9 s.
Interpret the results. In the paper, a peptide counts as a binder when its IC50 or Kd is below 500 nM.
Continue downstream. Use the predicted structures in other immunology tools on Tamarind, such as TCRmodel2 for TCR modeling.
Source: Baakman et al., A high-speed attention network for MHC-bound peptide identification and 3D modeling, Cell Reports Methods 6(4):101364 (2026). DOI: 10.1016/j.crmeth.2026.101364