Document
Datasets
We evaluated GTpMHC against existing baseline methods using three benchmark datasets: TranspMHC pMHC-I[1], SWING Class_I crossval, and SWING Class_II crossval[2].
1.TranspMHC pMHC-I dataset: The training dataset was collected from the Immune Epitope Database (IEDB)[3] up to 12 December 2023 and contains 96 MHC class I alleles with 193,952 samples. The testing dataset was collected from the in-house data of Anthem[4], which includes 4 MHC class I alleles with 19,504 samples.
2.SWING Class_I crossval dataset: This dataset was constructed from the NetMHCpan-4.1[5] evaluation set based on functional clustering of MHC molecules. The training set comprised eight alleles: HLA-A*02:11, HLA -A*25:01, HLA -A*11:01, HLA-B*07:02, HLA-B*15:03, HLA-B*40:01, HLA-C*08:02, and HLA-C*07:02. Six independent test sets were constructed, each corresponding to a single MHC allele: HLA-A*02:02, HLA-B*40:02, HLA-C*05:01, HLA-B*38:01, HLA-C*03:03, and HLA-A*32:01.
3.SWING Class_II crossval dataset: The dataset was also constructed based on functional clustering of MHC alleles using the NetMHCIIpan-4.2[6,7] training data. Five HLA-DR alleles (HLA-DRB1*01:01, HLA-DRB1*04:01, HLA-DRB1*03:01, HLA-DRB1*15:01, and HLA-DRB1*07:01) were used for model training. Generalization was assessed on two allele-specific test sets, consisting of HLA-DRB1*01:02 and HLA-DRB1*04:04, respectively.
TranspMHC pMHC-I dataset: train_pMHC-I.csv test_pMHC-I.csv
SWING Class_I crossval dataset: train_Class_I.csv test_HLA-C05_01.csv test_HLA-C03_03.csv test_HLA-B40_02.csv test_HLA-B38_01.csv test_HLA-A32_01.csv
SWING Class_II crossval dataset: train_class_II.csv test_DRB1_0404.csv test_DRB1_0102.csv
2. References
[1] Niu, R., Wang, J., Li, Y., et al. Attention-aware differential learning for predicting peptide-MHC class I binding and T cell receptor recognition. Briefings in Bioinformatics 2025;26(1). [2] Siwek, J.C., Omelchenko, A.A., Chhibbar, P., et al. Sliding Window Interaction Grammar (SWING): a generalized interaction language model for peptide and protein interactions. Nature Methods 2025;22(8):1707–1719. [3] Vita, R., Blazeska, N., Marrama, D., et al. The Immune Epitope Database (IEDB): 2024 update. Nucleic Acids Research 2025;53(D1):D436–D443. [4] Mei, S., Li, F., Xiang, D., et al. Anthem: a user customised tool for fast and accurate prediction of binding between peptides and HLA class I molecules. Briefings in Bioinformatics 2021;22(5). [5] Reynisson, B., Alvarez, B., Paul, S., et al. NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data. Nucleic Acids Research 2020;48(W1):W449–W454. [6] Nilsson, J.B., Kaabinejadian, S., Yari, H., et al. Machine learning reveals limited contribution of trans-only encoded variants to the HLA-DQ immunopeptidome. Communications Biology 2023;6(1):442. [7] Reynisson, B., Barra, C., Kaabinejadian, S., et al. Improved Prediction of MHC II Antigen Presentation through Integration and Motif Deconvolution of Mass Spectrometry MHC Eluted Ligand Data. Journal of Proteome Research 2020;19(6):2304–2315.