AI for T-Cell Antigen Identification: Data Resources, Computational Methods, and Benchmarking

AI-Driven Computational Methods, Data Resources, and Benchmarking for T-Cell Antigen Identification

Available Datasets for T-cell Antigen Identification

Peptide-MHC I Binding

Dataset name	Published in	Resources
SYFPEITHI	Immunogenetics, 1999	Code
MHCBN	Bioinformatics, 2003	Code
EPIMHC	Bioinformatics, 2005	Code
Abelin et al.	Immunity, 2017	Code
IEDB	Nucl. Acids Res., 2019	Code
Sarkizova et al.	Nat. Biotechnol., 2020	Code

Peptide-MHC II Binding

Dataset name	Published in	Resources
VDJdb	Nucl. Acids Res., 2018	Code
IEDB	Nucl. Acids Res., 2019	Code
Rappazzo et al.	Nat. Commun., 2020	Code
Strazar et al.	Immunity, 2023	Code

TCR-pMHC Binding

Dataset name	Published in	Resources
BindingDB	Nucl. Acids Res., 2016	Code
McPAS-TCR	Bioinformatics, 2017	Code
Dash et al.	Nature, 2017	Code
VDJdb	Nucl. Acids Res., 2018	Code
TetTCR-seq	Nat. Biotechnol., 2018	Code
IEDB	Nucl. Acids Res., 2019	Code
10x	Tech. Rep., 2019	Code
PIRD	Bioinformatics, 2020	Code
Heilkkila et al.	Mol. Immunol., 2020	Code
NeoTCR	Genomics Proteomics Bioinformatics, 2024	Code
ImmuneCODE	Front. Immunol., 2025	Code
TRAIT	Genomics Proteomics Bioinformatics, 2025	Code

Representative AI methods for T-cell antigen identification

Peptide-MHC I Binding Prediction

Model	Published in	Datasets used	N.G.	E.M.	Metrics	Resources
NetMHCpan-4.1	Nucl. Acids Res., 2020	IEDB	RS	BL	AUROC, PPVn	Resources
Anthem	Brief. Bioinform., 2021	IEDB,EPIMHC,MHCBN,SYFPEITHI	RS	BL	AUROC,Sensitivity,Specificity,Accuracy,MCC	Resources
TransPHLA	Nat. Mach. Intell., 2022	IEDB,EPIMHC,MHCBN,SYFPEITHI	RS	OR	AUROC,MCC,F1-Score,Accuracy	Resources
STMHCpan	Brief. Bioinform., 2023	IEDB	RS	OR	AUROC,Recall,Precision,F1-Score,Accuracy	Resources
MixMHCpred2.2	Cell Systems, 2023	Self-curated datasets from multiple public sources	RS	BL	AUROC,PPV	Resources
BigMHC	Nat. Mach. Intell., 2023	IEDB,NEPdb,TESLA,Neopepsee,MANAFEST	RS	OH	AUROC,AUPRC,PPVn	Resources
ImmuneApp	Nat. Commun., 2024	Self-curated datasets from multiple public sources	RS	BL	AUROC,AUPRC,PPVn	Resources
MixMHCpred3.0	Genome Med., 2025	Self-curated datasets from multiple public sources	RS	BL	AUROC,AUPRC	Resources
UniPMT	Nat. Mach. Intell., 2025	IEDB,TESLA,NEPdb,Neopepsee,MANAFEST	RS	LM	AUROC,AUPRC	Resources
UnifyImmun	Nat. Mach. Intell., 2025	IEDB,EPIMHC,MHCBN,SYFPEITHI	RS,NC	OR	AUROC,AUPRC,Accuracy,MCC,F1-Score	Resources
deepAntigen	Nat. Commun., 2025	IEDB,Sarkizova et al.,Abelin et al.,TESLA,Xu et al.	RS	OH	AUROC,AUPRC,Sensitivity,Specificity,Precision,NPCC	Resources

Peptide-MHC II Binding Prediction

Model	Published in	Datasets used	N.G.	E.M.	Metrics	Resources
DeepSeqPanII	IEEE/ACM Trans. Comput. Biol. Bioinform., 2021	IEDB	Not mentioned	OH&BL	AUROC,SRCC	Resources
DeepMHCII	Bioinformatics, 2022	IEDB	Not mentioned	OR	AUROC,PCC	Resources
MixMHC2pred2.0	Immunity, 2023	Self-curated datasets from multiple public sources	RS	OH&BL	AUROC	Resources
NetMHCIIpan4.2	Commun. Biol., 2023	Self-curated datasets from multiple public sources	RS	BL	AUROC,PPVn	Resources
NetMHCIIpan4.3	Sci. Adv., 2023	Self-curated datasets from multiple public sources	RS	BL	AUROC,PPVn	Resources
deepAntigen	Nat. Commun., 2025	Rappazzo et al.,Strazar et al.	RS	OH	AUROC,AUPRC,Sensitivity,Specificity,Precision,NPCC	Resources

TCR-pMHC Binding Prediction

Model	Published in	Datasets used	N.G.	E.M.	Metrics	Resources
ERGO-II	Front. Immunol., 2021	McPAS-TCR,VDJdb,Kanakry et al.	SH	OH	AUROC	Resources
NetTCR-2.0	Commun. Biol., 2021	IEDB,VDJdb,10x,MIRA	SH	PC	AUROC,PPVn	Resources
ImRex	Brief. Bioinform., 2021	VDJdb,McPAS-TCR,Dean et al.	BK,SH	PC,BL	AUROC,AUPRC	Resources
DLpTCR	Brief. Bioinform., 2021	TetTCR-seq,VDJdb,IEDB	BK	OH,PC	Recall,Precision,Accuracy,AUROC	Resources
pMTnet	Nat. Mach. Intell., 2021	PIRD,McPAS-TCR,VDJdb,10x,TetTCR-seq,Chen et al.	SH	BL	AUROC,AUPRC	Resources
DeepTCR	Nat. Commun., 2021	Dash et al.,10x,Glanville et al.,ImmunoMap,Chan et al.	Not mentioned	OH	AUROC,Recall,Precision,F1-Score	Resources
TITAN	Bioinformatics, 2021	VDJdb,ImmuneCODE,BindingDB	SH	BL	Accuracy,AUROC	Resources
PRIME2.0	Cell Systems, 2023	Self-curated datasets from multiple public sources	BK	BL	AUROC,PPV	Resources
TEINet	Brief. Bioinform., 2023	VDJdb,McPAS-TCR,Lu et al.	SH,BK	LM	AUROC,Accuracy,Precision,Recall	Resources
PanPep	Nat. Mach. Intell., 2023	IEDB,VDJdb,PIRD,McPAS-TCR,ImmuneCODE	BK	PC	AUROC,AUPRC	Resources
TEIM	Nat. Mach. Intell., 2023	IEDB,VDJdb,McPAS-TCR,ImmuneCODE	SH	BL	AUROC,MCC,AUPRC	Resources
PISTE	Nat. Mach. Intell., 2024	VDJdb,McPAS-TCR,Lu et al.	SH,BK	OR	AUROC,AUPRC,PPVn	Resources
MixTCRpred	Nat. Commun., 2024	VDJdb,McPAS-TCR,IEDB,10x	SH,BK	OR	AUROC	Resources
TPepRet	Bioinformatics, 2025	IEDB,VDJdb,McPAS-TCR	SH	PC	AUROC,AUPRC	Resources
UniPMT	Nat. Mach. Intell., 2025	IEDB,TetTCR-seq,VDJdb,McPAS-TCR,PIRD	SH	LM	AUROC,AUPRC	Resources
UnifyImmun	Nat. Mach. Intell., 2025	TetTCR-seq,VDJdb,IEDB,PIRD,Heilkkila et al.,10x,ImmuneCODE,BindingDB	SH,BK	OR	AUROC,AUPRC,Accuracy,MCC,F1-Score	Resources
TCRBagger	Cell System, 2025	IEDB,VDJdb,McPAS-TCR,PIRD	BK	PC	AUROC,AUPRC	Resources
deepAntigen	Nat. Commun., 2025	IEDB,VDJdb,McPAS-TCR,PIRD,ImmuneCODE,NeoTCR	BK	OH	AUROC,AUPRC,Sensitivity,Specificity,Precision,NPCC	Resources

N.G.=Negatives generation, E.M.=Embedding method, RS=Negatives by randomly sampling, NC=Negatives from non-cognate MHC alleles, SH = Negatives by shuffling, BK = Negatives from background data. OH = one-hot embedding, OR = ordinal embedding, PC = physicochemical property (or Atchley factors)-based embedding, KM=K-mer feature-based embedding, BL=BLOSUM-based embedding, LM=language model-based embedding.

Comprehensive Evaluation of 18 TCR-pMHC Binding Predictors

To provide a much-needed, unified assessment and establish a reproducible benchmark for the field, we conduct a comprehensive, head-to-head evaluation of 18 state-of-the-art TCR-pMHC binding prediction methods mentioned above. In order to make a fair comparison, all models are re-implemented and evaluated using a consistent framework, including unified data preprocessing (one-hot encoding for sequences and categorical attributes like V/J genes and MHC), negative sample generation, and training parameters (i.e., 40 epochs, batch size of 64, Adam optimizer with learning rate 0.0002 and weight decay 1e-5 on McPAS-TCR dataset; 80 epochs, batch size of 64, Adam optimizer with learning rate 0.0001 and weight decay 1e-5 on IEDB and VDJdb datasets). As some original implementations are not fully open-sourced or are based on different deep learning frameworks (e.g., Keras), we re-implement all models in PyTorch; while we strive to preserve the original performance, we cannot guarantee 100% replication of the original results. Evaluations are performed on three major public databases: the IEDB, McPAS-TCR, and VDJdb. The results are shown below.

To assess the clinical utility and robustness of the benchmarked methods, we conduct a critical out-of-distribution (OOD) generalization test on two independent Unseen Epitope Variant Datasets (Dataset I and Dataset II). These datasets, referenced from the stringent ePytope-TCR benchmark, are specifically designed to challenge models with novel peptide sequences, mirroring the difficulties encountered in real-world contexts where neoantigens frequently arise. The comprehensive performance evaluation is presented as follows.

Code Introduction

• data_preprocess.py: preprocess the raw downloaded dataset of IEDB, McPAS, and VDJdb.
• config.py: the file including the necessary settings.
• model.py: the file including all models.
• main.py: the main file to run all models.
• filter_data.py: filter the dataset to ensure the independence of two unseen OOD datasets.
• test_new_data.py: to evaluate the trained model on the two unseen OOD datasets.
• data: the folder including the raw downloaded datasets: IEDB, McPAS, and VDJdb. The size of the raw IEDB dataset is too large (>25MB) to upload to GitHub. You can download it by following the instructions in the guide file data_download.txt located in the data folder.
• unseen_data: the folder including the unseen dataset downloaded from ePytope-TCR benchmark.

License

MIT License