Table 1 |.

Method and feature comparison of BigMHC and prior works

		BigMHC	PRIME-2.0	MixMHCpred-2.2	TransPHLA	PRIME-1.0	MixMHCpred-2.1	MHCflurry-2.0	NetMHCpan-4.1	MHCnuggets-2.4.0	HLAthena
Publication year		2023	2023	2023	2022	2021	2020	2020	2020	2020	2020
Training	BA				X		X	X	X	X
	EL	X		X	X		X	X	X	X	X
	IM	X	X			X
Prediction	BA							X	X	X
	EL	X		X	X		X	X	X	X	X
	IM	X	X			X
Optional extra context								X			X
Retrainable		X						X		X	X
Transfer Learning		X
Open source		X	X	X	X	X	X	X		X	X
Pan-allele		X			X			X	X		X
Optional single-GPU		X			X			X		X
Optional multi-GPU		X
Has webserver			X	X	X	X	X		X		X
Min peptide Length		8	8	8	8	8	8	5	8	None	8
Max peptide Length		None	14	14	15	14	14	15	None	None	11
Allows wild-type amino acids		X			X				X	X

Cells with ‘X’ indicate that the method has the given feature. Training rows indicate the type of data on which models are trained, whereas prediction rows indicate what type of peptides the model explicitly predicts. Models that are provided with executables or source code for retraining on new data are considered retrainable. Pan-allele methods are those that encode the MHC sequence to generalize predictions across alleles rather than employing multiple allele-specific models. Optional extra context refers to any optional input, such as N-terminal and C-terminal flanking sequences or gene expression data. Models that can consume wild-type amino acids, are indicated in the final row. IM, immunogenicity.