Table 5.

Transfer learning methods for representative MHC alleles

MHC allel	SRCC	AUC	high affinity rate1	transfer learning method2	Improvement3
A*6802	0.499	0.817	0.9	fine-tune	0.61
	0.058	0.537	0.29	zero-shot
A*0206	0.458	0.778	0.75	fine-tune	0.08
	0.378	0.73	0.67	zero-shot
B*2705	0.701	0.929	0.92	fine-tune	0.92
	0.167	0.602	0	zero-shot
Mamu-A1*001:01	0.755	0.943	0.839	fine-tune	0.809
	0.256	0.65	0.03	zero-shot
A*0201	0.117	0.576	0.859	originally

^*The length of the peptides for training transfer learning is nine.

¹The high-affinity is scored by http://tools.iedb.org/mhci/.

²There are two types of transfer learning. Fine-tune indicates keeping training on the basis of the original model with a smaller learning rate ($\frac{1}{10}$ learning rate). Zero-shot indicates that direct transferring without further training. Originally indicates the model is the well-trained model that all the transfer learning is based on.

³the Improvement stands for the increase of high-affinity rate from zero-shot transfer method to fine-tune method