Table 2.

Benchmark results of MCC2 for SignalP 6.0, Fine-tuning ESM2-3B, and PEFT-SP models using different PEFT methods with the ESM2-3B backbone

Method/Backbone		Fine-tuning	Prompt tuning	Adapter tuning	LoRA
SP types	SignalP 6.0	ESM2-3B	ESM2-3B	ESM2-3B	ESM2-3B
Archaea Sec/SPI	0.793	0.771	0.777	0.825	0.783
Archaea Sec/SPII	0.825	0.864	0.509	0.783	0.730
Archaea Sec/SPIII^a	0.426	0.724	0.500	0.351	0.798
Archaea Tat/SPI	0.563	0.564	0.653	0.538	0.579
Archaea Tat/SPII^a	0.718	0.792	0.182	0.660	0.850
Eukarya Sec/SPI	0.958	0.948	0.954	0.954	0.960
Negative Sec/SPI	0.804	0.813	0.723	0.820	0.809
Negative Sec/SPII	0.929	0.946	0.886	0.950	0.945
Negative Sec/SPIII^a	0.902	0.982	0.970	0.899	0.919
Negative Tat/SPI	0.962	0.902	0.853	0.899	0.961
Negative Tat/SPII^a	0.486	0.358	0.325	0.405	0.520
Positive Sec/SPI	0.770	0.810	0.746	0.814	0.848
Positive Sec/SPII	0.882	0.908	0.833	0.911	0.939
Positive Sec/SPIII^a	0.902	1.000	0.951	0.969	1.000
Positive Tat/SPI	0.799	0.746	0.590	0.752	0.850
Positive Tat/SPII^a	0.786	0.603	0.148	0.669	0.783
Mean (MCC2)	0.781	0.796	0.663	0.762	0.830

Parameter-efficient fine-tuning on large protein language models improves signal peptide prediction