T4Seeker: a hybrid model for type IV secretion effectors identification

Table 3 Performance on an independent test set of different models

Models	SP	Precision	Recall	ACC	F1-score	MCC	AUC
Single feature models
DR	0.895	0.89	0.807	0.85	0.846	0.704	0.923
ESM-average	0.901	0.903	0.873	0.887	0.887	0.774	0.943
ESM-flatten-1024	0.851	0.858	0.866	0.858	0.851	0.717	0.925
LSTM	0.785	0.812	0.884	0.836	0.846	0.673	0.920
Concatenated features models
LSTM + ESM-average + DR	0.822	0.842	0.902	0.863	0.871	0.727	0.941
T4Seeker = LSTM + ESM-flatten-1024 + DR	0.944	0.945	0.92	0.932	0.932	0.864	0.970
Ablation studies
ESM-average + DR	0.92	0.92	0.846	0.882	0.880	0.767	0.937
ESM-flatten-1024 + DR	0.907	0.907	0.875	0.890	0.891	0.781	0.944
LSTM + DR	0.907	0.906	0.857	0.881	0.881	0.764	0.944
LSTM + ESM-average	0.916	0.913	0.839	0.877	0.874	0.756	0.961
LSTM + ESM-flatten-1024	0.907	0.907	0.866	0.886	0.886	0.773	0.961
Other models
Bastion4	0.963	0.958	0.821	0.89	0.885	0.79	0.892
T4SEpp	0.822	0.846	0.929	0.876	0.885	0.757	0.875
DeepSecEbd	0.972	0.970	0.848	0.909	0.904	0.825	0.910
T4SEfinder	0.495	0.647	0.884	0.694	0.747	0.413	0.690

ISSN: 1741-7007