A deep learning approach for orphan gene identification in moso bamboo (Phyllostachys edulis) based on the CNN + Transformer model

Table 3 The average BA and GM values under different CNN + Transformer structures

Method	BA	GM	Train time (min)	Test time (s)
E + FC_256	0.677	0.612	25	88
E + CNN_6 + FC_256	0.748	0.644	29	106
E + CNN_6 + CNN_3 + FC_256	0.773	0.784	28	99
E + Transformer + FC_256	0.844	0.832	57	390
E + CNN_6 + Transformer + FC_256	0.866	0.849	61	377
E + CNN_6 + CNN_3 + Transformer + FC_256	0.877	0.881	44	342
E + CNN_6 + CNN_3 + CNN_3 + Transformer + FC_256	0.853	0.837	48	366
E + CNN_6 + CNN_3 + Transformer + FC_256 + FC_256	0.871	0.865	55	404

Bold values indicate the highest values of different evaluation indicators
E: word embedding coding. CNN_6: multiscale convolution layer, with a convolution kernel size of {2, 3, 4, 5, 6, 7} for each scale. CNN_3: multiscale convolution layer, the convolution kernel size of each scale is {3, 6, 9}. Transformer: three-layer transformer neural network. FC_256: fully connected neural network with 256 neurons

ISSN: 1471-2105