Identifying cancer driver genes based on multi-view heterogeneous graph convolutional network and self-attention mechanism

Table 3 Ablation experiments

Methods	AUC	AUPRC
Gene–gene network	0.9101 ± 0.0002	0.8312 ± 0.0006
Gene–outlying gene network	0.9093 ± 0.0002	0.8325 ± 0.0006
Gene–miRNA network	0.9128 ± 0.0003	0.8346 ± 0.0007
Gene–gene and gene–outlying gene networks	0.9158 ± 0.0002	0.8407 ± 0.0006
Gene–gene and gene–miRNA networks	0.9174 ± 0.0002	0.8428 ± 0.0006
Gene–outlying gene and gene–miRNA networks	0.9187 ± 0.0002	0.8440 ± 0.0006
No pre-training on the gene–miRNA network	0.9177 ± 0.0002	0.8426 ± 0.0007
Removal of bilinear aggregation layer	0.9185 ± 0.0002	0.8440 ± 0.0006
Removal of the self-attention layer	0.9180 ± 0.0002	0.8438 ± 0.0006
Removal of logistic regression model	0.9178 ± 0.0002	0.8426 ± 0.0006
Using Random Forest as classifier	0.9055 ± 0.0003	0.8250 ± 0.0007
Using XGBoost as classifier	0.8990 ± 0.0003	0.8078 ± 0.0010
MRNGCN	0.9192 ± 0.0002	0.8446 ± 0.0006

Bold values indicates the best performance
Comparison of AUC and AUPRC values of MRNGCN and its variants on pan-cancer dataset

ISSN: 1471-2105