Table 2.

Average performance (F1, accuracy, precision, recall) of the different models in a supervised task scenario. Although our pathway-primed models are nearly ten times smaller (sparse), the performance is very close to the PPI-based NN. We report the mean for 100 iterations of train test splits

				F1			PRECISION			RECALL
Architecture	Number of nodes (2nd hidden layer)	Accuracy	Balanced accuracy	Macro	Micro	Weighted	Macro	Micro	Weighted	Macro	Micro	Weighted
Dense	–	0.825	0.788	0.748	0.825	0.802	0.769	0.825	0.844	0.788	0.825	0.825
Dense with pathways	–	0.810	0.781	0.743	0.810	0.783	0.763	0.810	0.823	0.781	0.810	0.810
Dense with PPI	–	0.802	0.770	0.730	0.802	0.774	0.753	0.802	0.817	0.770	0.802	0.802
Dense with PPI/GRN	–	0.800	0.777	0.735	0.800	0.771	0.757	0.800	0.815	0.777	0.800	0.800
Signaling pathways	–	0.813	0.781	0.743	0.813	0.790	0.764	0.813	0.834	0.781	0.813	0.813
	100	0.766	0.724	0.673	0.766	0.728	0.690	0.766	0.762	0.724	0.766	0.766