Table 4.

Performance evaluation on more realistic settings of KIBA datasets.

Scenario	Method	MSE $\downarrow$	CI $\uparrow$	$r_m^2\uparrow$
Cold drug	GraphDTA	0.471 (0.047)	0.713(0.002)	0.342(0.007)
	GEFA	0.464(0.032)	0.721(0.003)	0.346(0.006)
	FusionDTA	0.429(0.031)	0.748(0.005)	0.364(0.012)
	MgraphDTA	0.425(0.047)	0.746(0.002)	0.366(0.016)
	NHGNN(Ours)	0.385(0.029)	0.756(0.007)	0.400(0.015)
Cold target	GraphDTA	0.469(0.089)	0.610(0.035)	0.368(0.057)
	GEFA	0.462(0.091)	0.636(0.037)	0.362(0.052)
	FusionDTA	0.439(0.062)	0.685(0.032)	0.390(0.067)
	MgraphDTA	0.435(0.055)	0.674(0.028)	0.382(0.047)
	NHGNN(Ours)	0.382(0.071)	0.732(0.041)	0.452(0.054)
All cold	GraphDTA	0.676(0.113)	0.601(0.030)	0.149(0.067)
	GEFA	0.639(0.065)	0.628(0.047)	0.152(0.035)
	FusionDTA	0.587(0.086)	0.641(0.023)	0.193(0.053)
	MgraphDTA	0.590(0.094)	0.626(0.028)	0.182(0.012)
	NHGNN(Ours)	0.565(0.094)	0.649(0.037)	0.218(0.047)

Bold corresponds to the best performance for each metric, and underline indicates the second best.