Table 4.

Comparison of the proposed method with SOTA model in terms of three settings from the Davis dataset with 5-fold cross-validation

Model	Novel pair		Novel compound		Novel protein
	MSE	CI	MSE	CI	MSE	CI
DeepDTA (Öztürk et al., 2018)	0.631(±0.059)	0.533(±0.027)	0.482(±0.034)	0.613(±0.029)	0.701(±0.045)	0.759(±0.015)
DeepConvDTI (Lee et al., 2019)	0.598(±0.057)	0.546(±0.043)	0.512(±0.046)	0.681(±0.012)	0.789(±0.109)	0.714(±0.034)
TransformerCPI (Chen et al., 2020)	0.549(±0.038)	0.490(±0.032)	0.522(±0.027)	0.592(±0.026)	0.708(±0.032)	0.676(±0.005)
GraphDTA (GINs) (Nguyen et al., 2021)	0.846(±0.058)	0.459(±0.032)	0.452(±0.051)	0.670(±0.018)	0.970(±0.061)	0.660(±0.016)
HyperattentionDTI (Zhao et al., 2022)	0.671(±0.045)	0.517(±0.013)	0.506(±0.015)	0.578(±0.019)	0.784(±0.063)	0.674(±0.020)
Perceiver CPI (ours)	0.463(±0.013)	0.638(±0.028)	0.378(±0.010)	0.726(±0.017)	0.667(±0.018)	0.758(±0.010)

Note: The metrics are MSE (the lower, the better) and CI (the higher, the better) (± standard deviation)