ABT-MPNN: an atom-bond transformer-based message-passing neural network for molecular property prediction

Table 5 Ablation study results on classification (ClinTox) and regression (ESOL) tasks

No	Bond attention		Atom attention		Classification (ClinTox)	Regression (ESOL)
No	Transformer	Fastformer	No inter-atomic matrices	With inter-atomic matrices	Classification (ClinTox)	Regression (ESOL)
1					0.890 \(\pm\) 0.040	0.582 \(\pm\) 0.070
2	√				0.887 \(\pm\) 0.044	0.570 \(\pm\) 0.070
3		√			0.894 \(\pm\) 0.042	0.573 \(\pm\) 0.066
4			√		0.896 \(\pm\) 0.035	0.569 \(\pm\) 0.065
5				√	0.905 \(\pm\) 0.041	0.569 \(\pm\) 0.065
6		√	√		0.905 \(\pm\) 0.028	0.567 \(\pm\) 0.066
7		√		√	0.904 \(\pm\) 0.034	0.566 \(\pm\) 0.075

The evaluation metrics are represented as averaged values ± standard deviation from fivefold CV. The best performance values are highlighted in bold

ISSN: 1758-2946