Table 1.

Algorithm of ABT-MPNN

Initialization
1)	Given a molecular graph $G$, generate atom features $x_v$ and bond features $e_{\mathit{vw}}$ where $v\in Atom\left(G\right)$ and $w\in Neighbor\left(v\right)$; three inter-atomic matrices $M_{\mathit{Adjacency}},M_{\mathit{Distance}},M_{\mathit{Coulomb}}$; molecular descriptors $h_f$
2)	for each atom $v$ in molecule $G$:
3)	for each atom $w$ in molecule $Neighbor\left(v\right)$:
4)	$h_{\mathit{vw}}^0\leftarrow ReLU(W_{i}Concat\left(x_v,,,e_{\mathit{vw}}\right))$

Bond Embedding Phase
1)	Message-passing iteration:$t=1,2,\cdots ,T$
2)	while $1\le t\le T$:
3)	for each atom $v$ in molecule $G$:
4)	for each atom $w$ in molecule $Neighbor\left(v\right)$:
5)	$m_{\mathit{vw}}^t\leftarrow {\sum }_{k\in Neighbor\left(v\right)}{h}_{\mathit{kv}}^{t-1}-h_{\mathit{wv}}^{t-1}$
6)	$b_{\mathit{vw}}^t\leftarrow BondAttention\left(m_{\mathit{vw}}^t\right)+m_{\mathit{vw}}^t$
7)	$h_{\mathit{vw}}^t\leftarrow ReLU(h_{\mathit{vw}}^0+W_h{b}_{\mathit{vw}}^t)$

Atom Embedding Phase
1)	for each atom $v$ in molecule $G$:
2)	$m_v\leftarrow ReLU(W_{o}Concat(x_v,{\sum }_{w\in Neighbor\left(v\right)}{h}_{\mathit{vw}}^T))$
3)	$h_v\leftarrow AtomAttention\left(m_v,,,M_{\mathit{Adjacency}},,,M_{\mathit{Distance}},,,M_{\mathit{Coulomb}}\right)+m_v$

Molecule Embedding Phase
1)	$h\leftarrow {\sum }_{v\in G}{h}_v$
2)	$\widehat{y}\leftarrow FFN(Concat(h,h_f))$