Table 2.

Ablation studies on GraphBind with different settings^a

	Posb	PEc	rgd	rve	EUf	Ag	GRUh	NLi	Dj	Rec	Pre	F1	MCC	AUC
Base	C	T	20	10	T	S	T	4	128	0.676	0.537	0.598	0.558	0.926
(A)	SC									0.593	0.591	0.592	0.552	0.922
	CA									0.633	0.537	0.581	0.538	0.921
(B)		F								0.650	0.528	0.583	0.540	0.920
(C)			15							0.634	0.551	0.589	0.548	0.919
			25							0.656	0.540	0.593	0.551	0.923
			30							0.580	0.594	0.587	0.547	0.913
(D)				5						0.622	0.472	0.537	0.490	0.910
				13						0.663	0.540	0.595	0.555	0.923
(E)					F					0.570	0.483	0.523	0.474	0.899
(F)						M				0.561	0.407	0.472	0.418	0.875
(G)								2		0.630	0.524	0.573	0.529	0.914
								3		0.647	0.551	0.595	0.554	0.925
								5		0.647	0.545	0.592	0.550	0.925
								6		0.688	0.522	0.586	0.545	0.924
							F	2		0.670	0.523	0.587	0.546	0.925
							F	4		0.637	0.541	0.585	0.543	0.922
							F	6		0.669	0.504	0.575	0.533	0.922
(H)									64	0.593	0.527	0.558	0.513	0.910

^aOnly different settings are given and other settings (empty values) are the same as the base model. These metrics are calculated on the validation set of DNA-573_Train and the highest values are bolded.

^bPseudo-position of a residue: C, SC and CA stand for the centroid of residue, the centroid of residue side-chain and the position of alpha-C atom, respectively.

^cUse the relative distance from every node to the sphere center as position embeddings of nodes (T) or not (F).

^dRadius of the structural context: it defines the nodes belonging to a graph of a residue, and its unit is Å.

^eThe threshold of adjacent matrix: it binarizes a distance matrix to the adjacent matrix to define the adjacent edges belonging to a node, and its unit is Å.

^fUse the edge feature vectors (T) or not (F).

^gThe aggregation operation in the node update module and the graph update module. S and M stand for sum and max operation, respectively.

^hUse GRU (T) or not (F). If GRU is not used, the output ,, equal the intermediate output , , , respectively.

ⁱThe number of GNN-blocks.

^j D _e, D_v and D_u stand for the dimension of encoded edge feature vectors, the dimension of encoded node feature vectors and the dimension of encoded graph feature vectors, respectively. We set D_eD_vD_u.