Struct2Graph: a graph attention network for structure based predictions of protein–protein interactions

BMC Bioinformatics

Table 1 Fivefold cross-validation performance analysis of several machine learning methods on balanced dataset (1:1)

Method	Performance (%) on balanced training set—1:1
Method	Accuracy	Precision	Recall	Specificity
GaussianNB	72.14 ± 2.91	98.41 ± 0.51	45.05 ± 6.10	99.24 ± 0.30
QDA	78.66 ± 3.44	70.43 ± 3.41	99.42 ± 0.40	57.90 ± 7.12
k-NN	94.19 ± 0.56	99.49 ± 0.08	88.83 ± 1.10	99.54 ± 0.07
Decision trees	96.20 ± 0.43	97.59 ± 0.28	94.75 ± 0.99	97.66 ± 0.29
Random forest	98.86 ± 0.29	99.45 ± 0.19	98.27 ± 0.49	99.45 ± 0.19
Adaboost	97.85 ± 0.26	98.76 ± 0.18	96.92 ± 0.51	98.78 ± 0.18
SVC	98.49 ± 0.33	99.44 ± 0.18	97.53 ± 0.61	99.45 ± 0.18
DeepPPI	97.22 ± 0.44	98.26 ± 0.82	96.14 ± 0.88	98.29 ± 0.83
DeepFE-PPI	98.64 ± 0.32	99.16 ± 0.28	98.12 ± 0.51	99.17 ± 0.28
Struct2Graph	98.89 ± 0.24	99.50 ± 0.36	98.37 ± 0.34	99.45 ± 0.42

Method	MCC	F1-score	ROC-AUC	NPV
GaussianNB	52.69 ± 4.38	61.53 ± 6.00	95.24 ± 0.33	64.46 ± 2.37
QDA	63.06 ± 5.23	82.40 ± 2.27	78.66 ± 3.43	99.05 ± 0.58
k-NN	88.89 ± 1.02	93.86 ± 0.63	98.79 ± 0.21	89.92 ± 0.89
Decision trees	92.45 ± 0.84	96.15 ± 0.46	96.36 ± 0.30	95.23 ± 0.61
Random forest	97.74 ± 0.58	98.86 ± 0.30	99.63 ± 0.08	98.30 ± 0.46
Adaboost	95.72 ± 0.52	97.83 ± 0.27	99.20 ± 0.08	96.94 ± 0.50
SVC	97.01 ± 0.66	98.48 ± 0.34	99.63 ± 0.09	97.58 ± 0.59
DeepPPI	94.47 ± 0.87	97.19 ± 0.44	99.28 ± 0.11	96.23 ± 0.81
DeepFE-PPI	97.29 ± 0.64	98.64 ± 0.32	99.52 ± 0.09	98.14 ± 0.50
Struct2Graph	97.79 ± 0.49	98.94 ± 0.20	99.55 ± 0.16	98.24 ± 0.42

Bold face numbers indicate the best performance
Note that the proposed Struct2Graph method outperforms all other methods on the majority of metrics

ISSN: 1471-2105