Model performance and interpretability of semi-supervised generative adversarial networks to predict oncogenic variants with unlabeled data

Table 1 Performance comparison when using different number of labeled variants and different types of features. We trained SGAN models based on different size of labels (ranging from 250 to 4000 variants) and three groups of features: pathogenicity prediction scores, evidence-based scores, and ensemble scores (full features)

Features	Training size	Accuracy	Precision	Recall	Specificity	F1score	MCC	ROC AUC	PR AUC
Pathogenicity prediction scores	250	0.631	0.339	0.757	0.597	0.468	0.29	0.732	0.413
	500	0.643	0.346	0.748	0.614	0.473	0.299	0.741	0.421
	1000	0.625	0.337	0.774	0.584	0.469	0.294	0.734	0.412
	2000	0.642	0.346	0.751	0.613	0.474	0.299	0.737	0.413
	4000	0.631	0.34	0.767	0.594	0.471	0.297	0.734	0.406
Evidence-based scores	250	0.727	0.415	0.667	0.743	0.512	0.355	0.815	0.543
	500	0.75	0.447	0.706	0.762	0.548	0.406	0.814	0.627
	1000	0.729	0.428	0.778	0.716	0.552	0.416	0.848	0.676
	2000	0.705	0.406	0.81	0.677	0.541	0.404	0.849	0.678
	4000	0.78	0.492	0.781	0.78	0.604	0.486	0.832	0.647
Ensemble	250	0.541	0.306	0.897	0.444	0.456	0.29	0.786	0.507
	500	0.637	0.355	0.85	0.579	0.501	0.352	0.828	0.617
	1000	0.676	0.386	0.864	0.625	0.534	0.402	0.854	0.688
	2000	0.653	0.371	0.889	0.589	0.524	0.392	0.859	0.691
	4000	0.673	0.383	0.858	0.622	0.53	0.395	0.854	0.686

ISSN: 1471-2105