HetEnc: a deep learning predictive model for multi-type biological dataset

BMC Genomics

Table 2 Predictive performance (AUC) for the neuroblastoma dataset

	Model	RNA-seq			Microarray
	Model	FAV	OS_All	OS_HR	FAV	OS_All	OS_HR
Cross-validation	HetEnc	0.964 (0.009)	0.830 (0.019)	0.520 (0.044)	0.962 (0.011)	0.849 (0.024)	0.651 (0.044)
	HetEnc	0.969 (0.007)	0.854 (0.024)	0.592 (0.027)	0.948 (0.015)	0.825 (0.016)	0.569 (0.022)
	Raw-DNN^*	0.926 (0.043)	0.698 (0.058)	0.578 (0.03)	0.906 (0.054)	0.721 (0.035)	0.568 (0.031)
	FS-DNN^*	0.923 (0.052)	0.704 (0.046)	0.558 (0.028)	0.919 (0.056)	0.722 (0.047)	0.559 (0.025)
External Testing (on same testing set)	KNN	0.896 (0.032)	0.641 (0.032)	0.495 (0.048)	0.907 (0.035)	0.662 (0.031)	0.515 (0.041)
	NSC	0.901 (0.036)	0.700 (0.048)	0.499 (0.036)	0.921 (0.032)	0.713 (0.067)	0.510 (0.035)
	SVM	0.894 (0.043)	0.631 (0.024)	0.512 (0.050)	0.914 (0.035)	0.620 (0.034)	0.525 (0.047)
	RandomForest	0.905 (0.014)	0.740 (0.019)	0.563 (0.030)	0.912 (0.012)	0.727 (0.020)	0.560 (0.030)
	XGBoost	0.883	0.742	0.517	0.874	0.749	0.611
	Avg. of Best 60 SEQC Models	0.931 (0.02)	0.735 (0.072)	0.544 (0.052)	0.929 (0.02)	0.756 (0.082)	0.563 (0.038)

^*Raw-DNN used the raw 10,042 gene features as input of DNN model, FS-DNN further applied feature selection threshold (p < 0.05 for each endpoint) before entering the DNN model. The structure of DNN model used in Raw-DNN and FS-DNN are the same as the DNN used in HetEnc supervised learning step

ISSN: 1471-2164