Skip to main content
Fig. 2 | BMC Medical Genomics

Fig. 2

From: Predicting drug response of tumors from integrated genomic profiles by deep neural networks

Fig. 2

Model construction and evaluation using CCLE datasets. (a) Density plots of true (with missing values), imputed, and predicted IC50 data of CCLE and predicted data of TCGA. (b) Heatmaps of imputed and predicted IC50 data of CCLE. (c, d) Sample-wise Pearson and Spearman correlation between imputed and predicted IC50 data of CCLE samples. (e) Mean square errors of DeepDR and 4 other DNN-based designs. DeepDR was compared to a model with no TCGA pre-training (with encoders randomly initialized; abbreviated as Rand Init), with encoders substituted by PCAs, with Eenc only (no Menc), and with Menc only (no Eenc). Each model was trained for 100 times, each of which CCLE samples were randomly assigned into training, validation, and testing sets

Back to article page