Fig. 1: AUTOSurv workflow and key components illustration.

a KL-PMVAE was trained to conduct integration and dimension reduction on gene expression and miRNA expression data. b Latent features generated by KL-PMVAE will be combined with the demographic/clinical variables and fed into the LFSurv network. The output of LFSurv will be a prognostic index (\({PI}\)) for each patient that reflects the patient’s risk of death. \({{PI}}_{{med}}\): median prognostic index. c Illustration of KL-PMVAE. The VAE model consists of an encoder and a decoder. The encoder has one gene layer (each node represents a gene), one pathway layer (each node represents a pathway), and one miRNA layer (each node represents a miRNA) and learns a distribution estimate of the latent variables \(z\) (parameterized by means \(\mu\) and variances \({\sigma }^{2}\) which were stored in the latent bottleneck). The decoder takes a sample \(\hat{z}\) from the distribution estimate as input and outputs the reconstructed expression data \({\hat{x}}_{{miRNA}}\) and \({\hat{x}}_{{gene}}\). d Illustration of LFSurv. This network consists of an input layer, a hidden layer, and an output layer with only one node. The extracted latent features \(\mu\) were concatenated with the demographic/clinical variables. The network receives the concatenated features and outputs the prognostic index (\({PI}\)).