Deep Learning-Based Multi-Omics Data Integration Reveals Two Prognostic Subtypes in High-Risk Neuroblastoma

Abstract

High-risk neuroblastoma is a very aggressive disease, with excessive tumor growth and poor outcomes. A proper stratification of the high-risk patients by prognostic outcome is important for treatment. However, there is still a lack of survival stratification for the high-risk neuroblastoma. To fill the gap, we adopt a deep learning algorithm, Autoencoder, to integrate multi-omics data, and combine it with K-means clustering to identify two subtypes with significant survival differences. By comparing the Autoencoder with PCA, iCluster, and DGscore about the classification based on multi-omics data integration, Autoencoder-based classification outperforms the alternative approaches. Furthermore, we also validated the classification in two independent datasets by training machine-learning classification models, and confirmed its robustness. Functional analysis revealed that MYCN amplification was more frequently occurred in the ultra-high-risk subtype, in accordance with the overexpression of MYC/MYCN targets in this subtype. In summary, prognostic subtypes identified by deep learning-based multi-omics integration could not only improve our understanding of molecular mechanism, but also help the clinicians make decisions.

Variational autoencoders for cancer data integration: design principles and computational practice

“…More generally, our work relates to other approaches based on autoencoders for data integration on various tasks of cancer diagnosis and survival analysis. These include using Denoising Autoencoders for integrating various types of electronic health records (Miotto et al, 2016) as well as custom designed autoencoders for analyses of liver , bladder and neuroblastoma (Zhang et al, 2018) cancer types. CANCER DATA INTEGRATION In a broader context, our work is related to the long tradition of data integration approaches for addressing various challenges in cancer analyses.…”

“…Autoencoders have been deployed on a variety of tasks across different data types such as VARIATIONAL AUTOENCODERS FOR CANCER DATA INTEGRATION dimensionality reduction, data denoising, compression and data generation. In the context of cancer data integration, several studies highlighted their utility in combining data on different scales for identifying prognostic cancer traits such as liver , breast (Tan et al, 2015) and neuroblastoma cancer (Zhang et al, 2018) sub-types. The focus of these studies is to apply autoencoders to specific problems of cancer-data integration.…”

Uncovering Prognosis-Related Genes and Pathways by Multi-Omics Analysis in Lung Cancer

“…To predict lung-cancer-patient survival, we developed several supervised classifiers for which inputs were obtained from the unsupervised autoencoder. We first considered developing an SVM model because of the previously reported prediction success using multi-omics data from TCGA LIHC [12] and the neuroblastoma project combined from Therapeutically Applicable Research to Generate Effective Treatment (TARGET) with Sequencing Quality Control [13].…”

“…The result of confusion matrix is as shown in Table 2. Although the combination of the feature selection by ANOVA, followed by the development of an SVM model, gave the best performance of cancer-patient-survival prediction [13]. In this case, we investigated three additional classifiers, KNN performed with either a hyperparameter of Manhattan or Euclidean distance, RF with either a hyperparameter of Entropy or Gini impurity, and LR with either L1 or L2 regression.…”

Integrative Biology Approaches Applied to Human Diseases

“…In contrast, a top-down approach consists of the parallel clustering of different categories of data for automated and unified integration (92). Top-down methods consist of statistical and machine learning tools such as joint models (100), Bayesian analysis (101), factor analysis (102), multiple kernel learning (103), deep learning (104), and simultaneous clustering (105). There are many useful data integration methods, and the method selection depends on the nature of the data to be analyzed.…”