Abstract
In light of the rapid accumulation of large-scale omics datasets, numerous studies have attempted to characterize the molecular and clinical features of cancers from a multi-omics perspective. However, there are great challenges in integrating multi-omics using machine learning methods for cancer subtype classification. In this study, MoGCN, a multi-omics integration model based on graph convolutional network (GCN) was developed for cancer subtype classification and analysis. Genomics, transcriptomics and proteomics datasets for 511 breast invasive carcinoma (BRCA) samples were downloaded from the Cancer Genome Atlas (TCGA). The autoencoder (AE) and the similarity network fusion (SNF) methods were used to reduce dimensionality and construct the patient similarity network (PSN), respectively. Then the vector features and the PSN were input into the GCN for training and testing. Feature extraction and network visualization were used for further biological knowledge discovery and subtype classification. In the analysis of multi-dimensional omics data of the BRCA samples in TCGA, MoGCN achieved the highest accuracy in cancer subtype classification compared with several popular algorithms. Moreover, MoGCN can extract the most significant features of each omics layer and provide candidate functional molecules for further analysis of their biological effects. And network visualization showed that MoGCN could make clinically intuitive diagnosis. The generality of MoGCN was proven on the TCGA pan-kidney cancer datasets. MoGCN and datasets are public available at https://github.com/Lifoof/MoGCN. Our study shows that MoGCN performs well for heterogeneous data integration and the interpretability of classification results, which confers great potential for applications in biomarker identification and clinical diagnosis.
Highlights
Owing to the recent rapid developments in high-throughput sequencing technology, multi-omics research has strongly promoted the development of precision medicine
By applying MoGCN on the breast invasive carcinoma (BRCA) data in The Cancer Genome Atlas (TCGA, http://cancergenome.nih.gov/), we demonstrated that MoGCN could achieve the best performance in cancer subtype classification among the current algorithms
We used random forest as a benchmark classifier to compare the performance of different dimensionality reduction algorithms (Tables 2, 3)
Summary
Owing to the recent rapid developments in high-throughput sequencing technology, multi-omics research has strongly promoted the development of precision medicine. Chaudhary et al were the first to use a deep autoencoder (AE) (Hinton and Salakhutdinov, 2006) model to predict the survival of patients with hepatocellular carcinoma (Chaudhary et al, 2018); Chen et al designed a deep-learning framework, DeepType, that performs a joint model of supervised classification, unsupervised clustering, and dimensionality reduction to learn cancer-relevant data representation (Chen et al, 2020) These methods can handle large-scale datasets, but require substantial effort to interpret how specific features contribute to the predicted results. The non-Euclidean data integration approach trains models using the network topology data These methods can identify cancer subtypes by fusing the similarities derived from various omics data, such as similarity network fusion (SNF) (Wang et al, 2014), GrassmannCluster (Ding et al, 2019), and high-order path elucidated similarity (HOPES) (Xu et al, 2019). These networkbased processes are clinically intuitive, but existing studies have focused on the unsupervised integration of multi-omics datasets
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
