Integrative Data Analysis of Multi-Platform Cancer Data with a Multimodal Deep Learning Approach

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Cancer subtype identification is crucial for personalized therapy and understanding disease.
• High-throughput sequencing generates multi-platform genomic data (gene expression, miRNA, DNA methylation).
• Existing integrative clustering methods often fail to fully exploit intrinsic data properties and cross-modality correlations.

Purpose of the Study:

• To propose a novel machine learning model, the multimodal deep belief network (DBN), for integrative clustering of cancer patients using multi-platform genomic data.
• To develop a framework that captures both intra- and cross-modality correlations for robust cancer subtyping.

Main Methods:

• Developed a multimodal deep belief network (DBN) integrating gene expression, miRNA expression, and DNA methylation data.
• Employed a joint latent model to fuse common features across modalities.
• Utilized contrastive divergence (CD) for unsupervised parameter inference.

Main Results:

• The multimodal DBN effectively extracts unified latent features, capturing intra- and cross-modality correlations.
• Successfully identified meaningful cancer subtypes from multi-platform data.
• Identified key genes and miRNAs, including miR-29a, correlated with survival in ovarian cancer.

Conclusions:

• The multimodal DBN approach offers a powerful method for integrative analysis of multi-platform cancer data.
• This approach can reveal novel cancer subtypes and biomarkers for improved diagnostics and treatment strategies.
• Findings provide valuable insights for cancer pathogenesis research and personalized cancer therapy development.