Integration of Multimodal Data from Disparate Sources for Identifying Disease Subtypes

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Cancer progression and survival prediction are challenged by disease heterogeneity.
• Single-modality molecular data analysis is insufficient for accurate prognostication.
• Integrating multi-omics data offers a more robust approach to understanding cancer.

Purpose of the Study:

• To develop a deep learning-based multimodal data fusion system for enhanced cancer progression prediction.
• To improve the accuracy and robustness of predicting patient outcomes by integrating diverse molecular data.
• To address the challenge of missing data in multi-repository cancer studies.

Main Methods:

• An autoencoder-based multimodal data fusion system was designed.
• A fusion encoder was utilized to flexibly integrate information from multiple studies with partially coupled data.
• The system leveraged multiple data modalities including mRNA, DNA Methylation, and miRNA.

Main Results:

• The proposed data fusion pipeline successfully inferred missing data, outperforming baseline predictors.
• The system achieved high accuracy in differentiating short- and long-term survivors for glioblastoma multiforme (AUC 0.94), acute myeloid leukemia (AUC 0.75), and pancreatic adenocarcinoma (AUC 0.96).
• Controlled simulation studies validated the superiority of the multimodal approach.

Conclusions:

• Multimodal data fusion using deep learning significantly enhances the prediction of cancer progression and patient survival.
• The autoencoder-based system provides a robust framework for integrating heterogeneous molecular data from multiple sources.
• Accurate differentiation of cancer patient outcomes is achievable through advanced computational approaches.