MetaCancer: A deep learning-based pan-cancer metastasis prediction model developed using multi-omics data.

Area of Science:

• Computational biology
• Genomics
• Oncology

Background:

• Early prediction of cancer metastasis is crucial for timely treatment adjustments.
• Current computational methods often lack a pan-cancer perspective and focus on single genomic levels.
• There is a need for advanced models that integrate multi-omic data for improved metastasis prediction.

Purpose of the Study:

• To develop a deep learning (DL)-based model, MetaCancer, for pan-cancer metastasis status prediction.
• To evaluate the efficacy of integrating heterogeneous data layers (RNA-Seq, microRNA-Seq, DNA methylation) for metastasis prediction.
• To compare the performance of the proposed DL model against traditional machine learning methods.

Main Methods:

• Developed a DL model (convolutional variational autoencoder - CVAE) using multi-omic data from 400 cancer patients from The Cancer Genome Atlas (TCGA).
• Integrated RNA sequencing (mRNA), microRNA sequencing (microRNA-Seq), and DNA methylation data.
• Assessed feature importance and compared model performance using various metrics against alternative methods.

Main Results:

• Integrating multi-omic data significantly improved MetaCancer's performance compared to using mRNA data alone.
• mRNA-related features were identified as most significant for distinguishing primary tumors from metastatic ones.
• The proposed DL model demonstrated superior performance over a machine learning ensemble method.

Conclusions:

• The MetaCancer model offers a promising approach for early and accurate prediction of cancer metastasis across various cancer types.
• Multi-omic data integration is key to enhancing the predictive power of computational models for cancer metastasis.
• This study highlights the potential of deep learning in advancing precision oncology and improving patient outcomes.