Deep profiling of gene expression across 18 human cancers.

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Unsupervised deep learning offers potential for analyzing large gene expression datasets.
• Challenges in biological interpretability and methodological robustness have limited its application in cancer research.

Purpose of the Study:

• To develop and validate an unsupervised deep learning framework, DeepProfile, for analyzing gene expression data across multiple human cancers.
• To enhance biological interpretability and methodological robustness in the analysis of large-scale cancer transcriptomic data.

Main Methods:

• Developed DeepProfile, an unsupervised deep learning framework, to generate low-dimensional latent spaces from 50,211 transcriptomes across 18 human cancers.
• Compared DeepProfile's performance against traditional dimensionality reduction methods regarding biological interpretability.

Main Results:

• DeepProfile demonstrated superior biological interpretability compared to other methods.
• Identified universally important genes controlling immune cell activation and cancer-type-specific genes/pathways defining molecular subtypes.
• Linked latent variables to tumor characteristics, finding associations between mutation burden and cell-cycle genes.
• Discovered consistent associations between DNA-mismatch repair, MHC class II antigen presentation pathways, and patient survival.
• Identified tumor-associated macrophages as a source of survival-correlated MHC class II transcripts.

Conclusions:

• Unsupervised deep learning, exemplified by DeepProfile, can effectively extract biological insights from complex gene expression data.
• The framework facilitates the discovery of novel biomarkers and therapeutic targets by linking molecular features to clinical outcomes.