Selecting precise reference normal tissue samples for cancer research using a deep learning approach | JoVE Visualize

Area of Science:

Genomics
Bioinformatics
Cancer Research

Background:

Matched normal tissue samples are crucial for disease mechanism studies but are often difficult to obtain.
Cancer genomics resources like TCGA and TARGET lack matched normal tissues for all cancer types.
The Genotype-Tissue Expression (GTEx) project offers healthy tissue samples, but their utility as a reference for cancer research is unproven.

Purpose of the Study:

To systematically evaluate the feasibility of using GTEx normal tissue samples as a reference for cancer research.
To assess the consistency of disease expression signatures derived from GTEx normal tissues versus adjacent normal tissues from TCGA.
To identify computational methods for selecting appropriate normal samples from GTEx for disease signature analysis.

Main Methods:

Analyzed RNA-Seq data using a standardized computational pipeline.
Evaluated GTEx as a reference by comparing it against TCGA cancers with adjacent normal tissues.
Employed principal component analysis and autoencoder neural networks to correlate tumor and normal samples.
Assessed the ability of methods to predict tissue of origin and disease expression signature consistency.

Main Results:

18 out of 32 TCGA cancers had fewer than 10 matched adjacent normal tissue samples.
Autoencoder models demonstrated superior performance in predicting tissue of origin, correctly classifying 12 out of 14 cancers.
GTEx provides suitable normal samples for most cancers, but not all, with autoencoders aiding in selection.
Disease signatures derived from GTEx normal samples selected by autoencoders showed consistency with TCGA adjacent normal samples.
Using top 50 correlated samples, irrespective of tissue type, yielded comparable or improved results in some cancers.

Conclusions:

GTEx normal tissue samples are a viable reference for cancer studies, particularly when matched adjacent tissues are unavailable.
Deep learning approaches, such as autoencoders, are promising for selecting appropriate normal reference samples.
This study provides a computational framework for leveraging GTEx data in cancer research.