Identifying common transcriptome signatures of cancer by interpreting deep learning models

Area of Science:

• Genomics and Molecular Biology
• Cancer Research
• Bioinformatics

Background:

• Cancer is defined by uncontrolled cell growth and tissue invasion.
• While some cancer-associated genes are known, common transcriptomic features across diverse tumor types remain underexplored.
• Identifying novel, shared molecular alterations could reveal new therapeutic targets.

Purpose of the Study:

• To agnostically identify common transcriptomic features across various solid tumor types.
• To investigate whether these features are independent of known genetic alterations.
• To understand the functional and evolutionary significance of these shared transcriptomic signatures.

Main Methods:

• Utilized 13,461 RNA-sequencing samples from 19 normal and 18 solid tumor types.
• Trained three feed-forward neural networks analyzing protein-coding gene expression, lncRNA expression, and splice junction usage.
• Employed attribution analysis to identify key transcriptomic features distinguishing normal from tumor samples.

Main Results:

• All three models successfully identified consistent transcriptome signatures across different tumor types.
• Genes within these cancer signatures are under significant evolutionary and selective constraints.
• These commonly altered genes are infrequently affected by mutations or genomic alterations and differ from previously identified cancer genes.

Conclusions:

• Deregulation of RNA-processing genes and aberrant splicing are widespread in many solid tumors.
• These RNA-related alterations represent convergent pathways fundamental to core cancer processes.
• This highlights a novel set of molecular features critical for cancer biology, distinct from mutation-driven oncogenesis.