Tumor cell specific total mRNA expression informed neural networks predicts cancer progression

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Estimating tumor cell-specific mRNA expression (TmS) is crucial for understanding cancer phenotypes.
• Existing methods require matched DNA and RNA sequencing and are computationally intensive.

Purpose of the Study:

• To develop a deep learning framework (TmSNet) for predicting TmS from multi-omic data without matched DNA sequencing.
• To provide a scalable and computationally efficient alternative to current TmS estimation methods.

Main Methods:

• TmSNet integrates structured feature selection (gradient boosting, LASSO, elastic net) with neural networks.
• Input features include mRNA, DNA methylation, miRNA, and immune cell proportions.
• The model was trained and cross-validated on 12 TCGA cancer types and tested on external cohorts.

Main Results:

• TmSNet achieved high cross-validated performance (CCC = 0.93, R² = 0.88) in TCGA cohorts.
• The model generalized to external cohorts (SCAN-B: 0.54, FUSCC: 0.43).
• Predicted TmS values effectively stratified patients by risk and maintained known tumor subtype transcriptional profiles.

Conclusions:

• TmSNet accurately infers tumor molecular phenotypes and transcriptional activity from multi-omic data.
• This deep learning framework offers a scalable solution for analyzing heterogeneous cancer cohorts.
• TmSNet facilitates a deeper understanding of tumor biology without the need for matched DNA sequencing.