Multi-Run Concrete Autoencoder to Identify Prognostic lncRNAs for 12 Cancers

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Long non-coding RNAs (lncRNAs) are crucial in regulating gene expression and are implicated in cancer development.
• Identifying prognostic lncRNAs can significantly advance cancer diagnosis and therapeutic strategies.

Purpose of the Study:

• To develop a robust method for identifying key long non-coding RNAs (lncRNAs) capable of distinguishing between 12 different cancer types.
• To assess the prognostic value of identified lncRNAs in patient stratification.

Main Methods:

• Utilized a multi-run concrete autoencoder (mrCAE), a deep learning algorithm, for unsupervised feature selection on genome-wide lncRNA expression data from The Cancer Genome Atlas (TCGA).
• Analyzed 4768 samples across 12 cancer types to identify stable and informative lncRNAs.
• Compared mrCAE performance against single-run CAE, standard autoencoder (AE), and other feature selection methods.

Main Results:

• mrCAE demonstrated superior performance in feature selection compared to existing methods.
• Identified 128 top-ranking lncRNAs with 95% accuracy in distinguishing the origin of 12 different cancers.
• Survival analysis confirmed that 76 of these lncRNAs possess prognostic capabilities, differentiating patient risk groups.

Conclusions:

• The proposed mrCAE effectively identifies actual, reproducible features, outperforming AE in selecting latent features, making it valuable for precision medicine.
• The discovered set of prognostic lncRNAs holds potential for further investigation in developing targeted cancer therapies.