Survival stratification for colorectal cancer via multi-omics integration using an autoencoder-based model

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Prognosis stratification is crucial for tailoring colorectal cancer treatments.
• Addressing cancer heterogeneity requires advanced analytical methods.
• Multi-omics data integration offers potential for improved prognostic models.

Purpose of the Study:

• To develop and validate an autoencoder-based model for colorectal cancer prognosis prediction.
• To integrate multi-omics data (DNA methylation, RNA-seq, miRNA-seq) for enhanced feature extraction.
• To compare the autoencoder strategy with other dimensionality reduction and survival analysis methods.

Main Methods:

• Utilized The Cancer Genome Atlas (TCGA) database for multi-omics data.
• Implemented an autoencoder to transform and extract 175 survival-related features.
• Applied k-means clustering to stratify samples based on extracted features.
• Validated the model using hold-out sets and five external cohorts.

Main Results:

• The autoencoder strategy outperformed PCA, t-SNE, NMF, and Cox-PH in identifying survival-related features.
• Clustering revealed two distinct patient groups (G1 and G2) with significantly different survival rates.
• Significant variations in gene expression, miRNA, DNA methylation, and pathway profiles were observed between the prognosis groups.
• Constructed miRNA-mRNA networks involving key differentially expressed miRNAs and target genes.

Conclusions:

• The autoencoder-based framework accurately distinguishes between good and poor prognosis colorectal cancer patients.
• This computational approach enhances understanding of colorectal cancer molecular biology.
• The model facilitates personalized treatment strategies by improving prognosis stratification.