Multi-Omic Graph Transformers for Cancer Classification and Interpretation.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Next-generation sequencing generates large biological datasets, necessitating advanced computational methods for analysis.
• Integrating multi-omics data, such as microRNA (miRNA) and messenger RNA (mRNA), offers significant insights into disease classification and mechanisms.
• Current computational approaches struggle to effectively model and interpret complex miRNA-mRNA interactions within large-scale biological data.

Purpose of the Study:

• To develop a novel computational method for classifying cancer by modeling miRNA-mRNA interactions.
• To leverage graph-based representations with attention mechanisms for enhanced interpretability.
• To analyze multi-omics data for improved cancer classification and biomarker discovery.

Main Methods:

• Development of a Graph Transformer Network (GTN) utilizing message-passing and attention mechanisms to represent miRNA-mRNA interactions.
• Utilized patient-matched miRNA and mRNA expression data from The Cancer Genome Atlas (TCGA) for 12 cancer types.
• Incorporated miRNA targeting information from TargetScan and compared GTN performance against other machine learning models.

Main Results:

• The GTN achieved a 93.56% accuracy in classifying 12 different cancer types.
• Multi-omics models generally outperformed single-omics models in classification tasks.
• Attention analysis identified key targeting pathways and potential molecular biomarkers derived from integrated miRNA and mRNA data.

Conclusions:

• The developed GTN provides a highly interpretable method for cancer classification using integrated multi-omics data.
• The study highlights the utility of graph-based deep learning models for analyzing complex molecular interactions.
• The findings suggest that integrated miRNA-mRNA analysis can reveal crucial insights into cancer biology and identify potential diagnostic or therapeutic targets.