Amogel: a multi-omics classification framework using associative graph neural networks with prior knowledge for...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• High-throughput sequencing enables cancer subtype analysis and targeted treatments.
• Graph neural networks (GNNs) model complex biological systems and non-linear interactions in omics data.
• Current GNN approaches have limitations in simultaneously analyzing multi-omics data with prior biological knowledge.

Purpose of the Study:

• To propose a novel graph classification model, AMOGEL, for effective integration of multi-omics datasets and prior biological knowledge.
• To leverage GNNs combined with association rule mining (ARM) for enhanced cancer classification.
• To address limitations in current GNN models regarding simultaneous multi-omics data integration and prior knowledge utilization.

Main Methods:

• Developed AMOGEL, a graph classification model integrating multi-omics data (miRNA, mRNA, DNA methylation) and prior knowledge (PPIs, KEGG, GO).
• Employed early fusion with ARM to mine intra- and inter-omics relationships, creating a synthetic multi-omics graph.
• Introduced multi-dimensional edges and an attention-based gene ranking technique.

Main Results:

• AMOGEL demonstrated superior performance in classifying BRCA and KIPAN cancer subtypes compared to state-of-the-art models.
• Achieved higher classification accuracy, F1 score, and AUC score by effectively integrating diverse data types.
• Validated the model's capability to leverage multi-omics data and prior biological knowledge.

Conclusions:

• AMOGEL represents a significant advancement in integrating multi-omics data and prior knowledge for cancer subtype classification.
• The proposed model effectively utilizes the non-linear learning potential of GNNs and ARM.
• Findings pave the way for more accurate and personalized cancer diagnostics and treatments.