OmniCLIC: A Unified Omics Contrastive Learning Framework for Effective Integration and Classification of Multiomics...

Area of Science:

• Computational biology
• Bioinformatics
• Cancer research

Background:

• Multiomics data integration is crucial for cancer subtype classification but faces challenges like high dimensionality and heterogeneity.
• Existing methods often struggle with feature interpretability and generalization across diverse omics data.

Purpose of the Study:

• To develop a unified framework, OmniCLIC, for end-to-end multiomics integration, feature learning, and cancer subtype prediction.
• To improve the accuracy, robustness, and interpretability of cancer classification using multiomics data.

Main Methods:

• OmniCLIC utilizes OmniNet for omics-specific representation learning, a contrastive learning module for joint optimization, and OCDN for decision-level fusion.
• The framework incorporates feature-wise scaling and cross-modal correlation tensors to capture interomics relationships and enhance generalization.
• It was evaluated on four benchmark cancer datasets and extended to single-cell multiomics data (RNA+ATAC, RNA+ADT).

Main Results:

• OmniCLIC consistently outperformed state-of-the-art methods in accuracy and robustness on both binary and multiclass cancer datasets.
• The framework enabled biologically meaningful interpretation by identifying key molecular features and subtype-specific pathways.
• OmniCLIC demonstrated superior performance on single-cell multiomics data, confirming its generalizability.

Conclusions:

• OmniCLIC offers a powerful and interpretable solution for multiomics data integration in cancer subtype classification.
• The framework's ability to identify key molecular features and pathways provides valuable insights for cancer biology and precision medicine.
• OmniCLIC's generalizability across different data scales highlights its potential for broad application in multiomics research.