CrossAttOmics: multiomics data integration with cross-attention.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in healthcare

Background:

• High-throughput technologies provide diverse omics data, each offering a partial view of biological processes.
• Integrating multiple omics layers is crucial for accurate disease diagnosis, but requires methods to handle complex data relationships.
• Exploiting known regulatory links between omics can improve multimodal data representation.

Purpose of the Study:

• To introduce CrossAttOmics, a novel deep learning architecture for multiomics integration.
• To leverage cross-attention mechanisms to model interactions between different omics modalities.
• To enhance the accuracy of cancer type prediction by integrating multiomics data.

Main Methods:

• Each omics data type is encoded into a lower-dimensional space.
• A cross-attention mechanism is employed to compute interactions between modalities based on known regulatory links.
• The model architecture facilitates the construction of a comprehensive multimodal representation.

Main Results:

• CrossAttOmics accurately predicts cancer types by effectively utilizing multiomics interactions.
• The proposed model demonstrates superior performance compared to existing methods, particularly when training data is scarce.
• The integration of attribution methods like LRP allows for the identification of key interactions driving predictions.

Conclusions:

• CrossAttOmics offers a powerful framework for multiomics integration and cancer type prediction.
• The model's ability to leverage regulatory links enhances its predictive accuracy and interpretability.
• This approach holds promise for improving diagnostic capabilities in precision medicine.