AttOmics: attention-based architecture for diagnosis and prognosis from omics data

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• High-throughput omics data is increasingly available for personalized medicine.
• Deep learning models are crucial for analyzing omics data but struggle with high dimensionality and limited training sets.
• Current models fail to capture patient-specific molecular interactions within omics profiles.

Purpose of the Study:

• To introduce AttOmics, a novel deep-learning architecture designed for precision medicine.
• To address the limitations of existing deep learning models in handling omics data.
• To develop a model that can identify patient-specific molecular interactions.

Main Methods:

• Decomposition of omics profiles into feature groups.
• Application of the self-attention mechanism to these groups.
• Development of the AttOmics deep-learning architecture.

Main Results:

• AttOmics accurately predicts patient phenotypes.
• The model requires fewer parameters compared to traditional deep neural networks.
• Attention maps provide insights into phenotype-determining molecular groups.

Conclusions:

• AttOmics offers an effective approach for phenotype prediction using omics data.
• The self-attention mechanism enables the capture of patient-specific interactions.
• The model advances the application of deep learning in precision medicine.