COmic: convolutional kernel networks for interpretable end-to-end learning on (multi-)omics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Artificial Intelligence in Healthcare

Background:

• Omics datasets are rapidly growing, presenting opportunities for improved healthcare predictions.
• Current models often function as "black boxes," lacking transparency in high-stakes medical applications.
• Interpreting molecular factors influencing predictions is crucial for clinical trust and safety.

Purpose of the Study:

• To introduce Convolutional Omics Kernel Networks (COmic), an interpretable deep learning model for omics data analysis.
• To enable robust and transparent end-to-end learning on omics datasets of varying sizes.
• To facilitate the integration and analysis of multi-omics data.

Main Methods:

• Developed COmic, a novel artificial neural network combining convolutional kernel networks with pathway-induced kernels.
• Utilized pathway-induced Laplacian kernels to enhance model interpretability.
• Adapted COmic for both single-omics and multi-omics data integration.

Main Results:

• COmic demonstrated competitive or superior performance compared to existing models across six breast cancer cohorts.
• Models trained on multi-omics data using the METABRIC cohort showed robust results.
• Pathway-induced kernels successfully opened the "black-box" nature of neural networks, yielding intrinsically interpretable models.

Conclusions:

• COmic offers a powerful and interpretable solution for analyzing large-scale omics data in healthcare.
• The model's interpretability addresses safety and security concerns associated with black-box AI in clinical settings.
• COmic's adaptability to multi-omics data enhances its utility for comprehensive biological insights.