Patient-level proteomic network prediction by explainable artificial intelligence

Area of Science:

• Computational biology
• Bioinformatics
• Artificial intelligence in oncology

Background:

• Precision cancer therapy relies on understanding individual tumor protein networks.
• Current methods for network reconstruction from omics data often predict only average network features.
• Functional experiments are limited in scope for comprehensive network analysis.

Purpose of the Study:

• To apply explainable artificial intelligence (AI), specifically layer-wise relevance propagation (LRP), for inferring protein interaction networks at the individual patient level.
• To evaluate the performance of LRP in reconstructing both average and patient-specific networks from proteomic data.
• To identify oncogenic network features relevant for precision oncology.

Main Methods:

• Utilized Layer-Wise Relevance Propagation (LRP), an explainable AI technique.
• Applied LRP to proteomic profiling data from The Cancer Proteome Atlas.
• Compared LRP performance against state-of-the-art network prediction methods for individual tumors.

Main Results:

• LRP accurately reconstructed average (AUC 0.99) and individual (AUC 0.93) protein interaction networks.
• LRP outperformed existing methods for predicting networks in individual tumors.
• Identified known and novel oncogenic network features, including cancer-type specific and patient-specific variations.

Conclusions:

• Explainable AI (LRP) can effectively infer patient-level protein interaction networks from proteomic data.
• This approach enhances the potential for predictive diagnostics in precision oncology.
• The method facilitates the identification of "patient-level" oncogenic mechanisms for tailored therapies.