Pathway-Capsule Neural Networks

The progress made in Next-Generation Sequencing technology has resulted in many genetic databases, which offer the potential for more effective diagnostic tools. Recent literature highlights Artificial Intelligence (AI) as one of the most potent methods. Nonetheless, when it comes to analyzing genetic data, Machine Learning (ML) and Deep Learning (DL) face three significant challenges due to: (1) multi-modality analysis, (2) model explainability, and (3) extension to patient-specific biomarker identification. In this paper, we introduce Pathway-Capsule Neural Networks (PCNN) and its Bayesian version (BayesPCNN) to address the three limitations. We demonstrate the model using Big Data of 7, 372 samples with 57 protein expressions from the TCGA dataset. In the best setting, the proposed PCNN outperforms tree-based ML models and KNN with a significant gap of 10% − 30% in accuracy for Pan-Cancer detection; while only outperformed by kernel methods with under 2%. Another contribution of BayesPCNN is highlighted as Explainable AI (XAI), which reveals tumors with potential metastasis and patient-specific biomarkers.