Self-Normalizing Multi-Omics Neural Network for Pan-Cancer Prognostication.

Area of Science:

• Computational biology and bioinformatics
• Cancer genomics and precision oncology
• Machine learning in healthcare

Background:

• Extracting prognostic (overall survival, tertiary lymphoid structure ratios) and diagnostic (cancer type) signatures from sparse, high-dimensional multi-omics data is challenging due to heterogeneity and missingness.
• Existing methods struggle to integrate diverse omics data (gene expression, methylation, miRNA, mutations, protein) and clinical variables effectively.
• There is a need for robust models that can handle missing data modalities for comprehensive cancer analysis.

Purpose of the Study:

• To develop and validate SeNMo, a self-normalizing deep neural network for unified representation learning from heterogeneous omics data.
• To demonstrate SeNMo's capability in predicting overall survival, classifying primary cancer types, and predicting tertiary lymphoid structure ratios.
• To establish a modality-robust baseline model for diverse downstream tasks in multi-omics oncology.

Main Methods:

• Trained a self-normalizing deep neural network (SeNMo) on five omics layers (gene expression, DNA methylation, miRNA, somatic mutations, protein expression) and clinical variables.
• Utilized over 10,000 patient profiles from The Cancer Genome Atlas (TCGA) for training and internal validation.
• Performed external validation on the CPTAC lung squamous cell carcinoma cohort and an independent Moffitt Cancer Center cohort.

Main Results:

• Achieved a concordance index of 0.758 for overall survival prediction on a held-out TCGA test set.
• External validation showed concordance indices of 0.73 (CPTAC) and 0.66 (Moffitt).
• Classified primary cancer types with 99.8% accuracy and predicted tertiary lymphoid structure ratios, aligning with expert annotations (p < 0.05) and showing distinct survival outcomes.

Conclusions:

• SeNMo provides a robust, modality-agnostic deep learning framework for integrating multi-omics data in oncology.
• The model demonstrates strong performance in key clinical tasks, including survival prediction, cancer classification, and tertiary lymphoid structure ratio prediction.
• SeNMo holds significant translational potential for advancing precision oncology through comprehensive multi-omics analysis.