Tightly integrated multiomics-based deep tensor survival model for time-to-event prediction

Area of Science:

• Computational biology
• Bioinformatics
• Cancer genomics

Background:

• Multi-omics data offers valuable insights for predicting cancer patient survival.
• Integrating diverse omics data types presents significant analytical challenges.
• Existing methods struggle to fully capture complex patterns for accurate survival prediction.

Purpose of the Study:

• To develop a novel deep learning algorithm for integrating multi-omics cancer data.
• To improve the accuracy of cancer survival outcome prediction by leveraging high-dimensional omics information.
• To address the challenge of linking complex multi-omics patterns to survival.

Main Methods:

• A three-dimensional tensor was constructed to integrate multi-omics data.
• Tensor factorization was employed to derive latent factors.
• These latent factors were utilized in neural network-based survival models.
• The proposed algorithm was benchmarked against other multi-omics and individual omics approaches using TCGA data.

Main Results:

• The proposed tight integration framework demonstrated superior survival prediction performance.
• The algorithm outperformed models relying on individual genomic data.
• Comparative analysis utilized concordance index (C-index) and Integrated Brier Score (IBS) for evaluation.
• The method was validated on breast, colon, and rectal cancer datasets.

Conclusions:

• Deep learning-based integration of multi-omics data significantly enhances cancer survival prediction.
• The developed tensor factorization approach provides a powerful framework for multi-omics data integration.
• This method offers a promising advancement for personalized cancer treatment strategies and prognostic assessments.