A Multi-Modal Graph-Based Semi-Supervised Pipeline for Predicting Cancer Survival.

Area of Science:

• Bioinformatics
• Computational Biology
• Cancer Research

Background:

• Cancer survival prediction is crucial for optimizing treatment and improving patient quality of life.
• Gene expression profiling, particularly RNA sequencing (RNA-seq), offers rich data for identifying predictive biomarkers.
• High dimensionality of gene expression data and limited sample sizes pose significant challenges for accurate survival prediction.

Purpose of the Study:

• To develop and evaluate a computational pipeline for predicting cancer patient survival using multi-modal RNA-seq data.
• To address the challenge of high-dimensional transcriptomic data with limited samples.
• To explore the utility of manifold learning and graph-based semi-supervised learning for cancer survival prediction.

Main Methods:

• Utilized multiple data modalities from RNA-seq.
• Applied manifold learning to exploit input data structure.
• Employed Laplacian support vector machines, a graph-based semi-supervised learning (GSSL) paradigm, to leverage unlabeled samples.
• Implemented a stacked generalization strategy to fuse predictions from different models.

Main Results:

• The proposed pipeline demonstrated promising results in predicting cancer patient survival on two independent datasets.
• Fusion of multiple models through stacked generalization synergistically boosted predictive accuracy compared to single-modality approaches.
• The approach effectively handles the challenge of high-dimensional, low-sample-size transcriptomic data.

Conclusions:

• The developed pipeline offers a robust method for cancer survival prediction using RNA-seq data.
• Combining manifold learning and GSSL, along with model fusion, is an effective strategy for improving predictive performance.
• This approach is potentially applicable to other predictive tasks where labeled data is scarce or expensive to obtain.