PLUS: Predicting cancer metastasis potential based on positive and unlabeled learning

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Metastatic cancer causes over 90% of cancer deaths, making metastasis potential evaluation critical for patient outcomes.
• Accurately assessing metastasis potential computationally is challenging due to under-diagnosed metastasis events and biased classification labels.
• High dimensionality of molecular data, unknown metastasis prevalence, and limited confirmed cases further complicate predictive modeling.

Purpose of the Study:

• To develop a novel computational algorithm for accurate metastasis potential assessment.
• To address the challenges of under-detected metastasis events and unbalanced datasets in cancer research.
• To improve clinical decision-making and reduce metastasis-associated mortality.

Main Methods:

• Introduction of Positive and unlabeled Learning from Unbalanced cases and Sparse structures (PLUS), a novel algorithm.
• Utilizing a positive and unlabeled learning framework to handle under-detected metastasis events.
• Applying PLUS to The Cancer Genome Atlas (TCGA) Pan-Cancer gene expression data.

Main Results:

• PLUS demonstrates superior performance on synthetic datasets compared to existing state-of-the-art methods.
• Metastasis potential predictions generated by PLUS align well with clinical follow-up data.
• Identified predictive genes from PLUS analysis were validated using independent single-cell RNA-sequencing datasets.

Conclusions:

• PLUS is the first algorithm to effectively use a positive and unlabeled learning framework for metastasis prediction, accounting for under-detection.
• The algorithm successfully addresses unbalanced instance allocation and unknown metastasis prevalence, outperforming other methods.
• PLUS provides a robust computational tool for predicting cancer metastasis potential, with validated predictive genes offering new insights.