Discriminating early- and late-stage cancers using multiple kernel learning on gene sets.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Distinguishing early- and late-stage cancers is crucial for developing effective cancer therapies.
• Standard machine learning methods struggle with extracting biological insights from complex, correlated genomic data.

Purpose of the Study:

• To develop a machine learning approach for classifying early- and late-stage cancers using gene expression profiles.
• To leverage prior biological knowledge, specifically pathways and gene sets, to enhance cancer stage prediction and identify progression mechanisms.

Main Methods:

• A multiple kernel learning (MKL) formulation integrating pathways/gene sets was proposed.
• The MKL approach was compared against Random Forests and Support Vector Machines using gene expression data from 20 Cancer Genome Atlas cohorts.
• Performance was evaluated based on predictive accuracy and the ability to extract biologically relevant information.

Main Results:

• The proposed MKL method achieved statistically significant or comparable predictive performance to standard algorithms across most datasets.
• The MKL approach utilized fewer gene expression features compared to other methods.
• The algorithm successfully extracted meaningful, disease-specific information related to cancer progression mechanisms.

Conclusions:

• Multiple kernel learning incorporating gene sets offers an effective strategy for cancer stage classification.
• This method enhances predictive performance while providing insights into the molecular mechanisms of cancer progression.
• The developed algorithms and experimental scripts are publicly available for further research.