Knowledge-slanted random forest method for high-dimensional data and small sample size with a feature selection...

Area of Science:

• Computational Biology
• Machine Learning
• Genomics

Background:

• High-dimensional genetic and genomics data present challenges like the curse of dimensionality.
• Conventional random forest (RF) models can exhibit poor accuracy in high-dimensional settings, particularly with limited sample sizes.
• Integrating prior biological knowledge is a promising strategy to enhance machine learning model performance.

Purpose of the Study:

• To propose a novel knowledge-slanted random forest (RF) model.
• To improve the performance and explainability of gene selection in high-dimensional genomics data.
• To address limitations of conventional RF in scenarios with small sample sizes.

Main Methods:

• The knowledge-slanted RF integrates biological networks (e.g., protein-protein interaction networks) as prior knowledge.
• A random walk with restart algorithm determines gene relevance based on network topology.
• Gene relevance scores modify feature selection probabilities in the RF algorithm, enhanced by a modified Boruta algorithm.

Main Results:

• Knowledge-slanted RF demonstrated improved precision in outcome prediction compared to conventional RF and logistic lasso regression on simulated datasets.
• The method effectively identified more biologically relevant genes.
• Enhanced explainability was observed compared to standard RF approaches.

Conclusions:

• Knowledge-slanted RF offers a robust approach to handle high-dimensional genomics data, overcoming the curse of dimensionality.
• The integration of prior biological network knowledge significantly boosts model performance and interpretability.
• This method shows promise for identifying relevant genes in complex diseases, as validated in a case study of calcific aortic valve stenosis.