Biomarker signature identification in "omics" data with multi-class outcome

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Biomarker signature identification from "omics" data is crucial for predicting outcomes.
• Selecting minimal yet predictive molecular feature sets is challenging, especially for multi-class outcomes (e.g., cancer types or stages).
• Constraint-based, local causal discovery algorithms are effective but require robust conditional independence tests.

Purpose of the Study:

• To extend constraint-based, local causal discovery algorithms for continuous predictors and multi-class outcomes.
• To develop and implement a novel conditional independence test suitable for nominal and ordinal multi-class targets.
• To evaluate the performance of the new method against existing approaches on real-world gene-expression data.

Main Methods:

• Developed a conditional independence test based on multinomial logistic regression.
• Utilized the log-likelihood ratio test for model selection within the new test.
• Applied constraint-based, local causal discovery algorithms with the enhanced test to gene-expression datasets.

Main Results:

• The novel conditional independence test effectively handles continuous predictors and multi-class outcomes.
• Experimental evaluations on seven high-dimensional gene-expression datasets demonstrated superior performance.
• The new test identified smaller and more predictive biomarker signatures compared to alternative methods.

Conclusions:

• The developed conditional independence test is a valuable advancement for biomarker discovery in multi-class "omics" data.
• This method enhances the capability of causal discovery algorithms for complex biological outcomes.
• The findings suggest improved accuracy and efficiency in identifying predictive gene expression signatures.