Computing molecular signatures as optima of a bi-objective function: method and application to prediction in...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Gene selection methods for molecular signatures often lack explicit rationale for ranking genes.
• Existing approaches may not optimally balance signature size with predictive power.

Purpose of the Study:

• To develop a robust method for computing molecular signatures by optimizing a bi-objective function.
• To enhance prediction accuracy in oncogenomics while minimizing signature size.

Main Methods:

• Formulated molecular signature computation as a bi-objective optimization problem, minimizing signature size and maximizing interclass distance.
• Explored the solution space of all possible gene subsets to find optimal signatures.
• Applied the method to public oncology datasets using Diagonal Linear Discriminant Analysis (DLDA) classifier.

Main Results:

• Identified 'n' optimal non-empty signatures for 'n' genes, demonstrating a nested structure.
• Optimal signatures of size 'k' consisted of the top 'k' ranked genes contributing most to interclass distance.
• Achieved prediction performances comparable to or better than existing methods, with significantly smaller and robust signatures across five oncology datasets.
• Demonstrated superior performance in predicting chemotherapy response in breast cancer patients, yielding smaller gene signatures with known relevance.

Conclusions:

• Defining molecular signatures via bi-objective optimization (size and interclass distance) is a well-founded and efficient strategy for oncogenomic prediction.
• The method offers low computational complexity, yielding optimal signatures directly from gene rankings.
• Freely available software facilitates the application of this approach in research.