Nature-Inspired Multiobjective Cancer Subtype Diagnosis

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Oncology

Background:

• Cancer gene expression data is crucial for diagnosis and drug discovery.
• Existing computational methods often lack interpretability, are affected by noise, and yield suboptimal diagnostic quality.
• Addressing these limitations is essential for advancing cancer research and treatment.

Purpose of the Study:

• To introduce a novel multiobjective ensemble cuckoo search based on decomposition (MOECSA) algorithm.
• To simultaneously optimize feature selection, classification accuracy, and entropy-based relevance and redundancy measures.
• To enhance the diagnostic performance for cancer gene expression data.

Main Methods:

• Developed MOECSA, a multiobjective optimization algorithm.
• Implemented a novel binary encoding for gene subset selection and objective function calculation.
• Integrated an ensemble mechanism within the cuckoo search framework to balance convergence and diversity.
• Evaluated performance on 35 cancer gene expression datasets, 4 disease datasets, and 1 sequencing-based dataset.

Main Results:

• MOECSA demonstrated superior diagnostic performance across multiple levels compared to state-of-the-art algorithms.
• The algorithm effectively optimized four objectives: feature count, accuracy, relevance, and redundancy.
• Experimental results validated the algorithm's effectiveness and efficiency on diverse datasets.

Conclusions:

• MOECSA offers a significant advancement in classifying cancer gene expression data.
• The proposed method provides improved interpretability and diagnostic quality.
• This approach holds promise for enhanced cancer subtype diagnosis and drug discovery.