Integrative Analysis of Cancer Gene Expression Using Bio-Inspired Algorithms and Machine Learning: Identification of Key Genes

Background: understanding the genetic basis of cancer is essential for effective diagnosis, prognosis, and treatment. With an extensive array of gene expression data, pinpointing relevant genes is a significant challenge. This study employs a combination of Machine Learning (ML) and bio-inspired algorithms to isolate important genes within a Breast Cancer gene expression dataset. Methods: The research utilized a Breast Cancer dataset encompassing 24,481 Affymetrix probe IDs from 97 samples. The initial step involved using the Fisher scores method to select the top 1000 probe IDs. This list was further refined to the top 100 probe IDs using five bio-inspired algorithms: Salp Swarm Algorithm (SSA), Grey Wolf Optimizer (GWO), Artificial Bee Colony (ABC), Ant Colony Optimization (ACO), and Multi Objective Spotted Hyena Optimizer (MOSHO). These probe IDs were then evaluated using five classifiers: k-Nearest Neighbors (kNN), Support Vector Machine (SVM), Random Forest (RF), Naive Bayes (NB), and Nearest Centroid (NC). The final stage involved using BLAST for aligning the probe IDs and determining their corresponding gene names. Results: The study integrated results from all bio-inspired techniques and classifiers to identify frequently recurring genes. A new Breast Cancer dataset was created by merging gene selections from each algorithm. The genes were then analyzed based on their frequency of appearance across the optimizers, categorized into groups of 4, 3, and 2 occurrences. For genes consistently appearing across all algorithms and classifiers (with 2 occurrences), Recursive Feature Elimination (RFE) was applied. This process led to the successful pairing of 107 high-quality probe IDs using BLAST, with 62 of these IDs aligned to specific gene names. Conclusion: This in-depth and multi-layered approach identified 107 genes as potential key contributors to cancer progression. Notably, 17 of these genes have a direct link to cancer, while two may have potential associations. The study underscores the efficacy of combining bio-inspired algorithms with ML for genomic data analysis and suggests the need for further biological validation to establish the clinical significance of these genes in cancer.