Adaptive BMRFO: Optimizing Descriptor Selection for Improved QSAR Biodegradation Classification

This study presents an enhanced Adaptive Binary Manta Ray Foraging Optimization (ABMRFO) algorithm for feature selection in QSAR biodegradation classification. The proposed method integrates an adaptive Sigmoid transfer function for binary conversion and a dynamic somersault factor for improved search efficiency. Nine ABMRFO variants were tested on the QSAR Biodegradation dataset using holdout validation over 150 iterations, evaluated based on classification accuracy, convergence speed, fitness value, and computational efficiency. Among the evaluated variants, ABMRFO3 emerged as the top performer, achieving the highest classification accuracy of 90.38% while selecting an average of only 9.2 features. It demonstrated strong optimization capabilities with the lowest mean fitness value (0.0975), best fitness (0.0548), and worst fitness (0.1193). Its fast convergence was evidenced by an average computational time of 27.92 seconds. The Friedman test ranked ABMRFO3 first with a sum of ranks of 19, confirming its superior performance. Additionally, the Wilcoxon signed-rank test indicated statistically significant improvements of ABMRFO3 over other algorithms, further validating its effectiveness. Its adaptive mechanisms ensure exceptional search accuracy, computational efficiency, and solution stability, making it a robust solution for complex feature selection tasks in QSAR modeling.