An Optimized Genetic Algorithm-based BP Neural Network Model for the Classification of Human Pre-miRNA Subclasses

Pre-miRNAs, as the precursor forms of miRNAs, possess significant research potential due to their intricate structural features. Accurate classification of pre-miRNAs is crucial for comprehending their roles in post-transcriptional gene regulation processes. This paper focuses on two distinct subclasses of pre-miRNAs: canonical miRNAs and mirtrons. Canonical miRNAs exhibit high conservation across species, whereas mirtrons display species-specific roles and expression patterns. To address the classification task, our study employed a Backward Propagation Neural Network (BPNN) model. While BPNN is renowned for its ability to learn and adapt from non-linear data, it often encounters challenges when dealing with high-dimensional spaces, leading to local optima. To further enhance the BPNN model, our study integrated BPNN with Genetic Algorithm (GA), a type of optimization algorithm inspired by the process of natural selection and genetics. GA is adept at exploring large and complex search spaces, which helps find alternative weight parameters for the BPNN model and thus enhances the model's convergence capability. The original dataset utilized in this study consists of human pre-miRNA sequences accompanied by accurately annotated class labels, sourced from miRBase (v21). A set of 25 features was derived from the original dataset, out of which 10 features were selected using SHapley Additive exPlanations (SHAP) technique as the independent variables. GA-BPNN's model performance was evaluated using independent testing datasets split from the pre-processed dataset and compared with other common Machine Learning (ML) techniques, demonstrating improved prediction performances for pre-miRNAs.