Deep learning based attention enhanced phylogenetic radial basis function networks (AE-PRBFN) for genomic codon usage classification across species

Understanding codon usage patterns is essential for genomic classification and offers insights into the evolutionary and functional characteristics of species. This study focuses on classifying four agriculturally significant crop species Triticum aestivum (wheat), Oryza sativa (rice), Hordeum vulgare (barley), and Brachypodium distachyon using their codon usage biases. Traditional methods often struggle with high-dimensional genomic data, prompting the need for advanced deep learning techniques. Here, we introduce Attention Enhanced Phylogenetic Radial Basis Function Networks (AE-PRBFN), a novel neural architecture designed to improve classification accuracy and efficiency. We analyzed 253,076 high-quality coding sequences (131,656 wheat; 31,970 rice; 37,932 barley; 51,518 Brachypodium) using absolute codon frequencies (64 features) as input. AE-PRBFN achieved 100% accuracy, significantly outperforming standard RBFN, Multilayer Perceptron, Support Vector Machines, and Random Forest. Moreover, AE-PRBFN required only 2 training epochs (10-fold cross-validation) versus approximately 200 epochs for standard RBFN, demonstrating substantially faster convergence. AE-PRBFN captures subtle, species-specific codon usage signatures that distinguish these four Poaceae crops with perfect accuracy, establishing that absolute codon frequency encodes sufficient phylogenetic signal for taxonomic classification.