Deep learning decodes species-specific codon usage signatures in Brassica from coding sequences.

Plant species discrimination remains a significant challenge in modern genomics, particularly for closely related species with substantial agricultural importance. Current morphological and molecular approaches often lack the resolution needed for reliable differentiation, creating a pressing need for more sophisticated analytical methods. This study demonstrates how deep learning can address this gap by providing high-accuracy classification of four key Brassica species (B. juncea, B. napus, B. oleracea, and B. rapa) using genomic sequence data. We conducted a systematic comparison of seven neural network architectures, focusing on their ability to discriminate between these closely related species. Based on test data, the Multilayer Perceptron achieved 100% classification accuracy with equally high performance across all evaluation metrics (accuracy, precision, recall, F1-score, and MCC). Other architectures, including Leaky ReLU and Dropout Neural Networks, showed near-perfect performance (99.9% accuracy), while the Radial Basis Function Neural Network demonstrated more modest results (74.6% accuracy). These findings reveal important architectural considerations for genomic classification tasks. This work makes three key contributions to the field: (1) it establishes deep learning as a powerful approach for plant species classification, (2) provides comparative performance metrics across multiple network architectures, and (3) demonstrates that whole-genome sequence data can enable highly accurate discrimination without manual feature selection. Our results have immediate applications in crop improvement, biodiversity conservation, and agricultural biotechnology, while the methodology offers a template for similar classification challenges in other taxonomic groups.