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

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Accurate plant species discrimination is crucial for agriculture and conservation.
• Existing methods often lack the resolution for closely related species.
• Genomic data offers potential for improved species identification.

Purpose of the Study:

• To evaluate deep learning for classifying four key Brassica species using genomic data.
• To compare the performance of seven different neural network architectures for this task.
• To establish deep learning as a viable method for plant species classification.

Main Methods:

• Utilized whole-genome sequence data from four Brassica species.
• Systematically compared seven neural network architectures (Multilayer Perceptron, Leaky ReLU, Dropout, Radial Basis Function, etc.).
• Evaluated classification performance using accuracy, precision, recall, F1-score, and MCC.

Main Results:

• The Multilayer Perceptron achieved 100% classification accuracy.
• Other architectures like Leaky ReLU and Dropout Networks showed near-perfect performance (99.9%).
• Whole-genome data enabled high accuracy without manual feature selection.

Conclusions:

• Deep learning is a powerful tool for plant species classification using genomic data.
• Specific neural network architectures significantly impact classification performance.
• This methodology can be applied to other taxonomic groups and has implications for agriculture and conservation.