DeepTaxa: a hybrid CNN-BERT framework for 16S rRNA taxonomic classification.

Area of Science:

• Microbial genomics
• Bioinformatics
• Machine learning

Background:

• Accurate species-level classification of prokaryotic 16S rRNA sequences is challenging due to limitations in existing tools and incomplete databases.
• Deep learning methods have primarily focused on whole-genome metagenomics, leaving 16S rRNA taxonomy under-supported.

Purpose of the Study:

• To develop a novel deep learning framework for accurate species-level classification of prokaryotic 16S rRNA sequences.
• To improve upon existing taxonomic classification tools by leveraging a hybrid CNN-BERT architecture.

Main Methods:

• Developed DeepTaxa, a hybrid convolutional neural network (CNN) and BERT framework.
• Trained the model on the DNABERT-2 BPE vocabulary for parallel, rank-specific predictions across seven Linnean ranks.
• Utilized the Greengenes2 2024.09 test set for evaluation.

Main Results:

• DeepTaxa achieved 92.96% species-level accuracy and a 0.9212 F1 score on the Greengenes2 test set.
• The model demonstrated high F1 scores (>0.99) from domain through class and a low species-level expected calibration error (0.0242).
• DeepTaxa outperformed DADA2, QIIME 2, SINTAX, and Kraken 2 at the species rank, with a dedicated V3-V4 amplicon checkpoint achieving 87.55% species accuracy.

Conclusions:

• DeepTaxa offers a significant advancement in prokaryotic 16S rRNA sequence classification, particularly at the species level.
• The hybrid CNN-BERT framework provides robust and accurate taxonomic assignments across multiple ranks.
• Publicly available code, checkpoints, and datasets facilitate reproducible research and broader adoption in microbial genomics.