DeepTaxa

DeepTaxa predicts all seven Linnaean ranks from domain to species for prokaryotic 16S rRNA gene sequences in a single forward pass using a hybrid CNN-BERT deep learning framework.

Key Features:

• Hybrid architecture: A multiscale CNN paired with a transformer trained from scratch on the DNABERT-2 BPE vocabulary produces parallel rank-specific predictions.
• Multi-rank prediction: Simultaneous classification across all seven Linnaean ranks from domain to species in a single forward pass.
• Pre-trained checkpoints: Available for full-length 16S rRNA sequences and V3-V4 amplicons.
• Calibration: Species-level expected calibration error of 0.0242 on the Greengenes2 2024.09 test set.
• Robust performance: Cross-seed standard deviation less than or equal to 0.0008 F1 at every rank (3-seed mean).

Scientific Applications:

• 16S rRNA taxonomic classification: Species-level classification of prokaryotic 16S rRNA sequences without reliance on exact alignment, k-mer heuristics, or phylogenetic placement.
• Short amplicon analysis: Taxonomic assignment from approximately 420 bp V3-V4 amplicon windows with 87.55% species accuracy.
• Novel species assessment: Performance degrades smoothly with decreasing training-set similarity, enabling prediction confidence assessment for divergent sequences.

Methodology:

DeepTaxa pairs a multiscale CNN with a transformer trained from scratch on the DNABERT-2 BPE vocabulary to produce parallel rank-specific predictions. Trained and evaluated on the Greengenes2 2024.09 dataset.