DeepTaxa
A deep learning framework that classifies 16S rRNA gene sequences into all seven taxonomic ranks in a single pass, combining a convolutional network with a BERT-style transformer and reaching 92.95% species-level accuracy on a public benchmark.
DeepTaxa is a hierarchical taxonomic classifier for 16S rRNA gene sequences that predicts all seven ranks, from Domain to Species, in one forward pass. On 69,335 held-out sequences from the Greengenes2 2024.09 test set it reaches accuracies from 99.98% at the Domain rank down to 92.95% at the Species rank, averaged over five random seeds, with cross-seed standard deviation of at most 0.0004 F1 at every rank.
The default HybridCNNBERTClassifier fuses a multi-kernel convolutional network's local sequence-motif features with a small from-scratch BERT encoder's global context through learnable fusion weights and a residual connection. A CNN-only variant trains faster at slightly lower species accuracy, and a BERT-only variant is provided mainly for ablation, since it underperforms substantially on its own at the species rank. Three model families, full-length, V3-V4 and V4 regions, are published on Hugging Face, each with five seeds for ensembling plus a default single checkpoint.
Built with Python and PyTorch, including a QIIME 2 plugin. MIT licensed.