Releases · bioscan-ml/BarcodeBERT

This repository contains all the code accompanying the paper BarcodeBERT: Transformers for Biodiversity Analysis (Millan Arias et al., 2025)

BarcodeBERT is a BERT-style transformer model trained exclusively on a dataset of DNA barcode sequences extracted from a reference library of Canadian invertebrates. In addition to the full pretraining pipeline, you’ll find scripts and notebooks for evaluating BarcodeBERT (and several off-the-shelf DNA foundation models) in various downstream tasks:

Fine-tuning for supervised species-level classification.
Similarity retrieval for labelling rare or unseen species via nearest neighbour search in the embedding space.
BIN reconstruction, where BarcodeBERT embeddings are used to group sequences into putative Barcode Index Numbers.

What's Changed

BZSL implementation for BarcodeBERT: Transformers for Biodiversity Analysis by @atwang16 in #1
DOC: Fix paths to scripts shown in README by @scottclowe in #3
RF: Standardize requirements.txt into a single file by @scottclowe in #5
Bump transformers from 4.29.2 to 4.36.0 by @dependabot in #7
Bump black from 23.11.0 to 24.3.0 by @dependabot in #8
Bump transformers from 4.36.0 to 4.38.0 by @dependabot in #9
Bump scikit-learn from 1.3.0 to 1.5.0 by @dependabot in #10
MNT: Fix issues flagged by pre-commit by @scottclowe in #11
DOC: Update citation by @scottclowe in #12
[pre-commit.ci] pre-commit autoupdate by @pre-commit-ci in #13
[pre-commit.ci] pre-commit autoupdate by @pre-commit-ci in #15
Added conditional check for use_cuda before using torch.cuda by @NotMyLyfe in #14
Camera Ready Version for Bioinformatics by @millanp95 in #19