multimolecule / dnaberts

DNABERT-S

Pre-trained model on multi-species genome using a contrastive learning objective for species-aware DNA embeddings.

Disclaimer

This is an UNOFFICIAL implementation of the DNABERT-S: pioneering species differentiation with species-aware DNA embeddings by Zhihan Zhou, et al.

The OFFICIAL repository of DNABERT-S is at MAGICS-LAB/DNABERT_S .

The MultiMolecule team has confirmed that the provided model and checkpoints are producing the same intermediate representations as the original implementation.

The team releasing DNABERT-S did not write this model card for this model so this model card has been written by the MultiMolecule team.

Model Details

DNABERT-S is a bert -style model built upon DNABERT-2 and fine-tuned with contrastive learning for species-aware DNA embeddings. The model was trained using the proposed Curriculum Contrastive Learning (C²LR) strategy with the Manifold Instance Mixup (MI-Mix) training objective.

DNABERT-S shares the same architecture as DNABERT-2: it uses Byte Pair Encoding (BPE) tokenization, Attention with Linear Biases (ALiBi) instead of learned position embeddings, and incorporates a Gated Linear Unit (GeGLU) MLP and FlashAttention for improved efficiency.

Model Specification

Num Layers	Hidden Size	Num Heads	Intermediate Size	Num Parameters (M)	FLOPs (G)	MACs (G)	Max Num Tokens
12	768	12	3072	117.07	125.83	62.92	512

Links

• Code : multimolecule.dnaberts
• Data : GenBank
• Paper : DNABERT-S: pioneering species differentiation with species-aware DNA embeddings
• Developed by : Zhihan Zhou, Weimin Wu, Harrison Ho, Jiayi Wang, Lizhen Shi, Ramana V Davuluri, Zhong Wang, Han Liu
• Model type : BERT - MosaicBERT
• Original Repository : MAGICS-LAB/DNABERT_S

Usage

The model file depends on the multimolecule library. You can install it using pip:

pip install multimolecule

Direct Use

Feature Extraction

You can use this model directly with a pipeline for feature extraction:

import multimolecule  # you must import multimolecule to register models
from transformers import pipeline

predictor = pipeline("feature-extraction", model="multimolecule/dnaberts")
output = predictor("ATCGATCGATCG")

Downstream Use

Extract Features

Here is how to use this model to get the features of a given sequence in PyTorch:

from multimolecule import DnaBertSModel
from transformers import AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained("multimolecule/dnaberts")
model = DnaBertSModel.from_pretrained("multimolecule/dnaberts")

text = "ATCGATCGATCGATCG"
input = tokenizer(text, return_tensors="pt")

output = model(**input)

Sequence Classification / Regression

This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for sequence classification or regression.

Here is how to use this model as backbone to fine-tune for a sequence-level task in PyTorch:

import torch
from multimolecule import DnaBertSForSequencePrediction
from transformers import AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained("multimolecule/dnaberts")
model = DnaBertSForSequencePrediction.from_pretrained("multimolecule/dnaberts")

text = "ATCGATCGATCGATCG"
input = tokenizer(text, return_tensors="pt")
label = torch.tensor([1])

output = model(**input, labels=label)

Token Classification / Regression

This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for token classification or regression.

Here is how to use this model as backbone to fine-tune for a nucleotide-level task in PyTorch:

import torch
from multimolecule import DnaBertSForTokenPrediction
from transformers import AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained("multimolecule/dnaberts")
model = DnaBertSForTokenPrediction.from_pretrained("multimolecule/dnaberts")

text = "ATCGATCGATCGATCG"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), ))

output = model(**input, labels=label)

Contact Classification / Regression

This model is not fine-tuned for any specific task. You will need to fine-tune the model on a downstream task to use it for contact classification or regression.

Here is how to use this model as backbone to fine-tune for a contact-level task in PyTorch:

import torch
from multimolecule import DnaBertSForContactPrediction
from transformers import AutoTokenizer


tokenizer = AutoTokenizer.from_pretrained("multimolecule/dnaberts")
model = DnaBertSForContactPrediction.from_pretrained("multimolecule/dnaberts")

text = "ATCGATCGATCGATCG"
input = tokenizer(text, return_tensors="pt")
label = torch.randint(2, (len(text), len(text)))

output = model(**input, labels=label)

Training Details

DNABERT-S uses a two-phase Curriculum Contrastive Learning (C²LR) strategy. In phase I, the model is trained with Weighted SimCLR for one epoch. In phase II, the model is further trained with Manifold Instance Mixup (MI-Mix) for two epochs. The training starts from the pre-trained DNABERT-2 checkpoint.

Training Data

The DNABERT-S model was trained on pairs of non-overlapping DNA sequences from the same species, sourced from GenBank . The dataset consists of 47,923 pairs from 17,636 viral genomes, 1 million pairs from 5,011 fungi genomes, and 1 million pairs from 6,402 bacteria genomes. From the total of 2,047,923 pairs, 2 million were randomly selected for training and the rest were used as validation data. All DNA sequences are 10,000 bp in length.

Training Procedure

Pre-training

The model was trained on 8 NVIDIA A100 80GB GPUs.

• Temperature (τ): 0.05
• Hyperparameter (α): 1.0
• Epochs: 1 (phase I, Weighted SimCLR) + 2 (phase II, MI-Mix)
• Optimizer: Adam
• Learning rate: 3e-6
• Batch size: 48
• Checkpointing: Every 10,000 steps, best selected on validation loss
• Training time: ~48 hours

Citation

@article{zhou2025dnaberts,
  title={{DNABERT-S}: pioneering species differentiation with species-aware {DNA} embeddings},
  author={Zhou, Zhihan and Wu, Weimin and Ho, Harrison and Wang, Jiayi and Shi, Lizhen and Davuluri, Ramana V and Wang, Zhong and Liu, Han},
  journal={Bioinformatics},
  volume={41},
  pages={i255--i264},
  year={2025},
  doi={10.1093/bioinformatics/btaf188}
}

The artifacts distributed in this repository are part of the MultiMolecule project. If you use MultiMolecule in your research, you must cite the MultiMolecule project as follows:

@software{chen_2024_12638419,
  author    = {Chen, Zhiyuan and Zhu, Sophia Y.},
  title     = {MultiMolecule},
  doi       = {10.5281/zenodo.12638419},
  publisher = {Zenodo},
  url       = {https://doi.org/10.5281/zenodo.12638419},
  year      = 2024,
  month     = may,
  day       = 4
}

Contact

Please use GitHub issues of MultiMolecule for any questions or comments on the model card.

Please contact the authors of the DNABERT-S paper for questions or comments on the paper/model.

License

This model is licensed under the GNU Affero General Public License .

For additional terms and clarifications, please refer to our License FAQ .

SPDX-License-Identifier: AGPL-3.0-or-later