Awesome

DNABERT_S: Learning Species-Aware DNA Embedding with Genome Foundation Models

This Repo is the official implementatation of DNABERT_S: Learning Species-Aware DNA Embedding with Genome Foundation Models.

Contents

• 1. Introduction
• 2. Model and Data
• 3. Setup Environment
• 4. Quick Start
• 5. Training
• 6. Evaluation
• 7. Citation

1. Introduction

DNABERT-S is a foundation model based on DNABERT-2 specifically designed for generating DNA embedding that naturally clusters and segregates genome of different species in the embedding space, which can greatly benefit a wide range of genome applications, including species classification/identification, metagenomics binning, and understanding evolutionary relationships.

Results on species clustering.

2. Model and Data

2.1 Model

The pre-trained models is available at Huggingface as zhihan1996/DNABERT-S.

To download the model from command line:

# command line
gdown 1ejNOMXdycorDzphLT6jnfGIPUxi6fO0g # pip install gdown
unzip dnabert-s_train.zip  # unzip the data 

2.2 Data

The training data of DNABERT-S is available at

gdown 1p59ch_MO-9DXh3LUIvorllPJGLEAwsUp # pip install gdown
unzip dnabert-s_train.zip  # unzip the data 

The evaluation data is available at

gdown 1I44T2alXrtXPZrhkuca6QP3tFHxDW98c # pip install gdown
unzip dnabert-s_eval.zip  # unzip the data 

3. Setup environment

conda create -n DNABERT_S python=3.9
conda activate DNABERT_S

pip install -r requirements.txt
pip uninstall triton # this can lead to errors in GPUs other than A100

4. Quick Start

Our model is easy to use with the transformers package.

To load the model from huggingface:

import torch
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("zhihan1996/DNABERT-S", trust_remote_code=True)
model = AutoModel.from_pretrained("zhihan1996/DNABERT-S", trust_remote_code=True)

To calculate the embedding of a dna sequence

dna = "ACGTAGCATCGGATCTATCTATCGACACTTGGTTATCGATCTACGAGCATCTCGTTAGC"
inputs = tokenizer(dna, return_tensors = 'pt')["input_ids"]
hidden_states = model(inputs)[0] # [1, sequence_length, 768]

# embedding with mean pooling
embedding_mean = torch.mean(hidden_states[0], dim=0)
print(embedding_mean.shape) # expect to be 768

5. Training

Our code base expects pairs of DNA sequencesfor pre-training. We expect the training data to be a csv file with two columns and no header. Each row contains one pair of DNA sequences that you want to model to generate similar embedding for them. See data/debug_train.csv for an example.

Important arguments:

• resdir: dictionary to save model parameters
• datapath: dictionary of data
• train_dataname: the name of the training data file (e.g., "a.csv")
• val_dataname: the name of the validating data file (e.g., "a.csv")
• max_length: set it as 0.2 * DNA_length (e.g., 200 for 1000-bp DNA)
• train_batch_size: batch size for training data, change it to fit your GPU RAM
• con_method: contrastive learning method, including "same species", "dropout", "double_strand", "mutate"
• mix: whether use i-Mix method
• mix_layer_num: which layer to perform i-Mix, if the value is -1, it means manifold i-Mix
• curriculum: whether use curriculum learning
• Other arguments can also be adjusted.

For our curriculum contrastive learning method, you can use:

cd pretrain
export PATH_TO_DATA_DICT=/path/to/data
export TRAIN_FILE=debug_train.csv # use this for debug, for real training, please use train_2m.csv

python main.py \
    --resdir ./results/ \
    --datapath ${PATH_TO_DATA_DICT} \
    --train_dataname ${TRAIN_FILE} \
    --val_dataname val_48k.csv \
    --seed 1 \
    --logging_step 10000 \
    --logging_num 12 \
    --max_length 2000 \
    --train_batch_size 48 \
    --val_batch_size 360 \
    --lr 3e-06 \
    --lr_scale 100 \
    --epochs 3 \
    --feat_dim 128 \
    --temperature 0.05 \
    --con_method same_species \
    --mix \
    --mix_alpha 1.0 \
    --mix_layer_num -1 \
    --curriculum 

This training scripts expect 8 A100 80GB GPUs. If you are using other types of devices, please change the train_batch_size and max_length accordingly.

After model training, you will find the trained model at ./pretrain/results/$file_name

The file_name is automatically set based on the hyperparameters, and the code regularly save checkpoint.

It should be something like ./results/contrastive.HardNeg.epoch2.debug_train.csv.lr3e-06.lrscale100.bs48.maxlength200.tmp0.05.decay1.seed1.turn1/100

The best model after validating is saved in ./pretrain/results/$file_name/best/

Scripts for other experiments are all in ./pretrain/results

6. Evaluation

6.1 Prepare model

cd evaluate

6.1.1 Test pre-trained DNABERT-S

gdown 1ejNOMXdycorDzphLT6jnfGIPUxi6fO0g
unzip DNABERT-S.zip
export MODEL_DIR=/path/to/DNABERT-S (e.g., /root/Downloads/DNABERT-S)

6.1.2 Test you own model train with our code base

Copy the necessary files to the folder where the model is saved. This is a bug in Huggingface Transformers package. Sometimes the model file such as bert_layer.py are not automatically saved to the model directory together with the model weights. So we manually do it.

export MODEL_DIR=/path/to/the/trained/model # (e.g., /root/ICML2024/train/pretrain/results/epoch3.debug_train.csv.lr3e-06.lrscale100.bs24.maxlength2000.tmp0.05.seed1.con_methodsame_species.mixTrue.mix_layer_num-1.curriculumTrue/0)

cp model_codes/* ${MODEL_DIR}

6.2 Clustering and Classification

export DATA_DIR=/path/to/the/unziped/folders

# evaluate the trained model
python eval_clustering_classification.py --test_model_dir ${MODEL_DIR} --data_dir ${DATA_DIR} --model_list "test"

# evaluate baselines (e.g., TNF and DNABERT-2)
python eval_clustering_classification.py --data_dir ${DATA_DIR} --model_list "tnf, dnabert2"

6.3 Metagenomics Binning

export DATA_DIR=/path/to/the/unziped/folders
export MODEL_DIR=/path/to/the/trained/model

# evaluate the trained model
python eval_binning.py --test_model_dir ${MODEL_DIR} --data_dir ${DATA_DIR} --model_list "test"

# evaluate baselines (e.g., TNF and DNABERT-2)
python eval_binning.py --data_dir ${DATA_DIR} --model_list "tnf, dnabert2"

7. Citation

If you have any question regarding our paper or codes, please feel free to start an issue or email Zhihan Zhou (zhihanzhou2020@u.northwestern.edu).

If you use DNABERT-S in your work, please consider cite our papers:

DNABERT-S

@misc{zhou2024dnaberts,
      title={DNABERT-S: Learning Species-Aware DNA Embedding with Genome Foundation Models}, 
      author={Zhihan Zhou and Winmin Wu and Harrison Ho and Jiayi Wang and Lizhen Shi and Ramana V Davuluri and Zhong Wang and Han Liu},
      year={2024},
      eprint={2402.08777},
      archivePrefix={arXiv},
      primaryClass={q-bio.GN}
}

DNABERT-2

@misc{zhou2023dnabert2,
      title={DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genome}, 
      author={Zhihan Zhou and Yanrong Ji and Weijian Li and Pratik Dutta and Ramana Davuluri and Han Liu},
      year={2023},
      eprint={2306.15006},
      archivePrefix={arXiv},
      primaryClass={q-bio.GN}
}

DNABERT

@article{ji2021dnabert,
    author = {Ji, Yanrong and Zhou, Zhihan and Liu, Han and Davuluri, Ramana V},
    title = "{DNABERT: pre-trained Bidirectional Encoder Representations from Transformers model for DNA-language in genome}",
    journal = {Bioinformatics},
    volume = {37},
    number = {15},
    pages = {2112-2120},
    year = {2021},
    month = {02},
    issn = {1367-4803},
    doi = {10.1093/bioinformatics/btab083},
    url = {https://doi.org/10.1093/bioinformatics/btab083},
    eprint = {https://academic.oup.com/bioinformatics/article-pdf/37/15/2112/50578892/btab083.pdf},
}