DNABERT-2: Efficient Foundation Model and Benchmark for Multi-Species Genome

The repo contains:

1. The official implementation of DNABERT-2: Efficient Foundation Model and Benchmark for Multi-Species Genome
2. Genome Understanding Evaluation (GUE): a comprehensize benchmark containing 28 datasets for multi-species genome understanding benchmark.

Contents

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

Update (2024/02/14)

We publish DNABERT-S, 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. Please check it out here if you are interested.

1. Introduction

DNABERT-2 is a foundation model trained on large-scale multi-species genome that achieves the state-of-the-art performanan on $28$ tasks of the GUE benchmark. It replaces k-mer tokenization with BPE, positional embedding with Attention with Linear Bias (ALiBi), and incorporate other techniques to improve the efficiency and effectiveness of DNABERT.

2. Model and Data

The pre-trained models is available at Huggingface as zhihan1996/DNABERT-2-117M. Link to HuggingFace ModelHub. Link For Direct Downloads.

2.1 GUE: Genome Understanding Evaluation

GUE is a comprehensive benchmark for genome understanding consising of $28$ distinct datasets across $7$ tasks and $4$ species. GUE can be download here. Statistics and model performances on GUE is shown as follows:

3. Setup environment

# create and activate virtual python environment
conda create -n dna python=3.8
conda activate dna

# (optional if you would like to use flash attention)
# install triton from source
git clone https://github.com/openai/triton.git;
cd triton/python;
pip install cmake; # build-time dependency
pip install -e .

# install required packages
python3 -m pip install -r requirements.txt

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-2-117M", trust_remote_code=True)
model = AutoModel.from_pretrained("zhihan1996/DNABERT-2-117M", 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

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

5. Pre-Training

Codes for pre-training is coming soon.

6. Finetune

6.1 Evaluate models on GUE

Please first download the GUE dataset from here. Then run the scripts to evaluate on all the tasks.

Current script is set to use DataParallel for training on 4 GPUs. If you have different number of GPUs, please change the per_device_train_batch_size and gradient_accumulation_steps accordingly to adjust the global batch size to 32 to replicate the results in the paper. If you would like to perform distributed multi-gpu training (e.g., with DistributedDataParallel), simply change python to torchrun --nproc_per_node ${n_gpu}.

export DATA_PATH=/path/to/GUE #(e.g., /home/user)
cd finetune

# Evaluate DNABERT-2 on GUE
sh scripts/run_dnabert2.sh DATA_PATH

# Evaluate DNABERT (e.g., DNABERT with 3-mer) on GUE
# 3 for 3-mer, 4 for 4-mer, 5 for 5-mer, 6 for 6-mer
sh scripts/run_dnabert1.sh DATA_PATH 3

# Evaluate Nucleotide Transformers on GUE
# 0 for 500m-1000g, 1 for 500m-human-ref, 2 for 2.5b-1000g, 3 for 2.5b-multi-species
sh scripts/run_nt.sh DATA_PATH 0

6.2 Fine-tune DNABERT2 on your own datasets

Here we provide an example of fine-tuning DNABERT2 on your own datasets.

6.2.1 Format your dataset

First, please generate 3 csv files from your dataset: train.csv, dev.csv, and test.csv. In the training process, the model is trained on train.csv and is evaluated on the dev.csv file. After the training if finished, the checkpoint with the smallest loss on the dev.csv file is loaded and be evaluated on test.csv. If you do not have a validation set, please just make the dev.csv and test.csv the same.

Please see the sample_data folder for an sample of data format. Each file should be in the same format, with the first row as document head named sequence, label. Each following row should contain a DNA sequence and a numerical label concatenated by a , (e.g., ACGTCAGTCAGCGTACGT, 1 ).

Then, you are able to finetune DNABERT-2 on your own dataset with the following code:

cd finetune

export DATA_PATH=$path/to/data/folder  # e.g., ./sample_data
export MAX_LENGTH=100 # Please set the number as 0.25 * your sequence length. 
											# e.g., set it as 250 if your DNA sequences have 1000 nucleotide bases
											# This is because the tokenized will reduce the sequence length by about 5 times
export LR=3e-5

# Training use DataParallel
python train.py \
    --model_name_or_path zhihan1996/DNABERT-2-117M \
    --data_path  ${DATA_PATH} \
    --kmer -1 \
    --run_name DNABERT2_${DATA_PATH} \
    --model_max_length ${MAX_LENGTH} \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 16 \
    --gradient_accumulation_steps 1 \
    --learning_rate ${LR} \
    --num_train_epochs 5 \
    --fp16 \
    --save_steps 200 \
    --output_dir output/dnabert2 \
    --evaluation_strategy steps \
    --eval_steps 200 \
    --warmup_steps 50 \
    --logging_steps 100 \
    --overwrite_output_dir True \
    --log_level info \
    --find_unused_parameters False
    
# Training use DistributedDataParallel (more efficient)
export num_gpu=4 # please change the value based on your setup

torchrun --nproc-per-node=${num_gpu} train.py \
    --model_name_or_path zhihan1996/DNABERT-2-117M \
    --data_path  ${DATA_PATH} \
    --kmer -1 \
    --run_name DNABERT2_${DATA_PATH} \
    --model_max_length ${MAX_LENGTH} \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 16 \
    --gradient_accumulation_steps 1 \
    --learning_rate ${LR} \
    --num_train_epochs 5 \
    --fp16 \
    --save_steps 200 \
    --output_dir output/dnabert2 \
    --evaluation_strategy steps \
    --eval_steps 200 \
    --warmup_steps 50 \
    --logging_steps 100 \
    --overwrite_output_dir True \
    --log_level info \
    --find_unused_parameters False

7. Citation

If you have any question regarding our paper or codes, please feel free to start an issue or email Zhihan Zhou ([email protected]).

If you use DNABERT-2 in your work, please kindly cite our paper:

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},
}