Mistral-DNA: Mistral large language model for DNA sequences

Overview

Here is a repo to pretrain Mistral large language model for DNA sequences. Here the Mixtral model (Mixtral-8x7B-v0.1) was modified to significantly reduce the number of parameters mostly by removing layers, such that it could be trained on a GPU such as an RTX3090.

Requirements

If you have an Nvidia GPU, then you must install CUDA and cuDNN libraries. See:
https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html
https://developer.nvidia.com/cudnn
Be aware that you should check the compatibility between your graphic card and the versions of CUDA and cuDNN you want to install. This is a bit tricky and time consuming!

To know the version of your NVIDIA driver (if you use an NVIDIA GPU) and the CUDA version, you can type:

nvidia-smi

The versions that were used here were :

• NVIDIA-SMI 535.129.03
• Driver Version: 535.129.03
• CUDA Version: 12.2

The models were developed with python and transformers.

Before installing python packages, you need to install python3 (>=3.10.12) (if you don't have it):

sudo apt update
sudo apt install python3-dev python3-pip python3-venv

Make mistral-dna environment:

conda create -n mistral-dna python=3.8
conda activate mistral-dna

To install pytorch:

pip install torch>=1.13.0

Other python packages need to be installed:

pip install transformers>=4.37.0.dev0 numpy>=1.24.4 pandas>=1.4.4 sklearn==0.0 datasets>=2.14.4 peft>=0.7.2.dev0
pip install flash-attn==0.2.4
pip install accelerate>=0.21.0
pip install bitsandbytes>=0.37.0
pip install progressbar
pip install tensorboard
pip install torch-xla>=2.0

You might need to add this to your .bashrc:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/your_path/mistral-dna/lib/

The pretrained model is available here:

https://huggingface.co/RaphaelMourad/Mistral-DNA-v0.1

To generate the data, you need to first install R packages using the following command (to type inside R):

if (!require("BiocManager", quietly = TRUE))  
install.packages("BiocManager")  
BiocManager::install("BSgenome.Hsapiens.UCSC.hg38")  
BiocManager::install("GenomicRanges")
BiocManager::install("Biostrings")

Generate the data to pretrain the model

If you want to pretrain the model using the whole human genome, first use the R script:

• scriptR/script_generate_dna_sequences.R to generate the DNA sequences \

You will obtain the following file (too large to be stored on github):

• data/genome_sequences/hg38/sequences_hg38_200b.csv.gz (100% of the human genome)

Alternatively, you can skip this step and use smaller files stored on github:

• data/genome_sequences/hg38/sequences_hg38_200b_small.csv.gz (10% of the human genome)
• data/genome_sequences/hg38/sequences_hg38_200b_verysmall.csv.gz (1% of the human genome)

Pretraining the model

Second, in the python folder "scriptPython/", you'll find the jupyter notebook:

• script_pretrain_mistral-dna.ipynb to pretrain Mixtral model on DNA sequences. \

Select the data you want to pretrain the model on (full data, small data and very small data).

The script can be ran on Google Colab.

Fine-tuning the model for classification

Third, in the python folder "scriptPython/", you'll find the jupyter notebook:

• script_finetune.ipynb to finetune the pretrained Mixtral model on a specific classification task. \

To finetune the model, you must provide a dataset to train the model. In the notebook, we used the GUE dataset. The folder GUE should be put in the folder 'data'.

The script can be ran on Google Colab.

Prediction mutation effect with zero shot learning (using the pretrained model)

The script can be ran on Google Colab.

Generate artificial sequences (using the pretrained model)

The script can be ran on Google Colab.

Contact:

[email protected]