The first public available pytorch version alphafold2 is reproduced, and Columbia University is open source openfold, with more than 1000 stars

 Click on the above ↑↑↑“OpenCV School ” Source... Pay attention to me ： official account   Almost Human    to grant authorization 

Copy

AlphaFold2 yes 2021 year AI for Science The most dazzling star in the field . Now? , Someone is here. PyTorch It is reproduced in , And already in GitHub The open source . This recovery is now comparable in performance to the original AlphaFold2, And calculating force 、 Storage requirements are more public friendly .

just , Assistant professor of systems biology, Columbia University Mohammed AlQuraishi Announce on twitter , They trained a new one called OpenFold Model of , The model is AlphaFold2 Trainable PyTorch Duplicate version .Mohammed AlQuraishi Also said , This is the first one available to the public AlphaFold2 Reappear .

AlphaFold2 Protein structure can be predicted periodically with atomic accuracy , Technically, multi sequence alignment and deep learning algorithm are used to design , Combined with the physical and biological knowledge of protein structure, the prediction effect is improved . It has achieved 2/3 The outstanding achievement of protein structure prediction was listed on the 《 natural 》 The magazine . What's more surprising is ,DeepMind The team not only opened the model , Will also AlphaFold2 The forecast data is made into a free and open data set .

However , Open source doesn't mean you can use 、 To use . Actually ,AlphaFold2 The deployment of software system is very difficult , And high requirements for hardware 、 The data set download cycle is long 、 Large space , Every one of them makes ordinary developers flinch . therefore , The open source community has been working hard to achieve AlphaFold2 The available version of .

This time Columbia University Mohammed AlQuraishi Realized by professors and others OpenFold The total training time is about 100000 A100 Hours , But around 3000 It will be reached within hours 90% The accuracy of .

OpenFold With the original AlphaFold2 The accuracy of this method is quite , Even slightly better , May be because OpenFold Your training set is a little bigger ：

OpenFold The main advantage of is that the reasoning speed is significantly improved , For shorter protein sequences ,OpenFold The speed of reasoning can reach AlphaFold2 Twice as many . in addition , Due to the use of custom CUDA kernel ,OpenFold With less memory, you can infer longer protein sequences .

OpenFold Introduce

OpenFold It almost reproduces the original open source reasoning code (v2.0.1) All functions of , Except for those that tend to be eliminated 「 Model integration 」 function , This function is available in DeepMind I didn't perform well in my ablation test .

Whether or not DeepSpeed,OpenFold Can be used with full accuracy or bfloat16 Training . In order to achieve AlphaFold2 The original performance of , The team trained from scratch OpenFold, Model weights and training data have been published publicly . among , The training data contains approximately 400000 Share MSA and PDB70 Template file .OpenFold Also supports the use of AlphaFold The official parameters for protein reasoning .

Compared with other implementations ,OpenFold Has the following advantages ：

• Short sequence reasoning ： Accelerated in GPU The upper inference is less than 1500 The speed of the chain of amino acid residues ;
• Long sequence reasoning ： Low memory attention achieved through this study （low-memory attention） Reasoning about very long chains ,OpenFold Can be in a single A100 Up forecast exceed 4000 Sequence structure of residues , With the help of CPU offload Even longer sequences can be predicted ;
• Memory efficient during training and reasoning , stay FastFold Customization based on kernel modification CUDA The core of attention , The use of GPU The memory is better than the equivalent FastFold And existing PyTorch Less realization 4 Times and 5 times ;
• Efficiently align scripts ： The team uses the original AlphaFold HHblits/JackHMMER pipeline Or with MMseqs2 Of ColabFold, Millions of alignments have been generated .

Linux Installation and use of the system

The development team provides a local installation Miniconda、 establish conda A virtual environment 、 Install all Python Dependencies and download scripts for useful resources , Including two sets of model parameters .

Run the following command ：

scripts/install_third_party_dependencies.sh

Copy

Use the following command to activate the environment ：

source scripts/activate_conda_env.sh

Copy

Disable command ：

source scripts/deactivate_conda_env.sh

Copy

In the active environment , compile OpenFold Of CUDA kernel

python3 setup.py install

Copy

stay / usr/bin Install under path HH-suite：

# scripts/install_hh_suite.sh

Copy

Use the following command to download for training OpenFold and AlphaFold The database of ：

bash scripts/download_data.sh data/

Copy

If you want to use a set of DeepMind One or more sequences are reasoned by the pre training parameters of , You can run the following code ：

python3 run_pretrained_openfold.py \
    fasta_dir \
    data/pdb_mmcif/mmcif_files/ \
    --uniref90_database_path data/uniref90/uniref90.fasta \
    --mgnify_database_path data/mgnify/mgy_clusters_2018_12.fa \
    --pdb70_database_path data/pdb70/pdb70 \
    --uniclust30_database_path data/uniclust30/uniclust30_2018_08/uniclust30_2018_08 \
    --output_dir ./ \
    --bfd_database_path data/bfd/bfd_metaclust_clu_complete_id30_c90_final_seq.sorted_opt \
    --model_device "cuda:0" \
    --jackhmmer_binary_path lib/conda/envs/openfold_venv/bin/jackhmmer \
    --hhblits_binary_path lib/conda/envs/openfold_venv/bin/hhblits \
    --hhsearch_binary_path lib/conda/envs/openfold_venv/bin/hhsearch \
    --kalign_binary_path lib/conda/envs/openfold_venv/bin/kalign
    --config_preset "model_1_ptm"
    --openfold_checkpoint_path openfold/resources/openfold_params/finetuning_2_ptm.pt

Copy

For more details, see GitHub：https://github.com/aqlaboratory/openfold

Reference link ：

https://cloud.tencent.com/developer/article/1861192

https://twitter.com/MoAlQuraishi