The AlphaFold series has transformed protein structure prediction with remarkable accuracy, often matching experimental methods. While AlphaFold2 and AlphaFold-Multimer are open-sourced, facilitating rapid and reliable predictions, AlphaFold3 remains partially accessible and has not been open-sourced, restricting further development.
The PaddleHelix team is working on HelixFold3 to replicate the advanced capabilities of AlphaFold3. Insights from the AlphaFold3 paper inform our approach and build on our prior work with HelixFold, HelixFold-Single, HelixFold-Multimer, and HelixDock. Currently, HelixFold3's accuracy in predicting the structures of small molecule ligands, nucleic acids (including DNA and RNA), and proteins is comparable to that of AlphaFold3. We are committed to continuously enhancing the model's performance and rigorously evaluating it across a broader range of biological molecules. Please refer to our HelixFold3 technical report for more details.
Specific environment settings are required to reproduce the results reported in this repo,
- Python: 3.9
- CUDA: 12.0
- CuDNN: 8.4.0
- NCCL: 2.14.3
- Paddle: 2.6.1
Those settings are recommended as they are the same as we used in our A100 machines for all inference experiments.
HelixFold3 depends on PaddlePaddle. Python dependencies available through pip
is provided with pyproject.toml. kalign, the HH-suite and jackhmmer are
also needed to produce multiple sequence alignments. The download scripts require aria2c.
Locate to the directory of helixfold then run:
# Install py env
conda create -n helixfold -c conda-forge python=3.9
# activate the conda environment
conda activate helixfold
# adjust these version numbers as your situation
conda install -y cudnn=8.4.1 cudatoolkit=11.7 nccl=2.14.3 -c conda-forge -c nvidia
conda install -y -c bioconda aria2 hmmer==3.3.2 kalign2==2.04 hhsuite==3.3.0
conda install -y -c conda-forge openbabel
# install paddlepaddle
pip install paddlepaddle-gpu==2.6.1.post120 -f https://www.paddlepaddle.org.cn/whl/linux/mkl/avx/stable.html
# or lower version: https://paddle-wheel.bj.bcebos.com/2.5.1/linux/linux-gpu-cuda11.7-cudnn8.4.1-mkl-gcc8.2-avx/paddlepaddle_gpu-2.5.1.post117-cp39-cp39-linux_x86_64.whl
# downgrade pip
pip install --upgrade 'pip<24'
# edit configuration file at `./helixfold/config/helixfold.yaml` to set your databases and binaries correctly.
# install HF3 as a python library
pip install . --no-cache-dirNote: If you have a different version of python3 and cuda, please refer to here for the compatible PaddlePaddle dev package.
The conversion between .cif and .pdb relies on Maxit.
Download Maxit source code from https://sw-tools.rcsb.org/apps/MAXIT/maxit-v11.100-prod-src.tar.gz. Untar and follow
its README to complete installation. If you encouter error like your GCC version not support (9.4.0, for example), editing etc/platform.sh and reruning compilation again would make sense. See below:
# Check if it is a Linux platform
Linux)
# Check if it is GCC version 4.x
gcc_ver=`gcc --version | grep -e " 4\."` # edit `4\.` to `9\.`
if [[ -z $gcc_ver ]]In order to run HelixFold3, the genetic databases and model parameters are required.
The parameters of HelixFold3 can be downloaded here,
please place the downloaded checkpoint path in weight_path of helixfold/config/helixfold.yaml configuration file before install HF3 as a python module.
The script scripts/download_all_data.sh can be used to download and set up all genetic databases with the following configs:
-
By default:
scripts/download_all_data.sh ./data
will download the complete databases. The total download size for the complete databases is around 415 GB, and the total size when unzipped is 2.2 TB.
-
With
reduced_dbs:scripts/download_all_data.sh ./data reduced_dbs
will download a reduced version of the databases to be used with the
reduced_dbspreset. The total download size for the reduced databases is around 190 GB, and the total unzipped size is around 530 GB.
There are some demo input under ./data/ for your test and reference. Data input is in the form of JSON containing
several entities such as protein, ligand, nucleic acids, and iron. Proteins and nucleic acids inputs are their sequence.
HelixFold3 supports input ligand as SMILES, CCD id or small molecule files, please refer to /data/demo_6zcy_smiles.json and data/demo_p450_heme_sdf.json
for more details about SMILES input. Flexible input from small molecule is now supported. See obabel -L formats |grep -v 'Write-only'
A example of input data is as follows:
{
"entities": [
{
"type": "protein",
"sequence": "MDTEVYESPYADPEEIRPKEVYLDRKLLTLEDKELGSGNFGTVKKGYYQMKKVVKTVAVKILKNEANDPALKDELLAEANVMQQLDNPYIVRMIGICEAESWMLVMEMAELGPLNKYLQQNRHVKDKNIIELVHQVSMGMKYLEESNFVHRDLAARNVLLVTQHYAKISDFGLSKALRADENYYKAQTHGKWPVKWYAPECINYYKFSSKSDVWSFGVLMWEAFSYGQKPYRGMKGSEVTAMLEKGERMGCPAGCPREMYDLMNLCWTYDVENRPGFAAVELRLRNYYYDVVNHHHHHH",
"count": 1
},
{
"type": "ligand",
"ccd": "QF8",
"count": 1
}
]
}Another example of covalently modified input:
{
"entities": [
{
"type": "protein",
"sequence": "MDALYKSTVAKFNEVIQLDCSTEFFSIALSSIAGILLLLLLFRSKRHSSLKLPPGKLGIPFIGESFIFLRALRSNSLEQFFDERVKKFGLVFKTSLIGHPTVVLCGPAGNRLILSNEEKLVQMSWPAQFMKLMGENSVATRRGEDHIVMRSALAGFFGPGALQSYIGKMNTEIQSHINEKWKGKDEVNVLPLVRELVFNISAILFFNIYDKQEQDRLHKLLETILVGSFALPIDLPGFGFHRALQGRAKLNKIMLSLIKKRKEDLQSGSATATQDLLSVLLTFRDDKGTPLTNDEILDNFSSLLHASYDTTTSPMALIFKLLSSNPECYQKVVQEQLEILSNKEEGEEITWKDLKAMKYTWQVAQETLRMFPPVFGTFRKAITDIQYDGYTIPKGWKLLWTTYSTHPKDLYFNEPEKFMPSRFDQEGKHVAPYTFLPFGGGQRSCVGWEFSKMEILLFVHHFVKTFSSYTPVDPDEKISGDPLPPLPSKGFSIKLFPRP",
"count": 1
},
{
"type": "ligand",
"ccd": "HEM",
"count": 1
},
{
"type": "ligand",
"smiles": "CC1=C2CC[C@@]3(CCCC(=C)[C@H]3C[C@@H](C2(C)C)CC1)C",
"name": "lig",
"count": 1
},
{
"type": "bond",
"bond": "A,CYS,445,SG,B,HEM,1,FE,covale,2.3",
"_comment": "<chain-id>,<residue name>,<residue index>,<atom id>,<chain-id>,<residue name>,<residue index>,<atom id>,<bond type>,<bond length>",
"_another_comment": "use semicolon to separate multiple bonds",
"_also_comment": "For ccd input, use CCD key as residue name; for smiles and file input, use `name` identifier"
}
]
}EXPERIMENTAL Example input of residue replacements(Non Canonical Amino Acids):
{
"entities": [
{
"type": "protein",
"sequence": "MQIFVKTLTGKTITLEVESSDTIDNVKAKIQDKEGIPPDQQRLIFAGKQLEDGRTLADYNIQKESTLHLVLRLRGG",
"count": 1,
"_case_from": "https://www.rcsb.org/structure/5K9P",
"_note": "Ub"
},
{
"type": "ncaa",
"ccd": "SEP",
"_comment": "Register `SEP` as `sep`",
"_also_a_comment": "All ccd inputs will be processed as lowercase characters to avoid collisions with standard protein residues"
},
{
"type": "ncaa",
"name": "typ",
"mol2": "/repo/PaddleHelix/apps/protein_folding/helixfold3/data/typ.mol2",
"_comment": "Register this mol2 file as `typ`. Only MOL2 files are allowed. for NCAA, one must name all atoms with individual atom labels in the way that used by standard amino acid species"
},
{
"type": "modres",
"modres": "A,20,SER,sep",
"_note": "Ser20 phosphorylated ubiquitin",
"_comment": "Use `sep` to replace `SER`"
},
{
"type": "modres",
"modres": "A,59,TYR,typ",
"_note": "all phosphotyrosines, 59",
"_comment": "Use `typ` to replace `TYR`"
}
]
}The atom ID that bonded in covalent bond input is fatal. For seaking all atom ids in CCD database:
helixfold_check_ligand +ligand=HEMThis command outputs like the following:
[2024-08-31 10:55:10,131][absl][WARNING] - Using resolved obabel: /mnt/data/envs/conda_env/envs/helixfold/bin/obabel
[2024-08-31 10:55:10,132][absl][INFO] - Started Loading CCD dataset from /mnt/db/ccd/ccd_preprocessed_etkdg.pkl.gz
[2024-08-31 10:55:18,752][absl][INFO] - Finished Loading CCD dataset from /mnt/db/ccd/ccd_preprocessed_etkdg.pkl.gz in 8.620 seconds
[2024-08-31 10:55:18,752][absl][INFO] - CCD dataset contains 43488 entries.
[2024-08-31 10:55:18,752][absl][INFO] - Atoms in HEM: ['CHA', 'CHB', 'CHC', 'CHD', 'C1A', 'C2A', 'C3A', 'C4A', 'CMA', 'CAA', 'CBA', 'CGA', 'O1A', 'O2A', 'C1B', 'C2B', 'C3B', 'C4B', 'CMB', 'CAB', 'CBB', 'C1C', 'C2C', 'C3C', 'C4C', 'CMC', 'CAC', 'CBC', 'C1D', 'C2D', 'C3D', 'C4D', 'CMD', 'CAD', 'CBD', 'CGD', 'O1D', 'O2D', 'NA', 'NB', 'NC', 'ND', 'FE']
For seaking all atom ids in a given sdf/mol2/smiles:
# smiles
helixfold_check_ligand '+ligand="CNC(=O)c1nn(C)c2ccc(Nc3nccc(n3)n4cc(N[C@@H]5CCNC5)c(C)n4)cc12"'
# sdf file
helixfold_check_ligand +ligand=./60119277-3d.sdf
# mol2 file
helixfold_check_ligand +ligand=./60119277-3d.mol2
# multiple ccd ids:
helixfold_check_ligand +ligand=[SER,SEP,TYR,TYP,THR,THP]HF3 will output like the following:
helixfold_check_ligand '+ligand="CNC(=O)c1nn(C)c2ccc(Nc3nccc(n3)n4cc(N[C@@H]5CCNC5)c(C)n4)cc12"'
[2024-08-31 10:56:16,445][absl][WARNING] - Using resolved obabel: /mnt/data/envs/conda_env/envs/helixfold/bin/obabel
[2024-08-31 10:56:16,445][absl][INFO] - Guessed ligand input type: smiles
[2024-08-31 10:56:16,567][absl][INFO] - Started converting smiles to mol2: CNC(=O)c1nn(C)c2ccc(Nc3nccc(n3)n4cc(N[C@@H]5CCNC5)c(C)n4)cc12
[2024-08-31 10:56:16,567][absl][WARNING] - This takes a while ...
[2024-08-31 10:56:16,851][absl][INFO] - Finished converting smiles to mol2: CNC(=O)c1nn(C)c2ccc(Nc3nccc(n3)n4cc(N[C@@H]5CCNC5)c(C)n4)cc12 in 0.283 seconds
[2024-08-31 10:56:16,857][absl][INFO] - Atoms in CNC(=O)c1nn(C)c2ccc(Nc3nccc(n3)n4cc(N[C@@H]5CCNC5)c(C)n4)cc12 (smiles): {'UNK-': {'atom_symbol': ['C', 'N', 'C', 'O', 'C', 'N', 'N', 'C', 'C', 'C', 'C', 'C', 'N', 'C', 'N', 'C', 'C', 'C', 'N', 'N', 'C', 'C', 'N', 'C', 'C', 'C', 'N', 'C', 'C', 'C', 'N', 'C', 'C'], 'charge': [0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 'atom_ids': ['C1', 'N1', 'C2', 'O1', 'C3', 'N2', 'N3', 'C4', 'C5', 'C6', 'C7', 'C8', 'N4', 'C9', 'N5', 'C10', 'C11', 'C12', 'N6', 'N7', 'C13', 'C14', 'N8', 'C15', 'C16', 'C17', 'N9', 'C18', 'C19', 'C20', 'N10', 'C21', 'C22'], 'coval_bonds': [('C1', 'N1', 'SING'), ('N1', 'C2', 'SING'), ('C2', 'O1', 'DOUB'), ('C2', 'C3', 'SING'), ('C3', 'N2', 'AROM'), ('N2', 'N3', 'AROM'), ('N3', 'C4', 'SING'), ('N3', 'C5', 'AROM'), ('C5', 'C6', 'AROM'), ('C6', 'C7', 'AROM'), ('C7', 'C8', 'AROM'), ('C8', 'N4', 'SING'), ('N4', 'C9', 'SING'), ('C9', 'N5', 'AROM'), ('N5', 'C10', 'AROM'), ('C10', 'C11', 'AROM'), ('C11', 'C12', 'AROM'), ('C12', 'N6', 'AROM'), ('C9', 'N6', 'AROM'), ('C12', 'N7', 'SING'), ('N7', 'C13', 'AROM'), ('C13', 'C14', 'AROM'), ('C14', 'N8', 'SING'), ('N8', 'C15', 'SING'), ('C15', 'C16', 'SING'), ('C16', 'C17', 'SING'), ('C17', 'N9', 'SING'), ('N9', 'C18', 'SING'), ('C15', 'C18', 'SING'), ('C14', 'C19', 'AROM'), ('C19', 'C20', 'SING'), ('C19', 'N10', 'AROM'), ('N7', 'N10', 'AROM'), ('C8', 'C21', 'AROM'), ('C21', 'C22', 'AROM'), ('C3', 'C22', 'AROM'), ('C5', 'C22', 'AROM')], 'position': array([[ 2.8929, -1.0132, 1.1086],
[ 3.7391, -1.0171, -0.0475],
[ 4.0952, 0.1618, -0.6747],
[ 3.7451, 1.2639, -0.2632],
[ 4.9273, -0.0255, -1.8697],
[ 5.2505, -1.2502, -2.3477],
[ 6.0037, -0.9802, -3.4344],
[ 6.4721, -2.0524, -4.2816],
[ 6.2059, 0.3548, -3.6608],
[ 6.918 , 1.0385, -4.6439],
[ 6.8886, 2.4361, -4.5918],
[ 6.2178, 3.1502, -3.589 ],
[ 6.223 , 4.5552, -3.5106],
[ 7.2064, 5.4461, -3.9552],
[ 7.2928, 6.6137, -3.3032],
[ 8.3377, 7.4081, -3.6083],
[ 9.3053, 7.0807, -4.5284],
[ 9.1092, 5.8865, -5.1784],
[ 8.0557, 5.0867, -4.9217],
[ 10.0465, 5.4369, -6.1549],
[ 9.8984, 4.3439, -6.9834],
[ 10.9956, 4.3038, -7.8014],
[ 11.2377, 3.4349, -8.8185],
[ 10.147 , 2.7603, -9.5397],
[ 10.643 , 1.4653, -10.1679],
[ 11.0877, 1.8617, -11.5619],
[ 10.2177, 2.9614, -11.9789],
[ 9.6687, 3.5755, -10.7532],
[ 11.8347, 5.3372, -7.3132],
[ 13.2229, 5.6823, -7.7052],
[ 11.2358, 6.0634, -6.3575],
[ 5.4856, 2.4209, -2.6404],
[ 5.4943, 1.0102, -2.6657]], dtype=float32)}}
atom_ids is what we are looking for.
To run inference on a sequence or multiple sequences using HelixFold3's pretrained parameters, run e.g.:
LD_LIBRARY_PATH=$CONDA_PREFIX/lib/:$LD_LIBRARY_PATH \
helixfold \
input=./data/demo_8ecx.json \
output=. \
CONFIG_DIFFS.preset=allatom_demoLD_LIBRARY_PATH=$CONDA_PREFIX/lib/:$LD_LIBRARY_PATH \
helixfold --config-dir=. --config-name=myfold \
input=./data/demo_6zcy_smiles.json \
output=. \
CONFIG_DIFFS.preset=allatom_demoLD_LIBRARY_PATH=$CONDA_PREFIX/lib/:$LD_LIBRARY_PATH \
helixfold \
input=./data/demo_6zcy.json \
output=. \
CONFIG_DIFFS.preset=allatom_demo \
+CONFIG_DIFFS.model.global_config.subbatch_size=192 \
+CONFIG_DIFFS.model.num_recycle=10The descriptions of the above script are as follows:
LD_LIBRARY_PATH- This is required to load thelibcudnn.solibrary if you encounter issue likeRuntimeError: (PreconditionNotMet) Cannot load cudnn shared library. Cannot invoke method cudnnGetVersion.config-dir- The directory that contains the alterative configuration file you would like to use.config-name- The name of the configuration file you would like to use.input- Input data in the form of JSON or directory that contains such JSON file(s). For file input, check content pattern in./data/demo_*.jsonfor your reference.output- Model output path. The output will be in a folder named the same as your input json file under this path.CONFIG_DIFFS.preset- Adjusted model config preset name in./helixfold/model/config.py:CONFIG_DIFFS. The preset will be updated into final model configuration withCONFIG_ALLATOM.CONFIG_DIFFS.*- Override model any configuration inCONFIG_ALLATOM.
The outputs will be in a subfolder of output_dir, including the computed MSAs, predicted structures,
ranked structures, and evaluation metrics. For a task of inferring twice with diffusion batch size 3,
assume your input JSON is named demo_data.json, the output_dir directory will have the following structure:
<output_dir>/
└── demo_data/
├── demo_data-pred-1-1/
│ ├── all_results.json
│ ├── predicted_structure.pdb
│ └── predicted_structure.cif
├── demo_data-pred-1-2/
├── demo_data-pred-1-3/
├── demo_data-pred-2-1/
├── demo_data-pred-2-2/
├── demo_data-pred-2-3/
|
├── demo_data-rank[1-6]/
│ ├── all_results.json
| ├── predicted_structure.pdb
│ └── predicted_structure.cif
|
├── final_features.pkl
└── msas/
├── ...
└── ...
The contents of each output file are as follows:
final_features.pkl– Apicklefile containing the input feature NumPy arrays used by the models to predict the structures. If you need to re-run a inference without re-building the MSAs, delete this file.msas/- A directory containing the files describing the various genetic tool hits that were used to construct the input MSA.demo_data-pred-X-Y- Prediction results ofdemo_data.jsonin X-th inference and Y-thdiffusion batch, including predicted structures inciforpdband a JSON file containing all metrics' results.demo_data-rank*- Ranked results of a series of predictions according to metrics.
We suggest a single GPU for inference has at least 32G available memory. The maximum number of tokens is around
1200 for inference on a single A100-40G GPU with precision bf16. The length of inference input tokens on a
single V100-32G with precision fp32 is up to 1000. Inferring longer tokens or entities with larger atom numbers
per token than normal protein residues like nucleic acids may cost more GPU memory.
For samples with larger tokens, you can override model.global_config.subbatch_size in CONFIG_ALLATOM by using +CONFIG_DIFFS.model.global_config.subbatch_size=X on command runs, where X is a smaller number than 96, to save more GPU memory although this will cause a slower inference. Additionally, you can reduce the number of additional recycles by setting +CONFIG_DIFFS.model.num_recycle=Y, where Y is a smaller number than 3.
We are keen on support longer token inference, it will come in soon.
HelixFold3's code and model parameters are available under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License, license for non-commercial use by individuals or non-commercial organizations only. Please check the details in LICENSE before using HelixFold3.
[1] Abramson, J et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500. 10.1038/s41586-024-07487-w
[2] Jumper J, Evans R, Pritzel A, et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 577 (7792), 583–589. 10.1038/s41586-021-03819-2.
[3] Evans, R. et al. (2022). Protein complex prediction with AlphaFold-Multimer. Preprint at bioRxiv https://doi.org/10.1101/2021.10.04.463034
[4] Guoxia Wang, Xiaomin Fang, Zhihua Wu, Yiqun Liu, Yang Xue, Yingfei Xiang, Dianhai Yu, Fan Wang, and Yanjun Ma. Helixfold: An efficient implementation of alphafold2 using paddlepaddle. arXiv preprint arXiv:2207.05477, 2022
[5] Xiaomin Fang, Fan Wang, Lihang Liu, Jingzhou He, Dayong Lin, Yingfei Xiang, Kunrui Zhu, Xiaonan Zhang, Hua Wu, Hui Li, et al. A method for multiple-sequence-alignment-free protein structure prediction using a protein language model. Nature Machine Intelligence, 5(10):1087–1096, 2023
[6] Xiaomin Fang, Jie Gao, Jing Hu, Lihang Liu, Yang Xue, Xiaonan Zhang, and Kunrui Zhu. Helixfold-multimer: Elevating protein complex structure prediction to new heights. arXiv preprint arXiv:2404.10260, 2024.
[7] Lihang Liu, Donglong He, Xianbin Ye, Shanzhuo Zhang, Xiaonan Zhang, Jingbo Zhou, Jun Li, Hua Chai, Fan Wang, Jingzhou He, et al. Pre-training on large-scale generated docking conformations with helixdock to unlock the potential of protein-ligand structure prediction models. arXiv preprint arXiv:2310.13913, 2023.
If you use the code, data, or checkpoints in this repo, please cite the following:
@article{helixfold3,
title={Technical Report of HelixFold3 for Biomolecular Structure Prediction},
author={PaddleHelix Team},
year={2024}
}