Official implementation of "A unified pre-trained deep learning framework for cross-task reaction performance prediction and synthesis planning". Here, we provide a wiki document, which was auto-generated by DeepWiki based on latest commit. The correpsonding paper is available at Nature Machine Intelligence. Preprint is available at ResearchSquare.
Pytorch 2 support is under development (pytorch 2 branch).
- Unified Architecture: Integrates graph neural networks with Transformer to capture both intra and inter-molecular information.
- Pre-trained Model: A pre-trained model that learns chemical bond transformation patterns from 13 million reactions.
- Cross-task Reaction Prediction: Simultaneous handling of:
- Yield, regioselectivity and enantioselectivity estimation (regression)
- Forward- and retro-synthesis planning (sequence generation)
- Performance: RXNGraphormer achieves state-of-the-art performance across eight benchmark datasets for reactivity/selectivity prediction and forward-/retro-synthesis planning, as well as three external realistic datasets for reactivity and selectivity prediction.
conda create -n rxngraphormer python=3.8
conda activate rxngraphormer
# option 1: install pre-built package (recommended)
pip install rxngraphormer -i https://pypi.org/simple -f https://data.pyg.org/whl/torch-1.12.0+cu113.html --extra-index-url https://download.pytorch.org/whl/cu113
# option 2: install from source (development mode)
git clone https://github.com/licheng-xu-echo/RXNGraphormer.git
cd RXNGraphormer/
pip install -r requirements.txt -f https://data.pyg.org/whl/torch-1.12.0+cu113.html --extra-index-url https://download.pytorch.org/whl/cu113
pip install .Note: All codes were tested under Ubuntu 22.04.2 LTS
All model weights and preprocessed datasets are available via our figshare repository. Please put these files in model_path and dataset folders, like the following structure:
RXNGraphormer
|── rxngraphormer
├── model_path
│ ├── buchwald_hartwig/
| ├── C_H_func/
| ├── pretrained_classification_model/
| ├── suzuki_miyaura/
│ ├── thiol_addition/
| ├── USPTO_50k/
| ├── USPTO_480k/
| └── ...
|── dataset
│ ├── 50k_with_rxn_type/
| ├── external_validation_dataset/
│ ├── OOS/
│ ├── pretrain/
│ ├── USPTO_50k/
│ ├── USPTO_480k/
│ ├── USPTO_full/
│ └── ...
|── ...
Reaction performance prediction
from rxngraphormer.eval import reaction_prediction
# reactivity prediction
bh_model_path = "./model_path/buchwald_hartwig/seed0"
rxn_smiles_lst = ["CC(C)c1cc(C(C)C)c(-c2ccccc2P(C2CCCCC2)C2CCCCC2)c(C(C)C)c1.CN(C)C(=NC(C)(C)C)N(C)C.Cc1ccc(N)cc1.FC(F)(F)c1ccc(I)cc1.c1ccc(-c2ccon2)cc1>>Cc1ccc(Nc2ccc(C(F)(F)F)cc2)cc1",
"CC(C)c1cc(C(C)C)c(-c2ccccc2P(C2CCCCC2)C2CCCCC2)c(C(C)C)c1.CCOC(=O)c1cc(C)on1.CN(C)C(=NC(C)(C)C)N(C)C.Cc1ccc(N)cc1.FC(F)(F)c1ccc(Cl)cc1>>Cc1ccc(Nc2ccc(C(F)(F)F)cc2)cc1"]
bh_react_preds = reaction_prediction(bh_model_path,rxn_smiles_lst,task_type="reactivity")
# selectivity prediction
thiol_add_model_path = "./model_path/thiol_addition/seed0"
rxn_smiles_lst = ["COCc1cccc(-c2cc3c(c4c2OP(=O)(O)Oc2c(-c5cccc(COC)c5)cc5c(c2-4)CCCC5)CCCC3)c1.O=C(/N=C/c1ccccc1)c1ccccc1.Sc1ccccc1>>O=C(NC(Sc1ccccc1)c1ccccc1)c1ccccc1",
"COCc1cccc(-c2cc3ccccc3c3c2OP(=O)(O)Oc2c(-c4cccc(COC)c4)cc4ccccc4c2-3)c1.O=C(/N=C/c1cccc2ccccc12)c1ccccc1.SC1CCCCC1>>O=C(NC(SC1CCCCC1)c1cccc2ccccc12)c1ccccc1"]
thiol_add_sel_preds = reaction_prediction(thiol_add_model_path,rxn_smiles_lst,task_type="selectivity")
# It takes about 25 seconds to predict 100 reactions on a single AMD 7735HS CPU without GPU.Reaction synthesis planning prediction
from rxngraphormer.eval import reaction_prediction
# retro-synthesis planning
uspto_50k_model_path = "./model_path/USPTO_50k"
pdt_smiles_lst = ['COC(=O)[C@H](CCCCN)NC(=O)Nc1cc(OC)cc(C(C)(C)C)c1O',
'O=C(Nc1cccc2cnccc12)c1cc([N+](=O)[O-])c(Sc2c(Cl)cncc2Cl)s1',
'CCN(CC)Cc1ccc(-c2nc(C)c(COc3ccc([C@H](CC(=O)N4C(=O)OC[C@@H]4Cc4ccccc4)c4ccon4)cc3)s2)cc1']
rct_preds = reaction_prediction(uspto_50k_model_path, pdt_smiles_lst, task_type="retro-synthesis")
# forward-synthesis planning
uspto_480k_model_path = "./model_path/USPTO_480k"
rct_smiles_lst = ['C1CCOC1.CC(C)C[Mg+].CON(C)C(=O)c1ccc(O)nc1.[Cl-]',
'CN.O.O=C(O)c1ccc(Cl)c([N+](=O)[O-])c1',
'CCn1cc(C(=O)O)c(=O)c2cc(F)c(-c3ccc(N)cc3)cc21.O=CO']
pdt_preds = reaction_prediction(uspto_480k_model_path, rct_smiles_lst, task_type="forward-synthesis")
# It takes about 40 seconds to predict 100 reactions on a single AMD 7735HS CPU without GPU.Reaction embeddings generation
from rxngraphormer.rxn_emb import RXNEMB
# base on pretrained model
pretrain_model_path = "./model_path/pretrained_classification_model"
rxnemb_calc_pretrained = RXNEMB(pretrained_model_path=pretrain_model_path,model_type="classifier")
rxn_emb_pretrained = rxnemb_calc_pretrained.gen_rxn_emb(["C1CCCCC1.CCO.CS(=O)(=O)N1CCN(Cc2ccccc2)CC1.[OH-].[OH-].[Pd+2]>>CS(=O)(=O)N1CCNCC1",
"CCOC(C)=O.Cc1cc([N+](=O)[O-])ccc1NC(=O)c1ccccc1.Cl[Sn]Cl.O.O.O=C([O-])O.[Na+]>>Cc1cc(N)ccc1NC(=O)c1ccccc1",
"COc1ccc(-c2coc3ccc(-c4nnc(S)o4)cc23)cc1.COc1ccc(CCl)cc1F>>COc1ccc(-c2coc3ccc(-c4nnc(SCc5ccc(OC)c(F)c5)o4)cc23)cc1"])
# base on finetuned model
finetune_model_path = "./model_path/buchwald_hartwig/seed0"
rxnemb_calc_finetuneed = RXNEMB(pretrained_model_path=finetune_model_path,model_type="regressor")
rxn_emb_finetuned = rxnemb_calc_finetuneed.gen_rxn_emb(["CC(C)c1cc(C(C)C)c(-c2ccccc2P(C2CCCCC2)C2CCCCC2)c(C(C)C)c1.CN(C)C(=NC(C)(C)C)N(C)C.Cc1ccc(N)cc1.FC(F)(F)c1ccc(I)cc1.c1ccc(-c2ccon2)cc1>>Cc1ccc(Nc2ccc(C(F)(F)F)cc2)cc1",
"CCN=P(N=P(N(C)C)(N(C)C)N(C)C)(N(C)C)N(C)C.COc1ccc(OC)c(P([C@]23C[C@H]4C[C@H](C[C@H](C4)C2)C3)[C@]23C[C@H]4C[C@H](C[C@H](C4)C2)C3)c1-c1c(C(C)C)cc(C(C)C)cc1C(C)C.Cc1ccc(N)cc1.Ic1cccnc1.c1ccc(-c2ccon2)cc1>>Cc1ccc(Nc2cccnc2)cc1",
"CC(C)c1cc(C(C)C)c(-c2ccccc2P(C2CCCCC2)C2CCCCC2)c(C(C)C)c1.CCc1ccc(I)cc1.CN1CCCN2CCCN=C12.Cc1ccc(N)cc1.c1ccc2oncc2c1>>CCc1ccc(Nc2ccc(C)cc2)cc1"])Fictious reaction identification
from rxngraphormer.rxn_emb import RXNClassifier
pretrained_model_path = "./model_path/pretrained_classification_model"
classifier = RXNClassifier(pretrained_model_path,random_init=False)
# return predictions and confidences
preds,confs = classifier.rxn_pred(["BrCc1ccccc1.CCOC(=O)C1(c2ccc(-c3ccc(-c4onc(C)c4CCO)cc3)cc2)CC1>>COCC(=O)C1(c2ccc(-c3ccc(-c4onc(C)c4CCOCc4ccccc4)cc3)cc2)CC1",
"Brc1cc2c(cc1CCCI)OCO2.CCOC1=CC(=O)CCC1>>CC(O)C1=CC(=O)C(CCCc2cc3c(cc2Br)OCO3)CC1",
"B.CC#N.CCOC(=O)C(Cc1ccc(OS(=O)(=O)C(F)(F)F)cc1)NC(=O)C(OC)(c1ccccc1)C(F)(F)F.CNC.O=C([O-])[O-].[K+].[K+].[Pd].c1ccc(P(c2ccccc2)c2ccccc2)cc1.c1ccc(P(c2ccccc2)c2ccccc2)cc1.c1ccc(P(c2ccccc2)c2ccccc2)cc1.c1ccc(P(c2ccccc2)c2ccccc2)cc1>>CCOC(=O)C(Cc1ccccc1)NC(=O)C(OC)(c1ccccc1)C(F)(F)F",
"Br.CC[C@@H](c1cccc(OCc2ccccc2)c1)[C@@H](C)CO.ClCCl.O>>CC[C@@H](c1cccc(OCc2ccccc2)c1)[C@@H](C)CBr"]) Although the script automatically preprocesses the data when you start the model training program, we recommend that you perform data preprocessing separately to avoid errors when using multi-gpu training models. Multi-gpu is usually used for model pretraining and fine-tuning for reaction synthesis planning tasks.
In the pre-training task, multiple data sub-files are used to create training sets and validation sets. The file_num_trunck parameter in the json file can control how many sub-files are used. If it is set to 0, all sub-files are used.
python data_preprocess.py --config_json ./config/pretrain_parameters.json # preprocess pretraining data
CUDA_VISIBLE_DEVICES="0,1" python -m torch.distributed.launch --nproc_per_node 2 train_model.py --config_json ./config/pretrain_parameters.json # pretrain model with multi-gpu
python train_model.py --config_json ./config/C_H_func_parameters.json # finetune regression model
python train_model.py --config_json ./config/uspto_50k_parameters.json # finetune sequence genration modelpython eval_model.py --config_json ./config/buchwald_hartwig_eval.json # evaluate regression model
python eval_model.py --config_json ./config/uspto_50k_eval.json # evaluate sequence generation modelHere we provide a notebook to show how to generate fictitious reactions based on real reactions, how to generate delta-mol SMILES based on reaction SMILES, how to evaluate the regression performance of RXNGraphormer, use the embeddings generated by pre-trained RXNGraphormer to measure the distance between different reaction types in USPTO-50k, and use the embeddings generated by fine-tuned RXNGraphormer to examine the model's understanding of reaction structure-performance relationship.
Please cite the following paper: Xu, LC., Tang, MJ., An, J. et al. A unified pre-trained deep learning framework for cross-task reaction performance prediction and synthesis planning. Nat. Mach. Intell. 2025, 7, 1561.
If you have any questions, please contact us via email: [email protected] or [email protected].