This repository includes codes to run the model in paper

CLAIRE: A Contrastive Learning-based Predictor for EC number of chemical reactions

to predict EC numbers for chemical reactions.

1.Environment setup

In terminal

cd CLAIRE/
conda create -n claire python==3.10
conda activate claire
pip install -r requirements.txt

Install torch：You may install GPU or CPU version of torch.

conda install pytorch==1.11.0 cpuonly -c pytorch (CPU)
conda install pytorch==1.11.0 cudatoolkit=11.3 -c pytorch (GPU)

Run the following to install rxnfp:

bash rxnfp_env.sh

2.Data

You can download the (data) and place it under the CLAIRE/dev/ directory. The descriptions and purposes for the downloaded files are the following.

data/embedding: reaction embeddings from two schemes (DRFP and rxnfp), as well as the python scripts to obtain them;

data/pred_rxn_ECx: esm_emb (a dictionary for reaction SMILES and embeddings mapping), labels of testing and training sets; "x" here denotes different levels of EC numbers (first digit, two digits, three digits).

data/model_lookup_test.pkl: the featurized testing set (after embedding) in a matrix;

data/model_lookup_train.pkl: the featurized training set (after embedding) in a matrix [NOTE: this file is needed for predictions];

data/test_augmented.csv: testing set augmented samples in reaction SMILES format and their corresponding EC labels;

data/train_augmented.csv: training set augmented samples in reaction SMILES format and their corresponding EC labels;

data/predictable_EC.csv: EC numbers that are in the scope of our model. [NOTE: CLAIRE cannot predict EC numbers beyond this list].

3.How to use

(1). Run DRFP embeddings

Suppose you have three query reactions to be predicted (shown below), saved in a txt file ("my_rxn_smiles.txt"). Note that multiple reactants and products are seaparated by "."; reactants and products are separated by ">>".

NC(=O)c1ccc[n+]([C@@H]2O[C@H](COP(=O)(O)OP(=O)(O)OC[C@H]3O[C@@H](n4cnc5c(N)ncnc54)[C@H](O)[C@@H]3O)[C@@H](O)[C@H]2O)c1.NCCC=O.O>>NCCC(=O)O
C=C(C)CCOP(=O)([O-])OP(=O)([O-])[O-].CC(C)=CCOP(=O)(O)OP(=O)(O)O>>CC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCOP(=O)(O)OP(=O)(O)O
N.NC(=O)C1=CN([C@@H]2O[C@H](COP(=O)(O)OP(=O)(O)OC[C@H]3O[C@@H](n4cnc5c(N)ncnc54)[C@H](OP(=O)(O)O)[C@@H]3O)[C@@H](O)[C@H]2O)C=CC1.O=C([O-])CCC(=O)C(=O)[O-].[H+]>>N[C@@H](CCC(=O)[O-])C(=O)[O-]

Activate the claire environment:

cd CLAIRE/
conda activate claire

Run the following command to obtain DRFP embeddings and save it in "my_rxn_fps.pkl"

drfp my_rxn_smiles.txt my_rxn_fps.pkl -d 256

where -d is the dimension of the embeddings

(2). Run rxnfp embeddings

Activate the rxnfp environment:

In Python, import the relevant packages

from dev.prediction.inference_EC import inference
import pickle
import numpy as np
import pandas as pd
from rxnfp.transformer_fingerprints import (
    RXNBERTFingerprintGenerator, get_default_model_and_tokenizer, generate_fingerprints
)

compute for the rxnfp embeddings

model, tokenizer = get_default_model_and_tokenizer()
rxnfp_generator = RXNBERTFingerprintGenerator(model, tokenizer)
example_rxns = ["NC(=O)c1ccc[n+]([C@@H]2O[C@H](COP(=O)(O)OP(=O)(O)OC[C@H]3O[C@@H](n4cnc5c(N)ncnc54)[C@H](O)[C@@H]3O)[C@@H](O)[C@H]2O)c1.NCCC=O.O>>NCCC(=O)O", "C=C(C)CCOP(=O)([O-])OP(=O)([O-])[O-].CC(C)=CCOP(=O)(O)OP(=O)(O)O>>CC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCCC(C)=CCOP(=O)(O)OP(=O)(O)O", "N.NC(=O)C1=CN([C@@H]2O[C@H](COP(=O)(O)OP(=O)(O)OC[C@H]3O[C@@H](n4cnc5c(N)ncnc54)[C@H](OP(=O)(O)O)[C@@H]3O)[C@@H](O)[C@H]2O)C=CC1.O=C([O-])CCC(=O)C(=O)[O-].[H+]>>N[C@@H](CCC(=O)[O-])C(=O)[O-]"]
rxnfp = rxnfp_generator.convert_batch(example_rxns)
pickle.dump(rxnfp, open('rxnfp_emb.pkl', 'wb'))

(3). Concatenate the rxnfp and drfp embeddings

drfp = pickle.load(open('my_rxn_fps.pkl', 'rb'))
rxnfp = pickle.load(open('rxnfp_emb.pkl', 'rb'))
test_data = []

for ind, item in enumerate(rxnfp):
    rxn_emb = np.concatenate((np.reshape(item, (1,256)), np.reshape(drfp[ind], (1,256))), axis=1)
    test_data.append(rxn_emb)

test_data = np.concatenate(test_data,axis=0)

(4). Make predictions on the concatenated embeddings

Activate the claire environment:

train_data = pickle.load(open ('data/model_lookup_train.pkl', 'rb'))
train_labels = pickle.load(open ('data/pred_rxn_EC123/labels_train_ec3.pkl', 'rb')) #if you want 1-level EC or 2-level EC, change it to pred_rxn_EC1/labels_trained_ec1.pkl or pred_rxn_EC12/labels_trained_ec2.pkl, resepetively.
# input your test_labels
test_labels = None
test_tags = ['rxn_' + str(i) for i in range(len(test_data))]

# EC calling results using maximum separation
pretrained_model = '../results/model/pred_rxn_EC123/layer5_node1280_triplet2000_final.pth'
inference(train_data, test_data, train_labels, test_tags,test_labels, pretrained_model, evaluation=True, topk=3, gmm = '../gmm/gmm_ensumble.pkl')

The prediction results are saved in dev/test_prediction.csv.

This project uses part of codes (the gmm functions) from the CLEAN software developed by the Department of Chemical and Biomolecular Engineering at the University of Illinois Urbana-Champaign.