Stereo2vec

Stereo2vec is a word embedding method that addresses molecular discrimination by leveraging stereochemical data from molecules or by treating Isomeric SMILES notation as textual input in Natural Language Processing

Requirements

This project depends on the following Python (Python 3) libraries:

• RDKit version 2023.03.1 - pip install rdkit==2023.03.1
• NumPy version 1.24.3 - pip install numpy==1.24.3
• Pandas version 1.3.4 - pip install pandas==1.3.4
• DeepChem version 2.7.1 - pip install deepchem==2.7.1
• Tensorflow version 2.13.1 - pip install tensorflow==2.13.1
• Gensim version 4.1.2 - pip install gensim==4.1.2
• Mordred version 1.2.0 - pip install mordred==1.2.0
• MHFP version 1.9.6 - pip install mhfp==1.9.6
• Mol2vec version 0.1 - pip install mol2vec==0.1
• networkx version 2.5 - pip install networkx==2.5
• scikit-learn version 1.3.0 - pip install scikit-learn==1.3.0
• scipy version 1.8.1 - pip install scipy==1.8.1
• h5py version 3.10.0 - pip install h5py==3.10.0

Pretranied Models

Models are available at https://doi.org/10.5281/zenodo.10651454

Model Training

To train a model, the ./Evaluations/train.py script must be used. It takes 7 arguments, 4 of them are required:

• method_name: the selected method, "Mol2vec", "MolChiral2vec", "MolStereo2vec", "IsoString2vec", "IsoSymbol2vec" or "IsoOrder2vec" (required)
• base_path: the path that models will be saved (optional, by default it's "./examples/")
• file: the file that contains multiple SMILES (optional, by default it's "./examples/smiles_examples.txt", which has 1000 random SMILES)
• generate_corpus: the corpus generation option (required)
• gensim_method: the gensim method, "FastText" or "Word2vec" (required)
• train: the model training option (required)
• dimensions: the number of dimensions used to represent the molecule (optional, by default it's 300)

Example usage:

# To train a 300-dimensional MolChiral2vec model with FastText approach, with its own corpus generation
python train.py --method_name "MolChiral2vec" --gensim_method "FastText" --train "yes" --generate_corpus "yes"; 
# To train a 100-dimensional MolStereo2vec with Word2vec, with an existing corpus
python train.py --method_name "MolStereo2vec" --gensim_method "Word2vec" --train "yes" --generate_corpus "no" --dimensions 100;

Model Evaluation

To evaluate a model ./Evaluations/evaluate.py script must be used: it generates the numerical vectors for the selected model and selected file that contains multiple smiles and then compares the number of duplicated data. It uses prebuilt models. It takes 7 arguments, 1 of them are required:

• file: the file that contains multiple SMILES (optional, by default it's "./examples/smiles_examples.txt", which has 1000 random SMILES)
• method_name: the selected method (required)
  • Molecular Descriptors:
    • "RDKitDescriptors"
    • "MordredDescriptors"
  • Procedural Fingerprints:
    • "MACCS"
    • "RDKitFP"
    • "ECFP"
    • "ECFPChirality"
    • "MHFP"
  • Learning-based Fingerprints:
    • "GCF"
    • "Mol2vec"
  • Stereo2vec:
    • "MolChiral2vec"
    • "MolStereo2vec"
    • "IsoString2vec"
    • "IsoSymbol2vec"
    • "IsoOrder2vec"
• gensim_method: the gensim method, "FastText" or "Word2vec" (optional, by default it's FastText)
• dimensions: the number of dimensions used to represent the molecule (optional, by default it's 300)
• radius: the radius of atom environments considered in the fingerprint (optional, by default it's 1)
• large_set: to have a smaller molecules set for Mordred Descriptors to handle memory usage (optional, by default it's False, it will just work for Mordred Descriptors)
• batch_size: the number of molecules in each batch for Mordred Descriptors (optional, by default it's 5000, it will just work for Mordred Descriptors)

Example usage:

# If you want to use your own file, called filename, you can the following command:
python evaluate.py --method_name "MACCS" --file filename; 
# To test with the prebuilt file, you can use the following examples:
python evaluate.py --method_name "MACCS";
python evaluate.py --method_name "ECFPChirality" --dimensions 1024;
python evaluate.py --method_name "ECFPChirality" --dimensions 1024 --radius 2;
python evaluate.py --method_name "RDKitDescriptors";
python evaluate.py --method_name "MordredDescriptors" --large_set --batch_size 10000;
python evaluate.py --method_name "MolStereo2vec";
python evaluate.py --method_name "MolStereo2vec" --dimensions 200;
python evaluate.py --method_name "IsoSymbol2vec" --gensim_method "Word2vec" --dimensions 300;