The is the 6th revised version of initial https://github.com/bigchem/transformer-cnn presented in those pre-print / paper: https://arxiv.org/abs/1911.06603 & https://jcheminf.biomedcentral.com/articles/10.1186/s13321-020-00423-w
The code is adapted for terminal used within OCHEM and I add few protocols to run terminal call inside a jupyter notebook via instructions.
Data Format: a csv file with one columns as SMILES or smiles and the targets (can be multiple targets and with different type ie classes and regression tasks) can be sparse too.
This code was tested a batchsize 256 (use 64 for small datasets) on a RTX4090. It tasks 15 Go on the GPU for a run with 95k entries x n = 20.
python 3.11 (via conda see installation) pandas tensorflow==2.15.1 rdkit==2023.09.3 numpy==1.26.4 scikit-learn==1.3.2
conda create --name TCNN python=3.11
conda activate TCNN
pip install . In the installation folder run:
python transformer_cnn/augment_smiles.py -i fpk.csv -o transformer_cnn/augmentedFPK.csv -n 10 -s True -t Truepython transformer_cnn/runcv.py --data transformer_cnn/augmentedFPK.csv --output cv_results --naug 10python transformer_cnn/scoringregcv.py -d cv_results -t FPK -n 10 Options are available in the config.cfg file under the [Details] section:
first-line = True # Consider dataset with one line
n_epochs = 25
batch_size = 64
early-stopping = 0.9
learning_rate = 1.0E-4
chirality = True # Use chiral SMILES
retrain = False # Full retrain of the embedding transformer
fixed-learning-rate = True # Learning plateau decay
canonize = False
gpu = 0 # Define the GPU card number to use for the run
random = True
seed = 10666 # Default seed in OCHEM for reproducibility
augment = 20
lossmae = False # Change the loss to MAESuggestions:
- `lossmae can help for dataset with visible outliers.
To fully retrain the embedding transformer (learning to transform random SMILES into canonical SMILES), set:
retrain = TrueAdditionally, prepare a dataset with two columns, where each molecule's first line is structured as canonical,canonical other are randomsmiles,canonical:
canonicalsmile,canonicalsmile
randomsmiles,canonicalsmile
randomsmiles,canonicalsmile
...
If you already have a split dataset, follow these steps:
python augment_smiles.py -i train.csv -o trainaug.csv -n 10 -s True -t True
python augment_smiles.py -i apply.csv -o applynaug.csv -n 10 -s True -t Truepython run.py --mode train --data trainaug.csvpython run.py --mode apply --data applynaug.csvAugmentations are grouped using the "augid" from df.index with a modulo operator on the fly.
To ensure proper functioning, you must pass the correct -n augmentation number.
MIT License