This project is my internship project for my bachelor graduation. You can read my thesis here.
In this project, I aimed at using machine learning to solve the problem of retrosynthesis analysis, i.e. given a target organic molecule, a set of reactants compounds, which can react to form the target, are deduced. This is one of the main task in Computer-Assisted Synthesis Planning.
In particular, the models take a SMILES or SELFIES representations of a target molecule as inputs, and return another SMILES or SELFIES representing the reactants as output. Since SMILES and SELFIES are text strings, Sequence-to-sequence formalism, such as Recurrent Neural Network (RNN) or Transformer, was implemented.
The models were trained on a open-sources database called USPTO-50K, which was preprocessed by Liu et al. (2017)
All the scripts are written in Python 3.7. The following packages are required:
- TensorFlow 2.5.0
- RDKit 2020.09.1
- OpenNMT-tf 2.20.1
- SELFIES 1.0.4
- SmilesPE 0.0.3
To train a model, simply open the rnn_train_script.py or transformer_train_script.py, corresponding with RNN and Transformer architectures respectively, tune the hyper-parameters, save and run:
python rnn_train_script.py
if RNN, or
python transformer_train_script.py
if Transformer.
Checkpoints during training are saved in the model_checkpoints directory, along with the config of the models.
RNN and Transformer share a same infer_script.py file. Open the file and tune the hyper-parameters to your preference, save and run:
python infer_script.py
Inferring results are saved as a .npz file under the predictions_and_evaluation directory.
The content of the .npz file includes the following key words:
'ids'The inferred sequences representing the reactants in the form of character indices.'strings'The inferred sequences representing the reactants in the form of characters.'lengths'Lengths of the inferred sequences.'log_probs'Log probability of the every characters of the output sequences.'inference_time'Inference time in second.'attention'The attention matrices. This key word will not be included if no attention mechanism is used.
Evaluation is done by examining the .npz file from inference step. Open evaluate_script.py, tune the arguments and run:
python evaluate_script.py
The evaluation results are saved in a .csv within the same directory as the specified .npz file.