Multiobjective De Novo Drug Design with Recurrent Neural Networks and Nondominated Sorting

A cycle to generate and optimize molecules on many properties for pharmaceutical drug design.

Requirements

• BeautifulSoup
• MatPlotLib
• Numpy
• Pandas
• PyTorch
• RDKit
• Scikit-Learn
• Selenium

Data

A text file containing all 500,000 molecules used as training data is provided in data.zipx

Jupyter Notebooks

• WebScraper.ipynb: collects SMILES strings from the open source ChEMBL dataset of drug-like molecules
• DataPreprocessing.ipynb: processes training data to properly format the text file of SMILES strings
• LSTM.ipynb: trains an LSTM recurrent neural network model on the dataset of SMILES molecules
• Generator.ipynb: generates novel and valid molecules from the LSTM model
• DataPostprocessing.ipynb: tests generated molecules for validity and pharmacokinetic properties, then selects the most optimal with the nondominated sorting algorithm
• fineLSTM.ipynb: fine-tunes the LSTM model on the most optimal (selected) molecules from the previous iteration