This is the implementation of the paper - Generative Recurrent Networks for De Novo Drug Design

• Now support tensorflow >= 2.5.0

• Changed the code to use tensorflow 2.1.0 (tf.keras)

• Reimplimented all code to use tensorflow 2.0.0 (tf.keras)
• Changed data_loader to use generator to reduce memory usage
• Removed some unused atoms and symbols
• Changed directory layout

This model is built using Python 3.7. See Pipfile or requirements.txt for dependencies. I strongly recommend using GPU version of tensorflow.Learning this model with all the data is very slow in CPU mode (about 9 hrs / epoch). RDKit and matplotlib are used for SMILES cleanup, validation, and visualization of molecules and their properties. Recently, RDKit can be installed with pip, you don't have to use Anaconda! Scikit-learn is used for PCA.

Just run below. However, all the data is used according to the default setting. So please be careful, it will take a long time. If you don't have enough time, set data_length to a different value in base_config.json.

$ python train.py

After training, experiments/{exp_name}/{YYYY-mm-dd}/config.json is generated. It's a copy of base_config.json with additional settings for internal varibale. Since it is used for generation, be careful when rewriting.

See example_Randomly_generate_SMILES.ipynb.

See example_Fine-tuning_for_TRPM8.ipynb.

See base_config.json. If you want to change, please edit this file before training.

Download SQLite dump for ChEMBL25 (ftp://ftp.ebi.ac.uk/pub/databases/chembl/ChEMBLdb/latest/chembl_25_sqlite.tar.gz), which is 3.3 GB compressed, and 16 GB uncompressed.
Unpack it the usual way, cd into the directory, and open the database using sqlite console.

$ sqlite3 chembl_25.db
SQLite version 3.30.1 2019-10-10 20:19:45
Enter ".help" for usage hints.
sqlite> .output dataset.smi

You can get SMILES that annotated nM activities according to the following SQL query.

SELECT
  DISTINCT canonical_smiles
FROM
  compound_structures
WHERE
  molregno IN (
    SELECT
      DISTINCT molregno
    FROM
      activities
    WHERE
      standard_type IN ("Kd", "Ki", "Kb", "IC50", "EC50")
      AND standard_units = "nM"
      AND standard_value < 1000
      AND standard_relation IN ("<", "<<", "<=", "=")
    INTERSECT
    SELECT
      molregno
    FROM
      molecule_dictionary
    WHERE
      molecule_type = "Small molecule"
  );

You can get 556134 SMILES in dataset.smi. According to the paper, the dataset was preprocessed and duplicates, salts, and stereochemical information were removed, SMILES strings with lengths from 34 to 74 (tokens). So I made SMILES clean up script. Run the following to get cleansed SMILES. It takes about 10 miniutes or more. Please wait.

$ python cleanup_smiles.py datasets/dataset.smi datasets/dataset_cleansed.smi

You can get 438552 SMILES. This dataset is used for training.

The paper shows 5 TRPM8 antagonists for fine-tuning.

FC(F)(F)c1ccccc1-c1cc(C(F)(F)F)c2[nH]c(C3=NOC4(CCCCC4)C3)nc2c1
O=C(Nc1ccc(OC(F)(F)F)cc1)N1CCC2(CC1)CC(O)c1cccc(Cl)c1O2
O=C(O)c1ccc(S(=O)(=O)N(Cc2ccc(C(F)(F)C3CC3)c(F)c2)c2ncc3ccccc3c2C2CC2)cc1
Cc1cccc(COc2ccccc2C(=O)N(CCCN)Cc2cccs2)c1
CC(c1ccc(F)cc1F)N(Cc1cccc(C(=O)O)c1)C(=O)c1cc2ccccc2cn1

You can see this in datasets/TRPM8_inhibitors_for_fine-tune.smi.

Open the database using sqlite console.

$ sqlite3 chembl_25.db
SQLite version 3.30.1 2019-10-10 20:19:45
Enter ".help" for usage hints.
sqlite> .output known-TRPM8-inhibitors.smi

Then issue the following SQL query. I set maximum IC50 activity to 10 uM.

SELECT
  DISTINCT canonical_smiles
FROM
  activities,
  compound_structures
WHERE
  assay_id IN (
    SELECT
      assay_id
    FROM
      assays
    WHERE
      tid IN (
        SELECT
          tid
        FROM
          target_dictionary
        WHERE
          pref_name = "Transient receptor potential cation channel subfamily M member 8"
      )
  )
  AND standard_type = "IC50"
  AND standard_units = "nM"
  AND standard_value < 10000
  AND standard_relation IN ("<", "<<", "<=", "=")
  AND activities.molregno = compound_structures.molregno;

You can get 494 known TRPM8 inhibitors. As described above, clean up the TRPM8 inhibitor SMILES. Please use the -ft option to ignore SMILES strings (tokens) length restriction.

$ python cleanup_smiles.py -ft datasets/known-TRPM8-inhibitors.smi datasets/known_TRPM8-inhibitors_cleansed.smi

You can get 477 SMILES. I used this for mere visualization of the results of fine-tuning.