JANUS: Parallel Tempered Genetic Algorithm Guided by Deep Neural Networks for Inverse Molecular Design

This repository contains code for the paper: JANUS: Parallel Tempered Genetic Algorithm Guided by Deep Neural Networks for Inverse Molecular Design.

Originally by: AkshatKumar Nigam, Robert Pollice, Alán Aspuru-Guzik

Updated by: Gary Tom

Prerequsites:

Use Python 3.7 or up.

You will need to separately install RDKit version >= 2020.03.1. The easiest is to do this on conda.

JANUS uses SELFIES version 1.0.3. If you want to use a different version, pip install your desired version; this package will still be compatible. Note that you will have to change your input alphabets to work with other versions of SELFIES.

Major changes:

• Support the use of any version of SELFIES (please check your installation).
• Improved multiprocessing. Fitness function is not parallelized, in the case that the function already spawns multiple processes.
• GPU acceleration of neural networks.
• Early stopping for classifier.
• Included SMILES filtering option.
• Additional hyperparameters for controlling JANUS. Defaults used in paper are given in tests directory.

How to run:

Install JANUS using

pip install janus-ga

Example script of how to use JANUS is found in tests/example.py:

from janus import JANUS, utils
from rdkit import Chem, RDLogger
from rdkit.Chem import AllChem, RDConfig, Descriptors
RDLogger.DisableLog("rdApp.*")

import selfies

def fitness_function(smi: str) -> float:
    """ User-defined function that takes in individual smiles 
    and outputs a fitness value.
    """
    # logP fitness
    return Descriptors.MolLogP(Chem.MolFromSmiles(smi))

def custom_filter(smi: str):
    """ Function that takes in a smile and returns a boolean.
    True indicates the smiles PASSES the filter.
    """
    # smiles length filter
    if len(smi) > 81 or len(smi) == 0:
        return False
    else:
        return True

# all parameters to be set, below are defaults
params_dict = {
    # Number of iterations that JANUS runs for
    "generations": 200,

    # The number of molecules for which fitness calculations are done, 
    # exploration and exploitation each have their own population
    "generation_size": 5000,

    # Number of molecules that are exchanged between the exploration and exploitation
    "num_exchanges": 5,

    # Callable filtering function (None defaults to no filtering)
    "custom_filter": custom_filter,

    # Fragments from starting population used to extend alphabet for mutations
    "use_fragments": True,

    # An option to use a classifier as selection bias
    "use_classifier": True,
}

# Set your SELFIES constraints (below used for manuscript)
default_constraints = selfies.get_semantic_constraints()
new_constraints = default_constraints
new_constraints['S'] = 2
new_constraints['P'] = 3
selfies.set_semantic_constraints(new_constraints)  # update constraints

# Create JANUS object.
agent = JANUS(
    work_dir = 'RESULTS',                                   # where the results are saved
    fitness_function = fitness_function,                    # user-defined fitness for given smiles
    start_population = "./DATA/sample_start_smiles.txt",   # file with starting smiles population
    **params_dict
)

# Alternatively, you can get hyperparameters from a yaml file
# Descriptions for all parameters are found in default_params.yml
params_dict = utils.from_yaml(
    work_dir = 'RESULTS',  
    fitness_function = fitness_function, 
    start_population = "./DATA/sample_start_smiles.txt",
    yaml_file = 'default_params.yml',       # default yaml file with parameters
    **params_dict                           # overwrite yaml parameters with dictionary
)
agent = JANUS(**params_dict)

# Run according to parameters
agent.run()     # RUN IT!

Within this file are examples for: 1. A function for calculting property values (see function fitness_function). 2. Custom filtering of SMILES (see function custom_filter). 3. Initializing JANUS from dictionary of parameters. 4. Generating hyperparameters from provided yaml file (see function janus.utils.from_yaml).

You can run the file with provided test files

cd tests
python ./example.py

Important parameters the user should provide: - work_dir: directory for outputting results - fitness_function: fitness function defined for an input smiles that will be maximized - start_population: path to text file of starting smiles one each new line - generations: number if evolution iterations to perform - generation_size: number of molecules in the populations per generation - custom_filter: filter function checked after mutation and crossover, returns True for accepted molecules - use_fragments: toggle adding fragments from starting population to mutation alphabet - use_classifier: toggle using classifier for selection bias

See tests/default_params.yml for detailed description of adjustable parameters.

Outputs:

All results from running JANUS will be stored in specified work_dir.

The following files will be created: 1. fitness_explore.txt: Fitness values for all molecules from the exploration component of JANUS.
2. fitness_local_search.txt: Fitness values for all molecules from the exploitation component of JANUS. 3. generation_all_best.txt: Smiles and fitness value for the best molecule encountered in every generation (iteration). 4. init_mols.txt: List of molecules used to initialte JANUS. 5. population_explore.txt: SMILES for all molecules from the exploration component of JANUS. 6. population_local_search.txt: SMILES for all molecules from the exploitation component of JANUS. 7. hparams.json: Hyperparameters used for initializing JANUS.

Paper Results/Reproducibility:

Our code and results for each experiment in the paper can be found here: * Experiment 4.1: https://drive.google.com/file/d/1rscIyzpTvtyiEkoP1WsF-XtSHJGQStUU/view?usp=sharing * Experiment 4.3: https://drive.google.com/file/d/1tlIdfSWwzVeJ5kZ98l8G6osE9zf9wP1f/view?usp=sharing * GuacaMol: https://drive.google.com/file/d/1FqetwNg6VVc-C3eiPoosGZ4-47WpYBAt/view?usp=sharing

Questions, problems?

Make a github issue 😄. Please be as clear and descriptive as possible. Please feel free to reach out in person: (akshat[DOT]nigam[AT]mail[DOT]utoronto[DOT]ca, rob[DOT]pollice[AT]utoronto[DOT]ca)

License

Apache License 2.0