This repository containis the relevant code and experiments for the paper Assessing multi-objective optimization of molecules with genetic algorithms against relevant baselines.
The algorithm can be run by initialising a file similar Click.py, which is reproduced in part below. Most importantly, this involves defining a fitness function that outputs a tupple of properties, as well as defining properties, objectives, kindto denote maximization or minimization, as well as the type of scalarizer needed. Options available are Chimera, Hypervolume and WeightedSum.
A supplement variable is required for Chimera – where it denotes thresholds the optimization should aim to meet for each objective – and WeightedSum, wherein it denotes relative weights of objectives.
def fitness_function(smi):
"""
Parameters
----------
smi: SMILES string
For one molecule
Returns
-------
fitnesses: tuple(float)
Tuple of all fitness values computed
"""
mol = Chem.MolFromSmiles(smi)
sas_val = calculateScore(mol)
logp_val = MolLogP(mol)
qed_val = qed(mol)
fitnesses = (qed_val, logp_val, sas_val)
return fitnesses
def main():
print('name=main')
params = generate_params()
properties = ['qed','logp', 'sas']
objectives = ['max', 'max', 'min']
kind = 'Chimera'
supplement = [0.6,8,0]
agent = JANUS(
work_dir='RESULTS',
num_workers = 192,
fitness_function = fitness_function,
properties = properties,
objectives = objectives,
kind = kind,
supplement = supplement,
**params
)
agent.run()
print('done!')
if __name__ == '__main__':
main()