Magicmol - A light-weighted pipeline for drug-like molecule evolution and quick chemical space exploration

The flourishment of machine learning and deep learning methods has boosted
the development of cheminformatics, especially when it comes to the application
of drug discovery and new materials exploration. Lower time and space expenses
make it possible for scientists to search the enormous chemical space. Recently,
some work combines reinforcement learning strategies with an RNN-based
(Recurrent Neural Networks) model to optimize the property of generated
drug-like molecules, which notably improved a batch of critical factors for these
drug candidates. However, a common problem of these RNN-based methods is
that several generated molecules have difficulty in synthesizing even if owning
higher desired properties such as binding affinity. But still, the RNN-based
framework appears well in reproducing the molecule distribution among the
training set than other categories of models when it comes to the molecule
exploration tasks. Thus, to optimize the whole exploring process and make it
contribute to the optimization of specified molecules. In this paper, we devised a
light-weighted pipeline called - Magicmol with a re-mastered RNN network and
use SELFIES presentation instead of SMILES. Our backbone model achieve
extraordinary performance in evaluating metrics meanwhile reduced the training
cost and we devised reward truncate strategies to eliminate the ”model collapse”
problem. Also, adopting SELFIES presentation makes it possible for combining
STONED-SELFIES as a post-processing procedure for specified molecule
optimization and quick chemical space exploration