Research Progress on New Organic Molecules Design via Machine Learning

Low-cost and high-performance materials have become more and more important in past decades.It exhibits the technology level of a country.Chemists used to find the candidate material according to property regression and quantitative structure activity relationship (QSAR).Traditional methods focus on finding new molecule from prior knowledge with trial and error experiments.They are time-consuming and low efficiency on screening molecules.The appearance of machine learning (ML) changes this embarrassing situation in two ways.One is accelerating the property prediction process to prevent wasting time on worse candidates.The other is inverse molecule design which expands the imagination of human.Lots of researches show promising results using different inverse design method such as, variational auto-encoder (VAE), generative adversarial networks (GAN), reinforcement learning (RL), and recurrent neural network (RNN).They introduce uncertainty from different level to generate new structure candidates.In any method, molecule descriptor has a great impact on the result.The descriptor converts the 3D structures in real world to a vector or a notation string to feed into all kinds of ML models.Large number of descriptors have been developed in cheminformatic, bioinformatic, quantum chemistry and natural language process (NLP).Some classical descriptors are Coulomb matrix (CM), smooth overlap of atomic positions (SOAP), weighted graph (WG), simplified molecular input line entry specification (SMILES).They show different advantages and solving problems from different aspects.CM has clear definition and good result on energy regression.SOAP is good at reflecting local environment features of an atom.However, they are easy to encode but hard to decode.That is a reason why people prefer WG and SMILES in the structure inverse design tasks.WG and SMILES express structure as a graph (an atom as a node and a bond as an edge) or string to apply massive mature GNN or NLP algorithm on them.Nowadays, most of the ML applications on chemistry and molecule science are focus on developing new model to regress properties.However, it is thought that there is still large improving space on inverse design methods and traditional descriptors.In this paper, WG and SMILES are briefly introduced firstly.Then, four generative models are presented, including VAE, GAN, RL and RNN.Further, the current progress and challenges of inverse design methods are summarized case by case.Finally, some of the author՚s understanding and explorations are given out.It is proved that SMILES with BASE64 preprocessed shows some advantages on molecular reconstruction and worth to study deeply in future.