Deep learning methods in molecule design for drug discovery.
Access rightsNo access – contact firstname.lastname@example.org
Qian, Yuchen, 1989-
MetadataShow full item record
In the recent ten years, Deep Learning has undergone rapid development. It has been proved by countless success in different areas. Computer vision projects apply convolution neural networks to extract images' features. In the natural language processing field, recurrent neural networks are used to find the connections between words. Reinforcement learning researchers utilize algorithms to beat human players. However, it is challenging to apply deep learning models in some traditional areas, like drug discovery. Molecule design is the essential phase in a drug discovery cycle. In order to reduce the assays in vitro, the compound candidates should be delicately selected. The rational and fast compounds exploration in practical drug design projects is a tough problem that urgently needs to be solved. The conventional molecule design depends on the statistical algorithms and sophisticated chemist. Nevertheless, Deep Learning methods are expert at learning from data. The key advantages of Deep Learning are the efficient learning of a labeled dataset and the precise analysis for the hidden distribution. Although deep learning usually needs an enormous amount of data, the specific model and strategy can help algorithms learn useful information. This dissertation presents a molecule design system which is capable of learning from small datasets. In order to explain the DL modules in the system better, some simple basic DL structures are introduced in the first several chapters, along with applications in other fields. The molecule design system includes three main modules: a generator module, a predictor module, and a reinforcement learning module. Specifically, the generator's purpose is to offer compound candidates with desired properties. The predictor is designed to acquire information in the low-data condition. The reinforcement learning module plays as the connector between the generator and the predictor. With the assistance of the predictor and the reinforcement learning module, the generator can move the distribution of molecules' properties towards the desired region.