Structure generation and de novo design using reaction networks

This project is concerned with de novo molecular design whereby novel molecules are built in silico and evaluated against properties relevant to biological activity, such as physicochemical properties and structural similarity to active compounds. The aim is to encourage cost-effective compound design by reducing the number of molecules requiring synthesis and analysis.
One of the main issues in de novo design is ensuring that the molecules generated are synthesisable. In this project, a method is developed that enables virtual synthesis using rules derived from reaction sequences. Individual reactions taken from reaction databases were connected to form reaction networks. Reaction sequences were then extracted by tracing paths through the network and used to create ‘reaction sequence vectors’ (RSVs) which encode the differences between the start and end points of th esequences. RSVs can be applied to molecules to generate virtual products which are
based on literature precedents.
The RSVs were applied to structure-activity relationship (SAR) exploration using examples taken from the literature. They were shown to be effective in expanding the chemical space that is accessible from the given starting materials. Furthermore, each virtual product is associated with a potential synthetic route. They were then applied in de novo design scenarios with the aim of generating molecules that are predicted to be active using SAR models. Using a collection of RSVs with a set of small molecules as starting materials for de novo design proved that the method was capable of producing
many useful, synthesisable compounds worthy of future study.
The RSV method was then compared with a previously published method that is based on individual reactions (reaction vectors or RVs). The RSV approach was shown to be considerably faster than de novo design using RVs, however, the diversity of products was more limited.