King's College London

Researchers at King’s College London and Imperial College London have developed a new AI algorithm that can generate thousands of virtual drug-like molecules in a matter of seconds. The computer-based tool could be valuable in the development of new drugs tailored to specific drug targets.

Virtual-library screening is an important step in early drug discovery. It involves using computational tools to search through databases of existing compounds to find those with structures that are most likely to bind to a particular drug target, fitting together like two pieces of a puzzle. Once suitable compounds have been identified, they are optimised and tested in the laboratory in cell and animal models before entering clinical trials.

But a limitation of existing virtual libraries is that the search for suitable compounds is limited to those that already exist within the libraries, making it difficult to identify new potential drugs.

Researchers have developed a new AI algorithm, named DrugSynthMC (Drug Synthesis using Monte Carlo), that expands the diversity within drug libraries by generating the chemical structures of thousands of drug-like molecules per second, from scratch. The approach is outlined in the Journal of Chemical Information and Modeling.

“We showed that DrugSynthMC can expand the chemical diversity of compounds in available libraries, overcoming the limitations of existing drug collections,” Dr Prischi said.

DrugSynthMC uses a type of algorithm called Monte Carlo Tree Search – a mathematical technique that predicts all the possible outcomes based on a defined set of actions. In this case, DrugSynthMC builds the chemical structures of molecules in a simple text format by following a small set of instructions aimed at maximising the important features of orally available drugs. The team found that the algorithm was successful at generating a high proportion of compounds that are easy to synthesise, are soluble and non-toxic.

Importantly, DrugSynthMC can be customised to screen for molecules that are most likely to bind to a specific biological target. The team believes the main use of the AI algorithm in the future will be in the identification and optimisation of compounds against protein targets linked to specific diseases.

“I’m very excited. Even though this is a fairly simple algorithm, it’s far more efficient than anything more complex that has been tested or published out there, and will become very useful in AI-driven drug discovery for bespoke therapeutic targets,” said Dr Olivier Pardo, Reader in Cancer Cell Signalling at Imperial College London and co-senior author of the study.

The tool is publicly available for use by the research community.

The work was carried out in collaboration with Professor Tristan Cazenave and Milo Roucairol at the PRAIRIE Institute University Paris-Dauphine PSL, France. The work was supported in part by the French government under management of Agence Nationale de la Recherche as part of the “Investissements d’avenir” program, reference ANR19- P3IA-0001 (PRAIRIE 3IA Institute).