Generative AI Framework Could Help Create Better Medicine Faster
G2Retro, developed by investigators from Ohio State University, can produce hundreds of new reaction predictions in just a few minutes.
Drug development in the pharmaceutical industry is usually a process that takes several years, but new research published in the journal Communications Chemistry suggests that generative artificial intelligence (AI) could help significantly speed up the process.1
Most drug discovery is currently conducted by chemists who often use a technique called retrosynthesis. This process involves working backwards from an identified molecule and finding chemical reactions to make it. However, retrosynthesis can take a lot of time and be challenging even for experienced scientists.
In an effort to make the drug discovery process more efficient, a team of investigators from Ohio State University (OSU) developed a generative framework called G2Retro that accelerates the design of synthesis routes by producing reactions for any chemical.
“Using AI for things critical to saving human lives, such as medicine, is what we really want to focus on,” Xia Ning, PhD, lead author of the study and an associate professor of computer science and engineering at OSU, said in a statement.2 “Our aim was to use AI to accelerate the drug design process, and we found that it not only saves researchers time and money but provides drug candidates that may have much better properties than any molecules that exist in nature.”
The investigators employed a dataset that collected over 40,000 chemical reactions between 1976 and 2016 to train G2Retro. The generative framework learns from customized molecular graphs and then predicts potential reactions—similar to the process of retrosynthesis used by human chemists.
By developing new neural networks, G2Retro predicts each type of reaction and the associated atom changes. The top predicted reactions are then tested in synthon completion to produce different reactions. The framework prioritizes the most possible completion paths via a beam search strategy to avoid the generation of all possible reactions.
The researchers conducted a case study to test the AI’s effectiveness by seeing if the framework could accurately predict 4 new drugs on the market. The results showed that G2Retro correctly generated the same patented synthesis routes for all the treatments and also provided feasible alternative synthesis routes.
Additionally, investigators found that G2Retro was able to produce hundreds of new reaction predictions in just a few minutes once it was given a molecule.
However, the investigators noted that G2Retro has several limitations, including that it cannot cover all possible reaction center types, cannot cover bonds or rings that are attached at the reaction centers but do not appear in the training data, and that the atom-mapping it uses is not always available or of high quality.
“Our generative AI method G2Retro is able to supply multiple different synthesis routes and options, as well as a way to rank different options for each molecule,” said Ning. “This is not going to replace current lab-based experiments, but it will offer more and better drug options so experiments can be prioritized and focused much faster.”
