A Hong Kong-based startup, Insilico Medicine, published research showing that its deep learning system could identify potential treatments for fibrosis. That system, called generative tensorial reinforcement learning (GENTRL), was able to find six treatments in just 21 days, one of which showed promising results in an experiment involving mice. The research has been published in Nature Biotechnology, and the code for the model has been made available on Github.
“We’ve got AI [artificial intelligence] strategy combined with AI imagination,” says Insilico CEO Alex Zhavoronkov, who compares GENTRL to the AlphaGo machine learning system developed by Google’s Deepmind to challenge champion Go players.
Zhavoronkov founded the company in 2014. His original background was in computer science, but he became involved with biotechnology research, with an interest in research into slowing down the aging process. His PhD studies focused on using machine learning to look at the physics of molecular interactions in biological systems.
Insilico Medicine’s original philosophy was about using deep learning to train neural networks to go through large libraries of molecules to find drug targets. However, after becoming aware of Ian Goodfellow’s work in machine learning, Zhavoronkov decided to change tack.
The company wondered whether it was possible to make machines imagine new molecules with particular properties, instead of screening large vendor libraries. Screening molecules is the method used in the traditional drug discovery world, but Zhavoronkov wanted to see if this type of machine learning could speed up the process.
The current paper came about from a challenge put to the company by its colleagues in the chemistry world. They asked Insilico to use its system to develop potential drugs that can inhibit discoidin domain receptor 1 (DDR1) activity. DDR1 is an enzyme that is involved in fibrosis, and although it’s not yet clear if it regulates those processes, inhibiting its activity is being investigated as a potential therapy. The challenge was based on recently published research from a Genentech team, which had taken almost eight years to identify promising DDR1 kinase inhibitors.
Using GENTRL to design new drug candidates, Insilico then synthesized these and a leading candidate was successfully tested in mice. It took about 21 days for the AI system to design molecules, and the total time for design, synthesis, and validation was around 46 days.
Although none of the drugs designed by GENTRL appear to be more effective than inhibitors developed via the traditional research method, the traditional process to develop drug candidates took over eight years and millions of dollars to develop––compared to the handful of weeks and approximate $150,000 cost of Insilico’s method.
Although he cautions that the company still has much work ahead, Zhavoronkov says this research is an important breakthrough because it demonstrates the promise of AI for drug discovery.
Source: Forbes, September 2, 2019