In the classical model used by pharmaceutical companies, scientists in an R&D lab investigate naturally occurring molecules while searching for potential therapeutic properties. When they find a molecule that could be a candidate, they begin a series of tests to determine the treatment efficacy of the molecules (and also to receive FDA approval).
Rather than going forward through the process, Insilico works backwards. The company starts with an end objective — say stopping aging — and then uses a toolbox of deep learning algorithms to devise ideal molecules de novo. Those molecules may not exist anywhere in the world, but can be “manufactured” in the lab.
The key underlying technique for the company is what are known as GANs, or generative adversarial networks with reinforcement learning. At a high-level, GANs include a neural net “generator” that creates new products (in this case, molecules), and a discriminator that classifies the new product. Those neural nets then adapt over time in order to compete against each other more effectively.
GANs have been used to create fake photos that look almost photorealistic, but that no camera has ever taken. Zhavoronkov suggested to me that clinical patient data may one day be manufactured — providing far more data while protecting patient privacy.