New model based on combination of GANs with reinforcement learning for drug discovery

Insilico Medicine, a Baltimore-based next-generation artificial intelligence company specializing in the application of deep learning for target identification, drug discovery and aging research announces the publication of a new research paper in Molecular Pharmaceutics  journal titled "Adversarial Threshold Neural Computer for Molecular De Novo Design". The described Adversarial Threshold Neural Computer (ATNC) model based on the combination of Generative Adversarial Networks (GANs) with Reinforcement Learning (RL) is intended for the design of novel small organic molecules with the desired set of pharmacological properties.

"This is a proof of concept scratching the surface of what we have in house. Stay tuned for the cool experimental validation results to be announced this Summer. I hope that part of this work integrated into our pipeline will help make the world a better and healthier place and help make perfect molecules for specific targets and multiple targets that will have a much higher chance of becoming great drugs", said Evgeny Putin, the deep learning lead at Insilico Medicine.

The architecture of GANs was initially proposed by Ian Goodfellow in 2015, and since the inception, the GAN-based models have achieved the unprecedented accuracy in image, video and text generation. The fundamental principle of GANs is adversarial training based on the competition between the Generative and Discriminative networks that leads to joint evolution and highly accurate results with the desired properties. Insilico Medicine scientists pioneered the application of GANs and their conjunction with RL for drug discovery process and published the proof of concept.

"The GAN-RL architecture proposed by Putin in this paper demonstrated the ability to generate a substantial percentage of valid and unique molecular structures. This study is a proof of concept using string representations of molecular structure and internally we are using multiple integrated generators with reinforcement learning and the proprietary representation of molecular structure, which allows us to synthesize the exact molecules and link chemistry and biology", said Alex Zhavoronkov, the founder and CEO of Insilico Medicine.