This interview is part of our new AI in Healthcare series, where we interview the world's top thought leaders on the front lines of the intersections between AI and healthcare.
In this interview, we speak with Alex Zhavoronkov, Ph.D., CEO of Insilico Medicine, to understand how his company is using AI to transform healthcare, and what the future of the industry holds.
1. What’s the story behind Insilico Medicine? Why and how did you begin?
AZ: Insilico Medicine was founded in 2014 to utilize the power of deep learning to accelerate the pharmaceutical drug discovery process. Since our inception, we were moving in front of the technology train and publishing many firsts in this industry.
We have over 200 different collaborations all over the world and are constantly learning. When we just started transitioning from the advanced bioinformatics into deep learning in 2013-2014 and building the expertise in this area, even my team was very skeptical. In 2014, we got our differential pathway perturbation algorithms to work, and even that was too advanced for the big pharma companies. Our focus on the minute changes that transpire during aging did not help.
This was before TensorFlow and before the NVIDIA’s famous 2015 GTC conference, which followed right after PMWC 2015 and started the revolution and the hype in deep learning. We were also in the finals of the NVIDIA’s GTC, but a company which did style transfer on images using deep learning ultimately took the first prize. Only our president of European R&D, Alex Aliper, embraced deep learning and pushed full steam ahead. We started hiring the top deep learning talent all over the world through hackathons and then began training them in bioinformatics.
Our real breakthrough came at the end of 2015, when we got into a new technique in deep learning first proposed by Ian Goodfellow, called the Generative Adversarial Networks (GANs). We published the first proofs of concept and quickly became popular in the pharma world. Many of the top biotech investors and pharma executives wanted to talk to us at the JP Morgan Healthcare Conference 2016, where we met Jim Mellon, who invested in Insilico and then formed a separate company together with Greg Bailey. This team is now developing the molecules generated entirely using AI for age-related diseases. This team led our most recent round of funding and also took some of our molecules generated using AI into development.
We also started expanding into Asia by opening R&D centers in Taiwan, and a ROC and JV in Korea, where we focus on cosmetics applications. We are working on expanding into China, and have now moved our headquarters to Hong Kong to be closer to our main customer base.
2. Please describe your use case and how Insilico Medicine uses artificial intelligence:
AZ: Insilico Medicine is focusing on target identification and compound generation in one seamless pipeline. We use next-generation AI techniques to learn biology and chemistry in a very holistic way and learn to recognize the norm. Then we teach AI to recognize the changes between norm and disease and identify the key drivers behind this change. We also use “imaginative AI” to create new molecules for these targets with a list of specific medicinal chemistry properties.
About a quarter of our R&D is going into AI for age-related diseases. We described this work in the paper titled “Artificial intelligence for aging and longevity research: Recent advances and perspectives“. Here we show how to use one feature that every living being has, age, to groom together previously incompatible data types and identify the most important drivers of the process.
3. Could you share a specific customer/user that benefits from what you offer? What has your Insilico Medicine done for them?
AZ: We are actively partnering with major pharma companies. The generative adversarial networks (GANs) allow us to make the new chemical entities with certain new properties from scratch. When we add reinforcement learning into the process the system learns to generate the novel molecules with certain properties for a specific objective. We were the first to publish this approach in a peer-reviewed journal in 2016 and then took two years to get this technique to work and validate it experimentally. It helps us to design the molecules not present in nature and think in many more dimensions than a medicinal chemist. It provides us with a business model and also enables us to partner with pharma.
So far we managed to demonstrate that our deep learning approaches can be effectively used to discover novel molecular targets and our GAN-RL systems can be used to generate novel molecules in record time.