Insilico’s Alex Zhavoronkov Highlights Generative AI's Impact on Drug Discovery and Aging Research

By Allison Proffitt 

April 17, 2025 | At the Bio-IT World Conference & Expo earlier this month, Alex Zhavoronkov, CEO and founder of Insilico Medicine, shared the company's progress in leveraging generative AI to accelerate drug discovery, particularly focusing on aging research. Celebrating the company's 11th year, Zhavoronkov emphasized his singular mission: helping people live longer, healthier lives. 

"Currently, my problem in life is that there are so many ways to diagnose aging. You can look at yourself in the mirror or in the passport. But there are very, very few ways to intervene in aging," Zhavoronkov explained. His hypothesis drives Insilico's research: if a drug is powerful enough to target aging mechanisms, it should be effective against multiple diseases. 

Insilico Medicine has built an integrated AI platform that spans target discovery, small molecule design, biologics, and clinical trial outcome prediction. The company distinguishes itself by investing its own resources to validate its AI tools. "We actually bet a lot of money to actually see if the molecule works or not," Zhavoronkov noted, contrasting this approach with companies that rely solely on pharmaceutical partners to test their software. 

The company's business model involves developing novel drug candidates and then licensing them to pharmaceutical companies. Zhavoronkov highlighted an $80 million upfront deal with Exelixis for a USP1 inhibitor as an example of this strategy's success. Their most advanced program, targeting idiopathic pulmonary fibrosis (IPF), has completed Phase 2A clinical trials. 

“That's kind of the idea: to precook the cookies and sell them to the cookie makers, while testing your own cookie-making software and then making the software available to the community,” Zhavoronkov said.  

Validation Challenge  

Echoing sentiments shared by others in at the event, Zhavoronkov highlighted experimental validation as the “greatest challenge in drug discovery.” Unlike text or images, where humans can immediately judge the output quality, biological and chemical predictions require laboratory testing. Nevertheless, Insilico has dramatically accelerated the drug development timeline, requiring approximately 13 months to progress from initial concept to preclinical candidate compound—a milestone that includes efficacy data from multiple animal models and preliminary toxicology studies. 

To enable this rapid pace, the company operates globally with a significant presence in China, leveraging a network of over 40 contract research organizations that often run experiments in parallel, sometimes with intentional redundancy to ensure reliable results. 

“Those are realistic timelines,” Zhavoronkov said. “I think that currently we have already accelerated [drug discovery] to the limit. So unless we see some regulatory changes, it will be very difficult to do it faster.”  

Zhavoronkov shared several examples demonstrating generative AI's capabilities in drug discovery. In a 2019 Nature Biotechnology paper (DOI: 10.1038/s41587-019-0224-x), the company used generative tensorial reinforcement learning (GENTRL) to design molecules targeting DDR kinase. Within 21 days, they designed six molecules, four of which worked in binding studies, with one reaching animal testing in just 46 days. 

Another experiment addressed the challenge of designing compounds for "dark targets"—proteins without crystal structures. Using an early version of AlphaFold with molecular dynamics simulations, they developed compounds with 180 nanomolar potency within 50 days and published the results in Chemical Science (DOI: 10.1039/D2SC05709C). 

While they do not do drug repurposing “in house”, Zhavoronkov said, Insilico has explored drug repurposing applications to support other groups. Collaborating with the ALS Consortium, Insilico identified novel targets and repurposing candidates that were subsequently validated in fly models. One of these candidates was advanced to the equivalent of a Phase 2 investigator-initiated trial by 4B Therapeutics, demonstrating promising results within two and a half years from initial discovery. The manuscript is being drafted now.  

The Aging Emphasis 

A cornerstone of Insilico's approach is their "life models" technology. Like AI “world models,” which posit that you don't need to have a perfect understanding of the environment in order for agents to perform useful work, “life models” aim to understand biological processes "from cradle to grave." These models integrate multiple data types, including methylation, transcriptomics, and proteomics. “We realize that one universal feature that can unify those data types is age, because everybody in this room has age,” Zhavoronkov said.  

“We call it Precious because it's one model to rule the board, that can perform multiple tasks, that can predict age, generate biological data with ages and generations initially, annotate the results of omics experiments, etc.”  

The latest version, Precious3, is an open-source multimodal transformer trained on methylation, transcriptomics, proteomics, and text that has been made available to the research community. 

To validate these AI models, Insilico built a highly automated robotics laboratory in Suzhou during the COVID pandemic. This facility implements a complete reinforcement learning loop where AI predictions are tested experimentally, with results feeding back to improve the models. Zhavoronkov revealed they are now training humanoid robots to perform laboratory tasks that weren't designed for automation.