Insilico Medicine, a U.S.-based, Hong Kong–listed AI drug discovery company, is launching a new service that will train general-purpose large language models, like OpenAI’s GPT or Alibaba’s Qwen, to handle biology and chemistry tasks.
Generalist models “fail miserably” at the benchmarks used to measure how AI performs scientific tasks, Alex Zhavoronkov, Insilico’s founder and CEO, told Fortune. “You test it five times at the same task, and you can see that it’s so far from state of the art … It’s basically worse than random. It’s complete garbage.”
Far better are specialist AI models that are trained directly on chemistry or biology data. But these models often don’t allow a user to prompt them in plain language, the way someone can with the general purpose models, and they also lack the ability to complete tasks beyond specialized scientific functions.
Enter Insilico’s new “Science MMAI Gym,” designed to turn a generalist large language model into something that can perform as well as specialist models.
The gym is a pivot for Insilico, which calls it part of its “long-term road map toward pharmaceutical superintelligence.” The startup is part of a group of biotech companies trying to use machine learning and artificial intelligence to research and devise new drugs. But with the “gym,” Insilico is now targeting other biotech and pharmaceutical companies, offering to train new AI models for them.
Insilico will train models using a mix of domain-specific datasets, reward models, and reinforcement learning, and claims this process can improve model performance by up to 10 times against key benchmarks in chemistry and biology, and even approach the performance of models specifically designed for these scientific tasks.
But why would a company decide to train a general model, as opposed to using a specialist one? The reason is flexibility: A specialist model is very good at one thing—say, drug discovery—but can’t do other things; in contrast, a trained generalist model, even if it can’t quite match the performance of a specialist model, can maintain its ability to conduct many other tasks. That means a startup can rely on just one large model, as opposed to an array of specialist models.
“If the model is small, it starts forgetting some of the more primitive tasks that it was designed for,” Zhavoronkov says. “If the model is large, you don’t have that problem.”
Zhavoronkov admits that even generalist models that make it through Insilico’s gym still won’t perform as well as the best state-of-the-art specialized models. “For them to be able to reason in terms of molecular simulations, they need to understand and see the physics. The language is not really designed for that, so they’ll suck a little bit compared to frontier physics-based models,” he explains, though he expects that to improve in the next few years.
Yet as LLMs become more widespread—and as more startups adopt them—Zhavoronkov says he wants Insilico to become the “No. 1 trainer of those models.” Insilico has already been in conversation with potential clients about the training program, he says; while he didn’t share specific names, he said he reached out to “top frontier players in the U.S.”
Insilico, Hong Kong, and biotech
Founded in 2014, Insilico is rushing to be one of the first startups to get a wholly AI-designed drug through clinical trials and onto the market. One of the startup’s main efforts is a drug to tackle idiopathic pulmonary fibrosis, a condition where scar tissue forms in the lungs, making breathing difficult. The startup said it managed to get its drug to clinical trials in just 18 months, far shorter than the average of four years for more traditional biotech companies. Last year, the drug finished Phase II clinical trials, with researchers concluding that the results warranted “further investigation in larger-scale clinical trials of longer duration.”
Insilico is also targeting other conditions, like inflammatory bowel disease, as well as looking into new cancer and GLP-1 drugs.
In December, Insilico raised 2.3 billion Hong Kong dollars ($295 million) in its IPO, the largest biotech debut in the Chinese city in 2025. The IPO attracted companies like Eli Lilly, Tencent, and Oaktree as cornerstone investors.
The startup’s shares have skyrocketed since their trading debut on the Hong Kong Stock Exchange on Dec. 30. At 54.75 Hong Kong dollars ($7.02), as of Jan. 16, Insilico’s shares are now worth more than double their IPO offer price of 24.05 Hong Kong dollars ($3.08).
The Hang Seng Biotech Index, which tracks the 30 largest biotech companies listed in Hong Kong, has risen by 100% over the past 12 months, far ahead of the 37% gain recorded by the benchmark Hang Seng Index.
Insilico isn’t the only AI startup whose HK-listed shares have surged in recent weeks. Shares in MiniMax, a Chinese consumer AI startup, have risen by 160% since they started trading on Jan. 9. Chip designer Biren is also up by over 90% from its IPO offer price.
Still, investors in both the U.S. and in China are wondering whether the AI boom can last. While Zhavoronkov is keeping an eye on the possibility of an AI bubble forming in stock markets, he’s optimistic that AI drug discovery will be safer from a bursting bubble than other industries. “People can live without a conversational assistant, or AI-generated movies. But they cannot live without drugs.”
