The computer-powered molecule designer Insilico Medicine has been no stranger to generative AI—indeed, it was demonstrating that its networked methods could deliver gains in drug discovery well before programs like ChatGPT entered the international conversation—but now the company is taking things a step further with a series of initiatives that aim to make good on AI’s promises for the life sciences industry.
That includes updates to its established Pharma.AI platform, geared toward drugmakers—as well as a new AI assistant that will aid in writing up research papers, with a free trial version set to become available later this year.
Named DORA, the manuscript helper taps multiple AI and large language models to support the drafting of academic papers and case studies, in addition to applications for grants and patents. Insilico’s goal is to provide something of a digital wing man for early-career researchers—as well as non-native English speakers—who may be looking to contribute to the broader corpus of medical research.
“The reason why we’re doing this is to help enable people who, potentially, may not have exactly the same level of research capabilities or support as people in top-tier institutions,” Insilico founder and CEO Alex Zhavoronkov said in an interview. “For example, medical doctors in China—or medical doctors in general, who very often do not have the time to spend on research.”
“But,” Zhavoronkov said, “they can formulate a hypothesis. If they saw a case, they can annotate that, and then build that up in a consistent manner. Now they can unlock some of their knowledge by publishing or doing a preprint, or even just doing a blog post.”
DORA wouldn’t necessarily work the same way as other generative AI programs aimed at instant gratification—it’s not simply “enter prompt, get paper.”
By incorporating multiple LLMs and hundreds of interconnected agents, the overarching application takes time to go through several cycles of reinforcement learning and referencing while starting with the researchers’ submitted data as its bedrock. And, of course, human proofreading would still be necessary.
“In this case, we are not trying to give a snappy answer,” he said. “We are trying to prioritize accuracy, and there will be multiple rounds of generation and review by these agents. We cannot prevent hallucinations completely, but since we’re using multiple LLMs, they need to agree. And they usually agree to something as close to ground truth as possible.”
At the same time, concerns about data privacy and cybersecurity are top-of-mind among many healthcare providers and researchers—especially when it would come to submitting proprietary and protected patient information to a cloud-based AI. So Insilico is also wading into the hardware business, with plans to sell its analysis engines on fully offline servers.
The PandaOmics Box—sold without Wi-Fi or web connections—aims to provide secure, localized number-crunching. “So once a month, you take your USB stick, get the updates from the Internet, and then put this USB stick into the server and update the records,” Zhavoronkov said.
Finally, Insilico is also launching an open-source AI program with an eye toward research into aging and its related diseases. Precious-3 GPT is a multimodal model designed to ingest hard clinical data—such as compound structures, or results from various DNA and omics-based tests—as well as text-based reports and published research.
From there, it aims to explore virtual mouse, rat, monkey and human models, and work to predict factors such as drug sensitivity and cellular interactions—while accounting for how the body’s biology ages over time, down to the molecular level.
“The idea there was to try to find the relationships between the different preclinical models and humans,” Zhavoronkov said.
“You could potentially just talk to this model and say, ‘OK, which targets are implicated in aging and osteoarthritis of the knee in humans and monkeys? Show me those targets.’” he said. “Those are the types of questions that can be answered using this particular tool.”