Powered by massive datasets, Large Language Models (LLMs) have the potential to effectively accelerate genomics research, streamline clinical documentation, optimize real-time diagnostic workflows, support clinical decision-making, expedite drug discovery processes, and even generate synthetic data to facilitate experimental research.
However, as these models transform the landscape of biomedical research, they frequently encounter significant bottlenecks: beyond the structured data typically relied upon in healthcare, they often struggle to demonstrate efficacy in "edge cases"—such as rare or unusual diseases—primarily due to a critical lack of reliable and representative supporting data in these specific scenarios.
Mantis Biotech, a company headquartered in New York, USA, states that it is dedicated to developing a solution to bridge this gap in data availability. The platform developed by the company integrates data from diverse sources to generate synthetic datasets capable of constructing "digital twins" of the human body—models grounded in physics principles that can predict human anatomical structures, physiological functions, and behavioral patterns.
Mantis Biotech is actively championing this digital twin technology, applying it specifically to the fields of data aggregation and analysis. These digital twins offer broad utility, ranging from the research and testing of novel medical procedures and the training of surgical robots, to the simulation and prediction of medical conditions and even behavioral patterns. For instance, in a recent media interview, Mantis founder and CEO Georgia Witchel explained how a sports team could utilize this technology to predict the likelihood of an Achilles tendon injury in a specific NFL player, based on factors such as their recent performance, training load, dietary habits, and years of professional experience. To construct these digital twin models of the human body, Mantis’s platform first gathers data from a variety of sources—including textbooks, motion-capture cameras, biometric sensors, training logs, and medical imagery. Subsequently, the platform employs an LLM-based system to route, validate, and synthesize these diverse data streams. All this information is then fed into a physics engine to generate high-fidelity renderings of the dataset—renderings that can subsequently be used to train predictive models.
“We are able to gather all these disparate data sources and transform them into predictive models that forecast human performance. Consequently, whenever you wish to predict an individual’s performance, it represents an ideal application scenario for our technology,” Witchel stated.
In an interview with the media, Witchel explained that the physics engine layer constitutes the critical core of the entire technology. It enhances the realism of the generated synthetic data; by precisely simulating the physical properties of human anatomy, it helps the platform further augment the practical utility of the available information.
“If I were to ask you to estimate the hand pose of someone missing a finger, it would be an incredibly difficult task—simply because there are currently no publicly available, annotated datasets of hand positions for individuals with missing digits. We can generate such a dataset with great ease; we merely need to utilize our physics model, issue a command—such as ‘remove finger X’—and regenerate the model,” she noted.
Given that Mantis’s platform successfully bridges these gaps in data availability, Witchel believes the technology holds immense potential for widespread application within the biomedical sector. In the biomedical field, information regarding surgical procedures or specific patients is often difficult to acquire—frequently lacking structure or remaining fragmented across disparate data sources. She placed particular emphasis on the realm of special cases or rare diseases, noting that ethical and regulatory constraints often impede the inclusion of such patient data in public datasets or its utilization for training AI models—factors that significantly exacerbate the difficulty of obtaining such data.
“You know that scene where a three-year-old child runs around clutching a Barbie doll—grabbing it by the leg and slamming it down onto a table? I hope people will approach our digital twin models with that very same mindset,” she said. “I believe this will lead people to accept the idea that, when utilizing virtual avatars, it is permissible to conduct experiments using real human subjects. I feel that the prevailing mindset right now is quite the opposite—which is entirely understandable, as people’s privacy deserves to be respected. In fact, I maintain that people’s data should never be misused, particularly when digital twin models are involved.”
To date, Mantis Biotech has achieved considerable success within the realm of professional sports—a likely outcome given the market’s urgent need for sophisticated modeling and analysis of elite athletes. Witchel revealed that one of the startup’s key clients is an NBA team.
She explained: “We create digital models of athletes—models that essentially visualize an athlete’s jumping performance; and not just their performance today, but every single day over the past year. Furthermore, these models illustrate how their jumping mechanics fluctuate in response to factors such as increased sleep duration or the frequency with which they lift weights.”
The startup recently successfully closed a $7.4 million seed funding round. Led by Decibel VC, the round saw participation from Y Combinator, several angel investors, and Liquid 2. Reportedly, these funds will be allocated toward team expansion, advertising, marketing, and promotional initiatives.
Witchel stated that Mantis Biotech’s next strategic step involves continuously refining its technological framework with the ultimate goal of launching the platform to the general public, with a core focus on the field of preventive healthcare. Concurrently, the company is actively expanding its service offerings to provide support to pharmaceutical laboratories and researchers engaged in FDA clinical trials, aiming to furnish them with in-depth analytical reports regarding patient responses to various treatments.
