Large language models trained on extensive datasets can expedite genomics research, simplify clinical documentation, enhance real-time diagnostics, aid clinical decision-making, quicken drug discovery, and even produce synthetic data for experimental progress.
However, their potential to revolutionize biomedical research often meets a bottleneck: these models struggle with edge cases like rare diseases and atypical conditions due to a lack of reliable, representative data beyond structured healthcare data.
New York-based Mantis Biotech claims to bridge this data availability gap. The company’s platform combines different data sources to create synthetic datasets that can build “digital twins” of the human body: predictive models of anatomy, physiology, and behavior.
These digital twins are proposed for data aggregation and analysis, potentially to study and test new medical procedures, train surgical robots, and simulate or predict medical issues or even behavior patterns. For instance, a sports team could anticipate the likelihood of a specific NFL player developing an Achilles heel injury based on recent performance, training load, diet, and tenure, explained Mantis’ founder and CEO Georgia Witchel in a TechCrunch interview.
To create these twins, Mantis’ platform gathers data from sources like textbooks, motion capture cameras, biometric sensors, training logs, and medical imaging. An LLM-based system routes, validates, and synthesizes the inputs, and a physics engine generates high-fidelity renders of this dataset to train predictive models.
“We transform disparate data sources into predictive models of human performance. Any prediction of human performance is a good use of our technology,” Witchel stated.
The physics engine is crucial, Witchel told TechCrunch, by grounding the generated synthetic data and accurately modeling the physics of anatomy.
“If asked to do hand-pose estimation for someone missing a finger, it’s tough without public datasets of labeled hand positions for such cases. We can easily create those datasets by adjusting our physics model,” she said.
With Mantis’ platform bridging data source gaps, Witchel sees potential for widespread use in the biomedical sector, where procedure or patient information is often difficult to access, unstructured, or siloed. She highlighted edge cases or rare diseases where data is hard to gather due to ethical and regulatory constraints on using patients’ data.
“Think of a three-year-old with a Barbie, smashing it on the table—apply that mindset to our digital twins,” she said. “I want to shift the idea to testing humans virtually, while respecting privacy and not exploiting data.”
Currently, Mantis is successful in professional sports, modeling high-performing athletes. One major client is an NBA team.
“We create digital representations showing athletes’ jumps over time, linked to sleep or arm lift frequency,” Witchel explained.
Mantis recently raised $7.4 million in seed funding led by Decibel VC, with participation from Y Combinator, angel investors, and Liquid 2, funding hiring, advertising, and market functions.
Mantis plans to expand the tech and eventually release the platform for public preventive healthcare use. The company is also aiming to cater to pharmaceutical labs and researchers involved in FDA trials, providing insights into patient treatment responses.