Remember when IBM Watson was going to cure cancer? Around 2013, the tech giant partnered with Memorial Sloan Kettering, promising that machine learning would transform oncology by ingesting vast medical literature and patient records. The initiative quietly wound down years later, not because the AI wasn’t sophisticated enough, but because real-world medical data proved messier, sparser, and more fragmented than anyone anticipated. The problem wasn’t computational power—it was data availability and quality.
Now Mantis Biotech is taking a fundamentally different approach to this same challenge: instead of waiting for perfect datasets that may never materialize, they’re building digital twins of human biology to generate the data medicine desperately needs.
The Data Scarcity Paradox
Medical AI faces a peculiar contradiction. We generate enormous volumes of health data—electronic health records, genomic sequences, imaging studies—yet for any specific research question, usable data remains scarce. A rare disease might affect thousands globally, but getting standardized, longitudinal data from even a hundred patients proves nearly impossible. Privacy regulations, institutional silos, and inconsistent data collection create what I call “data deserts within data oceans.”
Traditional approaches try to solve this through data aggregation: federated learning, privacy-preserving computation, multi-institutional consortia. These help at the margins but don’t address the fundamental constraint that certain experiments simply cannot be run on human subjects, and certain patient populations will always be too small for statistical significance.
Digital Twins as Generative Models
Mantis Biotech’s digital twin approach represents a category shift in how we think about medical data. Rather than treating data scarcity as a collection problem, they’re framing it as a modeling problem. The core insight: if you can build sufficiently accurate computational models of human biological systems, you can generate synthetic data that captures the statistical properties and causal relationships of real patient populations.
This isn’t about creating simple statistical simulators. Modern digital twins integrate multiple modeling paradigms—mechanistic models of cellular processes, pharmacokinetic simulations, machine learning components trained on real patient data, and increasingly, agent-based models that capture individual variability. The goal is to create what amounts to a generative model of human physiology that respects known biological constraints while producing realistic variation.
The Validation Challenge
The critical question for any synthetic data approach: how do you validate that your digital twins actually reflect reality? This is where Mantis’s work gets technically interesting. You can’t simply compare synthetic outputs to real patient data—if you had enough real data for solid comparison, you wouldn’t need synthetic data in the first place.
Instead, validation requires a multi-layered approach. First, ensure that known biological relationships hold in the synthetic data—drug interactions, disease progressions, genetic associations. Second, test whether models trained on synthetic data generalize to real patients in prospective studies. Third, use the digital twins to make predictions about edge cases or rare scenarios, then validate those predictions as real-world data becomes available.
Where This Actually Helps
Digital twins won’t replace clinical trials or eliminate the need for real patient data. But they can address specific bottlenecks in medical research and drug development.
For rare diseases, where patient populations are inherently small, synthetic patients can help explore treatment protocols and identify promising drug candidates before committing to expensive trials. For personalized medicine, digital twins could simulate how a specific patient might respond to different treatments based on their genetic profile and medical history. For drug safety, synthetic populations can help identify potential adverse events in demographic groups underrepresented in clinical trials.
The recent news about AI helping solve labor issues in rare disease treatment connects directly to this. When you’re dealing with conditions that affect hundreds rather than millions, every efficiency gain in research and treatment development matters enormously.
The Architecture Implications
From an AI architecture perspective, medical digital twins represent a fascinating hybrid system. They combine physics-based simulation, causal modeling, and modern deep learning in ways that challenge our typical categorizations. The system needs to be interpretable enough that clinicians can understand and trust its outputs, yet flexible enough to capture the complexity of human biology.
This pushes us toward modular architectures where different components handle different aspects of biological modeling, with careful attention to how uncertainty propagates through the system. A digital twin that confidently produces wrong predictions is worse than useless—it’s dangerous.
The real test for Mantis and similar efforts will come in the next few years as these systems move from research tools to actual clinical decision support. The technology is promising, but medicine has seen many promising technologies fail at the implementation stage. The difference this time might be that we’re finally matching the right computational approach to the right problem: not trying to replace human judgment, but filling in the data gaps that have always limited it.
đź•’ Published: