They build virtual copies of human biology to create fake-but-realistic patient data so AI can learn from millions of cases that would take decades to collect in real life.

Mantis Biotechnology's synthetic data engine leverages physics-based digital twin simulations of biological systems—particularly cardiovascular models funded by their NIH NHLBI grant—to generate large-scale, statistically robust datasets for machine learning model training. Unlike purely statistical synthetic data generators (e.g., GANs trained on tabular clinical data), Mantis embeds first-principles physics (fluid dynamics, tissue mechanics, electrophysiology) into the generative process, producing data that respects biological constraints and edge-case pathophysiology. This is critical for rare disease modeling and medical device testing, where real-world datasets are prohibitively small or biased. The synthetic data preserves privacy compliance (no real patient data leakage), maintains full data lineage for regulatory traceability, and can simulate device-tissue interactions under conditions that would be unethical or impossible to replicate in clinical trials. The result is ML models that generalize better across patient populations and device configurations, dramatically compressing the R&D cycle from bench to FDA submission.

It's like a flight simulator for biology—instead of crashing real planes to train pilots, you crash virtual hearts to train AI doctors.

They built an AI that reads and writes FDA paperwork so biomedical engineers can ask plain-English questions and get instant, citation-backed answers instead of spending months buried in regulatory documents.

Mantis Biotechnology embeds large language models directly into its unified biomedical data platform to transform the regulatory submission process for medical device companies. The system ingests and indexes the full corpus of a customer's product development data—from CAD specifications and bench testing results to clinical trial datasets and prior FDA submissions—and makes it queryable via natural language. Engineers and regulatory affairs professionals can ask questions like "Show me all biocompatibility test results for devices with similar material composition to our current design" and receive instant, source-cited answers with full data lineage. Beyond querying, the LLM layer assists in drafting regulatory narratives (e.g., 510(k) substantial equivalence arguments, PMA clinical summaries) by automatically cross-referencing internal data against FDA predicate device databases and published clinical literature. The system maintains strict traceability—every generated claim links back to its source data—addressing the critical auditability requirement for regulatory submissions. This dramatically reduces the manual labor of document assembly while improving consistency and reducing human error in high-stakes filings.

It's like having a paralegal who has memorized every FDA filing ever made and can instantly draft your legal brief while showing their homework.

They create virtual versions of thousands of patients and run the clinical trial on computers first, so companies can fix problems in the trial design before spending hundreds of millions on the real thing.

Mantis Biotechnology applies its digital twin platform to clinical trial simulation by constructing patient-level virtual cohorts that model disease progression, treatment response, and adverse event profiles across diverse populations. Using a combination of physics-based biological models and ML trained on historical clinical data, the platform generates virtual patient populations that reflect realistic demographic, genetic, and comorbidity distributions. Trial sponsors can then simulate different trial designs—varying endpoints, inclusion/exclusion criteria, dosing regimens, randomization strategies, and sample sizes—against these virtual cohorts to predict statistical power, expected effect sizes, and failure modes before committing to real-world enrollment. The ML layer continuously learns from completed trials to improve predictive accuracy, while the physics engine ensures that simulated physiological responses remain biologically plausible. This approach is particularly powerful for rare diseases (where real patient recruitment is slow and expensive) and for adaptive trial designs where mid-trial modifications need to be pre-validated. The platform's data lineage capabilities ensure that all simulation assumptions and parameters are fully auditable, which is increasingly important as regulators like the FDA begin accepting in-silico evidence in device and drug submissions.

It's like playing SimCity but for clinical trials—you build the whole city of patients on your computer, watch what breaks, fix the zoning laws, and only then start construction in the real world.