AI will not cure all diseases without a verification layer.

AI can increasingly generate interventions faster than biology can validate them. Indeed, the cost of producing another therapeutic hypothesis or intervention in many modalities could soon approach zero. But this will not lead to cures until something can answer:

“What will this intervention actually do to a heterogeneous human biological system over time?”
‍

The missing piece is a scalable environment where models can test predictions against living human biology.

Public datasets, virtual cells, 2D assays, organoids, animals, and clinical trials are each limited in different ways: Clinical trials have exorbitant cost and take significant time, animal trials are costly and getting more expensive/supply constrained while being a weak proxy for humans, organoids are poorly standardized and not linked to organism-level outcomes, 2D assays and virtual cells hold even less signal since they cannot capture the exact multicellular interactions that matter, and public datasets have notorious data gaps.
‍

Polyphron’s bet is that human tissue is the right abstraction layer for this scalable environment; complex enough to contain the biology that matters, but tractable enough to manufacture, perturb, observe, and learn from.

Much of the significant AI progress toward superhuman performance in code generation and math has come from scaling reinforcement learning with verifiable rewards. Until now biology has lacked a human-relevant verification substrate that operates at the clock speed required to run RL and related techniques in a similar manner.

Engineering massively parallel arrays of small units of human tissue that capture the genotypic diversity of human biology and then training models to simulate them allow one to create such a substrate. Crucially, this substrate must combine a physical and an in silico component. The in silico tissue component will operate at or near the clock speed of inference and the physical tissue will operate at the clock speed of biology.

Tissue is the fastest, cheapest, most parallel verification substrate that still contains the biology that matters and we believe the path to biological superintelligence runs through closed-loop interaction with these living and simulated human systems. A sufficiently broad, calibrated physical-plus-simulated tissue system therefore becomes the sensorimotor interface through which an AI learns biology.
‍

A verification layer combining simulated and physical tissues.

We produce ultra-realistic proxies for human biology, both physically in the wet lab and in simulation, that redefine the Pareto-frontier between throughput and predictive validity against in vivo human outcomes. To make this possible, we're training frontier models that autonomously explore, understand, and reproduce human tissue biology. Our system self-improves to refine its own understanding of human biology and produce increasingly realistic human tissues.