Why tissue needs its own ESM

Protein language models changed what it means to design a protein. ESMFold showed that a model trained only on sequences learns enough structure to predict a protein's 3D shape from a single sequence (Lin et al., 2023). ESM3 went further and generated a working fluorescent protein far from anything in nature (Hayes et al., 2025).

Nothing comparable exists for tissue. This post explains why, and what it would take to build one.

What a protein model reads

A protein model learns from a simple, standardized object: a string of amino acids. There are hundreds of millions of them, written in the same alphabet, and evolution has already run the experiment of which ones work.

That is the quiet reason these models succeeded. The data was abundant, and it was consistent.

What a tissue model would have to read

To grow a tissue, you choose a cell source, a scaffold and its chemistry, its stiffness and porosity, the mechanical forces applied, the growth factors, their doses and the order and timing in which everything happens. Each of those choices interacts with the others.

We think of that set of choices as a recipe, and of a recipe as a sequence: cells, scaffold, mechanics, signals, timing. Written that way, tissue design looks like a sequence problem a model could learn.

The difficulty is that almost nobody writes recipes down that way.

The data problem

Tissue protocols live in methods sections, supplementary tables and lab notebooks. Units differ, conditions go unreported, and failed recipes are rarely published at all. A model trained on that record would learn the field's reporting habits as much as its biology.

Single-cell atlases help, and the cell foundation models built on them are real progress (Theodoris et al., 2023; Cui et al., 2024). But atlases describe what cells are, not how a tissue was grown. Even for the questions they do address, zero-shot benchmarks show these models do not always beat simpler baselines (Kedzierska et al., 2025).

So the bottleneck is not model architecture. It is the lack of a consistent, growing record that links each recipe to its outcome.

Why a lab has to be attached

If the data does not exist, it has to be generated, and generated in a form a model can learn from. That is the argument for a closed loop: the model proposes recipes, automated hardware runs them, standardized readouts measure what grew, and the results go straight back into training.

There is a published proof of concept. A RIKEN robotic system using Bayesian optimization searched about 200 million possible culture conditions by testing 143, and improved iPSC-derived retinal pigment epithelium production by 88% on pigmentation scores (Kanda et al., 2022). That was one cell type and one objective. The open question is whether the same loop, given a shared model across tissues, gets better at the second tissue because it learned from the first.

Why now

Two things changed. Open foundation models for proteins, cells and spatial data now exist to build on (Biohub, 2026; Tejada-Lapuerta et al., 2025). And demand moved: FDA's 2025 roadmap names organoids, organ-on-chip systems and AI models as alternatives to animal testing (FDA, 2025). That makes reproducible human tissue a requirement for drug developers, not just a research goal.

What we are building

Prodius is building that engine: a recipe model on top of open foundation models, connected to an automated lab that generates standardized data with every run. We will publish what works, and what does not, as we go.

If you have a tissue you cannot grow reliably, we would like to hear about it.