Prodius

A design engine that learns from the bench.

Four layers: public data and open models for breadth, a recipe model to propose, and an automated lab that tests the proposals and feeds the results back. This page describes what we are building.

The problem it is built for

Every design choice below answers one of these.

  • Protocols are found by hand

    Growing functional tissue is still mostly trial and error. A single differentiation protocol can take months or years of manual experiments to optimize. [11]

  • The search space is enormous

    Cell source, scaffold chemistry, stiffness, porosity, growth factors, dose, timing, oxygen and culture format all interact. One retinal culture protocol alone had about 200 million candidate condition combinations. [11]

  • The data is small and scattered

    Tissue outcomes sit in thousands of papers, recorded in incompatible ways. No general predictive model of tissue formation exists yet. [10]

  • Pharma needs models it can trust

    Regulators are moving away from mandatory animal testing, but organoids are still hard to reproduce from batch to batch and lab to lab. [18, 13]

Architecture

Read it bottom-up: data feeds models, models feed the recipe engine, the engine sends recipes to the lab, and the lab's results go back to the data.

  1. Layer 4

    Closed-loop lab

    Automated hardware that runs the recipes the engine proposes and measures what grows.

    Hardware

    • Automated liquid handling (Opentrons or equivalent)
    • Bioreactors
    • Organ-on-chip platforms

    Readouts

    • High-content imaging
    • Single-cell RNA-seq
    • Spatial transcriptomics (e.g. 10x Visium, Xenium)
    • Mechanical and functional assays

  2. Layer 3

    Recipe design engine

    The model that turns a target tissue into ranked recipes and chooses which experiments to run next.

    Model

    • Multimodal model over cell state, material properties and protocol parameters

    Experiment selection

    • Bayesian optimization (BoTorch, Ax) to pick the most informative next runs

  3. Layer 2

    Foundation models

    Open models that already understand proteins, cells and tissue neighborhoods, used as building blocks.

    Protein

    • Biohub ESMC and ESMFold2 (MIT-licensed), for growth factors, ECM-binding peptides and self-assembling peptide scaffolds

    Cell

    • Geneformer
    • scGPT
    • scFoundation
    • Universal Cell Embeddings

    Spatial and tissue

    • Nicheformer and spatial transcriptomics models for cell neighborhoods

  4. Layer 1

    Data foundation

    Public atlases for breadth, and our own standardized experiment data for depth.

    Cell atlases

    • CZ CELLxGENE (Census)
    • Human Cell Atlas
    • Tabula Sapiens
    • HuBMAP
    • Human Protein Atlas

    Protein data

    • UniProt
    • Protein Data Bank
    • ESM Atlas (Biohub, open data on AWS)

    Proprietary

    • Standardized experiment data from our own closed-loop runs
Layer 1 sits at the bottom. Each closed-loop run adds to the proprietary data in layer 1, which is what makes the next prediction better.

Data

Public atlases give breadth: what human cell types look like, which genes they express, and how proteins are built. We draw on CZ CELLxGENE Census, the Human Cell Atlas, Tabula Sapiens, HuBMAP, the Human Protein Atlas, UniProt, the Protein Data Bank and the ESM Atlas.

Atlases describe cells. They do not record how a tissue was grown. That data is small, scattered across papers and recorded inconsistently. So we generate our own: every closed-loop run produces standardized data in which recipe, conditions and outcome are linked. Over time, that dataset is the moat. [10, 4]

Models

We build on open foundation models rather than starting from scratch. Protein models (Biohub ESMC and ESMFold2) help design growth factors, ECM-binding peptides and self-assembling peptide scaffolds. Cell models (Geneformer, scGPT, scFoundation, Universal Cell Embeddings) represent cell state. Spatial models such as Nicheformer represent cell neighborhoods and tissue organization.

On top of them, we are building a multimodal recipe model that combines cell state, material properties and protocol parameters, and predicts which recipes are most likely to produce the target tissue. [4, 5, 6, 7, 9, 8]

Closed loop

A model trained only on literature will be wrong in ways it cannot see. The loop fixes that. Bayesian optimization (BoTorch and Ax) chooses the experiments that will teach the model the most, not just the ones it already rates highly.

Those experiments run on automated liquid handling, bioreactors and organ-on-chip platforms. They are read out by high-content imaging, single-cell RNA-seq, spatial transcriptomics (such as 10x Visium or Xenium), and mechanical and functional assays. The same loop design has worked for a single cell type in a third-party study. We are building it to generalize across tissues. [11, 14]

The Prodius closed loop1. Define the target. 2. Predict the recipe. 3. Test automatically. 4. Learn and improve. The loop then returns to the first step with a better model.1Define2Predict3Test4LearnEvery experimentimproves the next
  1. 1

    Define the target. Tissue type, cell composition, mechanics, function.

  2. 2

    Predict the recipe. Ranked recipes with confidence.

  3. 3

    Test automatically. Imaging, scRNA-seq, spatial, function.

  4. 4

    Learn and improve. Bayesian optimization picks the next runs.

Software stack

  • Python
  • PyTorch
  • scverse (Scanpy, Squidpy)
  • Cloud GPU inference
Partner with usRead the science