Sinewgrid Cell programs cells with sequence and formulation models.
Sinewgrid Cell is a foundation model family for cell therapy and mRNA medicines. It learns from single-cell measurements, perturbation screens and sequence libraries, then proposes candidates for the wet lab to test. Each result returns to training.
- Cell statePredict how a cell population responds to a signal, a dose or a genetic change.
- mRNA designPropose sequences and rank them for expression and stability before synthesis.
- DeliveryScore lipid nanoparticle formulations for uptake in a chosen cell type.
- ManufacturingModel expansion and activation conditions in cell therapy processes.
Why cells and mRNA need a model.
A cell therapy or mRNA medicine is designed in a space far too large to search by hand. An mRNA construct alone has a cap, two untranslated regions, a coding sequence and a tail, and each part trades off against the others: a codon choice that raises expression can weaken stability, and a sequence that works in one cell type can fail in another. Add the delivery formulation, the target cell type and the manufacturing conditions, and the number of combinations a lab could test is many times what it can afford to make. Sinewgrid Cell is built to narrow that space. It learns from single-cell measurements, perturbation screens and sequence libraries, predicts how a cell population will respond, and ranks candidate sequences and formulations before anything is synthesized, so that lab time goes to the few candidates with the best case.
We work in four programs: cell-state transitions, mRNA sequence design, delivery formulations and manufacturing conditions. The model is only as good as the loop around it. Public single-cell atlases are candidate sources for pretraining, each subject to a license review. A partner’s assay data stays in the partner’s own tenant, under the partner’s keys, and every candidate the lab tests becomes a labeled example for the next round of training. The compute follows the experiment calendar: pretraining is a long, steady run, while fine-tuning and ranking arrive in bursts after each assay batch, so we plan steady capacity for the first and on-demand capacity for the second. Sequence-design outputs are screened against biosecurity risk lists before release. This page describes a program in development, and it reports no results.
| Program | What the model predicts | Validation | Stage |
|---|---|---|---|
| Cell-state transitions | Which signals move a stem cell toward a target cell type. | Perturbation screens in partner labs | Planned Sequence 1 |
| mRNA sequence design | Expression and half-life of a candidate sequence. | In vitro expression assays | Planned Sequence 2 |
| LNP formulation | Uptake of a lipid mix in a target cell type. | Delivery assays | Planned Sequence 3 |
| Cell therapy process | Yield and quality under expansion and activation conditions. | Bench-scale runs | Planned Sequence 4 |
What the models learn from.
Order shows the intended sequence of work, not dates. Dates are set with the first design partner.
Single-cell and perturbation atlases
Large public single-cell resources, such as the Human Cell Atlas and Tabula Sapiens, are candidate pretraining sources, subject to a license review for each.
Assays from partner labs
Expression, half-life and delivery measurements for the constructs a partner cares about. Partner data stays in the partner's tenant, under the partner's keys.
Results that return to training
Every candidate the lab tests produces a labeled example. Those results become the next training set, which is the closed loop.
From a target to a tested candidate.
- Define the targetA cell type, a protein, a delivery route and the assay that will judge it.
- Train and proposeThe model ranks candidate sequences or formulations with an uncertainty estimate.
- Test in the labThe partner lab synthesizes and tests a small batch chosen to be informative.
- Return and retrainMeasured results update the model before the next batch.
Bursty, and tied to the experiment calendar
Pretraining is a steady run. Fine-tuning and ranking come in bursts after each assay batch. We buy steady capacity for the first and on-demand capacity for the second.
See the Cell reference designScreened before release
Sequence-design outputs are screened against biosecurity risk lists, access is tiered, and capability evaluations are reviewed by outside experts before any release.
Read the commitmentsPlanning a training run?
Tell us the model, the data and the schedule. We reply with a capacity plan and the evidence behind it.