Digital twins of human cells.
Twincyte builds AI models that predict how human cells respond to genetic perturbations, including in cell types where no perturbation has ever been measured. Run thousands of knockdowns in silico, then take the best few to the bench.
Taking part in the Arc Institute Virtual Cell Challenge 2026: zero-shot prediction across cell lines.
The problem
Biology has far more experiments than labs can run
The effect of turning a gene down depends on the cell it happens in. Measuring every gene in every relevant cell type is out of reach, so most decisions rest on data from the wrong cell type, or on no data at all.
Protein-coding genes
Each one can be knocked down, alone or in combination.
Cell types and states
The same knockdown can do very different things in different tissues.
Per screen
Single-cell CRISPR screens are powerful but slow and costly, and each covers one cell context at a time.
Product
Twincyte Virtual Cell
A predictive model of the cell you care about, built from nothing more than its unperturbed profile.
Zero-shot response prediction
Provide non-targeting control profiles of a cell line and get the predicted expression shift for any gene knockdown, with no perturbation data from that cell line required.
Transcriptome-wide readout
Predictions cover the whole transcriptome, with differentially expressed genes ranked by effect size and direction.
In-silico screens
Score thousands of knockdowns against a target signature, such as reversing a disease state, and get a ranked shortlist.
Confidence you can act on
Every prediction carries an uncertainty estimate, so you know which hypotheses deserve a wet-lab test.
Delivered as a web workspace and a Python API. Early access opens to academic labs first.
Approach
How it works
Generalizing to unseen cell types is the hard part. Our models are designed and evaluated for exactly that.
Learn from public perturbation data
Pretrain on large public single-cell atlases and CRISPR perturbation screens across many cell lines.
Represent genes with prior knowledge
Describe every gene with sequence-, pathway- and literature-derived features, so the model can reason about knockdowns it has never observed.
Condition on the cell's baseline
Characterize a new cell context only by its control profiles, and predict the change rather than just the state.
Benchmark on held-out cell lines
Score models only on cell lines and genes they never saw, with metrics that reward recovering the right differentially expressed genes, not just matching the average cell.
Use cases
Built for R&D teams
For computational biologists and drug-discovery teams who need to decide which experiments are worth running.
Target discovery
Rank knockdowns that push cells toward a desired state before committing to a screen.
Screen design
Choose which perturbations and cell lines to measure next, where a real experiment will teach the most.
Context transfer
Carry results from a screened cell line over to the cell type you actually care about.
Mechanism hypotheses
Match predicted knockdown signatures against compound and disease signatures to propose mechanisms of action.
Business model
Software that scales with your research
A free tier for academic research, usage-based pricing for companies, and private models for teams with their own perturbation data.
Free research tier
A monthly prediction quota for non-commercial research.
Usage-based API
Pay per prediction, with a shared workspace for biotech R&D teams.
Private models
Models fine-tuned on a customer's own perturbation data and deployed in their own cloud.
Pricing will be published with the public beta.
Roadmap
Where we are
Twincyte is at the pre-seed stage.
- Now · Q4 2026
First models and evaluation pipeline. Entry in the Arc Institute Virtual Cell Challenge 2026 (final submissions due November 5, 2026).
- Q1 2027
Private beta. Prediction API and workspace for academic labs.
- 2027
Beyond single knockdowns. Chemical perturbations and gene combinations; first biotech design partners.
Team
Founder
Maverick
Founder
Trained in pharmacy (B.S.), then spent more than ten years as a software and data/algorithm engineer in the tech industry. Started Twincyte to make predictive models of cell biology practical for everyday research.
Contact
Work with us
We're looking for early-access users, and for collaborators with perturbation data who want to see how far zero-shot prediction can go.