Introducing CT-Perturb: screening beyond the lab. A multimodal language model over tokenized biology that predicts how cells respond to compounds, gene knockouts, cytokines, and combinations. Here’s how it works ↓

Been waiting a long time to share this one. Proud of the team.

First question from every pharma team: does our data leave the building? Nope.

Still love how simple the core idea is. Cell as text, perturbation in the prompt, model tells you what changes

The silmitasertib result is still my favorite part of this work. It only boosted antigen presentation when interferon was around, which is the kind of thing a normal screen would never catch. Seeing that result hold up in the lab was especially satisfying.

A compound’s effect depends on the cell, dose, timing, and biological context. Add combinations and the number of experiments grows quickly. CT-Perturb helps explore that space and choose which predictions to test in the lab.

Built on CT1 and our C2S-Scale research, CT-Perturb learns from public and proprietary perturbational data. It encodes the cell, perturbation, and complex biological context to predict how gene expression and pathways change.

The research behind CT-Perturb has already uncovered context-dependent biology. C2S-Scale nominated silmitasertib as an amplifier of antigen presentation—but only with interferon present. Lab experiments confirmed it.

Can it predict combinations it hasn’t seen? On held-out compound–cell pairs, CT-Perturb outperformed all baselines we tested.

Measure a fraction. Predict the rest. With just 10% of the measurements for training, CT-Perturb reached 82% of the full-data model’s prediction score. Use that lab capacity to explore more cells, perturbations, and contexts.

CT-Perturb adapts to your screens inside your own environment—on-prem or in your controlled cloud. Your data stays with you. The model learns your biology. Explore the results and get in touch: