AlphaFold had the Protein Data Bank. The virtual cell now has consortia generating one — $500M+ in newly committed funding (CZ Biohub Virtual Biology Initiative, Apr 2026) building the open foundation atlas of cellular response.

What's still missing: mechanism on proprietary compounds, in the cell types pharma actually screens against. Open data cannot ingest pharma IP. Internal hERG and CiPA flag risk; they don't deliver mechanism. Generic CROs execute; they don't iterate.

Perturb is built to answer compound-specific cardiac safety questions for pharma drug-safety teams. Each engagement runs as a closed-loop research program in human iPSC-cardiomyocytes: AI-orchestrated experimental design, cloud-lab execution (Arctoris), multimodal mechanism profiling (Cell Painting + DRUG-seq + electrophysiology), active learning across 2–3 iterations within a 4–8 week window. NDA-protected.

The same engine extends to any molecular biology question that needs an answer fast.

What hERG can't tell you

Mechanism. Off-target liabilities. Pathway-level disruption. Multimodal readouts on your compound in iPSC-cardiomyocytes — not just a binary risk flag.

What generic CROs can't do

Iterate. Each round of data trains the model that picks the next round. 2–3 active-learning cycles inside a 4–8 week engagement.

What open data won't have

Your compound. NDA-protected from intake to delivery. Methods accrue to the engine; compound-specific findings stay yours.

Calibration POC for pharma cardiac safety
How it works
Week 0
Intake
NDA + scope. Compound, target endpoints, and the question handed off.
→
Active learning
2–3 cycles
Each round of data trains the model that picks the next round.
→
Week 4–8
Delivery
Mechanism profile on your compound. The cardiac-safety questions from intake, answered.
Inside each cycle
01 · DESIGN
AI selects
Next round's experimental design: perturbations, concentrations, conditions.
02 · EXECUTE
Arctoris runs
Plate runs end-to-end on the cloud lab. No manual handoffs.
03 · MEASURE
Multimodal
Cell Painting, DRUG-seq, and electrophysiology, co-measured.
04 · ANALYZE
Model updates
Each round of data trains the model that picks the next.
↻ Output of cycle N is input to cycle N+1. Most engagements run 2–3 cycles.

Each iteration compounds. The model learns which experiments tell you the most about your compound and picks the next round accordingly.

Founded by

Daniel Reda — two prior exits in life science data: CureTogether (acquired by 23andMe) and Redasoft (acquired by Hitachi). Background in Molecular Genetics.

Scientific Advisory Board forming.

Get in touch
daniel@perturb.bio