Example · Company: Recursion — industrialising phenotypic screening
Example post. A worked example of the company template, kept to show the format — replace it with your own notes.
What they sell
Nominally, drugs: a pipeline of internal clinical programs. Practically, for now, the platform — large pharma partnerships (Bayer, Roche/Genentech, Sanofi) that pay upfront and milestone fees for access to Recursion’s automated phenomics engine: robots perturb cells at enormous scale, microscopes image them, and ML models map which compounds and genes produce similar cellular phenotypes.
Why now
Three curves crossed: lab automation got cheap enough to run millions of experiments a week, GPU compute made it possible to learn from petabytes of cell images, and a decade of genetic screening data (CRISPR knockouts, Perturb-seq) gave supervision signal that didn’t exist before. The 2023–24 wave of foundation-model enthusiasm also reopened the funding window for capital-intensive platform stories — Recursion raised from NVIDIA and merged with Exscientia partly on that narrative.
Moat
The dataset is the honest answer: tens of petabytes of consistently generated, perturbation-labelled cell images that can’t be scraped or bought, plus the wet-lab/software integration to keep growing it. The counterargument is that a moat of data only matters if phenotypic similarity in HUVEC-type cell images actually translates into clinical success — a mapping nobody has proven yet.
Risks
- Translation risk: phenotypic hits in immortalised cell lines may not become differentiated drugs; several early internal programs have already been deprioritised after unimpressive readouts.
- Burn: the platform costs hundreds of millions a year; if partnerships slow before internal assets mature, the model inverts from “platform funds pipeline” to “pipeline hopes for platform”.
- Commoditisation: open-source perturbation atlases and cheaper imaging could erode the data advantage faster than expected.
What I’d want to know as an investor
What fraction of platform-nominated candidates survive into IND-enabling studies versus an industry baseline; the real unit economics of one phenomics “experiment”; how much partner revenue is recurring versus one-off upfronts; and whether any program in the clinic traces causally to a model prediction that a conventional screen would have missed.