MuseCell

MuseCell - a blinded AI scientist for drug screening Problem: phenotypic drug screens find hits fast, but working out how a hit acts - its mechanism of action - takes a biologist weeks of comparing images, ordering follow-up doses and reading papers. What it does: MuseCell is an agent that runs that loop on real cell images (BBBC021) under a blinded protocol. Muse Spark plans which 4 of ~24 microscope wells to image under a budget, measures every cell, looks at the cells with vision next to untreated controls, and commits to a mechanism without knowing the compound. Muse Wonder reviews the call as the PI and can send it back. A published nearest-neighbour vote over the agent's own wells makes the final call. Only then is the compound unblinded and checked against PubChem, ChEMBL and PubMed, and the run ends in a dossier that cites only papers it actually retrieved. Result: the right mechanism for 34/38 compounds, 18/19 on a held-out half, about $0.29 and 1.5 minutes per compound. A fixed "four highest doses" rule ties it: Muse Spark rediscovered the expert's strategy on its own, and we show that rule next to our number. Research tool for drug-screening scientists, not a medical product: it works on cell-line images from a public screen and never diagnoses patients or gives treatment advice. It is an agent that plans experiments, not an image classifier. Built during the hackathon: everything in the repository (repo created Oct 29, 2026; full commit history inside the event window). Not ours, used under their licenses: BBBC021 images (CC BY 4.0), ImageNet-pretrained ResNet50 weights, scikit-image, and the public PubChem, ChEMBL and PubMed APIs. Live demo (no install): https://aleksandrahodzzik.github.io/musecell/