# TransitionBio

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Kristofer Linton-Reid](https://cerebralvalley.ai/u/klintonreid)
- **GitHub:** https://github.com/transitionbio/phasepilot
- **Demo video:** https://youtu.be/RoKwOSsf1Rw
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4201

We built PhasePilot to tackle a problem we run into in the lab: working out how a protein behaves without using up the protein in the process.

At Transition Bio, we study how proteins and RNA assemble into condensates, and how drugs can change that behaviour. These assemblies help organise cells, and their disruption is implicated in cancer and neurodegeneration. To study them properly, we need to understand which conditions make them form and how that changes when we add a compound.

The problem is that all the variables interact. Protein concentration, RNA, salt, pH and other components can each change the result. Ten settings across six variables means a million combinations. Our PhaseScan platform already runs experiments in roughly 100-picolitre droplets, but even at that scale we can't comprehensively test every combination.

PhasePilot connects Meta's models to PhaseScan so the experiment can use incoming results to decide what to do next. A Bayesian planner suggests informative conditions around the phase boundary. Muse Spark reviews those suggestions alongside the results, image quality and remaining protein budget, then chooses whether to run another condition, repeat one, pause or finish. Segment Anything helps check droplet quality, and our existing classifier measures phase separation. The controller checks each command and protects the remaining protein reserve.

We demonstrated PhasePilot on our Beta3 instrument using G3BP1, polyA RNA and PEG, with live microscopy, droplet quality control and a record of Muse's decisions. In a separate three-variable replay benchmark against full scans, adaptive sampling used about 20x less protein (2.5 versus 52 microlitres) and 17x less active instrument time (about 3 versus 51 minutes), with reported boundary accuracy of at least 0.997. Those are replay results, and the instrument-time figure excludes incubation waits.

For us, the immediate use is developing assays for drug discovery with less protein and fewer unnecessary measurements. Longer term, we'd like to apply the same approach to other experiments where there are too many interacting variables to test everything, including cellular assays, formulation and enzyme engineering.

## More from Meta Global AI Developer Hackathon

- [Verity — Nnaemeka Eziamaka](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4198)
- [Ale + Jolly — Physical AI for Business](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4199)
- [Voni](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4200)
- [ZKward](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4202)
- [n Plus Alpha: GTM for SMB Creatives and soloprenuers](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4203)
- [Varun Aiyar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4204)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4201. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt