Skip to main content
← Work

Scientific computing · AI

Heat Protein Lab

The heat story told right: what heat does to the molecules that keep you alive, one real protein at a time.

2026 Google AIScientific vizGeminiStructural biology
Heat Protein Lab
17
Scientific skills wired into one AI-driven toolchain
HSF1
The heat-shock sensor at the center of the story
1
Protein at a time: depth over breadth

What it is

Heat Protein Lab is an interactive, browser-based lab for watching what heat does to proteins: denaturation, the heat-shock response, and the molecular machinery (HSF1 and friends) that cells use to survive it. You scroll down a temperature gradient, one “plate” at a time, and the molecules respond.

Open the lab → heat-protein-lab.pages.dev

Heat Protein Lab: Plate 0, 37 °C

The thing I built after the wrong thing

I didn’t build this at the Google I/O hackathon. At the hackathon I built Heat Threshold, an ambitious, multi-agent heat-safety dashboard that scheduled outdoor activity around wet-bulb risk. It worked, and it was the wrong thing: I’d answered a logistics question when the idea I actually cared about was the science, what heat does at the molecular level.

So afterward I rebuilt the heat idea from scratch, the way it should have been told the first time: not a swarm of agents, but a quiet, legible walk through one real protein.

How it works

It wires a set of small, independently-testable scientific skills, structure rendering, denaturation modelling, metrics, charting, into one coherent pipeline an AI agent can drive. Each step is traceable; the prose is held to the same standard as the code.

  • A denaturation view that animates structural unfolding as temperature climbs
  • Real structures, not cartoons, e.g. PDB 5D5U, HSF1 captured in contact with its DNA heat-shock element, with AlphaFold confidence shown honestly (pLDDT 61.31 for the flexible linker, not hidden)
  • A chart skill strict enough to catch its own design bugs

HSF1, the real structure, with an honest confidence score

Why it matters

Same thesis as the rest of my work: one person, paired with the right agent toolchain, can build a legible, inspectable system fast, even in a domain (structural biology) that isn’t my day job. But the real lesson is the pivot: I built the wrong thing first (Heat Threshold), recognized it, and used what I learned to build the right one.

It’s the third angle on a question I keep circling, what heat does, and how to stay ahead of it, after HeatCompass (from your pocket) and Heat Threshold (from the route). This one’s from the molecule.

Want something like this, built to ship?