GenomeCanvas

Explore the fixture in three dimensions

GenomeCanvas combines a Three.js protein scene with a small knowledge graph: 54 / 187 / 217. Backbone ribbons use the structures' confidence scores, and a detail view opens Mol*. The viewer is loaded separately, contributing 0 bytes to the recorded main JavaScript bundle; it still requires a runtime download.

The fixture contains 52 of 54. The two fallback structures are labelled substitutes, so their presence is visible instead of implying that every fixture has a real prediction.

The interface moves before it speaks

Ask for the proteins involved in a disease and the backend streams typed UI commands ahead of any prose: the universe dims, the graph re-roots and the camera re-frames before the first sentence arrives. The local grounded answer path works without a model key. Its recorded model calls: zero. Optional Claude narration makes external model calls. Grounding and typed commands constrain the workflow; they do not guarantee that generated prose is semantically correct.

Honest about where it stands

The deployment is configured but not provisioned, so the code is the artifact rather than a verified service URL. Visual neighbors in reduced-dimensional coordinates are spatial proximity, not proof of shared biological function or clinical utility. The data is a small fixture, and this update did not reproduce the project's tests or establish new scientific results.