The Immune Atlas¶
Open the atlas → Best on a desktop browser. Drag to orbit, scroll to zoom, click any molecule.
The pathway map explains the four colour-coded streams of the working disease model as a flowchart. This is the same model as a cell you can fly around inside: 217 molecules and 380 interactions across eleven compartments, from the nucleus and the mitochondrion down to a neighbouring lymphocyte.
The embedded view above is fully interactive, but the atlas is designed for a full window — use the button for the real thing.
What you can do with it¶
Enter a compartment. Double-click the mitochondrion and the camera flies inside, revealing detail that stays hidden at the whole-cell level: the VDAC1 macropore, the PNC1/PNC2 nucleotide carriers on the inner membrane, POLG at the mtDNA nucleoid, the respiratory complexes, and the mitophagy machinery that gets flagged and then jammed.
Isolate one arm. Turn every pathway layer off, then turn a single one back on. cGAS–STING alone is legible in a way it never is in a diagram with everything drawn at once.
Trace the coloured streams. The PURPLE, BLUE, RED, GOLD, ORANGE, TEAL and GREEN streams are the ones from The Mitochondrion Under Siege, carrying their original names and attack-vector numbers, so the atlas and the figure can be read side by side. Click one and it lights its whole chain, marking what feeds it and what it hands on.
Step through a cascade. This is the part that is hard to do on paper. Ask for every way the variant induces type-I interferon and the atlas enumerates the routes through the graph — 379 of them, within sixteen steps — then walks you along one reaction at a time, forwards and backwards, with the mechanism and its citation at each step.
Each route carries three things worth knowing:
- Net effect, the product of the polarities along it.
A565T ⊣ SAMHD1 ⊣ NF-κBis a double negative: losing a brake on a brake raises NF-κB. Some routes come out negative, and those are real negative feedback — the ones through IκBα, USP18 and A20 genuinely suppress their endpoint. - Evidence grade, taken from the weakest link on the route, because a chain is only as defensible as its softest step.
- Which feedback loops it passes through.
Compare treated against untreated. Pick a state in the left rail — wild type, the untreated variant, or any of ten study arms — and the board re-renders to what the model predicts that arm would do. Molecules running above wild type carry a warm halo, ones running below carry a cool one, and the arrows thicken or thin with the traffic through them. The rail names the biggest movers in words, measured against the untreated variant so you are reading what the drug changed rather than what the disease does.
It is worth seeing the two sensor arms side by side, because they are close to mirror images: MCC950 takes NLRP3 and IL-1β down while the interferon arm barely moves, and IMSB301 does the reverse. Neither leaves the other completely untouched, and that is not a rounding error — cGAS –STING primes NLRP3 through NF-κB, so blocking the first really should dim the second.
These are coarse directional expectations, not predicted fold-changes. The graph has no rate constants, no stoichiometry and no time axis, so the numbers say "this goes down, that does not" and nothing finer. Read them as arrows, not as measurements.
Send someone the exact view¶
Every view in the atlas is a URL. Set something up — a tour paused on the step that matters, a single traced route, one compartment with one layer left on — and the address bar already says so. ⧉ Copy link to this view in the left rail puts it on your clipboard, and the tour and cascade HUDs carry the same button so you can send the step you are looking at without leaving it.
Opening that link puts the reader exactly where you were. It is a few hundred bytes instead of a screen recording, it stays legible if they want to see what it does before clicking, and unlike a video they can carry on exploring from where you left them.
The grammar is plain enough to write by hand:
| Link | Opens on |
|---|---|
#tour=loop-a |
the Loop A tour, from the start |
#tour=loop-b&step=5 |
Loop B, step 5 — the therapeutic gap |
#node=vdac1 |
VDAC1, inspector open |
#stream=blue |
the BLUE mtDNA-escape stream, whole chain lit |
#all |
every downstream consequence of p.A565T |
#path=il1b |
every route from the variant to IL-1β |
#from=a565t&to=pem |
a trace between any two molecules you name |
#focus=mitochondrion&layers=mito.mitophagy |
inside the mitochondrion, two layers on |
#ev=S |
the whole map filtered to demonstrated evidence only |
#state=mcc950 |
the board as predicted under MCC950 (Arm 9) |
#state=vbit4&tour=loop-a |
the Loop A tour, rendered in the VBIT-4 state |
(All relative to /atlas/ — the full link is
https://glasscannonfoundation.org/atlas/#tour=loop-a and so on.)
A shared route carries its full chain of molecules rather than a position in the route list, so the link keeps meaning the same argument as the atlas grows. If a route is ever edited out from under a link, the atlas says so and shows the route list instead of quietly opening a different one.
Two things worth knowing before you send one to a patient or a colleague: the links contain no identifiers of any kind, and Pin the exact camera angle is off by default — tick it only if the specific angle is part of what you are trying to show, since each view already frames itself on arrival.
How confident is any of this?¶
Every molecule and every arrow carries an evidence grade, and the atlas will filter to show you only what survives at each level:
| meaning | |
|---|---|
| S | demonstrated in SAMHD1-deficient cells, animals, or AGS patients |
| G | well established in another immune or mitochondrial system, imported here |
| I | mechanistically consistent extrapolation, not yet tested in SAMHD1 models |
This is the same key used throughout the project's working documents. It is there to be used: set the evidence filter to S only and most of the picture disappears. What remains is the part that can be argued from directly. The rest is a hypothesis with its reasoning shown, which is the honest way to present a model that has not been tested yet.
What it is not¶
It is a qualitative map, not a simulation. An arrow means "acts on", not "acts on with magnitude x". There are no rate constants, no stoichiometry and no time axis, so it can tell you that a route exists and which direction it pushes — but not how much, or how fast, or which route dominates in a real cell.
Machine-Readable Model Downloads & Permalinks¶
The complete causal graph and semantic vector embeddings are available for computational biology pipelines, Cytoscape, and LLM agent RAG:
| Format | File & Permalinks | Description | Size |
|---|---|---|---|
| Pipe-Delimited Model | samhd1-model.txt |
248 nodes, 426 edges, 20 layers, S/G/I & L1–L6 evidence | ~230 KB |
| Cytoscape CX2 Standard | samhd1-innate-immune-cx2.json |
Direct import for Cytoscape Desktop, Cytoscape Web & NDEx | ~3.1 MB |
| Binary Vector Matrix | samhd1-model-embeddings.npz |
High-speed 1536-dim NumPy matrix for sub-millisecond semantic search | 0.86 MB |
| JSON Vector Embeddings | samhd1-model-embeddings.json |
Complete metadata + 1536-dim OpenAI embedding vectors | 9.8 MB |
Colophon¶
Built with three.js. No tracking, no analytics, no external requests — the page loads its own code and nothing else. Source and the full citation registry are in the project repository.