A free, source-graded 3D atlas of the human brain: every region, pathway, receptor, and drug is a graded, inspectable node on one map you can rotate, pull apart, search, and click through.
Important
**67% of the 1665 knowledge nodes are sourced or verified**in the shipped dataset, and every fact in the app carries a provenance grade you can inspect. This is a real programmatic count of the data (see Every node is graded).
Live at neurarium.olicorne.org.
Warning
Work in progress: it very likely contains mistakes. The anatomy (regions, shapes, projections, descriptions) is not yet reviewed or sourced and may contain model hallucinations; the drug data is machine-extracted from a single source (Stahl's Prescriber's Guide) and likewise unreviewed. Do not rely on any of it, and never use it for medical decisions.
neurarium takes facts that normally live scattered across atlases, pathway diagrams, receptor tables, and drug monographs, and lays them onto a single 3D model. The point is to make the relationships between them (which region projects where, which receptor sits in which structure, what a drug does and to what) visible at a glance rather than reconstructed in your head. Under the hood it is a graph of nodes, and because the dataset is machine-assembled, every node carries a source grade so you always know how much to trust it.
| Layer | What you see | What you can do with it |
|---|---|---|
| Anatomy | Cortical lobes, basal ganglia, diencephalon, limbic, and hindbrain as one procedural 3D mesh | Rotate, explode on a slider to reveal the deep nuclei, go transparent, peel away the near side, or isolate a single structure |
| Wiring | Neuron projections as directed arrows, colored by type (excitatory, dopaminergic, ...) or by excitatory/inhibitory sign | Click a pathway for its route, transmitter, and sources; play a named functional circuit as a traveling pulse |
| Receptors & targets | Receptors plus other molecular targets (transporters, enzymes, ion channels) | Focus one: the brain dims to the structures expressing it (scattered with glowing dots), beside its class, sign, and every drug acting on it |
| Drugs | Psychiatric drugs from Stahl's Prescriber's Guide | Focus one: effect-colored dots (boost / block / modulate) animate over the regions it touches, beads flow along the transmitter systems it works through, and the panel shows its structure, class, bindings, and each binding's source |
| Everything | One search box; fully URL-addressable state | Search regions, pathways, receptors, and drugs at once; pivot from a drug to its class or from a target to every drug that hits it; share any view as a deep link |
A node is any sourceable datum: a brain region, a projection between two regions, a functional circuit, a receptor, a receptor's expression in a given region, a drug, a single drug-to-target binding. Nodes interlink (a receptor links to the regions expressing it and the drugs acting on it), and a detail panel is a view of one node plus every node linked to it, so you explore outward from whatever you clicked.
Because the dataset is large and machine-assembled, the honest question for any node is how do we know this? Every source shown in a panel answers it inline with a colored provenance pill. The design goal is that every node carries a source, and the pill makes the gaps visible. From weakest to strongest:
- grey
?(LLM-only): produced by a model from memory, unchecked; may be a hallucination. - yellow
~(sourced): from the cited document, but the node itself was not quote-verified. - green
✓(verified): a model extracted a quote, it was programmatically confirmed present in the cited source, and a second model agreed it supports the node. Highest grade available, and still model-driven. - orange
NOSOURCE: no source or reference for that node yet.
The grade is part of the data, upgraded as each node is checked, so the coverage below is a real count:
67% of the 1665 knowledge nodes in the dataset are sourced or verified. A node is any sourceable datum (a region, a pathway, a receptor, a drug binding, ...). This is a programmatic count (tools/update_readme_stats.py, from the emitted data), not hand-typed:
| Node kind | Sourced or verified |
|---|---|
| Drug target bindings | 620 / 632 (98%) |
| Drug nomenclature (NbN) | 116 / 116 (100%) |
| Drug class | 156 / 158 (99%) |
| Neuron pathways | 99 / 103 (96%) |
| Functional circuits | 6 / 6 (100%) |
| Projection groups | 10 / 10 (100%) |
| Receptor classifications | 30 / 56 (54%) |
| Receptor expression regions | 0 / 383 (0%) |
| Target classifications | 21 / 25 (84%) |
| Target expression regions | 0 / 124 (0%) |
| Brain-region anatomy | 52 / 52 (100%) |
| Wikipedia reference links | 298 / 298 (100%) |
Drug bindings lead because they pass the full quote-verification gate; the anatomy, pathways, and expression regions are the current frontier. The same key and coverage bar live in the app's About panel.
On the sources. The reference works the data is checked against (Stahl's
Prescriber's Guide and the other psychopharmacology / neuroscience books) are
copyrighted, so only the tooling that uses them is committed, not the text. Anyone
holding a copy can reproduce the extraction and confirm every ✓-graded quote: drop
the Stahl PDF into sources/books/stahl/ and three committed scripts rebuild exactly
what the gate checks against:
uv run tools/pdf_to_pages.py # the PDF -> one Markdown file per page
uv run tools/build_index.py # the per-drug page index
python tools/check_data.py # re-verifies every quote is on its cited pageThe anatomy is plain structured data, kept deliberately separate from the
rendering. Under public/data/ it is split by node kind (structures, projections,
circuits, receptors, drugs; one JSON object per line) beside a self-describing
meta.json (colour and legend maps plus the sourcing tally) and one geometry file per
shape. It is generated from a single source of truth (tools/generate_data.py, with
the drug list in tools/drugs_data.json), so the plain JSONL/JSON is easy to consume
from another engine.
For the full data flow and module graph see ARCHITECTURE.md; for
the file-by-file map and how to extend the dataset see CLAUDE.md.
A sample of the planned directions, none fixed in order: more animation of activity and signal flow across the brain; more substances (LSD, MDMA, ketamine, nicotine, cannabis) with their commercial brand names; pathologies mapped onto regions, circuits, and transmitter systems; deeper pharmacology (CYP enzymatic interactions, second-order receptor effects); consistency checks that flag self-contradicting data; and toward full sourcing, lifting every node's grade from grey toward green as it is checked.
Found a bug, an anatomical or pharmacological inaccuracy, or have a feature request? Please open an issue on this repository. Corrections to the regions, projections, receptor, and drug data are especially welcome, as are ideas for what else belongs on a map like this. This began as a few-days demo during my medical residency and has kept absorbing new kinds of data more easily than expected, so suggestions for where to take it next genuinely help. You can also find other ways to reach me on my website.
The page loads its data with fetch(), so it must be served over HTTP (not opened
from disk). The served site is public/. From the repository root:
python tools/serve.py # serves public/ with caching disabled
# or: cd public && python -m http.server 8000Then open http://localhost:8000/.
Deliberately lightweight, with a small attack surface and no build step:
- Frontend: vanilla ES modules + three.js loaded via an
import map and vendored under
public/vendor/three, so the page runs no third-party script at runtime and works offline. No framework, bundler, ornode_modules. - Data:
tools/generate_data.py(Python standard library only) emits the anatomy aspublic/data/(meta.json+*.jsonl+shapes/*.json), fetched at runtime. - Serving: a hardened Caddy container (non-root, read-only rootfs, dropped capabilities, resource limits, strict Content-Security- Policy) behind a TLS-terminating reverse proxy.
- Debugging: an eruda on-screen console,
loaded only in dev or with
?debugso it never ships to normal visitors.
Built by Olivier Cornelis (developer and psychiatrist) with the help of Claude Code. Drug descriptions and molecular-structure images come from Wikipedia, used under CC BY-SA.