NeuroSTIAS

Neuroscience Spatial Transcriptomics Integrated Analysis System: a plug-and-play analysis system for neuroscience. It is built around 3D spatial transcriptomics and also covers brain cell-type atlasing, brain organoids, electrophysiology, neuron and circuit modelling, neuron morphology, and connectome analysis.

NeuroSTIAS grew out of 3D-STIAS V4.4.1, a general-purpose, Chinese-language, 12,000-line interactive script. It has been rebuilt as an English Python package with:

• Skills: every analysis is a self-describing plug-in (skill.yaml + a function). Put a new skill folder in skills/ and it is discovered automatically.
• Provenance by default: every run writes provenance.json with input checksums, git commit, parameters, seed, package versions and hardware.
• Replicate-aware statistics: animals, donors and organoids are the statistical unit, not spots.
• Reproducible benchmark experiments (experiments/) against established tools on public data.
• An optional LLM assistant plugin (plugins/llm_assistant/). It works with any local model (LM Studio, Ollama, llama.cpp) or, if you opt in, a cloud model. The core never needs it.

Install

./install.sh          # core + analysis extras (ephys, modelling, morphology, imaging)
./install.sh full     # + benchmark baselines + LLM plugin
source .venv/bin/activate
neurostias doctor

Use

neurostias skills list                         # what can it do?
neurostias skills show atlas.annotate          # inputs/outputs of one skill
neurostias run core.inspect input=my.h5ad      # run a skill -> runs/<skill>_<time>/
neurostias data list                           # declared datasets
neurostias experiment E01                      # rerun a benchmark experiment
neurostias report                              # rebuild paper tables/figures from results
neurostias menu                                # interactive menu
neurostias ask "Which layers are enriched for Sst interneurons?" --data my.h5ad   # optional plugin

Layout

Folder	Contents
neurostias/core	I/O conventions, provenance, skill registry and runner, replicate-aware statistics
neurostias/spatial3d	serial-section registration and 3D reconstruction, 3D graphs, domains, niches, communication
neurostias/atlas	curated brain marker panels, cell-type and cortical-layer annotation
neurostias/organoid	organoid fidelity and maturation against fetal references
neurostias/ephys	spike-train and LFP/EEG analysis
neurostias/modeling	Hodgkin–Huxley / LIF neurons, networks, virtual perturbations
neurostias/morphology	SWC morphology, Sholl analysis, morphometrics
neurostias/imaging	fMRI connectivity and graph metrics
neurostias/legacy	translated V4.4.1 modules (GRN, trajectory, stemness, ML, enrichment, ...)
neurostias/skills, skills/	built-in and drop-in skill manifests
plugins/	optional extras (LLM assistant)
experiments/	benchmark experiments, each with run.py, results and an auto-generated report
data/	manifest.yaml (sources, licences, checksums); data files themselves are not committed
paper/	landscape/gap analysis and auto-generated tables and figures

Data integrity

Results labelled real come from real measured data. Tests that apply a known perturbation to real data (for example, a known rigid transform to check registration) are labelled semi_synthetic everywhere they appear. Report numbers are generated from result files and are never typed by hand.

Benchmark experiments

ID	Question	Data	Compared with
E01	Cortical-layer (domain) recovery	DLPFC, 12 Visium sections, 3 donors	Scanpy Leiden, k-means, BANKSY, SpaGCN, GraphST, V4.4.1
E02	Do 3D cross-section graphs beat 2D graphs?	DLPFC adjacent pairs, GSE214349 (5 mice), MERFISH (12 sections)	per-section 2D graphs incl. degree-matched 2D; z-weight sweep
E03	Registration accuracy (semi-synthetic)	real DLPFC / GSE214349 sections with known transforms	PASTE, centroid alignment
E04	Brain cell-type and layer annotation	MERFISH (author classes), DLPFC (manual layers)	V4.4.1 panels, argmax, CellTypist
E05	False positives from pseudoreplication	DLPFC null section splits; GSE214349 design audit	spot-level Wilcoxon / t-test (V4.4.1)
E06	Neurotransmitter-aware communication	MERFISH, DLPFC	generic V4.4.1 database, LIANA
E07	Organoid maturation and regional identity	HNOCA time-course subset	V4.4.1 stemness, genes-detected (CytoTRACE principle)
E08	Spike-train / EEG validation	DANDI 000006, EEGBCI	Elephant, MNE CSP+LDA
E09	Simulator correctness	HH, LIF, Brunel network	Brian2, closed form
E10	Morphometrics and cell-class classification	NeuroMorpho (Allen Cell Types)	NeuroM
E11	Functional connectome	nilearn development fMRI (155 subjects)	nilearn reference pipeline, motion-only baseline, motion-regressed connectome
E12	(optional, not run) LLM assistant routing and grounding	154 pre-registered questions	local Qwen 9B / 27B

Each experiment folder holds run.py, results/ (tables, figures, summary.json, provenance.json) and an auto-generated REPORT.md. neurostias report gathers them into paper/.