Autonima is a large language model (LLM)-guided framework for neuroimaging meta-analysis. It automates article screening against expert-defined eligibility criteria, parses heterogeneous coordinate tables into the distinct analyses that produced them, and selects individual analyses for quantitative synthesis.
Full documentation: https://neurostuff.github.io/autonima/
A project is one YAML file: a PubMed query, article-level inclusion and exclusion criteria, retrieval sources, parsing settings, and one set of contrast-specific criteria per target.
- Search — PubMed through the Entrez API.
- Abstract screening — an LLM judges each record against the article-level criteria, returning a decision, a criterion-by-criterion assessment and its reasoning. Abstract criteria are usually more permissive, since an abstract carries incomplete information.
- Full-text retrieval — PubMed Central via pubget, plus any local corpus registered as a full-text source, such as HTML already downloaded from publisher text-mining APIs. Autonima does not call publisher APIs itself. Records with no usable text are marked unavailable rather than rejected: a retrieval failure is not an eligibility decision.
- Full-text screening — the same procedure against the complete criteria, with the full text in context.
- Coordinate parsing — heuristics identify candidate tables; an LLM reads each table with its caption and footnotes and separates it into the distinct statistical analyses it reports, keyed on contrast, direction, group, condition, session or any other explicitly labelled dimension.
- Analysis selection — every parsed analysis is evaluated against each target's criteria in the context of its article, producing an analysis × target inclusion matrix. One analysis may serve several targets.
- Meta-analysis — selected coordinates are written as a NiMADS studyset and submitted to NiMARE (MKDA, ALE or KDA; FWE or FDR correction).
Stage outputs are cached, so re-running resumes rather than repeating paid API calls.
Both drive the same pipeline and produce identical outputs.
The CLI runs one config, once, into one folder. Use it for scripted and reproducible work.
autonima run config.yamlThe web UI manages many projects over time — live progress, cancellation, cloning a project to make a variant, browsing meta-analysis artifacts, and storing API credentials. Use it while developing criteria.
autonima ui --workspace .The CLI has no memory between invocations; the UI keeps a workspace. A project created in one can be run from the other. See the Web UI guide.
The simplest way to get going is to use uv:
uv tool install 'git+https://github.com/neurostuff/autonima.git[llm,meta,ui]'This will install autonima into ~/.local/bin, so make sure that is in your PATH.
If you're going to be hacking on autonima, clone the repository and use uv sync:
git clone https://github.com/neurostuff/autonima.git
cd autonima
uv sync --extra llm,meta,uiTo run autonima from the source tree, use uv run autonima instead of autonima.
If you'd prefer not to use uv, you can still use pip install.
Autonima needs Python 3.11 or newer (tested on 3.11 through 3.15).
git clone https://github.com/neurostuff/autonima.git
cd autonima
pip install -e ".[llm,meta,ui]"That is the full install. The base install (pip install -e .) has no LLM
client, so it runs autonima run-search but not screening. Each extra adds one
capability; keep the quotes, which zsh needs:
pip install -e ".[llm]" # screening, parsing and annotation (OpenAI-compatible API)
pip install -e ".[meta]" # `autonima meta`
pip install -e ".[ui]" # `autonima ui` local web app
pip install -e ".[readability]" # optional enhanced HTML extraction; needs Node.js
pip install -e ".[docs]" # local docs buildLLM stages read your key from OPENAI_API_KEY, and NCBI asks for a contact
address in the config's search.email.
autonima create-sample-config > config.yaml # a starting config
autonima validate config.yaml # check it before spending anything
autonima run config.yaml # run the pipeline
autonima meta config/outputs # meta-analyse the NiMADS outputOmitting OUTPUT_FOLDER derives it from the config filename stem, so
projects/cue_reactivity/default.yaml writes to
projects/cue_reactivity/default/. Pass one explicitly to override:
autonima run config.yaml runs/my_reviewrun-search and run-abstract execute the pipeline only as far as those stages,
which is useful for checking a query or a criteria set cheaply before committing
to full-text retrieval.
search:
database: "pubmed"
query: "schizophrenia AND working memory AND fMRI"
max_results: 100
email: "your.email@example.com" # contact address for NCBI
retrieval:
sources:
- pubget
load_excluded: false
screening:
abstract:
model: "gpt-5-mini-2025-08-07"
objective: "Identify fMRI studies of working memory in schizophrenia"
inclusion_criteria:
- Human participants
- fMRI neuroimaging
fulltext:
model: "gpt-5-mini-2025-08-07"
objective: "Identify fMRI studies of working memory in schizophrenia"
inclusion_criteria:
- Human participants
- fMRI neuroimaging
parsing:
parse_coordinates: false
coordinate_model: "gpt-4o-mini"
output:
directory: "results"
annotation:
enabled: falseModel identifiers and API endpoints are configurable, so any OpenAI-compatible provider can be used.
Autonima v0.1.0 is the version evaluated in the AutoNIMA manuscript. See releases.
