BioNexus start

BioNexus is a Scientific Reliability Layer for AI-Assisted Biology, not a generic bioinformatics toolbox. Its Capability Plane provides reference implementations to demonstrate and test evidence boundaries. It stops at numeric clusters + marker tables and does not annotate cell types.

First laboratory use

For completed single-cell differential expression, use single-cell-de-audit. Start with existing DE results, a sample sheet and the proposed claim. The core table-review path does not require scientific backends or a new analysis run. bionexus audit-de --demo --bundle review-demo demonstrates a synthetic review; it deliberately returns a non-pass and never establishes laboratory benefit.

For other tasks, preserve the explicit user scope and use the routes below.

Before backend execution

python scripts/doctor.py

Honor tier, ready.scverse_ready / scvi_ready / spatial_ready, allowed_next_actions, and forbidden_claims.

Install: pip install -e . (kernel). scRNA gold chain: pip install -e ".[goldchain]". Spatial: pip install -e ".[spatial]". Full scVI: pip install -e ".[scverse]".

Capability Plane: Reference Implementations

The capability plane provides reference execution pipelines (not an all-in-one bioinformatics suite):

Priority	Tier	Skills (Reference Implementations)	When
1	core	single-cell-rna-qc, spatial-transcriptomics (squidpy), scvi-tools, nextflow-development	Default for real data
2	wrapper	Allotrope, provenance	Named lab-ops jobs
3	heuristic (not auto-discovered)	biologics, pLM, ACMG combiner, structure, multiome	Only if user asked and accepts grade C
4	outline	start, problem-selection	Planning only

Heuristic skills live as SKILL.legacy.md. Do not open them for a generic “analyze my data” request. To opt in, rename that file back to SKILL.md.

Core scRNA gold chain

python scripts/doctor.py
python skills/single-cell-rna-qc/scripts/scrna_inspect.py raw.h5ad
python skills/single-cell-rna-qc/scripts/scrna_convert.py 10x_dir/ -o raw.h5ad
python skills/single-cell-rna-qc/scripts/scrna_pipeline.py raw.h5ad -o clustered.h5ad
python skills/single-cell-rna-qc/scripts/scrna_plot.py clustered.h5ad -o figures/
python skills/single-cell-rna-qc/scripts/scrna_scrublet.py raw.h5ad -o raw_scrub.h5ad
python skills/single-cell-rna-qc/scripts/scrna_pseudobulk.py clustered.h5ad -o pb.csv --by sample condition --design pb_design.tsv
python skills/single-cell-rna-qc/scripts/scrna_deseq.py pb.csv --design pb_design.tsv --condition condition --reference control --contrast-level treated -o de.csv

Core spatial gold chain (squidpy)

python skills/spatial-transcriptomics/scripts/spatial_inspect.py visium.h5ad
python skills/spatial-transcriptomics/scripts/spatial_pipeline.py visium.h5ad -o spatial_out.h5ad

Endpoint: clustered .h5ad + markers/SVG CSV. Clusters are numbers. Do not invent cell types.

When not to use a core skill

User has	Do not use	Use instead
Only FASTQs / need nf-core	scrna_pipeline.py	nextflow-development
Already-clustered object, just plots	full gold chain	scrna_plot.py / spatial_inspect.py
Technical batch that Harmony cannot fix	Harmony-only	scvi-tools on counts
Spatial without coordinates	spatial gold chain	refuse; do not invent obsm['spatial']
“What cell type is this?”	this plugin	stop; clusters stay numeric

MCP

Local server defaults to the BioNexus compatibility surface (UniProt, Ensembl, gnomAD, PDB, AF, Reactome, STRING, GEO, GTEx). Prefer dedicated Literature/Database peer plugins when the host provides them; BioNexus does not bundle those peers. Set BIONEXUS_LOCAL_HOSTED_FALLBACKS=1 only for disaster recovery.