BioMapper

biomapper

Python client for the BioMapper2 API — map biological entity names to standardized knowledge-graph identifiers (CHEBI, HMDB, PubChem, RefMet, and more).

from biomapper import map_entity

result = map_entity("L-Histidine")
print(result.primary_curie)     # RM:0129894
print(result.confidence_tier)   # high
print(result.ids_for("CHEBI"))  # ['15971']
print(result.equivalent_ids_for("HMDB"))  # ['HMDB0000177']

---

Installation

# Core (async HTTP client + Pydantic models)
pip install biomapper

# With the external benchmark suite
pip install 'biomapper[benchmarks]'

The core install includes the MCP server for agents (Claude Code, Claude Desktop, other MCP clients); launch it with uvx biomapper (see Use from an agent (MCP)).

Use 1.6.2 or later for uvx biomapper. 1.6.0 and 1.6.1 shipped the server behind an mcp extra (uvx --from 'biomapper[mcp]' biomapper-mcp); that line still works on 1.6.2.

Use 1.5.6 or later for harmonize_cohorts() and ambiguous-link flags (see Harmonization).

Use 1.5.4 or later to call the hosted deployment without a key (see API keys).

Use 1.5.3 or later for the benchmark suite. Earlier published releases, up to and including 1.4.0, ship neither biomapper.benchmarks nor the benchmarks extra, so pip install 'biomapper[benchmarks]' against them warns that the extra is unknown and installs only the core client. 1.5.2 contains the suite but its extras were published empty, so the same command installs the core client without pandas, requests, rdkit, openpyxl or defusedxml and the suite fails on first import; install those by hand if you are pinned to 1.5.2. 1.5.3 is the first release whose benchmarks extra actually carries its dependencies. The changelog records which versions reached the index; 1.5.0 and 1.5.1 are tagged (not published).

---

Notebooks

• notebooks/biomapper_tutorial.ipynb: the harmonization protocol end to end, keyless and build-pinned: resolve two cohort panels, link them by identifier-only CURIE sets, and read a certified and a refused structural certificate.
• notebooks/monti_overlap_benchmark.ipynb: BioMapper vs Monti et al. 2026 on the NECS cross-cohort overlap, as linked-count coverage for four cohorts plus a structural certification of the NECS to Arivale links, replayed from a pinned run with a small live check.
• notebooks/ukbb_arivale_harmonization_sop.ipynb: a standard operating procedure for harmonizing two cohorts with this package, worked on UK Biobank × Arivale (metabolites, proteins, clinical labs): inclusion rules, map, match, verify, and results asserted against the Activity 2 milestone, replayed offline from shipped panel metadata.

---

API keys (optional)

No API key is needed for the hosted deployment (https://biomapper.expertintheloop.io). From 1.5.4 the client is keyless by default: with no key configured it sends no X-API-Key header, and every entry point, including the synchronous wrappers, works as shown below.

A key is only for a self-hosted or authenticated deployment. Pass it as api_key=, or set it in your environment:

export BIOMAPPER_API_KEY=your-key-here

or in a .env file in your project root:

BIOMAPPER_API_KEY=your-key-here

A configured key is sent unless anonymous=True is passed. To force a keyless call while BIOMAPPER_API_KEY is set, use BioMapperClient(anonymous=True). If a deployment does require a key and none was sent, the first request raises BioMapperAuthError saying so.

Releases before 1.5.4 raise BioMapperConfigError when no key is configured, and their synchronous wrappers cannot reach a keyless deployment; upgrade, or use BioMapperClient(anonymous=True) directly.

---

Quick start

Single lookup (synchronous)

from biomapper import map_entity

result = map_entity("L-Histidine")

print(result.resolved)          # True
print(result.primary_curie)     # RM:0129894
print(result.chosen_kg_id)      # CHEBI:15971
print(result.chosen_kg_id_review)  # None  (or 'divergent_refmet' when a ChEBI conflict is flagged for review)
print(result.confidence_score)  # 2.489
print(result.confidence_tier)   # high  (≥2.0)
print(result.ids_for("CHEBI"))  # ['15971']
print(result.ids_for("refmet_id"))  # ['RM0129894']

# KG equivalent IDs — all identifiers from the resolved knowledge graph node
print(result.kg_equivalent_ids)            # {'CHEBI': ['15971', '44637'], 'HMDB': ['HMDB0000177'], ...}
print(result.equivalent_ids_for("HMDB"))   # ['HMDB0000177']

Batch mapping (synchronous)

from biomapper import map_entities, summarize

records = [
    {"name": "L-Histidine"},
    {"name": "Glucose", "identifiers": {"HMDB": "HMDB00122"}},
    {"name": "Sphinganine"},
]

results = map_entities(records, progress=True)  # tqdm bar with [notebook]
summary = summarize(results)

print(f"{summary.resolved}/{summary.total_queried} resolved")
print(f"Resolution rate: {summary.resolution_rate:.1%}")
print(summary.vocabulary_coverage)

Inputs are auto-chunked at 1000 entities per request against the native POST /map/batch endpoint, so 10,000 records cost 10 round-trips.

Dataset upload (synchronous)

For larger inputs, hand the server a TSV/CSV file directly and stream results back. The server processes the file row-by-row over the POST /map/dataset/stream endpoint:

from pathlib import Path
from biomapper import map_dataset_file_sync

result = map_dataset_file_sync(
    Path("compounds.tsv"),
    name_column="name",
    provided_id_columns=["hmdb_id"],
    progress=True,         # tqdm bar
    total_hint=1000,       # optional; enables % progress
)
result.raise_for_error()   # opt-in: raise BioMapperError if the stream truncated
print(f"resolved {sum(1 for r in result.results if r.resolved)} of {len(result.results)}")

name_column and provided_id_columns are required — the server uses them to map your file's columns to entity names and identifier hints. For per-result streaming into a UI or custom processing, use the async BioMapperClient.map_dataset_file_iter method (see the tutorial notebook in notebooks/).

Discovering what the API supports

from biomapper import list_annotators, list_vocabularies, list_entity_types

for a in list_annotators():
    print(f"{a.slug:30s} {a.name}")

# 300+ supported vocabularies (CHEBI, HMDB, PubChem, …)
vocabs = list_vocabularies()
print(f"{len(vocabs)} vocabularies supported")

# Biolink entity types with aliases and default vocabulary prefixes
for et in list_entity_types():
    print(f"{et.type}: {', '.join(et.aliases)}")
    if et.default_prefixes:
        print(f"  prefixes: {', '.join(et.default_prefixes)}")

Tuning resolution

The mapping calls accept the API's resolution options as keyword arguments. An option you do not pass is omitted from the request, so the server's own default applies and the payload is unchanged for callers who ignore them.

Coverage is uneven, so check this before reaching for one. Passing an option to a call that does not accept it raises TypeError locally, before any request:

array_delimiters is absent from the dataset routes because the API does not accept it there. The four missing from map_dataset_file_sync are a gap in that wrapper rather than an API limitation: the async map_dataset_file_iter underneath it does accept them, so use that directly if you need them on a file-based run.

from biomapper import map_entity  # map_entity / map_entities accept all six

result = map_entity(
    "PC 34:1",
    vocab="refmet",             # restrict to one vocabulary (or a list)
    prefer_human=False,         # gene/protein: allow a non-human ortholog to win
    prefer_canonical=False,     # non-gene: allow a non-canonical-namespace node to win
    candidate_limit=20,         # candidates each Kestrel search annotator retrieves (1..100)
    array_delimiters=["|"],     # how delimited ID strings are split
    kestrel_top_n=5,            # opt in to raw Kestrel passthrough rows (1..100)
)

kestrel_top_n populates result.kestrel_results with the raw rows Kestrel returned, exactly as returned, for each search endpoint the pipeline used. Those rows are untrusted external data: treat every field as unescaped and unverified.

Resolution certificates and refusal

A mapping result carries the structural certificate the API computed for chosen_kg_id, which is how you tell "we could not check this" apart from "we checked and disagree".

result = map_entity("cortisone")

cert = result.certificate
if cert is not None:
    print(cert.state)                    # corroborated | uncorroborated | contradicted |
                                         # unavailable | not_applicable
    print(cert.independent_source)       # registry consulted, e.g. "pubchem"
    print(cert.independent_of_selection)  # False => corroboration would be circular
    print(cert.refmet_availability)      # voted | no_match | unavailable | not_queried

# Why the resolver declined, when it did. Distinct from `error`:
# an error means the call failed; a refusal means the pipeline ran and declined to assert.
print(result.refusal_reason)

state="contradicted" means a human should look, not that the resolver is wrong: a name lookup at an external registry can itself return a related-but-different compound. state="unavailable" means there was nothing to check against, which is unverifiable rather than wrong.

Async usage

import asyncio
from biomapper import BioMapperClient

async def main() -> None:
    async with BioMapperClient() as client:
        # Verify connectivity
        health = await client.health_check()
        print(health)  # {'status': 'healthy', ...}

        # Single
        result = await client.map_entity(
            "L-Histidine",
            identifiers={"HMDB": "HMDB00177"},
        )

        # Batch — auto-chunked at 1000 entities per request
        results = await client.map_entities(
            [{"name": "L-Histidine"}, {"name": "Glucose"}],
            progress=True,
        )

        # Stream from a file — per-result as they arrive
        from pathlib import Path
        async for r in client.map_dataset_file_iter(
            Path("compounds.tsv"),
            name_column="name",
            provided_id_columns=["hmdb_id"],
        ):
            print(r.query_name, r.primary_curie)

asyncio.run(main())

map_dataset_file_iter is the primitive for UIs and custom processing that want per-result reactivity. Callers needing a blocking, fully-collected result should use map_dataset_file_sync instead (see above).

Jupyter notebooks

Apply nest_asyncio before using sync helpers inside a running event loop:

import nest_asyncio
nest_asyncio.apply()  # required in Jupyter

from biomapper import map_entities
results = map_entities([{"name": "L-Histidine"}], progress=True)

The worked notebooks are listed under Notebooks. Both run against the keyless hosted deployment, so no API key is needed, and both print the exact KRAKEN build that answered.

Preprocessing functions

from biomapper.extras.metabolon import clean_compound_name, extract_hmdb_id

# Strip quotes and collision-energy suffixes
clean_compound_name('"1,3-Diphenylguanidine_CE45"')  # '1,3-Diphenylguanidine'
clean_compound_name('"4,6-DIOXOHEPTANOIC ACID"')     # '4,6-DIOXOHEPTANOIC ACID'
clean_compound_name('L-Histidine')                   # 'L-Histidine'  (unchanged)

# Extract HMDB accessions from ms1_compound_name format
extract_hmdb_id('HMDB:HMDB03349-2257 L-Dihydroorotic acid')  # 'HMDB03349'
extract_hmdb_id('HMDB00177')                                  # 'HMDB00177'
extract_hmdb_id(None)                                         # None

---

Harmonization (cross-dataset equivalence)

biomapper.harmonize links two already-resolved datasets locally. It runs entirely on the client: no extra requests, no knowledge-graph access, so it works offline and is fully testable without a network. Two entities, one per cohort, link when any of these holds, and each link records which one as its basis (strongest first):

Name links (from 1.5.5, opt-in with link_by_name=True) catch entities that resolution put on different nodes, such as acetylcarnitine (c2) and acetylcarnitine (C2). They are currently intended for small-molecule / metabolite panels; other entity types are pending review of case-sensitive naming conventions (gene and protein symbols, for example). Rules:

• Only entities that resolved on both sides link by name; an unresolved one stays a refusal candidate. Names are compared on query_name, even when you pass a custom key=; links are reported under the keys.
• No fuzzy matching and no punctuation stripping, only whitespace normalization.
• A casefold-only match where either name is NAME_CASEFOLD_MIN_LENGTH_EXCLUSIVE (4) characters or fewer is withheld, not linked, and listed in report.name_match_withheld for review: Co (cobalt) and CO (carbon monoxide) are different things.
• Pass results of one entity type per harmonize() call. The results carry no type to check, so this is the caller's responsibility; a cross-type collision such as cAMP (metabolite) vs CAMP (gene) is out of scope.
• Off by default: harmonize(...) links by node and identifier only, exactly as before 1.5.5.

report.summary() counts links by basis (links_by_basis) and reports n_name_only_links (links with no identifier evidence) and n_name_match_withheld, so a report can say "X linked by identifier, Y by name only". Every link also carries bases, the full set of bases that apply.

Ambiguous links (from 1.5.6). Every link records a_partners and b_partners, how many entities each side links to. link.ambiguous is true when either is above one: several distinct entities landed on one entry, so the pair is a review candidate rather than a confident match. report.one_to_one_links and report.ambiguous_links split the links, and summary() adds n_one_to_one_links and n_ambiguous_links. All links are still returned; nothing is dropped. On NECS x Xu et al. 2022 scored against Metabolon compound IDs, one-to-one links were 397 of 398 correct (99.7%) versus 439 of 575 (76.3%) for all links.

from biomapper import map_entities
from biomapper.harmonize import harmonize

ukbb    = map_entities([{"name": "Glucose"},   {"name": "Urea"}])
arivale = map_entities([{"name": "D-glucose"}, {"name": "X-12345"}])

report = harmonize(ukbb, arivale, a_label="ukbb", b_label="arivale")

report.n_links          # 1
report.links[0].basis   # 'node'
report.links[0].shared  # frozenset({'CHEBI:17234', 'KEGG:C00031'}): the identifiers shared
report.b_unresolved     # ('X-12345',) — a refusal candidate, never silently dropped
report.summary()

Two rules the linker is built around:

1. No structure identifiers. INCHIKEY, INCHI and SMILES are excluded. Linking on a structure hash would make any downstream structural certificate circular and would make precision 100% by construction.
2. Prefix synonyms normalize. KEGG.COMPOUND:C00031 equals KEGG:C00031; genuinely different identifier spaces such as KEGG.GLYCAN stay distinct.

An entity that resolved to nothing is a refusal candidate, not a link. It is named in a_unresolved / b_unresolved, counted in summary(), and left out of the link-rate denominator so non-resolution is never scored as non-equivalence. An entity whose mapping call errored is tracked separately again in a_errors / b_errors.

Import it as from biomapper.harmonize import harmonize. The name is deliberately not bound on the package root, where it would shadow the submodule.

Harmonizing two cohorts in one call: harmonize_cohorts()

harmonize_cohorts() runs the whole protocol: it reads both cohort tables, maps them, links them with harmonize(), and writes a pinned report. It exists because the obvious manual path can quietly make results worse. When only one cohort publishes an identifier type, supplying it moves that cohort onto code-specific entries the other cohort never reaches by name: on the UK Biobank x Arivale clinical labs, supplying Arivale's LOINC codes cut matched pairs from 31 to 2.

from biomapper import harmonize_cohorts

report = harmonize_cohorts(
    "ukbb_labs.tsv",                  # a DataFrame, a list of dicts, or a TSV/CSV path
    "arivale_labs.tsv",               # ('#' comment lines are skipped)
    entity="labs",                    # -> biolink:ClinicalMeasurement
    a_name_column="field_name",
    b_name_column="Display Name",
    b_vocabularies={"LOINC": ["Labcorp LOINC ID", "Quest LOINC ID"]},
    a_label="ukbb",
    b_label="arivale",
)
# UserWarning: LOINC is declared only for 'arivale' ... not supplied as mapping input
report.summary()["arms"]["names_only"]   # pairs, one-to-one/ambiguous, link bases, ...
report.review_queue                      # codes whose entry disagrees with the name's entry
# harmonize_cohorts: wrote results to .../biomapper_runs/harmonize_cohorts_20261008T190000Z

What it does:

1. Names-only arm. Both cohorts are mapped by name with the resolved category and no identifiers. entity takes metabolites (SmallMolecule), proteins (Protein), genes (Gene), labs (ClinicalMeasurement, where lab tests live in Biolink) or any raw biolink: category; category= overrides an alias. Name linking is on for SmallMolecule only, as in harmonize().
2. Identifier arm, shared vocabularies only. You declare identifier columns per vocabulary and cohort ({"LOINC": [...]}); names are normalized, so KEGG.COMPOUND equals KEGG. Only vocabularies both cohorts declare are supplied as mapping input, and only rows carrying one are re-mapped; the rest reuse their names-only result. With nothing shared the arm is skipped and the report says why; with no identifier columns only the names-only arm runs.
3. Review queue for one-sided codes. Each code of a vocabulary only one cohort declares is resolved on its own (annotation_mode="none", one request per row and code) and compared with the row's names-only entry by identifier-set intersection. Every code gets a status (agree, disagree, code_unresolved, name_unresolved_code_resolved, errored); report.review_queue lists everything but agree, with both entries. On the labs this lists Arivale "Glucose", whose name lands on one entry while its LOINC code 2345-7 says another.
4. Report. Both HarmonizationResults (report.names_only, report.identifier), their diff (pairs found only by names, only with identifiers, by both), one-to-one and ambiguous link counts per arm (n_one_to_one_links / n_ambiguous_links, as in harmonize()), link bases, warnings, excluded rows (blank names), and pins. Each arm's link table carries a_partners, b_partners and a one_to_one flag, so ambiguous links can be reviewed.

Output by default. Results are written to biomapper_runs/harmonize_cohorts_<UTC stamp>/ under the current directory (override with output_dir=; save=False to skip; report.write(path) any time): mapping results per arm and cohort, written as each arm finishes so a failure keeps finished work, link tables per arm, diff.tsv, review_queue.tsv, settings.json and summary.json. Nothing beyond what you passed in is written, and input files are recorded by name and SHA-256, never by absolute path.

Pins. The package version; the API's /health status and self-reported version (labelled "self-reported, known stale": the API does not expose its engine release, which is recorded as unavailable); Kestrel's /health kestrel_version and kg_build together with the Kestrel URL used (kestrel_url=); and a UTC timestamp. probe_pins=False skips the two reads.

Requests. The default mapper sends batches of 10 with a 300 s timeout (batch_size=, timeout=, progress=True to see them); a failed batch becomes errored rows that are counted, not an exception. Pass mapper= (any (records, *, entity_type, annotation_mode) -> list[MappingResult] callable) to replay recorded results offline; the tests replay the UK Biobank x Arivale labs this way (tests/fixtures/cohorts_labs_cm/, KRAKEN kg 2.3.0 / 3dd08a5b: 37 pairs, 21 one-to-one by name).

---

Use from an agent (MCP)

biomapper ships a local stdio MCP server (the biomapper command, alias biomapper-mcp), so an agent can resolve entities through the same public, keyless API as the Python client. It runs on your machine; nothing is hosted. Launch it with one command (needs uv):

uvx biomapper

Claude Code (the server sees your project folder through MCP roots):

claude mcp add --scope user biomapper -- uvx biomapper

Claude Desktop (claude_desktop_config.json). Desktop does not inherit your shell PATH, so use the absolute path that which uvx prints, and list the folders map_file may use, because Desktop shares no project roots:

{
  "mcpServers": {
    "biomapper": {
      "command": "/absolute/path/to/uvx",
      "args": ["biomapper"],
      "env": { "BIOMAPPER_MCP_ROOTS": "/Users/you/data" }
    }
  }
}

Tools

Every result says resolved, unresolved, refused, or error. An error is a failed call, never a confirmed no-match. Results carry a compact provenance block (client version, API URL, KG build per Kestrel /health, labeled assumed, not verified, and kg_verified).

Throughput, and why the limits exist

The hosted deployment is one shared instance that answers about 4-5 s per entity and slows sharply under concurrent load, so the server is deliberately polite:

• One request in flight per server process. Interactive calls and file-job chunks share one gate; a call that cannot get it within ~10 s returns "busy, retry shortly".
• Every tool call returns in under 60 s (Claude Desktop cancels at ~60 s), which is why map_batch stops at 5 and long work runs as a job.
• Loop guard: more than 20 map_entity/map_batch calls in 5 minutes are refused with a pointer to map_file. Discovery and status tools do not count.

File jobs

map_file(path, name_column, entity_type, identifier_columns?, output_path?) validates the file (readable, columns present, rows counted, output not already there, /health answering) and queues it. identifier_columns maps a column to its vocabulary, e.g. {"hmdb_id": "HMDB"}.

• Output: <input>.biomapper.tsv next to the input plus <input>.biomapper.provenance.json (full pins: API, Kestrel, KG build). It never overwrites a file. Progress is written to <output>.partial until the job completes.
• Resumable: rows are mapped in chunks of 5 and cached as they land. If the agent session ends, the job is paused; the next session in the same folder resumes it automatically, without re-mapping cached rows. Jobs idle for more than 7 days become expired and resume only through resume_job.
• Honest failure: three whole-chunk failures in a row, or one chunk slower than 240 s, move the job to stalled and it stops sending requests; resume_job retries after a health check. A row that errors three times is written as error. A changed input file, or a KG build that differs from the cached rows, fails the job; rerun=true starts over.
• Folders: map_file only reads and writes inside the client's MCP roots plus BIOMAPPER_MCP_ROOTS, with symlinks resolved. With neither, it refuses and says how to set them; the interactive tools still work.

Environment

On HPC or shared hosts whose home directory is on NFS/SMB, set BIOMAPPER_MCP_HOME to local disk: job ownership uses OS file locks, which network filesystems may not honor across machines.

---

API reference

MappingResult

result.ids_for("CHEBI")        # ['15971']
result.ids_for("refmet_id")    # ['RM0129894']
result.ids_for("PUBCHEM.COMPOUND")  # []

# KG equivalent IDs — all identifiers from the resolved knowledge graph node
result.equivalent_ids_for("HMDB")  # ['HMDB0000177']
result.equivalent_ids_for("LM")   # ['ST01010001', 'ST01010093']

DatasetMappingResult

Return type of map_dataset_file_sync. Captures per-row results plus an opt-in error signal for partial runs.

result.raise_for_error()   # raises BioMapperError if .error is set; else no-op

raise_for_error mirrors httpx.Response.raise_for_status and turns the partial-result contract into an explicit caller opt-in — silent consumption of a truncated run (using .results without checking .error) is the footgun this model is designed to prevent.

Note: confidence_score on dataset-stream results is always None — the /map/dataset/stream endpoint emits a slimmer per-row payload than /map/batch and does not include the annotator assigned_ids block. Use map_entity / map_entities if you need confidence tiers.

Confidence tiers

Error handling

from biomapper import (
    BioMapperError,       # base class
    BioMapperAuthError,   # 401/403: a key is required, or the key was rejected
    BioMapperRateLimitError,  # 429 — throttled
    BioMapperServerError,     # 5xx
    BioMapperTimeoutError,    # request timeout
    BioMapperConfigError,     # invalid config, e.g. api_key= together with anonymous=True
)

try:
    result = map_entity("Glucose")
except BioMapperAuthError as e:
    # Only for a deployment that requires a key; the message says whether one was
    # missing or rejected.
    print(e)
else:
    if result.error:
        print(f"mapping failed: {result.error}")

An authentication failure (401/403) raises BioMapperAuthError from every entry point, sync or async, because it is systemic: every later request would fail the same way. The synchronous map_entity and map_entities go through the batch endpoint, where transient failures (429, 5xx, timeouts, a malformed response) and per-entity problems do not raise: each affected record comes back as MappingResult(error=...) rather than aborting the batch. Check .error on every result. The single-entity call on the async client, await client.map_entity(...), raises the typed exceptions (BioMapperRateLimitError, BioMapperServerError, BioMapperTimeoutError) directly.

Dataset streaming (map_dataset_file_sync) uses a two-tier contract:

• Initial-request errors (401, 422, 500, connect timeout) raise as typed exceptions — these happen before any row is processed, so partial results don't exist to preserve.
• Mid-stream transport failures are captured into DatasetMappingResult.error with the partial results preserved in .results. Call .raise_for_error() to get exception semantics, or inspect .error directly for "accept partial, log the rest" workflows.

Callback exceptions raised from on_result propagate unwrapped and replace the return value — partial results collected up to that point are lost. For UI consumers with failure-prone callbacks, wrap the callback body in your own try/except if you want partial data to survive.

---

External benchmark suite

An API-only reproduction harness for the benchmarks the BioMapper preprint reports. It runs against a deployment, not a local engine checkout, which matters for two reasons: running in-process against a checkout measures a library rather than the deployed service, and pointing at a deployment pins provenance to the backend that actually served the answers.

pip install 'biomapper[benchmarks]'

python -m biomapper.benchmarks list                 # the 11 arms and the 2 deliberate skips
python -m biomapper.benchmarks all                  # all arms, production endpoint
python -m biomapper.benchmarks arm hajjar           # a single arm
python -m biomapper.benchmarks --endpoint dev all   # dev, for testing

The public KRAKEN endpoint is keyless, and a BioMapper2 deployment with no keys configured is open, so the suite runs unauthenticated by default. Set BIOMAPPER_API_KEY if your deployment requires one; the key is never accepted on the command line, because argv is visible to other processes and lands in shell history.

The arms

Read the Reportable as column before quoting anything. Four of the graph's ingested sources (lipidmaps, refmet, loinc, ncbigene) are gold sources for arms above, so those arms measure coverage, not independent accuracy: the gold identifier and BioMapper's answer come from the same place. Each run labels every arm from the build's own source list rather than from a static note, so the label tracks the graph.

Two further reporting rules the suite enforces rather than documents:

• Gene arms report accuracy per target namespace. The any-namespace roll-up is emitted flagged quotable: false, because the namespaces perform very differently and the roll-up has been quoted as though it described all of them.
• A skipped arm is not a zero and not a pass. It carries a reason into the manifest.

Provenance

Every run records what actually served it, read from Kestrel /health (keyless) rather than hardcoded: kestrel_version, kg_version, biolink_version, build_timestamp, git_commit, the full source_versions map, the package version, each dataset's source SHA, and a run id. The build is read once before the arms start and re-read afterwards; if it moved mid-suite, the manifest says so, because the pins would no longer describe every number.

Results are saved by default to a timestamped directory (override with --out). There is no flag that discards them: the expensive part of a run is live API traffic.

What stayed in the engine repo

The CI regression gates (gate.py, conflation_gate.py, test_regression_gate.py, kg-regression.yml) guard merges and need engine internals. The --resolver-mode {weighted,vote} A/B toggles a resolver constructor argument that is deliberately not on the API surface.

---

Development

git clone https://github.com/Phenome-Health/biomapper
cd biomapper
poetry install --with dev --extras all

make check          # format → lint → type-check → test
make test           # tests only
make coverage       # HTML coverage report

docs/solutions/ holds documented solutions to past problems — bugs, best practices, and workflow patterns — organized by category with YAML frontmatter (module, tags, problem_type). Relevant when implementing or debugging in an area something has already been written about.

---

License

MIT — see LICENSE.

---

Related

• BioMapper2 API: https://biomapper.expertintheloop.io