Network datasets

A dataset is a named selection over the OmniPath interaction record: LIANA is the ligand-receptor slice, MetaLinksDB is the metabolite-protein one. The service holds the definitions, so this client does not carry a list of dataset names. It reads them.

Point the client at a service

The datasets are new, so they are on a development deployment first. Set the base URL once, and every later call uses it:

import omnipath_client as op

op.set_base_url('https://dev3.omnipathdb.org/api')
op.base_url()

'https://dev3.omnipathdb.org/api'

Set OMNIPATH_BASE_URL in the environment to do the same without a call, for a script that must not name a deployment in its source.

See what the service carries

op.datasets.names()

['metalinksdb', 'liana']

These names come from the service, so tab completion after op.datasets. offers what this deployment actually serves.

Get a dataset as a DataFrame

liana = op.datasets.liana
df = liana.get(limit = 5)

df.select(['source_label', 'target_label', 'interaction_type', 'resources'])

┌──────────────┬──────────────┬──────────────────┬──────────────────┐
│ source_label ┆ target_label ┆ interaction_type ┆ resources        │
╞══════════════╪══════════════╪══════════════════╪══════════════════╡
│ col5a1       ┆ ddr1         ┆ ligand_receptor  ┆ connectomedb2025 │
│ col5a1       ┆ sdc3         ┆ ligand_receptor  ┆ connectomedb2025 │
│ col5a1       ┆ ddr2         ┆ ligand_receptor  ┆ connectomedb2025 │
│ col5a1       ┆ itga2        ┆ ligand_receptor  ┆ connectomedb2025 │
│ col5a1       ┆ itgb1        ┆ ligand_receptor  ┆ connectomedb2025 │
└──────────────┴──────────────┴──────────────────┴──────────────────┘

The frame is polars by default, or the first backend installed. Ask for another with backend:

op.datasets.metalinksdb.get(limit = 5, backend = 'pandas')

The service pages, so limit is the size of one request. Page further with the cursor a page returns:

page = liana.get(limit = 1000)

Ask what a dataset is made of

liana.resources()
liana.attributes()

['connectomedb2025']
['endpoints', 'label', 'references', 'evidence']

info() gives the whole registry row: the contributing resources, the interaction classes the dataset is restricted to, the attributes it projects, and — for a dataset assembled from parts, as MetaLinksDB is — the composition it is built by.

op.datasets.metalinksdb.info()['curation']

{'chembl_curation': 'mechanism_of_action',
 'chemical_class_gate': 'metabolite',
 'excluded_from_combined': ['bindingdb']}

Ask how much there is, without fetching it

stats = liana.stats()
stats['total'], stats['total_is_estimate']

(27161, True)

Nothing stores the collapsed dataset, so a total is the cost governor's estimate unless you ask for an exact count. total_is_estimate always says which one you got. Restrict the scope to count a part of it:

liana.stats(organism = '9606')

Every summary describes the scope you asked for

Source counts, references and the sign and direction flags on each row describe the resources your query admits — not every resource in the build. Restricting to one resource therefore changes them, which is the point.

Past the datasets

A dataset is a parameter, not a special case, so anything a dataset accessor does not reach is reached through the general functions:

op.interactions(
    filters = {
        'datasets': 'liana',
        'exclude_resources': 'connectomedb2025',
    },
    limit = 10,
)

op.interaction_stats(resources = 'signor', exact_total = True)
op.interaction_parameters(organism = '9606')

op.interaction_parameters() reports the values each parameter can still take under the scope you have narrowed to — so a filter that would leave nothing is visible before you run the query.