stars pypi cran

Introduction

LaminDB is an open-source data framework to enable learning at scale in computational biology. It lets you track data transformations, curate datasets, manage metadata, and query a built-in database for biological entities & data structures.

Why?

Reproducing analytical results or understanding how a dataset or model was created can be a pain. Leave alone training models on historical data, orthogonal assays, or datasets generated by other teams.

Biological datasets are typically managed with versioned storage systems (file systems, object storage, git, dvc), UI-focused community or SaaS platforms, structureless data lakes, rigid data warehouses (SQL, monolithic arrays), and data lakehouses for tabular data.

LaminDB goes further with a lakehouse that models biological datasets beyond tables with enough structure to enable queries and enough freedom to keep the pace of R&D high.

For data structures like DataFrame, AnnData, .zarr, .tiledbsoma, etc., LaminDB tracks the rich context that collaborative biological research requires and uses it to validate and index datasets to enable queries. In particular, you get

  • data lineage: data sources and transformations; scientists and machine learning models

  • domain knowledge and experimental metadata: the features and labels derived from domain entities

In this blog post, we discuss a breadth of data management problems of the field.

LaminDB specs

The Python & R packages lamindb & laminr share almost the same API (. → $).

Manage data & metadata with a unified API (“lakehouse”).

  • Use a built-in SQLite/Postgres database to organize files, folders & arrays across any number of storage locations

  • Query & search across data & metadata: filter, search

  • Model entities as an ORM which their own registry: Record

  • Model files and folders as datasets & models via one class: Artifact

  • Slice large array stores: open → guide

  • Cache & load artifacts: cache, load

  • Manage features & labels: Feature, Schema, ULabel

  • Use array formats in memory & storage: DataFrame, AnnData, MuData, tiledbsoma, … backed by parquet, zarr, tiledb, HDF5, h5ad, DuckDB, …

  • Create iterable & queryable collections of artifacts with data loaders: Collection

  • Version artifacts, collections & transforms: IsVersioned

Track data lineage across notebooks, scripts, pipelines & UI.

  • Track scripts & notebooks with a simple method call: track()

  • Track functions with a decorator: tracked()

  • A unified registry for all your notebooks, scripts & pipelines: Transform

  • A unified registry for all data transformation runs: Run

  • Manage execution reports, source code and Python environments for notebooks & scripts

  • Integrate with workflow managers: redun, nextflow, snakemake

Manage registries for experimental metadata & in-house ontologies, import public ontologies.

Validate, standardize & annotate.

Organize and share data across a mesh of LaminDB instances.

  • Create & connect to instances with the same ease as git repos: lamin init & lamin connect

  • Zero-copy transfer data across instances

Integrate with analytics tools.

Zero lock-in, scalable, auditable.

  • Zero lock-in: LaminDB runs on generic backends server-side and is not a client for “Lamin Cloud”

    • Flexible storage backends (local, S3, GCP, https, HF, R2, anything fsspec supports)

    • Two SQL backends for managing metadata: SQLite & Postgres

  • Scalable: metadata registries support 100s of millions of entries, storage is as scalable as S3

  • Plug-in custom schema modules & manage database schema migrations

  • Auditable: data & metadata records are hashed, timestamped, and attributed to users (full audit log to come)

  • Secure: embedded in your infrastructure

  • Tested, typed, idempotent & ACID

LaminHub is a data collaboration hub built on LaminDB similar to how GitHub is built on git.

LaminHub specs

Explore at lamin.ai/explore.

Secure & intuitive access management.

Rather than configuring storage & database permissions directly on AWS or GCP, LaminHub allows you to manage access similar to how you’d do it on GitHub or Notion. See Access management.

A UI to work with LaminDB instances.

See an overview of all datasets, models, code, and metadata in your instance.

See validated datasets in context of ontologies & experimental metadata.

Query & search.

See scripts, notebooks & pipelines with their inputs & outputs.

Track pipelines, notebooks & UI transforms in one place.

Quickstart

For setup, install the lamindb Python package and connect to a LaminDB instance.

pip install 'lamindb[jupyter,bionty]'  # support notebooks & biological ontologies
lamin login  # <-- you can skip this for local & self-hosted instances
lamin connect account/instance  # <-- replace with your instance
I don’t have write access to an instance.

Here’s how to create a local instance.

lamin init --storage ./mydata --modules bionty

In a Python session, transfer an scRNA-seq dataset from the laminlabs/cellxgene instance, compute marker genes with Scanpy, and save results.

import lamindb as ln

# Access inputs -------------------------------------------

ln.track()  # track your run of a notebook or script
artifact = ln.Artifact.using("laminlabs/cellxgene").get("7dVluLROpalzEh8m")  # query the artifact https://lamin.ai/laminlabs/cellxgene/artifact/7dVluLROpalzEh8m
adata = artifact.load()[:, :100]  # load into memory or sync to cache: filepath = artifact.cache()

# Your transformation -------------------------------------

import scanpy as sc  # find marker genes with Scanpy

sc.pp.normalize_total(adata)
sc.pp.log1p(adata)
sc.tl.rank_genes_groups(adata, groupby="cell_type")

# Save outputs --------------------------------------------

ln.Artifact.from_anndata(adata, key="my-datasets/my-result.h5ad").save()  # save versioned output
ln.finish()  # finish the run, save source code & run report

For setup, install the laminr and lamindb packages and connect to a LaminDB instance.

install.packages("laminr", dependencies = TRUE)  # install the laminr package from CRAN
laminr::install_lamindb(extra_packages = c("bionty"))  # install lamindb & bionty for use via reticulate
laminr::lamin_login()  # <-- you can skip this for local & self-hosted instances
laminr::lamin_connect("<account>/<instance>")  # <-- replace with your instance
I don’t have write access to an instance.

Here’s how to create a local instance.

laminr::lamin_init(storage = "./mydata", modules = c("bionty"))

In an R session, transfer an scRNA-seq dataset from the laminlabs/cellxgene instance, compute marker genes with Seurat, and save results.

library(laminr)
ln <- import_module("lamindb")  # instantiate the central object of the API

# Access inputs -------------------------------------------

ln$track()  # track your run of a notebook or script
artifact <- ln$Artifact$using("laminlabs/cellxgene")$get("7dVluLROpalzEh8m")  # query the artifact https://lamin.ai/laminlabs/cellxgene/artifact/7dVluLROpalzEh8m
adata <- artifact$load()  # load the artifact into memory or sync to cache via filepath <- artifact$cache()

# Your transformation -------------------------------------

library(Seurat)  # find marker genes with Seurat
seurat_obj <- CreateSeuratObject(counts = as(Matrix::t(adata$X), "CsparseMatrix"), meta.data = adata$obs)
seurat_obj[["RNA"]] <- AddMetaData(GetAssay(seurat_obj), adata$var)
Idents(seurat_obj) <- "cell_type"
seurat_obj <- NormalizeData(seurat_obj)
markers <- FindAllMarkers(seurat_obj, features = Features(seurat_obj)[1:100])
seurat_path <- tempfile(fileext = ".rds")
saveRDS(seurat_obj, seurat_path)

# Save outputs --------------------------------------------

ln$Artifact(seurat_path, key = "my-datasets/my-seurat-object.rds")$save()  # save versioned output
ln$Artifact$from_df(markers, key = "my-datasets/my-markers.parquet")$save()  # save versioned output
ln$finish()  # finish the run, save source code & run report

If you did not use RStudio’s notebook mode, create an html export and then run the following.

laminr::lamin_save("my-analyis.Rmd")  # save source code and html report for a `.qmd` or `.Rmd` file

The script produced the following data lineage.

artifact.view_lineage()

Explore data lineage interactively here.

Track notebooks & scripts

LaminDB provides a framework to transform datasets into more useful representations: validated & queryable datasets, machine learning models, and analytical insights. The transformations can be notebooks, scripts, pipelines, or functions.

The metadata involved in this process are stored in a LaminDB instance, a database that manages datasets in storage. For the following walk through LaminDB’s core features, we’ll be working with a local instance.

lamin init --storage ./lamin-intro --modules bionty

library(laminr)
lamin_init(storage = "./laminr-intro", modules = c("bionty"))
Hide code cell content 
!lamin init --storage ./lamin-intro --modules bionty

→ initialized lamindb: anonymous/lamin-intro
What else can I configure during setup?

  1. You can pass a cloud storage location to --storage (S3, GCP, R2, HF, etc.)

    --storage s3://my-bucket

  2. Instead of the default SQLite database pass a Postgres connection string to --db:

    --db postgresql://<user>:<pwd>@<hostname>:<port>/<dbname>

  3. Instead of a default instance name derived from the storage location, provide a custom name:

    --name my-name

  4. Mount additional schema modules:

    --modules bionty,wetlab,custom1

For more info, see Install & setup.

If you decide to connect your instance to the hub, you will see data & metadata in a UI.

Let’s now track the notebook that’s being run.

import lamindb as ln

ln.track()  # track the current notebook or script

library(laminr)
ln <- import_module("lamindb")  # instantiate the central `ln` object of the API

ln$track()  # track a run of your notebook or script
Hide code cell content 
import lamindb as ln

ln.track()  # track the current notebook or script

→ connected lamindb: anonymous/lamin-intro

→ created Transform('ZTGPZqWSil700000'), started new Run('rSjKb6jC...') at 2025-04-03 08:48:38 UTC

→ notebook imports: anndata==0.11.4 bionty==1.2.1 lamindb==1.3.2

By calling track(), the notebook gets automatically linked as the source of all data that’s about to be saved! You can see all your transforms and their runs in the Transform and Run registries.

ln.Transform.df()

ln$Transform$df()
Hide code cell content 
ln.Transform.df()
uid key description type source_code hash reference reference_type space_id _template_id version is_latest created_at created_by_id _aux _branch_code
id
1 ZTGPZqWSil700000 introduction.ipynb Introduction notebook None None None None 1 None None True 2025-04-03 08:48:38.074000+00:00 1 None 1 
ln.Run.df()

ln$Run$df()
Hide code cell content 
ln.Run.df()
uid name started_at finished_at reference reference_type _is_consecutive _status_code space_id transform_id report_id _logfile_id environment_id initiated_by_run_id created_at created_by_id _aux _branch_code
id
1 rSjKb6jCHjWm3sjAPQBu None 2025-04-03 08:48:38.080656+00:00 None None None None 0 1 1 None None None None 2025-04-03 08:48:38.081000+00:00 1 None 1
What happened under the hood?

  1. The full run environment and imported package versions of current notebook were detected

  2. Notebook metadata was detected and stored in a Transform record with a unique id

  3. Run metadata was detected and stored in a Run record with a unique id

The Transform registry stores data transformations: scripts, notebooks, pipelines, functions.

The Run registry stores executions of transforms. Many runs can be linked to the same transform if executed with different context (time, user, input data, etc.).

How do I track a pipeline instead of a notebook?

Leverage a pipeline integration, see: Pipelines – workflow managers. Or manually add code as seen below.

transform = ln.Transform(name="My pipeline")
transform.version = "1.2.0"  # tag the version
ln.track(transform)
Why should I care about tracking notebooks?

Because of interactivity & humans are in the loop, most mistakes happen when using notebooks.

track() makes notebooks & derived results reproducible & auditable, enabling to learn from mistakes.

This is important as much insight generated from biological data is driven by computational biologists interacting with it. An early blog post on this is here.

Is this compliant with OpenLineage?

Yes. What OpenLineage calls a “job”, LaminDB calls a “transform”. What OpenLineage calls a “run”, LaminDB calls a “run”.

Manage artifacts

The Artifact class manages datasets & models that are stored as files, folders, or arrays. Artifact is a registry to manage search, queries, validation & storage access.

You can register data objects (DataFrame, AnnData, …) and files or folders in local storage, AWS S3 (s3://), Google Cloud (gs://), Hugging Face (hf://), or any other file system supported by fsspec.

Manage Dataframes

Let’s first look at an exemplary dataframe.

df = ln.core.datasets.small_dataset1(with_typo=True)
df

df <- ln$core$datasets$small_dataset1(otype = "DataFrame", with_typo = TRUE)
df
Hide code cell content 
df = ln.core.datasets.small_dataset1(with_typo=True)
df
ENSG00000153563 ENSG00000010610 ENSG00000170458 perturbation sample_note cell_type_by_expert cell_type_by_model assay_oid concentration treatment_time_h donor
sample1 1 3 5 DMSO was ok B cell B cell EFO:0008913 0.1% 24 D0001
sample2 2 4 6 IFNJ looks naah CD8-positive, alpha-beta T cell T cell EFO:0008913 200 nM 24 D0002
sample3 3 5 7 DMSO pretty! 🤩 CD8-positive, alpha-beta T cell T cell EFO:0008913 0.1% 6 None

This is how you create an artifact from a dataframe.

artifact = ln.Artifact.from_df(df, key="my_datasets/rnaseq1.parquet").save()
artifact.describe()

artifact <- ln$Artifact$from_df(df, key = "my_datasets/rnaseq1.parquet")$save()
artifact$describe()
Hide code cell content 
artifact = ln.Artifact.from_df(df, key="my_datasets/rnaseq1.parquet").save()
artifact.describe()

Artifact .parquet/DataFrame
└── General
    ├── .uid = 'fPTOf1qAg8pNJqRH0000'
    ├── .key = 'my_datasets/rnaseq1.parquet'
    ├── .size = 9012
    ├── .hash = 'ZHlfaXCXxza090J-PA1nCg'
    ├── .n_observations = 3
    ├── .path = /home/runner/work/lamin-docs/lamin-docs/docs/lamin-intro/.lamindb/fPTOf1qAg8pNJqRH0000.parquet
    ├── .created_by = anonymous
    ├── .created_at = 2025-04-03 08:48:38
    └── .transform = 'Introduction'

And this is how you load it back into memory.

artifact.load()

artifact$load()
Hide code cell content 
artifact.load()
ENSG00000153563 ENSG00000010610 ENSG00000170458 perturbation sample_note cell_type_by_expert cell_type_by_model assay_oid concentration treatment_time_h donor
sample1 1 3 5 DMSO was ok B cell B cell EFO:0008913 0.1% 24 D0001
sample2 2 4 6 IFNJ looks naah CD8-positive, alpha-beta T cell T cell EFO:0008913 200 nM 24 D0002
sample3 3 5 7 DMSO pretty! 🤩 CD8-positive, alpha-beta T cell T cell EFO:0008913 0.1% 6 None

Trace data lineage

You can understand where an artifact comes from by looking at its Transform & Run records:

artifact.transform

artifact$transform
Hide code cell content 
artifact.transform

Transform(uid='ZTGPZqWSil700000', is_latest=True, key='introduction.ipynb', description='Introduction', type='notebook', space_id=1, created_by_id=1, created_at=2025-04-03 08:48:38 UTC)

artifact.run

artifact$run
Hide code cell content 
artifact.run

Run(uid='rSjKb6jCHjWm3sjAPQBu', started_at=2025-04-03 08:48:38 UTC, space_id=1, transform_id=1, created_by_id=1, created_at=2025-04-03 08:48:38 UTC)

Or visualize deeper data lineage with the view_lineage() method. Here we’re only one step deep.

artifact.view_lineage()

artifact$view_lineage()
Hide code cell content 
artifact.view_lineage()

! calling anonymously, will miss private instances
_images/f0644ff8632dd1d0856d74842a0d5606a02844d23820c0aaef2dab453ddebea4.svg
Show me a more interesting example, please!

Explore and load the notebook from here.

Explore data lineage interactively here.

I just want to see the transforms.
artifact.transform.view_lineage()  # Python only

Data lineage also helps to understand what a dataset is being used for. Many datasets are being used over and over for different purposes.

Once you’re done, at the end of your notebook or script, call finish(). Here, we’re not yet done so we’re commenting it out.

# ln.finish()  # mark run as finished, save execution report, source code & environment

# ln$finish()  # mark run as finished, save execution report & source code

If you did not use RStudio’s notebook mode, you have to render an HTML externally.

  1. Render the notebook to HTML via one of:

    • In RStudio, click the “Knit” button

    • From the command line, run

      Rscript -e 'rmarkdown::render("introduction.Rmd")'

    • Use the rmarkdown package in R

      rmarkdown::render("introduction.Rmd")

  2. Save it to your LaminDB instance via one of:

    • Using the lamin_save() function in R

      lamin_save("introduction.Rmd")

    • Using the lamin CLI

      lamin save introduction.Rmd
Here is how a notebook looks on the hub.

Explore.

To create a new version of a notebook or script, run lamin load on the terminal, e.g.,

$ lamin load https://lamin.ai/laminlabs/lamindata/transform/13VINnFk89PE0004
→ notebook is here: mcfarland_2020_preparation.ipynb

Manage versioning

Just like transforms, artifacts are versioned. Let’s create a new version by revising the dataset.

# keep the dataframe with a typo around - we'll need it later
df_typo = df.copy()

# fix the "IFNJ" typo
df["perturbation"] = df["perturbation"].cat.rename_categories({"IFNJ": "IFNG"})

# create a new version
artifact = ln.Artifact.from_df(df, key="my_datasets/rnaseq1.parquet").save()

# see all versions of an artifact
artifact.versions.df()

# keep the dataframe with a typo around - we'll need it later
df_typo <- df

# fix the "IFNJ" typo
levels(df$perturbation) <- c("DMSO", "IFNG")
df["sample2", "perturbation"] <- "IFNG"

# create a new version
artifact <- ln$Artifact$from_df(df, key = "my_datasets/rnaseq1.parquet")$save()

# see all versions of an artifact
artifact$versions$df()
Hide code cell content 
# keep the dataframe with a typo around - we'll need it later
df_typo = df.copy()

# fix the "IFNJ" typo
df["perturbation"] = df["perturbation"].cat.rename_categories({"IFNJ": "IFNG"})

# create a new version
artifact = ln.Artifact.from_df(df, key="my_datasets/rnaseq1.parquet").save()

# see all versions of an artifact
artifact.versions.df()

→ creating new artifact version for key='my_datasets/rnaseq1.parquet' (storage: '/home/runner/work/lamin-docs/lamin-docs/docs/lamin-intro')
uid key description suffix kind otype size hash n_files n_observations _hash_type _key_is_virtual _overwrite_versions space_id storage_id schema_id version is_latest run_id created_at created_by_id _aux _branch_code
id
2 fPTOf1qAg8pNJqRH0001 my_datasets/rnaseq1.parquet None .parquet dataset DataFrame 9012 iBiiWBkIitgFtLcru2CLyA None 3 md5 True False 1 1 None None True 1 2025-04-03 08:48:39.019000+00:00 1 None 1
1 fPTOf1qAg8pNJqRH0000 my_datasets/rnaseq1.parquet None .parquet dataset DataFrame 9012 ZHlfaXCXxza090J-PA1nCg None 3 md5 True False 1 1 None None False 1 2025-04-03 08:48:38.599000+00:00 1 None 1
Can I also create new versions without passing key?

That works, too, you can use revises:

artifact_v1 = ln.Artifact.from_df(df, description="Just a description").save()
# below revises artifact_v1
artifact_v2 = ln.Artifact.from_df(df_updated, revises=artifact_v1).save()

The good thing about passing revises: Artifact is that you don’t need to worry about coming up with naming conventions for paths.

The good thing about versioning based on key is that it’s how all data versioning tools are doing it.

Manage files & folders

Let’s look at a folder in the cloud that contains 3 sub-folders storing images & metadata of Iris flowers, generated in 3 subsequent studies.

# we use anon=True here in case no aws credentials are configured
ln.UPath("s3://lamindata/iris_studies", anon=True).view_tree()

# we use anon=True here in case no aws credentials are configured
ln$UPath("s3://lamindata/iris_studies", anon = True).view_tree()
Hide code cell content 
# we use anon=True here in case no aws credentials are configured
ln.UPath("s3://lamindata/iris_studies", anon=True).view_tree()

3 sub-directories & 151 files with suffixes '.csv', '.jpg'
s3://lamindata/iris_studies
├── study0_raw_images/
│   ├── iris-0337d20a3b7273aa0ddaa7d6afb57a37a759b060e4401871db3cefaa6adc068d.jpg
│   ├── iris-0797945218a97d6e5251b4758a2ba1b418cbd52ce4ef46a3239e4b939bd9807b.jpg
│   ├── iris-0f133861ea3fe1b68f9f1b59ebd9116ff963ee7104a0c4200218a33903f82444.jpg
│   ├── iris-0fec175448a23db03c1987527f7e9bb74c18cffa76ef003f962c62603b1cbb87.jpg
│   ├── iris-125b6645e086cd60131764a6bed12650e0f7f2091c8bbb72555c103196c01881.jpg
│   ├── iris-13dfaff08727abea3da8cfd8d097fe1404e76417fefe27ff71900a89954e145a.jpg
│   ├── iris-1566f7f5421eaf423a82b3c1cd1328f2a685c5ef87d8d8e710f098635d86d3d0.jpg
│   ├── iris-1804702f49c2c385f8b30913569aebc6dce3da52ec02c2c638a2b0806f16014e.jpg
│   ├── iris-318d451a8c95551aecfde6b55520f302966db0a26a84770427300780b35aa05a.jpg
│   ├── iris-3dec97fe46d33e194520ca70740e4c2e11b0ffbffbd0aec0d06afdc167ddf775.jpg
│   ├── iris-3eed72bc2511f619190ce79d24a0436fef7fcf424e25523cb849642d14ac7bcf.jpg
│   ├── iris-430fa45aad0edfeb5b7138ff208fdeaa801b9830a9eb68f378242465b727289a.jpg
│   ├── iris-4cc15cd54152928861ecbdc8df34895ed463403efb1571dac78e3223b70ef569.jpg
│   ├── iris-4febb88ef811b5ca6077d17ef8ae5dbc598d3f869c52af7c14891def774d73fa.jpg
│   ├── iris-590e7f5b8f4de94e4b82760919abd9684ec909d9f65691bed8e8f850010ac775.jpg
│   ├── iris-5a313749aa61e9927389affdf88dccdf21d97d8a5f6aa2bd246ca4bc926903ba.jpg
│   ├── iris-5b3106db389d61f4277f43de4953e660ff858d8ab58a048b3d8bf8d10f556389.jpg
│   ├── iris-5f4e8fffde2404cc30be275999fddeec64f8a711ab73f7fa4eb7667c8475c57b.jpg
│   ├── iris-68d83ad09262afb25337ccc1d0f3a6d36f118910f36451ce8a6600c77a8aa5bd.jpg
│   ├── iris-70069edd7ab0b829b84bb6d4465b2ca4038e129bb19d0d3f2ba671adc03398cc.jpg
│   ├── iris-7038aef1137814473a91f19a63ac7a55a709c6497e30efc79ca57cfaa688f705.jpg
│   ├── iris-74d1acf18cfacd0a728c180ec8e1c7b4f43aff72584b05ac6b7c59f5572bd4d4.jpg
│   ├── iris-7c3b5c5518313fc6ff2c27fcbc1527065cbb42004d75d656671601fa485e5838.jpg
│   ├── iris-7cf1ebf02b2cc31539ed09ab89530fec6f31144a0d5248a50e7c14f64d24fe6e.jpg
│   ├── iris-7dcc69fa294fe04767706c6f455ea6b31d33db647b08aab44b3cd9022e2f2249.jpg
│   ├── iris-801b7efb867255e85137bc1e1b06fd6cbab70d20cab5b5046733392ecb5b3150.jpg
│   ├── iris-8305dd2a080e7fe941ea36f3b3ec0aa1a195ad5d957831cf4088edccea9465e2.jpg
│   ├── iris-83f433381b755101b9fc9fbc9743e35fbb8a1a10911c48f53b11e965a1cbf101.jpg
│   ├── iris-874121a450fa8a420bdc79cc7808fd28c5ea98758a4b50337a12a009fa556139.jpg
│   ├── iris-8c216e1acff39be76d6133e1f549d138bf63359fa0da01417e681842210ea262.jpg
│   ├── iris-92c4268516ace906ad1ac44592016e36d47a8c72a51cacca8597ba9e18a8278b.jpg
│   ├── iris-95d7ec04b8158f0873fa4aab7b0a5ec616553f3f9ddd6623c110e3bc8298248f.jpg
│   ├── iris-9ce2d8c4f1eae5911fcbd2883137ba5542c87cc2fe85b0a3fbec2c45293c903e.jpg
│   ├── iris-9ee27633bb041ef1b677e03e7a86df708f63f0595512972403dcf5188a3f48f5.jpg
│   ├── iris-9fb8d691550315506ae08233406e8f1a4afed411ea0b0ac37e4b9cdb9c42e1ec.jpg
│   ├── iris-9ffe51c2abd973d25a299647fa9ccaf6aa9c8eecf37840d7486a061438cf5771.jpg
│   ├── iris-a2be5db78e5b603a5297d9a7eec4e7f14ef2cba0c9d072dc0a59a4db3ab5bb13.jpg
│   ├── iris-ad7da5f15e2848ca269f28cd1dc094f6f685de2275ceaebb8e79d2199b98f584.jpg
│   ├── iris-bc515e63b5a4af49db8c802c58c83db69075debf28c792990d55a10e881944d9.jpg
│   ├── iris-bd8d83096126eaa10c44d48dbad4b36aeb9f605f1a0f6ca929d3d0d492dafeb6.jpg
│   ├── iris-bdae8314e4385d8e2322abd8e63a82758a9063c77514f49fc252e651cbd79f82.jpg
│   ├── iris-c175cd02ac392ecead95d17049f5af1dcbe37851c3e42d73e6bb813d588ea70b.jpg
│   ├── iris-c31e6056c94b5cb618436fbaac9eaff73403fa1b87a72db2c363d172a4db1820.jpg
│   ├── iris-ca40bc5839ee2f9f5dcac621235a1db2f533f40f96a35e1282f907b40afa457d.jpg
│   ├── iris-ddb685c56cfb9c8496bcba0d57710e1526fff7d499536b3942d0ab375fa1c4a6.jpg
│   ├── iris-e437a7c7ad2bbac87fef3666b40c4de1251b9c5f595183eda90a8d9b1ef5b188.jpg
│   ├── iris-e7e0774289e2153cc733ff62768c40f34ac9b7b42e23c1abc2739f275e71a754.jpg
│   ├── iris-e9da6dd69b7b07f80f6a813e2222eae8c8f7c3aeaa6bcc02b25ea7d763bcf022.jpg
│   ├── iris-eb01666d4591b2e03abecef5a7ded79c6d4ecb6d1922382c990ad95210d55795.jpg
│   ├── iris-f6e4890dee087bd52e2c58ea4c6c2652da81809603ea3af561f11f8c2775c5f3.jpg
│   └── meta.csv
├── study1_raw_images/
│   ├── iris-0879d3f5b337fe512da1c7bf1d2bfd7616d744d3eef7fa532455a879d5cc4ba0.jpg
│   ├── iris-0b486eebacd93e114a6ec24264e035684cebe7d2074eb71eb1a71dd70bf61e8f.jpg
│   ├── iris-0ff5ba898a0ec179a25ca217af45374fdd06d606bb85fc29294291facad1776a.jpg
│   ├── iris-1175239c07a943d89a6335fb4b99a9fb5aabb2137c4d96102f10b25260ae523f.jpg
│   ├── iris-1289c57b571e8e98e4feb3e18a890130adc145b971b7e208a6ce5bad945b4a5a.jpg
│   ├── iris-12adb3a8516399e27ff1a9d20d28dca4674836ed00c7c0ae268afce2c30c4451.jpg
│   ├── iris-17ac8f7b5734443090f35bdc531bfe05b0235b5d164afb5c95f9d35f13655cf3.jpg
│   ├── iris-2118d3f235a574afd48a1f345bc2937dad6e7660648516c8029f4e76993ea74d.jpg
│   ├── iris-213cd179db580f8e633087dcda0969fd175d18d4f325cb5b4c5f394bbba0c1e0.jpg
│   ├── iris-21a1255e058722de1abe928e5bbe1c77bda31824c406c53f19530a3ca40be218.jpg
│   ├── iris-249370d38cc29bc2a4038e528f9c484c186fe46a126e4b6c76607860679c0453.jpg
│   ├── iris-2ac575a689662b7045c25e2554df5f985a3c6c0fd5236fabef8de9c78815330c.jpg
│   ├── iris-2c5b373c2a5fd214092eb578c75eb5dc84334e5f11a02f4fa23d5d316b18f770.jpg
│   ├── iris-2ecaad6dfe3d9b84a756bc2303a975a732718b954a6f54eae85f681ea3189b13.jpg
│   ├── iris-32827aec52e0f3fa131fa85f2092fc6fa02b1b80642740b59d029cef920c26b3.jpg
│   ├── iris-336fc3472b6465826f7cd87d5cef8f78d43cf2772ebe058ce71e1c5bad74c0e1.jpg
│   ├── iris-432026d8501abcd495bd98937a82213da97fca410af1c46889eabbcf2fd1b589.jpg
│   ├── iris-49a9158e46e788a39eeaefe82b19504d58dde167f540df6bc9492c3916d5f7ca.jpg
│   ├── iris-4b47f927405d90caa15cbf17b0442390fc71a2ca6fb8d07138e8de17d739e9a4.jpg
│   ├── iris-5691cad06fe37f743025c097fa9c4cec85e20ca3b0efff29175e60434e212421.jpg
│   ├── iris-5c38dba6f6c27064eb3920a5758e8f86c26fec662cc1ac4b5208d5f30d1e3ead.jpg
│   ├── iris-5da184e8620ebf0feef4d5ffe4346e6c44b2fb60cecc0320bd7726a1844b14cd.jpg
│   ├── iris-66eee9ff0bfa521905f733b2a0c6c5acad7b8f1a30d280ed4a17f54fe1822a7e.jpg
│   ├── iris-6815050b6117cf2e1fd60b1c33bfbb94837b8e173ff869f625757da4a04965c9.jpg
│   ├── iris-793fe85ddd6a97e9c9f184ed20d1d216e48bf85aa71633eff6d27073e0825d54.jpg
│   ├── iris-850229e6293a741277eb5efaa64d03c812f007c5d0f470992a8d4cfdb902230c.jpg
│   ├── iris-86d782d20ef7a60e905e367050b0413ca566acc672bc92add0bb0304faa54cfc.jpg
│   ├── iris-875a96790adc5672e044cf9da9d2edb397627884dfe91c488ab3fb65f65c80ff.jpg
│   ├── iris-96f06136df7a415550b90e443771d0b5b0cd990b503b64cc4987f5cb6797fa9b.jpg
│   ├── iris-9a889c96a37e8927f20773783a084f31897f075353d34a304c85e53be480e72a.jpg
│   ├── iris-9e3208f4f9fedc9598ddf26f77925a1e8df9d7865a4d6e5b4f74075d558d6a5e.jpg
│   ├── iris-a7e13b6f2d7f796768d898f5f66dceefdbd566dd4406eea9f266fc16dd68a6f2.jpg
│   ├── iris-b026efb61a9e3876749536afe183d2ace078e5e29615b07ac8792ab55ba90ebc.jpg
│   ├── iris-b3c086333cb5ccb7bb66a163cf4bf449dc0f28df27d6580a35832f32fd67bfc9.jpg
│   ├── iris-b795e034b6ea08d3cd9acaa434c67aca9d17016991e8dd7d6fd19ae8f6120b77.jpg
│   ├── iris-bb4a7ad4c844987bc9dc9dfad2b363698811efe3615512997a13cd191c23febc.jpg
│   ├── iris-bd60a6ed0369df4bea1934ef52277c32757838123456a595c0f2484959553a36.jpg
│   ├── iris-c15d6019ebe17d7446ced589ef5ef7a70474d35a8b072e0edfcec850b0a106db.jpg
│   ├── iris-c45295e76c6289504921412293d5ddbe4610bb6e3b593ea9ec90958e74b73ed2.jpg
│   ├── iris-c50d481f9fa3666c2c3808806c7c2945623f9d9a6a1d93a17133c4cb1560c41c.jpg
│   ├── iris-df4206653f1ec9909434323c05bb15ded18e72587e335f8905536c34a4be3d45.jpg
│   ├── iris-e45d869cb9d443b39d59e35c2f47870f5a2a335fce53f0c8a5bc615b9c53c429.jpg
│   ├── iris-e76fa5406e02a312c102f16eb5d27c7e0de37b35f801e1ed4c28bd4caf133e7a.jpg
│   ├── iris-e8d3fd862aae1c005bcc80a73fd34b9e683634933563e7538b520f26fd315478.jpg
│   ├── iris-ea578f650069a67e5e660bb22b46c23e0a182cbfb59cdf5448cf20ce858131b6.jpg
│   ├── iris-eba0c546e9b7b3d92f0b7eb98b2914810912990789479838807993d13787a2d9.jpg
│   ├── iris-f22d4b9605e62db13072246ff6925b9cf0240461f9dfc948d154b983db4243b9.jpg
│   ├── iris-fac5f8c23d8c50658db0f4e4a074c2f7771917eb52cbdf6eda50c12889510cf4.jpg
│   └── meta.csv
└── study2_raw_images/
    ├── iris-01cdd55ca6402713465841abddcce79a2e906e12edf95afb77c16bde4b4907dc.jpg
    ├── iris-02868b71ddd9b33ab795ac41609ea7b20a6e94f2543fad5d7fa11241d61feacf.jpg
    ├── iris-0415d2f3295db04bebc93249b685f7d7af7873faa911cd270ecd8363bd322ed5.jpg
    ├── iris-0c826b6f4648edf507e0cafdab53712bb6fd1f04dab453cee8db774a728dd640.jpg
    ├── iris-10fb9f154ead3c56ba0ab2c1ab609521c963f2326a648f82c9d7cabd178fc425.jpg
    ├── iris-14cbed88b0d2a929477bdf1299724f22d782e90f29ce55531f4a3d8608f7d926.jpg
    ├── iris-186fe29e32ee1405ddbdd36236dd7691a3c45ba78cc4c0bf11489fa09fbb1b65.jpg
    ├── iris-1b0b5aabd59e4c6ed1ceb54e57534d76f2f3f97e0a81800ff7ed901c35a424ab.jpg
    ├── iris-1d35672eb95f5b1cf14c2977eb025c246f83cdacd056115fdc93e946b56b610c.jpg
    ├── iris-1f941001f508ff1bd492457a90da64e52c461bfd64587a3cf7c6bf1bcb35adab.jpg
    ├── iris-2a09038b87009ecee5e5b4cd4cef068653809cc1e08984f193fad00f1c0df972.jpg
    ├── iris-308389e34b6d9a61828b339916aed7af295fdb1c7577c23fb37252937619e7e4.jpg
    ├── iris-30e4e56b1f170ff4863b178a0a43ea7a64fdd06c1f89a775ec4dbf5fec71e15c.jpg
    ├── iris-332953f4d6a355ca189e2508164b24360fc69f83304e7384ca2203ddcb7c73b5.jpg
    ├── iris-338fc323ed045a908fb1e8ff991255e1b8e01c967e36b054cb65edddf97b3bb0.jpg
    ├── iris-34a7cc16d26ba0883574e7a1c913ad50cf630e56ec08ee1113bf3584f4e40230.jpg
    ├── iris-360196ba36654c0d9070f95265a8a90bc224311eb34d1ab0cf851d8407d7c28e.jpg
    ├── iris-36132c6df6b47bda180b1daaafc7ac8a32fd7f9af83a92569da41429da49ea5b.jpg
    ├── iris-36f2b9282342292b67f38a55a62b0c66fa4e5bb58587f7fec90d1e93ea8c407a.jpg
    ├── iris-37ad07fd7b39bc377fa6e9cafdb6e0c57fb77df2c264fe631705a8436c0c2513.jpg
    ├── iris-3ba1625bb78e4b69b114bdafcdab64104b211d8ebadca89409e9e7ead6a0557c.jpg
    ├── iris-4c5d9a33327db025d9c391aeb182cbe20cfab4d4eb4ac951cc5cd15e132145d8.jpg
    ├── iris-522f3eb1807d015f99e66e73b19775800712890f2c7f5b777409a451fa47d532.jpg
    ├── iris-589fa96b9a3c2654cf08d05d3bebf4ab7bc23592d7d5a95218f9ff87612992fa.jpg
    ├── iris-61b71f1de04a03ce719094b65179b06e3cd80afa01622b30cda8c3e41de6bfaa.jpg
    ├── iris-62ef719cd70780088a4c140afae2a96c6ca9c22b72b078e3b9d25678d00b88a5.jpg
    ├── iris-819130af42335d4bb75bebb0d2ee2e353a89a3d518a1d2ce69842859c5668c5a.jpg
    ├── iris-8669e4937a2003054408afd228d99cb737e9db5088f42d292267c43a3889001a.jpg
    ├── iris-86c76e0f331bc62192c392cf7c3ea710d2272a8cc9928d2566a5fc4559e5dce4.jpg
    ├── iris-8a8bc54332a42bb35ee131d7b64e9375b4ac890632eb09e193835b838172d797.jpg
    ├── iris-8e9439ec7231fa3b9bc9f62a67af4e180466b32a72316600431b1ec93e63b296.jpg
    ├── iris-90b7d491b9a39bb5c8bb7649cce90ab7f483c2759fb55fda2d9067ac9eec7e39.jpg
    ├── iris-9dededf184993455c411a0ed81d6c3c55af7c610ccb55c6ae34dfac2f8bde978.jpg
    ├── iris-9e6ce91679c9aaceb3e9c930f11e788aacbfa8341a2a5737583c14a4d6666f3d.jpg
    ├── iris-a0e65269f7dc7801ac1ad8bd0c5aa547a70c7655447e921d1d4d153a9d23815e.jpg
    ├── iris-a445b0720254984275097c83afbdb1fe896cb010b5c662a6532ed0601ea24d7c.jpg
    ├── iris-a6b85bf1f3d18bbb6470440592834c2c7f081b490836392cf5f01636ee7cf658.jpg
    ├── iris-b005c82b844de575f0b972b9a1797b2b1fbe98c067c484a51006afc4f549ada4.jpg
    ├── iris-bfcf79b3b527eb64b78f9a068a1000042336e532f0f44e68f818dd13ab492a76.jpg
    ├── iris-c156236fb6e888764485e796f1f972bbc7ad960fe6330a7ce9182922046439c4.jpg
    ├── iris-d99d5fd2de5be1419cbd569570dbb6c9a6c8ec4f0a1ff5b55dc2607f6ecdca8f.jpg
    ├── iris-d9aae37a8fa6afdef2af170c266a597925eea935f4d070e979d565713ea62642.jpg
    ├── iris-dbc87fcecade2c070baaf99caf03f4f0f6e3aa977e34972383cb94d0efe8a95d.jpg
    ├── iris-e3d1a560d25cf573d2cbbf2fe6cd231819e998109a5cf1788d59fbb9859b3be2.jpg
    ├── iris-ec288bdad71388f907457db2476f12a5cb43c28cfa28d2a2077398a42b948a35.jpg
    ├── iris-ed5b4e072d43bc53a00a4a7f4d0f5d7c0cbd6a006e9c2d463128cedc956cb3de.jpg
    ├── iris-f3018a9440d17c265062d1c61475127f9952b6fe951d38fd7700402d706c0b01.jpg
    ├── iris-f47c5963cdbaa3238ba2d446848e8449c6af83e663f0a9216cf0baba8429b36f.jpg
    ├── iris-fa4b6d7e3617216104b1405cda21bf234840cd84a2c1966034caa63def2f64f0.jpg
    ├── iris-fc4b0cc65387ff78471659d14a78f0309a76f4c3ec641b871e40b40424255097.jpg
    └── meta.csv

Let’s create an artifact for the first sub-folder.

artifact = ln.Artifact("s3://lamindata/iris_studies/study0_raw_images").save()
artifact

artifact = ln$Artifact("s3://lamindata/iris_studies/study0_raw_images")$save()
artifact
Hide code cell content 
artifact = ln.Artifact("s3://lamindata/iris_studies/study0_raw_images").save()
artifact

Artifact(uid='hhECqGAHiBTEnU8M0000', is_latest=True, key='iris_studies/study0_raw_images', suffix='', size=658465, hash='IVKGMfNwi8zKvnpaD_gG7w', n_files=51, space_id=1, storage_id=2, run_id=1, created_by_id=1, created_at=2025-04-03 08:48:40 UTC)

As you see from path, the folder was merely registered in its present storage location without copying it.

artifact.path

artifact$path
Hide code cell content 
artifact.path

S3QueryPath('s3://lamindata/iris_studies/study0_raw_images')

LaminDB keeps track of all your storage locations.

ln.Storage.df()

ln$Storage$df()
Hide code cell content 
ln.Storage.df()
uid root description type region instance_uid space_id run_id created_at created_by_id _aux _branch_code
id
2 OH8uNPcWtyac s3://lamindata None s3 us-east-1 None 1 None 2025-04-03 08:48:39.987000+00:00 1 None 1
1 pLjWpK154J0j /home/runner/work/lamin-docs/lamin-docs/docs/l... None local None 3MepSh2Col3I 1 None 2025-04-03 08:48:35.176000+00:00 1 None 1

To cache the cloud folder locally, call cache().

artifact.cache()

artifact$cache()
Hide code cell content 
artifact.cache()

PosixUPath('/home/runner/.cache/lamindb/lamindata/iris_studies/study0_raw_images')

If the data is large, you might not want to download but stream it via open(). For more on this, see: Slice arrays.

How do I update or delete an artifact?
artifact.description = "My new description"  # change description
artifact.save()  # save the change to the database
artifact.delete()  # move to trash
artifact.delete(permanent=True)  # permanently delete
How do I create an artifact for a local file or folder?

Source path is local:

ln.Artifact("./my_data.fcs", key="my_data.fcs")
ln.Artifact("./my_images/", key="my_images")

Upon artifact.save(), the source path will be copied or uploaded into your instance’s current storage, visible & changeable via ln.settings.storage.

If the source path is remote or already in a registered storage location (one that’s registered in ln.Storage), artifact.save() will not trigger a copy or upload but register the existing path.

ln.Artifact("s3://my-bucket/my_data.fcs")  # key is auto-populated from S3, you can optionally pass a description
ln.Artifact("s3://my-bucket/my_images/")  # key is auto-populated from S3, you can optionally pass a description

You can use any storage location supported by `fsspec`.
Which fields are populated when creating an artifact record?

Basic fields:

  • uid: universal ID

  • key: a (virtual) relative path of the artifact in storage

  • description: an optional string description

  • storage: the storage location (the root, say, an S3 bucket or a local directory)

  • suffix: an optional file/path suffix

  • size: the artifact size in bytes

  • hash: a hash useful to check for integrity and collisions (is this artifact already stored?)

  • hash_type: the type of the hash

  • created_at: time of creation

  • updated_at: time of last update

Provenance-related fields:

  • created_by: the User who created the artifact

  • run: the Run of the Transform that created the artifact

For a full reference, see Artifact.

What exactly happens during save?

In the database: An artifact record is inserted into the Artifact registry. If the artifact record exists already, it’s returned.

In storage:

  • If the default storage is in the cloud, .save() triggers an upload for a local artifact.

  • If the artifact is already in a registered storage location, only the metadata of the record is saved to the artifact registry.

How does LaminDB compare to a AWS S3?

LaminDB provides a database on top of AWS S3 (or GCP storage, file systems, etc.).

Similar to organizing files with paths, you can organize artifacts using the key parameter of Artifact.

However, you’ll see that you can more conveniently query data by entities you care about: people, code, experiments, genes, proteins, cell types, etc.

Are artifacts aware of array-like data?

Yes.

You can make artifacts from paths referencing array-like objects:

ln.Artifact("./my_anndata.h5ad", key="my_anndata.h5ad")
ln.Artifact("./my_zarr_array/", key="my_zarr_array")

Or from in-memory objects:

ln.Artifact.from_df(df, key="my_dataframe.parquet")
ln.Artifact.from_anndata(adata, key="my_anndata.h5ad")

You can open large artifacts for slicing from the cloud or load small artifacts directly into memory via:

artifact.open()

Query & search registries

To get an overview over all artifacts in your instance, call df.

ln.Artifact.df()

ln$Artifact$df()
Hide code cell content 
ln.Artifact.df()
uid key description suffix kind otype size hash n_files n_observations _hash_type _key_is_virtual _overwrite_versions space_id storage_id schema_id version is_latest run_id created_at created_by_id _aux _branch_code
id
3 hhECqGAHiBTEnU8M0000 iris_studies/study0_raw_images None None None 658465 IVKGMfNwi8zKvnpaD_gG7w 51.0 NaN md5-d False True 1 2 None None True 1 2025-04-03 08:48:40.058000+00:00 1 None 1
2 fPTOf1qAg8pNJqRH0001 my_datasets/rnaseq1.parquet None .parquet dataset DataFrame 9012 iBiiWBkIitgFtLcru2CLyA NaN 3.0 md5 True False 1 1 None None True 1 2025-04-03 08:48:39.019000+00:00 1 None 1
1 fPTOf1qAg8pNJqRH0000 my_datasets/rnaseq1.parquet None .parquet dataset DataFrame 9012 ZHlfaXCXxza090J-PA1nCg NaN 3.0 md5 True False 1 1 None None False 1 2025-04-03 08:48:38.599000+00:00 1 None 1

LaminDB’s central classes are registries that store records (Record objects). If you want to see the fields of a registry, look at the class or auto-complete.

ln.Artifact

ln$Artifact
Hide code cell content 
ln.Artifact

Artifact
  Simple fields
    .uid: CharField
    .key: CharField
    .description: CharField
    .suffix: CharField
    .kind: CharField
    .otype: CharField
    .size: BigIntegerField
    .hash: CharField
    .n_files: BigIntegerField
    .n_observations: BigIntegerField
    .version: CharField
    .is_latest: BooleanField
    .created_at: DateTimeField
    .updated_at: DateTimeField
  Relational fields
    .space: Space
    .storage: Storage
    .run: Run
    .schema: Schema
    .created_by: User
    .ulabels: ULabel
    .input_of_runs: Run
    .feature_sets: Schema
    .collections: Collection
    .references: Reference
    .projects: Project
  Bionty fields
    .organisms: bionty.Organism
    .genes: bionty.Gene
    .proteins: bionty.Protein
    .cell_markers: bionty.CellMarker
    .tissues: bionty.Tissue
    .cell_types: bionty.CellType
    .diseases: bionty.Disease
    .cell_lines: bionty.CellLine
    .phenotypes: bionty.Phenotype
    .pathways: bionty.Pathway
    .experimental_factors: bionty.ExperimentalFactor
    .developmental_stages: bionty.DevelopmentalStage
    .ethnicities: bionty.Ethnicity

Each registry is a table in the relational schema of the underlying database. With view(), you can see the latest changes to the database.

ln.view()

ln$view()
Hide code cell content 
ln.view()

****************

* module: core *

****************

Artifact
uid key description suffix kind otype size hash n_files n_observations _hash_type _key_is_virtual _overwrite_versions space_id storage_id schema_id version is_latest run_id created_at created_by_id _aux _branch_code
id
3 hhECqGAHiBTEnU8M0000 iris_studies/study0_raw_images None None None 658465 IVKGMfNwi8zKvnpaD_gG7w 51.0 NaN md5-d False True 1 2 None None True 1 2025-04-03 08:48:40.058000+00:00 1 None 1
2 fPTOf1qAg8pNJqRH0001 my_datasets/rnaseq1.parquet None .parquet dataset DataFrame 9012 iBiiWBkIitgFtLcru2CLyA NaN 3.0 md5 True False 1 1 None None True 1 2025-04-03 08:48:39.019000+00:00 1 None 1
1 fPTOf1qAg8pNJqRH0000 my_datasets/rnaseq1.parquet None .parquet dataset DataFrame 9012 ZHlfaXCXxza090J-PA1nCg NaN 3.0 md5 True False 1 1 None None False 1 2025-04-03 08:48:38.599000+00:00 1 None 1 
Run
uid name started_at finished_at reference reference_type _is_consecutive _status_code space_id transform_id report_id _logfile_id environment_id initiated_by_run_id created_at created_by_id _aux _branch_code
id
1 rSjKb6jCHjWm3sjAPQBu None 2025-04-03 08:48:38.080656+00:00 None None None None 0 1 1 None None None None 2025-04-03 08:48:38.081000+00:00 1 None 1 
Storage
uid root description type region instance_uid space_id run_id created_at created_by_id _aux _branch_code
id
2 OH8uNPcWtyac s3://lamindata None s3 us-east-1 None 1 None 2025-04-03 08:48:39.987000+00:00 1 None 1
1 pLjWpK154J0j /home/runner/work/lamin-docs/lamin-docs/docs/l... None local None 3MepSh2Col3I 1 None 2025-04-03 08:48:35.176000+00:00 1 None 1 
Transform
uid key description type source_code hash reference reference_type space_id _template_id version is_latest created_at created_by_id _aux _branch_code
id
1 ZTGPZqWSil700000 introduction.ipynb Introduction notebook None None None None 1 None None True 2025-04-03 08:48:38.074000+00:00 1 None 1 
******************

* module: bionty *

******************

Source
uid entity organism name in_db currently_used description url md5 source_website space_id dataframe_artifact_id version run_id created_at created_by_id _aux _branch_code
id
53 5Xov8Lap bionty.Disease all mondo False False Mondo Disease Ontology http://purl.obolibrary.org/obo/mondo/releases/... None https://mondo.monarchinitiative.org 1 None 2024-02-06 None 2025-04-03 08:48:35.263000+00:00 1 None 1
54 69lnSXfR bionty.Disease all mondo False False Mondo Disease Ontology http://purl.obolibrary.org/obo/mondo/releases/... None https://mondo.monarchinitiative.org 1 None 2024-01-03 None 2025-04-03 08:48:35.263000+00:00 1 None 1
55 4ss2Hizg bionty.Disease all mondo False False Mondo Disease Ontology http://purl.obolibrary.org/obo/mondo/releases/... None https://mondo.monarchinitiative.org 1 None 2023-08-02 None 2025-04-03 08:48:35.263000+00:00 1 None 1
56 Hgw08Vk3 bionty.Disease all mondo False False Mondo Disease Ontology http://purl.obolibrary.org/obo/mondo/releases/... None https://mondo.monarchinitiative.org 1 None 2023-04-04 None 2025-04-03 08:48:35.263000+00:00 1 None 1
57 UUZUtULu bionty.Disease all mondo False False Mondo Disease Ontology http://purl.obolibrary.org/obo/mondo/releases/... None https://mondo.monarchinitiative.org 1 None 2023-02-06 None 2025-04-03 08:48:35.263000+00:00 1 None 1
58 7DH1aJIr bionty.Disease all mondo False False Mondo Disease Ontology http://purl.obolibrary.org/obo/mondo/releases/... None https://mondo.monarchinitiative.org 1 None 2022-10-11 None 2025-04-03 08:48:35.263000+00:00 1 None 1
59 4kswnHVF bionty.Disease human doid False True Human Disease Ontology http://purl.obolibrary.org/obo/doid/releases/2... None https://disease-ontology.org 1 None 2024-05-29 None 2025-04-03 08:48:35.263000+00:00 1 None 1
Which registries have I already learned about? 🤔

  • Artifact: datasets & models stored as files, folders, or arrays

  • Transform: transforms of artifacts

  • Run: runs of transforms

  • User: users

  • Storage: local or cloud storage locations

Every registry supports arbitrary relational queries using the class methods get and filter. The syntax for it is Django’s query syntax.

Here are some simple query examples.

# get a single record (here the current notebook)
transform = ln.Transform.get(key="introduction.ipynb")

# get a set of records by filtering for a directory (LaminDB treats directories like AWS S3, as the prefix of the storage key)
ln.Artifact.filter(key__startswith="my_datasets/").df()

# query all artifacts ingested from a transform
artifacts = ln.Artifact.filter(transform=transform).all()

# query all artifacts ingested from a notebook with "intro" in the title
artifacts = ln.Artifact.filter(
    transform__description__icontains="intro",
).all()

# get a single record (here the current notebook)
transform <- ln$Transform$get(key = "introduction.Rmd")

# get a set of records by filtering for a directory (LaminDB treats directories like AWS S3, as the prefix of the storage key)
ln$Artifact$filter(key__startswith = "my_datasets/")$df()

# query all artifacts ingested from a transform
artifacts <- ln$Artifact$filter(transform = transform)$all()

# query all artifacts ingested from a notebook with "intro" in the title
artifacts <- ln$Artifact$filter(
  transform__description__icontains = "intro",
)$all()
Hide code cell content 
# get a single record (here the current notebook)
transform = ln.Transform.get(key="introduction.ipynb")

# get a set of records by filtering for a directory (LaminDB treats directories like AWS S3, as the prefix of the storage key)
ln.Artifact.filter(key__startswith="my_datasets/").df()

# query all artifacts ingested from a transform
artifacts = ln.Artifact.filter(transform=transform).all()

# query all artifacts ingested from a notebook with "intro" in the title
artifacts = ln.Artifact.filter(
    transform__description__icontains="intro",
).all()
What does a double underscore mean?

For any field, the double underscore defines a comparator, e.g.,

  • name__icontains="Martha": name contains "Martha" when ignoring case

  • name__startswith="Martha": name starts with "Martha

  • name__in=["Martha", "John"]: name is "John" or "Martha"

For more info, see: Query & search registries.

Can I chain filters and searches?

Yes: ln.Artifact.filter(suffix=".jpg").search("my image")

The class methods search and lookup help with approximate matches.

# search artifacts
ln.Artifact.search("iris").df().head()

# search transforms
ln.Transform.search("intro").df()

# look up records with auto-complete
ulabels = ln.ULabel.lookup()

# search artifacts
ln$Artifact$search("iris")$df()

# search transforms
ln$Transform$search("intro")$df()

# look up records with auto-complete
ulabels = ln$ULabel$lookup()
Show me a screenshot

For more info, see: Query & search registries.

Features & labels

Features & labels make it easier to find datasets and help standardizing them so that they’re re-usable by analysts and machine learning models alike. Features are measurement dimensions (e.g. "species", "temperature") and labels are measured values (e.g. "human", "mouse"). In stats, a feature is a variable while a label is a category.

Can you give me examples for what findability and usability means?

  1. Findability: Which datasets measured expression of cell marker CD14? Which characterized cell line K562? Which have a test & train split? Etc.

  2. Usability: Are there typos in feature names? Are there typos in labels? Are types and units of features consistent? Etc.

Let’s annotate an artifact with a ULabel, a built-in universal label ontology.

# create & save a typed label
experiment_type = ln.ULabel(name="InVitroStudy", is_type=True).save()
my_experiment = ln.ULabel(name="My experiment", type=experiment_type).save()

# annotate the artifact with a label
artifact.ulabels.add(my_experiment)

# describe the artifact
artifact.describe()

# create & save a typed label
experiment_type = ln$ULabel(name="InVitroStudy", is_type=True)$save()
my_experiment = ln$ULabel(name="My experiment", type=experiment_type)$save()

# annotate the artifact with a label
artifact$ulabels$add(my_experiment)

# describe the artifact
artifact$describe()
Hide code cell content 
# create & save a typed ulabel
experiment_type = ln.ULabel(name="InVitroStudy", is_type=True).save()
my_experiment = ln.ULabel(name="My experiment", type=experiment_type).save()

# annotate the artifact with a ulabel
artifact.ulabels.add(my_experiment)

# describe the artifact
artifact.describe()

Artifact 
├── General
│   ├── .uid = 'hhECqGAHiBTEnU8M0000'
│   ├── .key = 'iris_studies/study0_raw_images'
│   ├── .size = 658465
│   ├── .hash = 'IVKGMfNwi8zKvnpaD_gG7w'
│   ├── .n_files = 51
│   ├── .path = s3://lamindata/iris_studies/study0_raw_images
│   ├── .created_by = anonymous
│   ├── .created_at = 2025-04-03 08:48:40
│   └── .transform = 'Introduction'
└── Labels
    └── .ulabels                    ULabel                     My experiment

This is how you query artifacts by ulabels.

ln.Artifact.filter(ulabels=my_experiment).df()

ln$Artifact$filter(ulabels=my_experiment)$df()
Hide code cell content 
ln.Artifact.filter(ulabels=my_experiment).df()
uid key description suffix kind otype size hash n_files n_observations _hash_type _key_is_virtual _overwrite_versions space_id storage_id schema_id version is_latest run_id created_at created_by_id _aux _branch_code
id
3 hhECqGAHiBTEnU8M0000 iris_studies/study0_raw_images None None None 658465 IVKGMfNwi8zKvnpaD_gG7w 51 None md5-d False True 1 2 None None True 1 2025-04-03 08:48:40.058000+00:00 1 None 1

You can also annotate with labels from other registries, e.g., the biological ontologies in bionty.

import bionty as bt

# create a cell type label from the source ontology
cell_type = bt.CellType.from_source(name="effector T cell").save()

# annotate the artifact with a cell type
artifact.cell_types.add(cell_type)

# describe the artifact
artifact.describe()

bt <- import_module("bionty")

# create a cell type label from the source ontology
cell_type <- bt$CellType$from_source(name = "effector T cell")$save()

# annotate the artifact with a cell type
artifact$cell_types$add(cell_type)

# describe the artifact
artifact$describe()
Hide code cell content 
import bionty as bt

# create a cell type label from the source ontology
cell_type = bt.CellType.from_source(name="effector T cell").save()

# annotate the artifact with a cell type
artifact.cell_types.add(cell_type)

# describe the artifact
artifact.describe()

Artifact 
├── General
│   ├── .uid = 'hhECqGAHiBTEnU8M0000'
│   ├── .key = 'iris_studies/study0_raw_images'
│   ├── .size = 658465
│   ├── .hash = 'IVKGMfNwi8zKvnpaD_gG7w'
│   ├── .n_files = 51
│   ├── .path = s3://lamindata/iris_studies/study0_raw_images
│   ├── .created_by = anonymous
│   ├── .created_at = 2025-04-03 08:48:40
│   └── .transform = 'Introduction'
└── Labels
    └── .cell_types                 bionty.CellType            effector T cell                          
        .ulabels                    ULabel                     My experiment

If you want to annotate by non-categorical metadata or indicate the feature for a label, annotate via features.

# define the "temperature" & "experiment" features
ln.Feature(name="temperature", dtype=float).save()
ln.Feature(name="experiment", dtype=ln.ULabel).save()

# annotate the artifact
artifact.features.add_values({"temperature": 21.6, "experiment": "My experiment"})

# describe the artifact
artifact.describe()

# define the "temperature" & "experiment" features
ln$Feature(name = "temperature", dtype = "float")$save()
ln$Feature(name = "experiment", dtype = ln$ULabel)$save()

# annotate the artifact
artifact$features$add_values(
  list("temperature" = 21.6, "experiment" = "My experiment")
)

# describe the artifact
artifact$describe()
Hide code cell content 
# define the "temperature" & "experiment" features
ln.Feature(name="temperature", dtype=float).save()
ln.Feature(name="experiment", dtype=ln.ULabel).save()

# annotate the artifact
artifact.features.add_values({"temperature": 21.6, "experiment": "My experiment"})

# describe the artifact
artifact.describe()

Artifact 
├── General
│   ├── .uid = 'hhECqGAHiBTEnU8M0000'
│   ├── .key = 'iris_studies/study0_raw_images'
│   ├── .size = 658465
│   ├── .hash = 'IVKGMfNwi8zKvnpaD_gG7w'
│   ├── .n_files = 51
│   ├── .path = s3://lamindata/iris_studies/study0_raw_images
│   ├── .created_by = anonymous
│   ├── .created_at = 2025-04-03 08:48:40
│   └── .transform = 'Introduction'
├── Linked features
│   └── experiment                  cat[ULabel]                My experiment                            
│       temperature                 float                      21.6                                     
└── Labels
    └── .cell_types                 bionty.CellType            effector T cell                          
        .ulabels                    ULabel                     My experiment

Curate datasets

You already saw how to ingest datasets without validation. This is often enough if you’re prototyping or working with one-off studies. But if you want to create a big body of standardized data, you have to invest the time to curate your datasets.

Let’s define a Schema to curate a DataFrame.

# define valid labels
perturbation_type = ln.ULabel(name="Perturbation", is_type=True).save()
ln.ULabel(name="DMSO", type=perturbation_type).save()
ln.ULabel(name="IFNG", type=perturbation_type).save()

# define the schema
schema = ln.Schema(
    name="My DataFrame schema",
    features=[
        ln.Feature(name="perturbation", dtype=ln.ULabel).save(),
        ln.Feature(name="cell_type_by_model", dtype=bt.CellType).save(),
    ],
).save()

With a Curator, we can save an annotated & validated artifact with a single line of code.

# create a curator
curator = ln.curators.DataFrameCurator(df, schema)

# save curated artifact
artifact = curator.save_artifact(key="my_curated_dataset.parquet")  # calls .validate()

# see the parsed annotations
artifact.describe()

# query for a ulabel that was parsed from the dataset
ln.Artifact.get(ulabels__name="IFNG")
Hide code cell output 
• saving validated records of 'cell_type_by_model'

✓ added 1 record from public with CellType.name for "cell_type_by_model": 'B cell'

✓ "perturbation" is validated against ULabel.name

✓ "cell_type_by_model" is validated against CellType.name

→ returning existing artifact with same hash: Artifact(uid='fPTOf1qAg8pNJqRH0001', is_latest=True, key='my_datasets/rnaseq1.parquet', suffix='.parquet', kind='dataset', otype='DataFrame', size=9012, hash='iBiiWBkIitgFtLcru2CLyA', n_observations=3, space_id=1, storage_id=1, run_id=1, created_by_id=1, created_at=2025-04-03 08:48:39 UTC); to track this artifact as an input, use: ln.Artifact.get()

! key my_datasets/rnaseq1.parquet on existing artifact differs from passed key my_curated_dataset.parquet

✓ 2 unique terms (18.20%) are validated for name

! 9 unique terms (81.80%) are not validated for name: 'ENSG00000153563', 'ENSG00000010610', 'ENSG00000170458', 'sample_note', 'cell_type_by_expert', 'assay_oid', 'concentration', 'treatment_time_h', 'donor'

✓ loaded 2 Feature records matching name: 'perturbation', 'cell_type_by_model'

! did not create Feature records for 9 non-validated names: 'ENSG00000010610', 'ENSG00000153563', 'ENSG00000170458', 'assay_oid', 'cell_type_by_expert', 'concentration', 'donor', 'sample_note', 'treatment_time_h'

→ returning existing schema with same hash: Schema(uid='AioylZkuGqgR3cHF86OQ', name='My DataFrame schema', n=2, itype='Feature', is_type=False, hash='9yxkOigtfQfBCwDkqGYxkA', minimal_set=True, ordered_set=False, maximal_set=False, space_id=1, created_by_id=1, run_id=1, created_at=2025-04-03 08:48:43 UTC)

! updated otype from None to DataFrame

Artifact .parquet/DataFrame
├── General
│   ├── .uid = 'fPTOf1qAg8pNJqRH0001'
│   ├── .key = 'my_datasets/rnaseq1.parquet'
│   ├── .size = 9012
│   ├── .hash = 'iBiiWBkIitgFtLcru2CLyA'
│   ├── .n_observations = 3
│   ├── .path = /home/runner/work/lamin-docs/lamin-docs/docs/lamin-intro/.lamindb/fPTOf1qAg8pNJqRH0001.parquet
│   ├── .created_by = anonymous
│   ├── .created_at = 2025-04-03 08:48:39
│   └── .transform = 'Introduction'
├── Dataset features/.feature_sets
│   └── columns • 2                 [Feature]                                                           
│       cell_type_by_model          cat[bionty.CellType]       B cell, T cell                           
│       perturbation                cat[ULabel]                DMSO, IFNG                               
└── Labels
    └── .cell_types                 bionty.CellType            T cell, B cell                           
        .ulabels                    ULabel                     DMSO, IFNG                               

Artifact(uid='fPTOf1qAg8pNJqRH0001', is_latest=True, key='my_datasets/rnaseq1.parquet', suffix='.parquet', kind='dataset', otype='DataFrame', size=9012, hash='iBiiWBkIitgFtLcru2CLyA', n_observations=3, space_id=1, storage_id=1, run_id=1, schema_id=1, created_by_id=1, created_at=2025-04-03 08:48:39 UTC)

If we feed a dataset with an invalid dtype or typo, we’ll get a ValidationError.

curator = ln.curators.DataFrameCurator(df_typo, schema)

# validate the dataset
try:
    curator.validate()
except ln.errors.ValidationError as error:
    print(str(error))
Hide code cell output 
• mapping "perturbation" on ULabel.name

!   1 term is not validated: 'IFNJ'
    → fix typos, remove non-existent values, or save terms via .add_new_from("perturbation")

✓ "cell_type_by_model" is validated against CellType.name

1 term is not validated: 'IFNJ'
    → fix typos, remove non-existent values, or save terms via .add_new_from("perturbation")

For more info, see Curate datasets.

Manage biological registries

The generic Feature and ULabel registries will get you pretty far.

But let’s now look at what you do can with a dedicated biological registry like Gene.

Every bionty registry is based on configurable public ontologies (>20 of them).

import bionty as bt

cell_types = bt.CellType.public()
cell_types
Hide code cell output 
PublicOntology
Entity: CellType
Organism: all
Source: cl, 2024-08-16
#terms: 2959

cell_types.search("gamma-delta T cell").head(2)
Hide code cell output
name definition synonyms parents
ontology_id
CL:0000798 gamma-delta T cell A T Cell That Expresses A Gamma-Delta T Cell R... gamma-delta T-cell|gamma-delta T lymphocyte|ga... [CL:0000084]
CL:4033072 cycling gamma-delta T cell A(N) Gamma-Delta T Cell That Is Cycling. proliferating gamma-delta T cell [CL:4033069, CL:0000798]

Define an AnnData schema.

# define var schema
var_schema = ln.Schema(
    name="my_var_schema",
    itype=bt.Gene.ensembl_gene_id,
    dtype=int,
).save()

# define composite schema
anndata_schema = ln.Schema(
    name="my_anndata_schema",
    otype="AnnData",
    components={"obs": schema, "var": var_schema},
).save()

Validate & annotate an AnnData.

import anndata as ad
import bionty as bt

# store the dataset as an AnnData object to distinguish data from metadata
adata = ad.AnnData(df.iloc[:, :3], obs=df.iloc[:, 3:])

# save curated artifact
curator = ln.curators.AnnDataCurator(adata, anndata_schema)
artifact = curator.save_artifact(key="my_datasets/my_rnaseq1.h5ad")
artifact.describe()
Hide code cell output 
✓ created 1 Organism record from Bionty matching name: 'human'

• saving validated records of 'columns'

✓ added 3 records from public with Gene.ensembl_gene_id for "columns": 'ENSG00000153563', 'ENSG00000010610', 'ENSG00000170458'

✓ "perturbation" is validated against ULabel.name

✓ "cell_type_by_model" is validated against CellType.name

• path content will be copied to default storage upon `save()` with key 'my_datasets/my_rnaseq1.h5ad'

✓ storing artifact 'Knm0RCqX26Glfk010000' at '/home/runner/work/lamin-docs/lamin-docs/docs/lamin-intro/.lamindb/Knm0RCqX26Glfk010000.h5ad'

✓ 3 unique terms (100.00%) are validated for ensembl_gene_id

✓ 2 unique terms (25.00%) are validated for name

! 6 unique terms (75.00%) are not validated for name: 'sample_note', 'cell_type_by_expert', 'assay_oid', 'concentration', 'treatment_time_h', 'donor'

✓ loaded 2 Feature records matching name: 'perturbation', 'cell_type_by_model'

! did not create Feature records for 6 non-validated names: 'assay_oid', 'cell_type_by_expert', 'concentration', 'donor', 'sample_note', 'treatment_time_h'

→ returning existing schema with same hash: Schema(uid='AioylZkuGqgR3cHF86OQ', name='My DataFrame schema', n=2, itype='Feature', is_type=False, hash='9yxkOigtfQfBCwDkqGYxkA', minimal_set=True, ordered_set=False, maximal_set=False, space_id=1, created_by_id=1, run_id=1, created_at=2025-04-03 08:48:43 UTC)

! updated otype from None to DataFrame

✓ saved 1 feature set for slot: 'var'

Artifact .h5ad/AnnData
├── General
│   ├── .uid = 'Knm0RCqX26Glfk010000'
│   ├── .key = 'my_datasets/my_rnaseq1.h5ad'
│   ├── .size = 30232
│   ├── .hash = 'TWUlDi-V-Try1uCxWJ65nQ'
│   ├── .n_observations = 3
│   ├── .path = /home/runner/work/lamin-docs/lamin-docs/docs/lamin-intro/.lamindb/Knm0RCqX26Glfk010000.h5ad
│   ├── .created_by = anonymous
│   ├── .created_at = 2025-04-03 08:48:46
│   └── .transform = 'Introduction'
├── Dataset features/.feature_sets
│   ├── var • 3                     [bionty.Gene]                                                       
│   │   CD8A                        int                                                                 
│   │   CD4                         int                                                                 
│   │   CD14                        int                                                                 
│   └── obs • 2                     [Feature]                                                           
│       cell_type_by_model          cat[bionty.CellType]       B cell, T cell                           
│       perturbation                cat[ULabel]                DMSO, IFNG                               
└── Labels
    └── .cell_types                 bionty.CellType            T cell, B cell                           
        .ulabels                    ULabel                     DMSO, IFNG

Query for typed features.

# get a lookup object for human genes
genes = bt.Gene.filter(organism__name="human").lookup()
# query for all feature sets that contain CD8A
feature_sets = ln.Schema.filter(genes=genes.cd8a).all()
# write the query
ln.Artifact.filter(feature_sets__in=feature_sets).df()
Hide code cell output
uid key description suffix kind otype size hash n_files n_observations _hash_type _key_is_virtual _overwrite_versions space_id storage_id schema_id version is_latest run_id created_at created_by_id _aux _branch_code
id
4 Knm0RCqX26Glfk010000 my_datasets/my_rnaseq1.h5ad None .h5ad dataset AnnData 30232 TWUlDi-V-Try1uCxWJ65nQ None 3 md5 True False 1 1 3 None True 1 2025-04-03 08:48:46.247000+00:00 1 None 1

Update ontologies, e.g., create a cell type record and add a new cell state.

# create an ontology-coupled cell type record and save it
neuron = bt.CellType.from_source(name="neuron").save()

# create a record to track a new cell state
new_cell_state = bt.CellType(
    name="my neuron cell state", description="explains X"
).save()

# express that it's a neuron state
new_cell_state.parents.add(neuron)

# view ontological hierarchy
new_cell_state.view_parents(distance=2)
Hide code cell output 
✓ created 1 CellType record from Bionty matching name: 'neuron'

✓ created 3 CellType records from Bionty matching ontology_id: 'CL:0002319', 'CL:0000404', 'CL:0000393'
_images/8f9ed16972ab4f7fcc9549176de33247e2eb3f308e7778bd32d148c2df631926.svg

Scale learning

How do you integrate new datasets with your existing datasets? Leverage Collection.

# a new dataset
df2 = ln.core.datasets.small_dataset2(otype="DataFrame")
adata = ad.AnnData(df2.iloc[:, :3], obs=df2.iloc[:, 3:])
curator = ln.curators.AnnDataCurator(adata, anndata_schema)
artifact2 = curator.save_artifact(key="my_datasets/my_rnaseq2.h5ad")
Hide code cell output 
• saving validated records of 'columns'

✓ added 1 record from public with Gene.ensembl_gene_id for "columns": 'ENSG00000004468'

✓ "perturbation" is validated against ULabel.name

✓ "cell_type_by_model" is validated against CellType.name

• path content will be copied to default storage upon `save()` with key 'my_datasets/my_rnaseq2.h5ad'

✓ storing artifact 'tcTI5lBFjjxMSrUD0000' at '/home/runner/work/lamin-docs/lamin-docs/docs/lamin-intro/.lamindb/tcTI5lBFjjxMSrUD0000.h5ad'

✓ 3 unique terms (100.00%) are validated for ensembl_gene_id

✓ 2 unique terms (100.00%) are validated for name

→ returning existing schema with same hash: Schema(uid='AioylZkuGqgR3cHF86OQ', name='My DataFrame schema', n=2, itype='Feature', is_type=False, hash='9yxkOigtfQfBCwDkqGYxkA', minimal_set=True, ordered_set=False, maximal_set=False, space_id=1, created_by_id=1, run_id=1, created_at=2025-04-03 08:48:43 UTC)

! updated otype from None to DataFrame

✓ saved 1 feature set for slot: 'var'

Create a collection using Collection.

collection = ln.Collection([artifact, artifact2], key="my-RNA-seq-collection").save()
collection.describe()
collection.view_lineage()
Hide code cell output 
Collection 
└── General
    ├── .uid = 'nQfbfkXVP3pr6O8v0000'
    ├── .key = 'my-RNA-seq-collection'
    ├── .hash = 'diHtydJXZXAzmmM4xVyXsg'
    ├── .created_by = anonymous
    ├── .created_at = 2025-04-03 08:48:49
    └── .transform = 'Introduction'
_images/0feb03979b75bf7155f62c8fbb72bcd0cdacad008577bcebe21964bdf7129b80.svg
# if it's small enough, you can load the entire collection into memory as if it was one
collection.load()

# typically, it's too big, hence, open it for streaming (if the backend allows it)
# collection.open()

# or iterate over its artifacts
collection.artifacts.all()

# or look at a DataFrame listing the artifacts
collection.artifacts.df()
Hide code cell output
uid key description suffix kind otype size hash n_files n_observations _hash_type _key_is_virtual _overwrite_versions space_id storage_id schema_id version is_latest run_id created_at created_by_id _aux _branch_code
id
4 Knm0RCqX26Glfk010000 my_datasets/my_rnaseq1.h5ad None .h5ad dataset AnnData 30232 TWUlDi-V-Try1uCxWJ65nQ None 3 md5 True False 1 1 3 None True 1 2025-04-03 08:48:46.247000+00:00 1 None 1
5 tcTI5lBFjjxMSrUD0000 my_datasets/my_rnaseq2.h5ad None .h5ad dataset AnnData 21224 HGR5w3RPp7HiFaM6wInRPw None 3 md5 True False 1 1 3 None True 1 2025-04-03 08:48:48.951000+00:00 1 None 1

Directly train models on collections of AnnData.

# to train models, batch iterate through the collection as if it was one array
from torch.utils.data import DataLoader, WeightedRandomSampler
dataset = collection.mapped(obs_keys=["cell_medium"])
sampler = WeightedRandomSampler(
    weights=dataset.get_label_weights("cell_medium"), num_samples=len(dataset)
)
data_loader = DataLoader(dataset, batch_size=2, sampler=sampler)
for batch in data_loader:
    pass

Read this blog post for more on training models on distributed datasets.