Phenomics Imaging · recursionpharma
rxrx3-core
Curated phenomics dataset of Cell Painting images and embeddings for benchmarking drug-target interactions.
What rxrx3-core is
RxRx3-core is a curated phenomics dataset released by Recursion to benchmark drug-target interactions and train representation models using high-content microscopy. Designed as a standardized challenge resource, it pairs genetic knockouts with small-molecule compound treatments across varying concentrations.
The dataset contains six-channel Cell Painting images and precomputed embeddings spanning 222,601 microplate wells in human umbilical vein endothelial cells (HUVEC). It captures 735 genetic knockouts and 1,674 small-molecule perturbations, linking compound-induced phenotypic shifts with specific target gene profiles.
By compressing a multi-condition concentration-response screen into a distribution of under 18 GB, RxRx3-core lowers the computational barrier to evaluating cellular representations, phenotypic mechanism-of-action mapping, and phenomic foundation models.
How it was made
RxRx3-core is sampled from the larger RxRx3 repository. Wells receive either CRISPR-mediated gene perturbations or small-molecule chemical treatments tested across multiple concentrations in micromolar units.
Each well is imaged across six fluorescence microscopy channels standard to Cell Painting assays. The dataset provides tabular metadata detailing well locations, canonical SMILES, compound concentrations, gene targets, and control designations, alongside precomputed image embeddings generated by OpenPhenom, MAE-L/8, and MAE-G/8 foundation models stored in Parquet format.
Key facts
- Cell type
- Human umbilical vein endothelial cells (HUVEC)
- Number of wells
- 222,601
- Genetic knockouts
- 735
- Compound perturbations
- 1,674
- Assay modality
- 6-channel Cell Painting fluorescence microscopy
- Precomputed embeddings
- OpenPhenom, MAE-L/8, MAE-G/8
- Storage footprint
- Under 18 GB
- File formats
- Parquet, CSV
What you can use it for
- Drug-target interaction benchmarking
- Evaluate algorithms on their ability to match small-molecule phenotypic signatures to the corresponding CRISPR-induced target gene phenotypes across concentrations.
- Foundation model evaluation
- Test and compare visual representation learning architectures, such as masked autoencoders and phenomic models, using standardized embeddings and benchmark suites.
- Mechanism-of-action prediction
- Classify the biological targets and pathways modulated by chemical compounds by analysing morphological shifts in high-content cellular imaging.
Explore the data
No licence declared
The authors have not published a licence, which means no permission to redistribute the data has been granted to anyone. The rows are not served here; read them on the original dataset page.How the columns are distributed
Over 381,606 rows of default/train
__key__
string_text- min
- 28
- median
- 29
- mean
- 29.02
- max
- 30
jp2
image- min
- 512
- median
- 512
- mean
- 512
- max
- 512
Strengths and limits
Good at
- Provides concentration-response phenotypic profiling paired with CRISPR knockouts at multi-thousand perturbation scale.
- Maintains an accessible download footprint of under 18 GB despite containing data from 222,601 wells.
- Includes precalculated embeddings from OpenPhenom, MAE-L/8, and MAE-G/8 to allow computational analysis without raw image processing overhead.
Watch out for
- Biological observations are restricted to a single cellular context, human umbilical vein endothelial cells (HUVEC).
- Contains a curated selection of 735 genes and 1,674 compounds rather than whole-genome or broad-chemical-library coverage.
Questions
Can I look inside rxrx3-core without downloading it?
Not here: it declares no licence, so no permission to redistribute its rows has been granted. Its description and statistics are still shown.
What licence is rxrx3-core under?
Its authors have not declared a licence, which means no permission to reuse or redistribute it has been granted.
Which cell line is profiled in RxRx3-core?
All perturbation experiments in RxRx3-core are conducted using human umbilical vein endothelial cells (HUVEC).
What imaging channels are provided?
The microscopy data consists of six-channel images collected using the Cell Painting assay protocol.
Which models were used to generate the precomputed embeddings?
The repository includes embeddings extracted using OpenPhenom, MAE-L/8, and MAE-G/8 vision models.
What information is included in the metadata?
The metadata provides well identifiers, experiment and plate numbers, well coordinates, gene names, compound names, canonical SMILES strings, concentrations in micromolar, perturbation types, cell type, and experimental control classifications.
How to cite it
RxRx3-core: Benchmarking drug-target interactions in High-Content Microscopy, LMRL Workshop at ICLR 2025. arXiv:2503.20158
Written from rxrx3-core's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.