README.md

---
license: mit
language:
- en
tags:
- biology
- genomics
- yeast
- transcription-factors
- callingcards
- transposon
- binding
- gene-expression
pretty_name: "Calling Cards Transcription Factor Binding Dataset"

experimental_conditions:
  temperature_celsius: room
  media:
    name: synthetic_complete_minus_ura_his_leu
    carbon_source:
      - compound: D-galactose
        concentration_percent: 2
    nitrogen_source:
      - compound: amino_acid_dropout_mix
        concentration_percent: unspecified
        specifications:
          - minus_ura
          - minus_his
          - minus_leu

citation: Mateusiak, C, Erdenebaatar, Z, Jia, E, Plaggenberg, JN, Wang, Y, Shively, C, Liao, G, Mitra, RD, Brent, MR. 2026. Functional synergy partially explains why most transcription factor binding is non-functional. bioRxiv 2026.
doi: https://doi.org/10.64898/2026.01.19.700460

features:
- applies_to:
  - genome_map_meta
  - annotated_feature_meta
  - annotated_feature_combined_meta
  fields:
  - name: condition
    dtype:
      class_label:
        names: [
          "standard", "rapa", "starvation", "glu_1_gal_1",
          "del_MET28", "glu_1_gal_2", "del_FKH2", "del_TYE7"
        ]
    description: >-
      Experimental condition of the sample, including standard growth, rapamycin treatment,
      nutrient starvation, mixed carbon source conditions, and gene deletion strains
    role: experimental_condition
    definitions:
      standard:
        media:
          name: synthetic_complete
          carbon_source:
            - compound: D-glucose
              concentration_percent: 2
      rapa:
        perturbation_method:
          type: chemical_treatment
          compound: rapamycin
          description: Rapamycin treatment to inhibit TORC1 signaling
      starvation:
        description: "Nutrient starvation condition - specific media composition not defined in source"
      glu_1_gal_1:
        media:
          carbon_source:
            - compound: D-glucose
              concentration_percent: 1
            - compound: D-galactose
              concentration_percent: 1
      glu_1_gal_2:
        media:
          carbon_source:
            - compound: D-glucose
              concentration_percent: 1
            - compound: D-galactose
              concentration_percent: 2
      del_MET28:
        genotype:
          deletions:
            - gene: MET28
              description: MET28 deletion strain
      del_FKH2:
        genotype:
          deletions:
            - gene: FKH2
              description: FKH2 deletion strain
      del_TYE7:
        genotype:
          deletions:
            - gene: TYE7
              description: TYE7 deletion strain

- applies_to:
  - annotated_feature_reprocess_intergenic
  - annotated_feature_reprocess_intergenic_analysis
  fields:
  - name: ir_name
    dtype: string
    description: >-
      Unique identifier of the intergenic region. See
      yeast_genome_resources/intergenic_regions_metadata_5_1.csv for details on
      the region (location, etc). Note that these intergenic regions are defined
      as the region between the end of one ORF and the start of the next, and
      are named according to the locus tags of the flanking ORFs (e.g.,
      YAL001C-YAL002W). A intergenic  region is assigned to a promoter only when
      the 5' end is continuous with the region.

- applies_to:
  - annotated_feature
  - annotated_feature_meta
  - genome_map
  - genome_map_meta
  - annotated_feature_reprocess_yiming
  - annotated_feature_reprocess_mindel
  - annotated_feature_reprocess_start_codon_500
  - annotated_feature_reprocess_intergenic
  fields:
  - name: batch
    dtype: string
    description: Experimental batch identifier for controlling batch effects (partition key)
    role: experimental_condition

- applies_to:
  - annotated_feature_meta
  - annotated_feature_combined_meta
  - genome_map_meta
  - 2026_analysis_set
  - annotated_feature_reprocess_mindel_analysis
  - annotated_feature_reprocess_start_codon_500bp_analysis
  - annotated_feature_reprocess_intergenic_analysis
  fields:
  - name: regulator_locus_tag
    dtype: string
    description: Systematic gene identifier for the transcription factor
    role: regulator_identifier
  - name: regulator_symbol
    dtype: string
    description: Standard gene symbol for the transcription factor
    role: regulator_identifier

- applies_to:
  - annotated_feature
  - annotated_feature_combined
  - 2026_analysis_set
  - annotated_feature_reprocess_yiming
  - annotated_feature_reprocess_mindel
  - annotated_feature_reprocess_mindel_analysis
  - annotated_feature_reprocess_start_codon_500
  - annotated_feature_reprocess_start_codon_500bp_analysis
  - annotated_feature_reprocess_intergenic
  - annotated_feature_reprocess_intergenic_analysis
  fields:
  - name: target_locus_tag
    dtype: string
    description: Systematic gene identifier for the target gene
    role: target_identifier
  - name: target_symbol
    dtype: string
    description: Standard gene symbol for the target gene
    role: target_identifier

- applies_to:
  - annotated_feature
  - annotated_feature_combined
  - 2026_analysis_set
  - annotated_feature_reprocess_yiming
  - annotated_feature_reprocess_mindel
  - annotated_feature_reprocess_mindel_analysis
  - annotated_feature_reprocess_start_codon_500
  - annotated_feature_reprocess_start_codon_500bp_analysis
  - annotated_feature_reprocess_intergenic
  - annotated_feature_reprocess_intergenic_analysis
  fields:
  - name: experiment_hops
    dtype: float64 
    description: Number of transposon insertion events (hops) at target locus in the experimental sample
    role: quantitative_measure
  - name: background_hops
    dtype: float64
    description: Number of transposon insertion events (hops) at target locus in the background control
    role: quantitative_measure
  - name: callingcards_enrichment
    dtype: float64
    description: Enrichment score calculated as ratio of normalized experimental to background hops
    role: quantitative_measure
  - name: poisson_pval
    dtype: float64
    description: P-value from Poisson test for statistical significance of binding enrichment
    role: quantitative_measure

- applies_to:
  - annotated_feature
  - annotated_feature_combined
  - 2026_analysis_set
  fields:
  - name: background_total_hops
    dtype: float64
    description: Total number of background hops across all loci in the control sample
    role: quantitative_measure
  - name: experiment_total_hops
    dtype: float64
    description: Total number of experimental hops across all loci in the experimental sample
    role: quantitative_measure

- applies_to:
  - annotated_feature_reprocess_yiming
  - annotated_feature_reprocess_mindel
  - annotated_feature_reprocess_mindel_analysis
  - annotated_feature_reprocess_start_codon_500
  - annotated_feature_reprocess_start_codon_500bp_analysis
  - annotated_feature_reprocess_intergenic
  - annotated_feature_reprocess_intergenic_analysis
  fields:
  - name: total_background_hops
    dtype: float64
    description: Total number of background hops across all loci in the control sample
    role: quantitative_measure
  - name: total_experiment_hops
    dtype: float64
    description: Total number of experimental hops across all loci in the experimental sample genomic (not mito) chromosomes
    role: quantitative_measure
  - name: log_poisson_pval
    dtype: float64
    description: Log-transformed Poisson p-value. This has greater numeric resolution for significant loci
    role: quantitative_measure
  - name: poisson_qval
    dtype: float64
    description: FDR-adjusted q-value from Poisson test (multiple testing correction)
    role: quantitative_measure
  - name: hypergeometric_pval
    dtype: float64
    description: P-value from hypergeometric test for statistical significance of binding enrichment
    role: quantitative_measure
  - name: log_hypergeometric_pval
    dtype: float64
    description: Log-transformed hypergeometric p-value
    role: quantitative_measure
  - name: hypergeometric_qval
    dtype: float64
    description: FDR-adjusted q-value from hypergeometric test (multiple testing correction)
    role: quantitative_measure
    
configs:
- config_name: annotated_feature
  description: >-
    This is data that was originally processed through
    https://github.com/cmatKhan/callingCardsTools/ and stored (including some more
    processing) in https://github.com/cmatKhan/yeastregulatorydb. It is the data that
    was used for the QC and filtering decisions in the 2026 modeling paper. In general,
    unless you are trying to exactly replicate the 2026 modeling paper, you should use
    the 2026_analysis_set for analysis that uses the published results. Or, to use data
    that can be reproduced directly from the genome_map data,
    `annotated_feature_reprocess_*`. The suffix indicates which promoter set was used 
    to generate the results from the genome_map data.
  dataset_type: annotated_features
  genome_resources:
    region_sets:
      Kang:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/yiming_promoters.bed
        join_column: target_locus_tag
  data_files:
  - split: train
    path: annotated_feature/*/*.parquet
  partitioning:
    enabled: true
    partition_by: ["batch"]
    path_template: "annotated_feature/batch={batch}/*.parquet"
  dataset_info:
    features:
    - name: id
      dtype: int64 
      description: Unique identifier for each binding measurement
    - name: hypergeometric_pval
      dtype: float64
      description: P-value from hypergeometric test for statistical significance of binding enrichment
      role: quantitative_measure

- config_name: annotated_feature_meta
  description: Metadata for the annotated_features dataset.
  dataset_type: metadata
  applies_to: ["annotated_feature"]
  data_files:
  - split: train
    path: annotated_feature_meta.parquet
  dataset_info:
    features:
    - name: id
      dtype: float64
      description: Unique identifier for the metadata record
      role: sample_id
    - name: genome_map_id
      dtype: float64
      description: >-
        Genome map identifier linking to the genome_map and genome_map_meta dataset
      role: secondary_sample_id
    - name: pss_id
      dtype: string
      description: >-
        Identifier from a defunct database (promoter set sig id)
      role: secondary_sample_id
    - name: binding_id
      dtype: string
      description: >-
        Identifier from a defunct database (binding id)
      role: secondary_sample_id
    - name: data_usable
      dtype: string
      description: Indicator of whether the data is suitable for analysis
    - name: analysis_set
      dtype: bool
      description: >-
        TRUE if this record is to be used for analysis. FALSE otherwise.
        This was determined in 2025. Replicates needed `>=`3k hops and
        DTO `<=` 0.01 in either kemmeren or hackett

- config_name: annotated_feature_combined
  description: >-
    For the 2026 modeling paper, we labeled replicates passing if it has `>=`3k hops
    and DTO `<=` 0.01 in either kemmeren or hackett. For a TF with more than 1 passing
    replicate, a combined sample is created by summing the hops across the passing
    replicates. This is the data that is used for the 2026 modeling paper as predictors.
    It is retained here for replication and transparency, but we do not recommend
    using it for new analysis. Instead, to use the published results, use the
    `2026_analysis_set` which includes the same combined samples, but also includes
    the passing single replicates. Otherwise, the annotated_feature_reprocess_*_analysis
    datasets are more directly reproducible from the genome_map data, using a specified
    promoter set, and have combined samples using the same logic.
  dataset_type: annotated_features
  genome_resources:
    region_sets:
      Kang:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/yiming_promoters.bed
        join_column: target_locus_tag
  data_files:
  - split: train
    path: annotated_feature_combined/*/*.parquet
  dataset_info:
    partitioning:
      enabled: true
      partition_by: ["genome_map_id_set"]
      path_template: "annotated_feature_combined/genome_map_id_set={genome_map_id_set}/*.parquet"
    features:
    - name: genome_map_id_set
      dtype: string
      description: >-
        Hyphen-delimited set of genome map IDs corresponding to the combined replicates for this
        regulator (partition key)
    - name: hypergeometric_pval
      dtype: float64
      description: P-value from hypergeometric test for statistical significance of binding enrichment
      role: quantitative_measure

- config_name: annotated_feature_combined_meta
  description: Metadata for the annotated_feature_combined dataset.
  dataset_type: metadata
  applies_to: ["annotated_feature_combined"]
  data_files:
  - split: train
    path: annotated_feature_combined_meta.parquet
  dataset_info:
    features:
    - name: genome_map_id_set
      dtype: string
      description: Hyphen-delimited set of genome map IDs used as the partition key in annotated_feature_combined
    - name: pss_id
      dtype: string
      description: Passing sample set identifier grouping replicates used in this combined analysis
    - name: binding_id
      dtype: string
      description: Unique identifier for this combined binding measurement record
    - name: analysis_set
      dtype: bool
      description: >-
        For a TF with more than 1 passing replicate, a combined samples is created.
        This is based on the QC done in 2025 for the modeling paper. See the
        annotated_feature_meta for more details

- config_name: 2026_analysis_set
  description: >-
    This dataset is the dataset that was used in the 2026 modeling paper. A passing 
    replicate has >=3000 hops had a dto empirical pvalue < 0.01 against either
    kemmeren or hackett. Where a given regulator had multiple passing replicates,
    those replicates were combined (see annotated_feature_combined). This dataset
    should be used when you want to use the published results from the 2026 modeling
    paper. If you want to use data that can be reproduced directly from the genome_map
    data included in this repo, especially when called against different promoter
    sets, then use the annotated_feature_reprocess_*_analysis datasets.
  default: true
  genome_resources:
    region_sets:
      Kang:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/yiming_promoters.bed
        join_column: target_locus_tag
  dataset_type: annotated_features
  metadata_fields: ["gm_id","regulator_locus_tag","regulator_symbol", "experiment_total_hops", "background_total_hops"]
  data_files:
  - split: train
    path: 2026_analysis_set.parquet
  dataset_info:
    features:
    - name: gm_id
      dtype: string
      description: >-
        genome_map id. If the sample is a combination of multiple samples, then it is a
        hyphen-delimited set of genome map IDs corresponding to the combined replicates for this
        regulator.

- config_name: genome_map
  description: >-
    This is the raw binding data (qbeds) from the nf-core/callingcards pipeline. It
    can be processed into annotated_feature datasets suing the scripts/quantify_regions.R script. You can use your own promoter definitions (bed format) to do this, or those provided in BrentLab/yeast_genome_resources
  dataset_type: genome_map
  data_files:
  - split: train
    path: genome_map/*/*.parquet
  dataset_info:
    features:
    - name: id
      dtype: int64
      description: Unique identifier for each genomic interval
      role: sample_id
    - name: chr
      dtype: string
      description: Chromosome name (e.g., chrI, chrII, etc.)
    - name: start
      dtype: int64
      description: Start position of genomic interval
    - name: end
      dtype: int64
      description: End position of genomic interval
    - name: depth
      dtype: int64
      description: Number of transposon insertion events (read depth) in this genomic interval
    - name: strand
      dtype: string
      description: Strand information (+ or -) for the genomic interval
    partitioning:
      enabled: true
      partition_by: ["batch"]
      path_template: "genome_map/batch={batch}/*.parquet"

- config_name: genome_map_meta
  description: Metadata for genome map datasets including regulator information and experimental details
  dataset_type: metadata
  applies_to: ["genome_map", "annotated_feature_reprocess_yiming", "annotated_feature_reprocess_mindel", "annotated_feature_reprocess_start_codon_500", "annotated_feature_reprocess_intergenic"]
  data_files:
  - split: train
    path: genome_map_meta.parquet
  dataset_info:
    features:
    - name: id
      dtype: float64
      description: Unique identifier for the metadata record
    - name: binding_id
      dtype: string
      description: current django managed database identifier for the dataset to the 'binding' table
    - name: replicate
      dtype: float64
      description: Biological replicate number, within batch
    - name: notes
      dtype: string
      description: Additional notes or comments about the experiment

- config_name: annotated_feature_reprocess_yiming
  description: >-
    Calling Cards annotated features reprocessed from the genome_map data
    using scripts/quantify_regions.R against the yiming promoters in
    BrentLab/yeast_genome_resources. This is very nearly exactly the same as 
    annotated_features, though there may be some differences around the boundaries 
    (intentional), and this includes higher numeric resolution in the most significant
    promoters by using hte log argument in the poisson distribution function.
  dataset_type: annotated_features
  data_files:
  - split: train
    path: annotated_feature_reprocess_yiming/*/*.parquet
  genome_resources:
    region_sets:
      Kang:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/yiming_promoters.bed
        join_column: target_locus_tag
  partitioning:
    enabled: true
    partition_by: ["batch"]
    path_template: "annotated_feature_reprocess_yiming/batch={batch}/*.parquet"
  dataset_info:
    features:
    - name: id
      dtype: int64
      description: >-
        Genome map identifier linking to the genome_map and genome_map_meta dataset

- config_name: annotated_feature_reprocess_mindel
  description: >-
    This is the genome_map data quantified against the Mindel promoters
    (see BrentLab/yeast_genome_resources) using scripts/quantify_regions.R.
  dataset_type: annotated_features
  data_files:
  - split: train
    path: annotated_feature_reprocess_mindel/*/*.parquet
  genome_resources:
    region_sets:
      Mindel:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/mindel_promoters.csv.gz 
        join_column: target_locus_tag
  partitioning:
    enabled: true
    partition_by: ["batch"]
    path_template: "annotated_feature_reprocess_mindel/batch={batch}/*.parquet"
  dataset_info:
    features:
    - name: genome_map_id
      dtype: int64
      description: >-
        Genome map identifier linking to the genome_map and genome_map_meta dataset

- config_name: annotated_feature_reprocess_start_codon_500
  description: >-
    This is the genome_map data quantified against the promoters defined
    as 500bp upstream of the start codon for each gene 
    (see BrentLab/yeast_genome_resources) using
    scripts/quantify_regions_500bp_intergenic.R.
  dataset_type: annotated_features
  data_files:
  - split: train
    path: annotated_feature_reprocess_start_codon_500/*/*.parquet
  genome_resources:
    region_sets:
      start_codon_500bp:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/start_codon_500bp_upstream_promoters.bed
        join_column: target_locus_tag
  partitioning:
    enabled: true
    partition_by: ["batch"]
    path_template: "annotated_feature_reprocess_start_codon_500/batch={batch}/*.parquet"
  dataset_info:
    features:
    - name: genome_map_id
      dtype: int64
      description: >-
        Genome map identifier linking to the genome_map and genome_map_meta dataset

- config_name: annotated_feature_reprocess_intergenic
  description: >-
    This is the genome_map data quantified against the promoters defined
    as the full intergenic region upstream of each gene (see
    BrentLab/yeast_genome_resources) using scripts/quantify_regions_500bp_intergenic.R.
  dataset_type: annotated_features
  data_files:
  - split: train
    path: annotated_feature_reprocess_intergenic/*/*.parquet
  genome_resources:
    region_sets:
      intergenic:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/intergenic_regions_metadata_5_1.csv
        join_column: ir_name
  partitioning:
    enabled: true
    partition_by: ["batch"]
    path_template: "annotated_feature_reprocess_intergenic/batch={batch}/*.parquet"
  dataset_info:
    features:
    - name: genome_map_id
      dtype: int64
      description: >-
        Genome map identifier linking to the genome_map and genome_map_meta dataset

- config_name: annotated_feature_reprocess_mindel_analysis
  description: >-
    This is the analysis set for the mindel data. It is generated using the same logic as the 2026_analysis_set, but using the results from the reprocessing against the mindel promoters. A passing replicate has >=3000 hops had a dto empirical pvalue < 0.01 against either kemmeren or hackett. Where a given regulator had multiple passing replicates, those replicates were combined (see annotated_feature_combined). This dataset should be used when you want to use data that can be reproduced directly from the genome_map data included in this repo, using the mindel promoter definitions. See scripts/quantify_regions.R as well as BrentLab/yeast_comparative_analysis/scripts for details of how this was conducted.
  dataset_type: annotated_features
  metadata_fields: ["combined_id","regulator_locus_tag","regulator_symbol", "total_experiment_hops", "total_background_hops"]
  data_files:
  - split: train
    path: annotated_feature_reprocess_mindel_analysis.parquet
  genome_resources:
    region_sets:
      Mindel:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/mindel_promoters.csv.gz 
        join_column: target_locus_tag
  dataset_info:
    features:
    - name: combined_id
      dtype: string
      description: >-
        The genome map identifier of either a single passing, or multiple
        passing replicates (hyphen-delimited) that were combined for the analysis set. replicates were combined if the DTO empirical
        pvalue was <= 0.01 in either kemmeren or hackett. After combining,
        if they have more than 3k hopes (single or combined), then they are
        included in the analysis set.
      role: sample_id

- config_name: annotated_feature_reprocess_start_codon_500bp_analysis
  description: >-
    This is the analysis set for promoters created with 500bp regions upstream
    of the start codon. It uses the same passing replicates as the 2025_analysis_set. see R/scripts/quantify_regions_500bp_intergenic.R for details of how this was conducted.
  dataset_type: annotated_features
  metadata_fields: ["combined_id","regulator_locus_tag","regulator_symbol", "total_experiment_hops", "total_background_hops"]
  data_files:
  - split: train
    path: annotated_feature_reprocess_start_codon_500bp_analysis.parquet
  genome_resources:
    region_sets:
      start_codon_500bp:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/start_codon_500bp_upstream_promoters.bed
        join_column: target_locus_tag
  dataset_info:
    features:
    - name: combined_id
      dtype: string
      description: >-
        The genome map identifier of either a single passing, or multiple
        passing replicates (hyphen-delimited) that were combined for the
        2026 analysis set.
      role: sample_id

- config_name: annotated_feature_reprocess_intergenic_analysis
  description: >-
    This is the analysis set for promoters created with the full intergenic 
    region upstream of a given target. It uses the same passing replicates as the 2025_analysis_set. see R/scripts/quantify_regions_500bp_intergenic.R for details of how this was conducted.
  dataset_type: annotated_features
  metadata_fields: ["combined_id","regulator_locus_tag","regulator_symbol", "total_experiment_hops", "total_background_hops"]
  data_files:
  - split: train
    path: annotated_feature_reprocess_intergenic_analysis.parquet
  genome_resources:
    region_sets:
      intergenic:
        path: https://huggingface.co/datasets/BrentLab/yeast_genome_resources/blob/main/intergenic_regions_metadata_5_1.csv
        join_column: ir_name
  dataset_info:
    features:
    - name: combined_id
      dtype: string
      description: >-
        The genome map identifier of either a single passing, or multiple
        passing replicates (hyphen-delimited) that were combined for the
        2026 analysis set.
      role: sample_id
---
# Calling Cards

This is data produced in both the Brent Lab and Mitra Lab at Washington University.

## Accessing Data

The examples below require
[labretriever](https://github.com/cmatKhan/labretriever#installation)
(`pip install labretriever`) and/or the
[HuggingFace Hub client](https://huggingface.co/docs/huggingface_hub/installation)
(`pip install huggingface_hub`).

### Accessing Data with labretriever

This repository is part of a collection configured as a unified database using
[labretriever.VirtualDB](https://cmatkhan.github.io/labretriever/virtual_db_configuration/).
Download the
[collection config](https://github.com/BrentLab/tfbpshiny/blob/main/tfbpshiny/brentlab_yeast_collection.yaml)
and use it to query the data directly in Python, or with an AI assistant using the
[labretriever plugin](https://cmatkhan.github.io/labretriever/mcp_server/#quick-install-claude-code-plugin).

```python
from labretriever.virtual_db import VirtualDB
from labretriever.datacard import DataCard

# Citation and metadata
card = DataCard("BrentLab/callingcards")
print([c.config_name for c in card.configs])  # list available datasets

# print citation
info = card.info()
print(info["citation"])

# path to the downloaded brentlab_yeast_collection.yaml
vdb = VirtualDB("/path/to/brentlab_yeast_collection.yaml")

print(vdb.get_dataset_description("callingcards"))
vdb.query("SELECT * FROM callingcards LIMIT 5")
```

### Direct parquet access

The repository contains more data than what is exposed through the collection
configuration. Use `DataCard.info()` to inspect available files, then download
and query with DuckDB.

Some files are single parquet files (e.g. metadata files); others are
partitioned datasets. Download a metadata file first to identify relevant
partitions before fetching the full data.

Single parquet file example:

```python
from huggingface_hub import snapshot_download
import duckdb

repo_path = snapshot_download(
    repo_id="BrentLab/callingcards",
    repo_type="dataset",
    allow_patterns="annotated_feature_meta.parquet",
)
conn = duckdb.connect()
# returns a pandas DataFrame with the first 5 rows
conn.execute(
    "SELECT * FROM read_parquet(?) LIMIT 5",
    [f"{repo_path}/annotated_feature_meta.parquet"],
).df()
```

Partitioned dataset example (the `annotated_feature` directory):

```python
repo_path = snapshot_download(
    repo_id="BrentLab/callingcards",
    repo_type="dataset",
    allow_patterns="annotated_feature/**",
)
conn.execute(
    "SELECT * FROM read_parquet(?) LIMIT 5",
    [f"{repo_path}/annotated_feature/**/*.parquet"],
).df()
```

### Accessing using R

Clone the repository and read parquet files directly with
[arrow](https://arrow.apache.org/docs/r/):

```r
# install.packages("arrow")
arrow::read_parquet("annotated_feature_meta.parquet")
```