Taykhoom/lenti-mpra-glm-tang

GLM-Tang Task 1: lentiMPRA Activity

This dataset contains the exact HepG2 and K562 lentiMPRA inputs used for Task 1 of Tang et al. The task is single-target regression: predict orientation-specific reporter activity from a 230 nt synthesized sequence.

Each model input is:

15-nt adapter + 200-nt regulatory element + 15-nt adapter

The adapters are part of the released benchmark input. sequence therefore stores the complete 230 nt reporter rather than only the 200 nt element.

Loading

from datasets import load_dataset

dataset = load_dataset("Taykhoom/lenti-mpra-glm-tang")
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

Filter by cell_line to reproduce the separate HepG2 and K562 tasks:

hepg2_train = train.filter(lambda row: row["cell_line"] == "HepG2")
k562_train = train.filter(lambda row: row["cell_line"] == "K562")

Dataset size

Split	Rows
train	511,328
validation	63,916
test	63,936
Total	639,180

The full dataset contains 245,852 HepG2 rows and 393,328 K562 rows.

The split follows the released archive and code:

• folds 1–8: train;
• fold 9: validation;
• fold 10: test.

This is approximately 80/10/10. The Tang paper's Methods text describes 70/10/20, but the distributed HDF5 files and released preprocessing code use the fold assignment above.

Sequence and target contract

sequence is the exact archived reporter orientation and must be used unchanged for benchmark reproduction.

label is the source HDF5 y value: orientation-specific transcriptional activity. Models in the released code use mean-squared error and report Pearson correlation separately for HepG2 and K562.

aux_target_tang_normalized_activity_average is Tang's source HDF5 y_avg value averaged across the two assayed element orientations. aux_target_table_s6_observed_log2_rna_dna is the exact Supplementary Table S6 Observed log2(RNA/DNA) value joined by fold and element name. Both are target-derived metadata and must not be used as input features.

The join covers every reporter row. Tang y_avg is an affine normalization of the Table S6 value: slope/intercept 0.7291221908/-0.2190302783 for HepG2 and 0.4847236142/-0.0979363992 for K562, with maximum residual below 2e-7.

Each canonical regulatory element contributes four rows:

1. one assayed element orientation;
2. its complete reporter reverse complement;
3. the opposite assayed element orientation;
4. its complete reporter reverse complement.

Use:

• reverse_complement_group_id for the two rows sharing one label;
• cre_id for the four rows from one cell-specific element;
• global_cre_id to group the same canonical element across cell lines.

Columns

Column	Description
split	Released benchmark partition: train, val, or test.
sequence	Exact 230 nt reporter input in the archived orientation.
label	Orientation-specific transcriptional activity.
example_id	Stable unique row identifier: cell line, source fold, and row index.
reverse_complement_group_id	Shared by a reporter and its full reverse complement.
source_index	Zero-based row index inside the source fold.
source_fold	Original fold number, 1–10.
cell_line	HepG2 or K562.
cre_id	Cell-line-specific canonical regulatory-element identifier.
global_cre_id	Canonical regulatory-element identifier shared across cell lines.
reverse_complement	Whether the full reporter is the non-canonical member of its reverse-complement pair.
core_reverse_complement	Whether the 200 nt element is in the non-canonical orientation.
adapter_orientation	forward or reverse adapter orientation.
aux_target_tang_normalized_activity_average	Tang-normalized orientation-averaged HDF5 y_avg; target metadata, not a feature.
aux_target_table_s6_observed_log2_rna_dna	Exact Table S6 observed log2(RNA/DNA) activity; target metadata, not a feature.
element_name	Supplementary Table S3 element name.
element_category	potential enhancer or promoter.
chrom, start, end, strand	Author-supplied zero-based, half-open hg38 element coordinates and reporter-core orientation.
genome_build	Always GRCh38/hg38.
sequence_matches_hg38	Whether sequence[15:215] exactly equals the oriented hg38 interval.
sequence_edit_distance	Minimum substitutions/insertions/deletions between the assayed 200 nt element and hg38 interval.
coordinates_author_assigned	Always true; coordinates come from Supplementary Table S3 rather than sequence inference.
cross_cell_split_leakage	Whether the same global_cre_id has different split assignments across HepG2 and K562.

Genomic coordinates

The reporter remains exactly 230 nt even when the corresponding hg38 interval is longer than 200 bases. Assembly insertions can make an author-supplied interval span up to 254 hg38 bases. The assay sequence is never silently replaced by the reference interval.

Across all rows:

• 638,016 reporter cores exactly match hg38;
• 1,164 differ;
• the affected rows require 2,040 total edit operations.

Important limitations

Tang did not release the code that constructed the fold HDF5 files from the source workbooks. This release is exact from the archived HDF5 boundary forward; it does not claim to recreate the unpublished selection and orientation-expansion step from raw workbooks alone.

cross_cell_split_leakage affects 21,560 rows, but the original benchmark trains separate models per cell line. Within each cell line, the four-row CRE groups remain split-consistent.

Processing and verification

The pinned inputs, fold repacking, Table S3 reconciliation, exact row-order checks, coordinate validation, and manifest contract are documented at stable repository links:

• Processing repository: https://github.com/TaykhoomDalal/GLM-Tang-Processing/tree/main
• Task documentation: https://github.com/TaykhoomDalal/GLM-Tang-Processing/blob/main/lenti-mpra/README.md
• Reproducibility manifest: https://github.com/TaykhoomDalal/GLM-Tang-Processing/blob/main/manifest.json

Sources and citation

Tang, Z., Somia, N., Yu, Y. & Koo, P. K. Evaluating the representational power of pre-trained DNA language models for regulatory genomics. Genome Biology 26, 203 (2025). https://doi.org/10.1186/s13059-025-03674-8

Agarwal, V. et al. Massively parallel characterization of transcriptional regulatory elements. Nature (2025). https://doi.org/10.1038/s41586-024-08430-9

License

The Tang Zenodo archive and Agarwal supplementary release are distributed under CC BY 4.0. Cite both publications when using this benchmark.