Taykhoom/lentimpra-enhancer-activity

LentiMPRA Enhancer Activity

This dataset contains the large-scale and joint lentiMPRA libraries released with Agarwal et al. Each row is one designed construct with its 200 nt assayed insert, source annotations, replicate measurements, and published modeling targets where available.

The cell-specific config has three biological splits:

Split	Rows
hepg2	164,307
k562	243,780
wtc11	75,542

The joint config contains 60,000 constructs measured across all three cell lines. These names describe assay libraries, not machine-learning train/test partitions.

Loading

from datasets import load_dataset

hepg2 = load_dataset(
    "Taykhoom/lentimpra-enhancer-activity",
    "cell-specific",
    split="hepg2",
)
joint = load_dataset(
    "Taykhoom/lentimpra-enhancer-activity",
    "joint",
    split="joint",
)
performance = load_dataset(
    "Taykhoom/lentimpra-enhancer-activity",
    "published-performance",
    split="large_scale",
)

Sequence and measurements

The source oligos are 230 nt:

AGGACCGGATCAACT + 200-nt insert + CATTGCGTGAACCGA

sequence contains only the 200 nt insert. Reverse-orientation constructs already contain the assayed reverse-complemented insert and should not be reverse-complemented again.

Coordinate columns use zero-based, half-open, BED-like interval semantics as released by the source tables. Most genomic intervals are 200 nt, but the exceptional source interval widths are preserved rather than forced to 200 nt. There are 345 non-200 nt hg38 intervals: 153 promoters, 44 potential enhancers, and 148 Inoue et al. controls; no hg19 interval is non-200 nt. The hg19_strand and hg38_strand columns are source genomic strand; orientation is construct orientation in the assay, and the two concepts must not be substituted for one another. Assembly claims are limited to the source labels hg19 and hg38; no patch release or sequence accession is inferred.

The three cell-specific files share these column groups:

Columns	Meaning
name, element_id, category, sequence	Construct identity, library annotation, and exact 200 nt assayed insert.
hg19_*, hg38_*, orientation	Released genomic coordinates, strand, and assay orientation where available.
replicate_{1,2,3}_log2_rna_dna	Full-precision replicate activity.
replicate_{1,2,3}_num_observed_barcodes	Source barcode counts.
mean_log2_rna_dna, observed_log2_rna_dna	Published QC-filtered mean and orientation-averaged standardized target.
active	Nullable binary call derived from the cell-line shuffled-control threshold.
fold	Published 10-fold assignment for modeled elements.

The joint file uses the same design columns and repeats the replicate, mean, active, source ESS, and observed ESS columns with hepg2_, k562_, and wtc11_ prefixes. Its fold column is shared across the three targets.

All source design rows are retained. Null values mark constructs excluded from the corresponding assay, QC, or modeling subset.

Known sequence shortcuts

In the joint library, 54,000 potential-enhancer rows have complete three-cell activity labels; 53,151 also have a non-null published fold. Exploratory checks reported here use the latter modeling cohort. GC/CpG content, dinucleotide composition, and source library all predict activity. For example, GC content alone reached ROC-AUC 0.685 for WTC11, and a 6-mer nearest-centroid baseline reached 0.322 balanced accuracy on the eight activity profiles versus 0.125 chance.

The signal is not entirely explained by those shortcuts. After matching always-active (111) and always-inactive (000) enhancers by source library, GC content, and dinucleotide composition, the 6-mer baseline reached 0.724 balanced accuracy. Dinucleotide-preserving sequence shuffling reduced it to 0.517. There were no exact or reverse-complement duplicate sequences.

These checks suggest a mixture of low-order composition and higher-order sequence signal. Model results should therefore use the published folds and include simple composition baselines rather than treating an unconstrained zero-shot sequence score as purely biological signal.

Published model performance

The published-performance config transcribes the Benchmarking results sheets from Supplementary Tables S8 and S12. Each row contains dataset, cell_line, test_fold, model, and the paper-reported held-out-fold pearson_r.

The large_scale split has 180 rows: 3 cell lines, 10 folds, and 6 models. The joint split has 150 rows: 3 cell lines, 10 folds, and 5 models. These are the published aggregate correlations, not row-level predictions.

Mean Pearson correlation across the ten folds:

Library	Cell line	Biochemical	MPRAnn	MPRALegNet	Sei	Enformer
Large-scale	HepG2	0.7320	0.7665	0.8333	0.7914	0.8306
Large-scale	K562	0.7235	0.8082	0.8665	0.8150	0.8497
Large-scale	WTC11	0.7124	0.7572	0.7946	0.8212	0.8335
Joint	HepG2	0.6610	0.7255	0.7795	0.7891	0.8191
Joint	K562	0.5738	0.6780	0.7465	0.7457	0.7811
Joint	WTC11	0.6057	0.7139	0.7668	0.7820	0.8011

The large-scale source sheet also includes the universal-feature biochemical model; its complete per-fold values are retained in the Parquet.

Processing

The downloader, pinned archive checksum, workbook reconciliation, adapter trimming, activity thresholds, target reconstruction, and output verification are available at:

• Processing repository: https://github.com/TaykhoomDalal/LentiMPRA-Enhancer-Processing/tree/main
• Task documentation: https://github.com/TaykhoomDalal/LentiMPRA-Enhancer-Processing/blob/main/LentiMPRA/README.md
• Reproducibility manifest: https://github.com/TaykhoomDalal/LentiMPRA-Enhancer-Processing/blob/main/manifest.json

Citation

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

License

The article and supplementary data are distributed under CC BY 4.0. Cite the source publication when using these measurements.