---
pretty_name: Malinois/Gosai MPRA Regression
task_categories:
- tabular-regression
tags:
- biology
- genomics
- dna
- mpra
- carbon
size_categories:
- 100K<n<1M
---

# Malinois/Gosai MPRA Regression

This dataset preprocesses the Gosai et al. 2024 supplementary MPRA table used by the
Malinois benchmark for supervised DNA-to-activity regression. Each row contains a DNA
sequence and three cell-type-specific activity targets: `K562_log2FC`,
`HepG2_log2FC`, and `SKNSH_log2FC`.

No new license is asserted by this preprocessing. Users should follow the terms of
the source publication and supplementary data.

## Source

- Publication: Gosai et al., *Machine-guided design of cell-type-targeting
  cis-regulatory elements*, Nature 2024.
- Source table: `41586_2024_8070_MOESM4_ESM.txt`.
- Source URL: `https://static-content.springer.com/esm/art%3A10.1038%2Fs41586-024-08070-z/MediaObjects/41586_2024_8070_MOESM4_ESM.txt`.

## Splits

Chromosome splits match the Carbon fine-tuning experiments and the public Malinois
setup we used:

| Split | Chromosomes | Rows | Rows with all SE <= 1.0 |
|---|---:|---:|---:|
| train | all except validation/test chromosomes | 668,946 | 627,661 |
| validation | 19, 21, X | 62,406 | 58,811 |
| test | 7, 13 | 66,712 | 62,582 |

Total rows after filtering finite targets/standard errors and nonempty sequences:
798,064.

## Columns

- `id`: original row identifier from the source table.
- `split`: train, validation, or test.
- `chromosome`: normalized chromosome label.
- `data_project`, `oligo`, `variant_class`: source metadata.
- `sequence`: uppercase DNA sequence.
- `reverse_complement`: reverse complement of `sequence`.
- `forward_rc_concat`: `<dna>sequence</dna><dna>reverse_complement</dna>`,
  matching the best Carbon fine-tuning recipe.
- `K562_log2FC`, `HepG2_log2FC`, `SKNSH_log2FC`: raw regression targets.
- `K562_lfcSE`, `HepG2_lfcSE`, `SKNSH_lfcSE`: target standard errors.
- `*_train_zscore`: target standardized using train-split mean/std.
- `all_se_le_1`: true when all three SE columns are `<= 1.0`; this was the
  main reported validation/test metric filter.
- `any_log2fc_gt_0_5`: true when any target is greater than `0.5`; this was used
  for optional high-activity training upsampling.

Train z-score statistics:

| Target | Mean | Std |
|---|---:|---:|
| K562_log2FC | 0.49943020 | 1.17725282 |
| HepG2_log2FC | 0.46267671 | 1.05124023 |
| SKNSH_log2FC | 0.41405871 | 1.16609108 |

## Usage

```py
from datasets import load_dataset

ds = load_dataset("HuggingFaceBio/malinois-mpra-regression")
train = ds["train"]
validation_metric = ds["validation"].filter(lambda row: row["all_se_le_1"])

example = train[0]
sequence = example["forward_rc_concat"]
labels = [
    example["K562_log2FC_train_zscore"],
    example["HepG2_log2FC_train_zscore"],
    example["SKNSH_log2FC_train_zscore"],
]
```

To recreate the dataset:

```sh
python create_dataset.py --repo-id HuggingFaceBio/malinois-mpra-regression --push
```