Shenzhi-Chen/DeepSTARR-Mouse-dataset

Dataset Summary

This dataset was used for training the DeepSTARR-Mouse model for predicting chromatin accessibility and enhancer activity.

Source Data

Sequences were derived from Gorkin et al., Nature (2020) (https://doi.org/10.1038/s41586-020-2093-3) and the VISTA Enhancer Browser (https://enhancer.lbl.gov/). The data were further processed to generate tissue-specific activity and accessibility annotations.

License

CC BY 4.0 (Attribution required). For research use only.

Intended Use

This dataset is provided for scientific and reproducibility purposes accompanying our publication in Nature.

Citation

Please cite our paper: [citation once published]

Accessibility Model Dataset

The Accessibility_model_dataset provides the training data for the sequence-to-accessibility model. It includes data from three tissues — heart, limb, and midbrain(CNS) — each organized into three cross-validation folds. For each tissue and fold, the dataset contains three subsets: training, validation, and test. Each subset includes:

a FASTA file containing DNA sequences, and a TXT file containing the corresponding chromatin accessibility values. In total, there are 18 files per tissue (3 folds × 3 subsets × 2 file types).

Activity Model Dataset

The Activity_model_dataset provides the training data for the sequence-to-activity model. It follows the same structure as the accessibility dataset, including data for heart, limb, and midbrain(CNS), each with three cross-validation folds. For each fold, there are training, validation, and test subsets. Each subset contains:

a FASTA file with DNA sequences, and a TXT file with the corresponding enhancer activity values. This organization results in 18 files per tissue (3 folds × 3 subsets × 2 file types).

Test dataset

The test dataset provides the randomly 600k generated sequences which used for enhancer activity model evaluation. all_vista_seq.fa are all vista sequences which were used to randomly generate design starting sequences.