# Datasets The datasets described below are necessary for different model training, sequence design, and data analysis programs included in this repository. The `download_data.py` script included in this folder can download each dataset individually. ## Data for training accessibility models Processed data necessary for model training and some analysis will be downloaded into the [`dhs_index`](./dhs_index/) subfolder. The [`process_data.ipynb`](./dhs_index/process_data.ipynb) notebook reproduces our data processing workflow starting from raw published sources, but re-running this should not be necessary in most cases. ## Cell line and mouse retina MPRA results Processed MPRA data will be downloaded into the [`mpra`](./mpra/) subfolder. The downloaded file `enhancer_mpra_processed.tsv` contains annotations and processed MPRA measurements of tested enhancers. Annotation columns include: - Sequence ID - Enhancer sequence - Source (i.e. genome, Fast SeqProp, DEN) - Category (e.g. the value `single_fsp` corresponds to enhancers targing a single biosample designed with Fast SeqProp) - Intended target biosample(s) - Information about sequences designed via motif embedding (columns that start with `motif_`) - Measured accessibility for genome-sourced sequences (columns starting with `dhs_signal_`) - DHS64-predicted accessibilities (columns starting with `dhs_pred_`). MPRA result columns include: - Sequencing read counts in columns `{cell_type}_reads_r{i}`, where `i` corresponds to the replicate number. - DESeq2 outputs for each assayed cell type, most importantly log2(RNA/DNA) estimates in columns `log2FC_{cell_type}`. The downloaded file `mpra_data_splits.json` contains MPRA sequence IDs partitioned into 7 folds, used for finetuning the DHS64-MPRA model. The notebook [`make_data_splits.ipynb`](./mpra/make_data_splits.ipynb) can regenerate this file.