SaumyaGupta-99/spliceformer-data

Spliceformer — training and evaluation data

Datasets for Spliceformer: splice-site prediction, masked-nucleotide pretraining, and the m6A / ADAR / RBP fine-tuning tasks.

python scripts/download_assets.py --data splice     # splice HDF5s
python scripts/download_assets.py --all             # everything

Contents

Folder	Size	What
splice/	~36 GB	Splice-site HDF5, GENCODE and GTEx labels, 10 kb and 400 nt
mlm/	3.5 GB	merged_masked_split.h5 — masked-nucleotide pretraining corpus
m6a/	56 MB	22 TSVs, 11 tissues / cell lines
rbp/	3.9 GB	250 ENCODE eCLIP TSVs, per-nucleotide binding labels
adar/	19 GB	Slic-seq A-to-I editing, balanced + full + pileups

Splice HDF5 layout

Chunked with paired keys X{i} / Y{i}:

X0  (2179, 15000, 4)   int8   5000 target nt + 5000 nt flank on each side
Y0  (1, 2179, 5000, 3) int8   one-hot [neither, acceptor, donor]

Note the leading singleton axis on Y. Indexing it with the sample index selects that axis rather than the sample — the loader in the repo handles this; a naive f["Y0"][i] does not.

The 400 nt files use X = (N, 5400, 4) with the same Y.

⚠️ GENCODE and GTEx are not interchangeable

At a given context length the GENCODE and GTEx files contain byte-identical sequences but different labels — the same genomic windows annotated from GENCODE versus derived from GTEx tissue expression. They were originally both named dataset_train_all.h5 / dataset_test_0.h5 in separate directories, which made it easy to score a model against the wrong labels. They are now distinguished by filename:

GENCODE	GTEx
dataset_train_all.h5	dataset_gtex_train_all.h5
dataset_test_0.h5	dataset_gtex_test_0.h5

TSV schemas

m6A ({tissue}_{train,test}.tsv) — sequence name, binary label, sequence. One label per sequence. Tissues: A549, CD8T, HCT116, HEK293, HEK293T, HeLa, HepG2, MOLM13, brain, kidney, liver.

RBP (ENCFF*.tsv) — id, sequence (1000 nt), labels (1000-char digit string, one binary label per nucleotide), split, cls. The outer 100 nt each side are cropped before the loss so scored positions have full receptive-field context.

ADAR (balanced/adar_dataset.tsv) — split, label, seq. Sequences are 15 001 bp, reverse-complemented where needed, with the candidate adenosine at index 7500. balanced/ is the training set; full/ is the unbalanced superset.

Related

• Code: https://github.com/NNeuralDynamics/Spliceformer
• Weights: SaumyaGupta-99/spliceformer-models
• GTEx tissues: SaumyaGupta-99/spliceformer-gtex-tissues
• HAEC training data: mrunyan1/haec-training-data

License

MIT