jing-z/leaf1-pretraining-cfdna

LEAF-1 pretraining corpus — cfDNA fragments (hg38)

This dataset contains the cell-free DNA (cfDNA) portion of the pretraining corpus of LEAF-1. The sample manifest and checksums of this dataset are also archived on Zenodo (10.5281/zenodo.23195009). The other portions of the pretraining corpus: scATAC-seq on Zenodo (10.5281/zenodo.23176597); TCGA ATAC-seq is controlled access, its manifest and processing code are on Zenodo (10.5281/zenodo.23195360).

2,301 plasma cfDNA samples from four published studies, one BED file per sample, 211 GB compressed (984 GB uncompressed), 34.0 billion fragments in total.

Contents

file	n	description
EE<id>.hg38.frag_filtered.bed.gz	2,301	filtered fragment BED, bgzip-compressed
indexes.tar	1	the 2,301 EE<id>.hg38.frag_filtered.bed.index sparse line-offset indexes (see below)
samples.tsv	1	one row per sample: FinaleDB id, study, label, LEAF-1 train/test split, fragment count
checksums.tsv	1	md5 and size of every file in this dataset
README.md	1	this file

Usage

• To reproduce LEAF-1 pretraining: decompress every file first (ls *.bed.gz | xargs -P 16 -n 1 bgzip -d; ~984 GB uncompressed in total). The LEAF-1 data loader reads the uncompressed .bed files together with the .bed.index sidecars from indexes.tar, which must sit next to the corresponding .bed file with the same basename; the samples.tsv split column reproduces the train/test partition.
• For any other use: no decompression is needed. The files are plain gzip and can be streamed with zcat / gzip.open(), or indexed with tabix -p bed <file> for region queries.

Download everything with the Hugging Face CLI:

hf download jing-z/leaf1-pretraining-cfdna --repo-type dataset --local-dir leaf1-cfdna

BED format

Five tab-separated columns, no header, one fragment per line:

chr1	792519	792775	30	-

chrom, start (0-based), end (exclusive), mapq, strand. Coordinates are on GRCh38/hg38. This is the FinaleDB .frag layout; the mapq and strand columns are carried over unchanged and are not used by LEAF-1.

Compression

Files are compressed with bgzip (htslib). They are standard gzip streams — gunzip, zcat and Python's gzip.open() all work. Because bgzip is block-based, each file can also be indexed with tabix for random access without full decompression.

The .bed.index files in indexes.tar are the sparse indexes used by the LEAF-1 data loader: one line per 1,000 fragments, line_number cumulative_byte_offset. These offsets refer to the uncompressed .bed file and are only valid after bgzip -d / gunzip. They are not offsets into the compressed stream. If you do not use the LEAF-1 loader you can ignore them and use tabix instead.

Verification

md5sum <file>          # compare with column 2 of checksums.tsv

Every compressed file was verified at packaging time: the md5 of the decompressed stream was compared with the md5 of the original filtered BED, and all 2,301 matched.

Provenance

Upstream source

All fragments were downloaded from FinaleDB (Zheng et al., 2021, Bioinformatics 37(16):2502–2503, doi:10.1093/bioinformatics/btaa919; http://finaledb.research.cchmc.org/), which provides uniformly reprocessed, hg38-aligned fragment files for published cfDNA sequencing studies. Sample identifiers (EE<id>) are FinaleDB entry ids. The 2,301 samples come from:

study	n	samples
Adalsteinsson et al., 2017, Nat Commun 8:1324	1,508	breast and prostate cancer
Cristiano et al., 2019, Nature 570:385–389	537	seven cancer types, healthy
Jiang et al., 2015, PNAS 112:E1317–E1325	225	liver cancer, hepatitis B, cirrhosis, healthy
Sun et al., 2019, Genome Res 29:418–427	31	colorectal cancer, liver transplant

Please cite FinaleDB and the original studies when reusing these files.

Processing

Starting from the FinaleDB EE<id>.hg38.frag.gz files, the following filters were applied, in order:

1. mapping quality ≥ 30 (column 4);
2. fragment length ≥ 20 bp;
3. no overlap with the ENCODE hg38 blacklist v2 (bedtools intersect -v).

No other transformation was applied. The filtered BEDs are the direct input to the LEAF-1 pretraining pipeline, which retrieves flanking genomic context from the hg38 FASTA at training time. samples.tsv records which samples were in the pretraining train (2,070) and held-out test (231) partitions.

License and citation

Data: CC BY 4.0.

If you use this dataset, please cite the LEAF-1 paper, FinaleDB, and the original studies listed above.