EvoLenTokenizer/analysis-data

EvoLen — token analysis input data

Derived interval files needed to reproduce the token analyses in Section 4 of EvoLen: Evolution-Guided Tokenization for DNA Language Model (arXiv:2604.08698).

Analysis code lives in the evolen repository under analysis/.

Contents

region_beds/
  source/                  INPUT to the P4 enrichment analysis -- the four genomic
                           regions, merged and cleaned:
                           promoters_2kb.clean.merged.bed   (28,251 intervals)
                           enhancers_dels.clean.merged.bed  (1,464,531)
                           exon.clean.merged.bed            (402,955)
                           intron.clean.merged.bed          (150,128)
  conservation_crossed/    OUTPUT of that analysis, provided for checking: the four
                           regions crossed with conservation category
                           {promoter,enhancer,exon,intron}_{conserved,neutral,accelerated}.bed
                           conservation_{conserved,neutral,accelerated}.bed
  simple/                  the same four regions without the conservation split
ccre_classes/              ENCODE SCREEN cCRE classes as BED, one file per class
                           CA, CA-CTCF, CA-H3K4me3, CA-TF, PLS, TF, dELS, pELS
                           *_balanced.bed are downsampled to the smallest class (26,102)
motifs/motifs.txt          JASPAR 2024 vertebrate motifs, thresholded to consensus
                           sequences (PWM positions at 0.5, wildcards trimmed, <= 12 bp)

region_beds/source/ is what analysis/enrichment/enrichment_heatmap.py reads; it generates the conservation split and the crossed BEDs itself, so region_beds/conservation_crossed/ is included only so results can be compared without re-running. ccre_classes/ backs the Multi-SCREEN task construction; motifs/ backs the P1 motif preservation analysis (Figure 2A).

Not included — fetch these yourself

Two inputs are public reference data and are not mirrored here.

hg38 reference genome (~3.3 GB):

wget https://hgdownload.soe.ucsc.edu/goldenPath/hg38/bigZips/hg38.fa.gz
gunzip hg38.fa.gz && samtools faidx hg38.fa

phyloP conservation scores. The analysis reads per-chromosome bedGraph, which is a mechanical conversion of the public bigWig (~70 GB expanded, so it is regenerated rather than distributed):

wget https://hgdownload.soe.ucsc.edu/goldenPath/hg38/phyloP100way/hg38.phyloP100way.bw
# convert per chromosome with UCSC bigWigToBedGraph
bigWigToBedGraph -chrom=chr1 hg38.phyloP100way.bw chr1.bedGraph

Point process_bedgraph_all.py --bedgraph_dir at the directory of resulting .bedGraph files to produce the {chrom}_phylop_segment.csv files that drive both tokenizer construction and the phyloP analyses.

Usage

export EVOLEN_ROOT=/path/to/your/data_root     # analysis scripts resolve paths from this

Related: token_evaluation hosts the phyloP analysis outputs (per-token aggregates used for Figure 2C).