Recipe: DNase-seq¶
This recipe trains a Cherimoya model on DNase-seq, which measures chromatin accessibility via DNase I hypersensitivity. DNase-seq is typically single-end and is most commonly modeled as a single unstranded output track, although stranded variants exist.
If your data is paired-end ATAC-seq instead, see Recipe: ATAC-seq.
Inputs¶
Reference genome FASTA (e.g.
hg38.fa).A BAM file of aligned single-end DNase-seq reads.
A motif database in MEME format.
Optional: a BED of peak coordinates. If not provided, MACS3 will call peaks.
Generate the pipeline JSON¶
cherimoya pipeline-json \
-s hg38.fa \
-i dnase.bam \
-m JASPAR_2024.meme \
-n dnase_experiment \
-o dnase.pipeline.json \
-u
Flag-by-flag:
-u— unstranded output (single signal track). The most common setup for DNase-seq.No
-pe(single-end reads).No
-f(the input is a BAM of aligned reads, not a fragment file).No shift by default. If your protocol calls for a DNase I-specific shift (some pipelines apply +1 / 0 to mark cut sites), set
-ps/-nshere.
For a stranded variant, omit -u. Cherimoya will then produce two
output tracks (+ and - strand) and the trained model will have
signal_groups=[2] (one stranded group).
Run the pipeline¶
cherimoya pipeline -p dnase.pipeline.json
Steps invoked, in order:
MACS3 peak calling on the BAM with no controls (output:
dnase_experiment_peaks.narrowPeak).bam2bwconverts the BAM to an unstranded bigWig (dnase_experiment.bw).GC-matched negative sampling (
dnase_experiment.negatives.bed).Train a 9-layer 128-filter Cherimoya model with
signal_groups=[1]andn_control_tracks=0.Compute count attributions via saturation mutagenesis on the validation chromosomes.
Call seqlets, annotate with tomtom-lite.
Run TF-MoDISco motif discovery.
Marginalize each motif at the center of negative loci.
Notes¶
DNase-seq has historically been processed by many slightly different protocols. If you trust the upstream pipeline that produced the BAM, you can leave the shift parameters at 0; if you produced the BAM yourself and want footprint-resolution predictions, a small
+1 / 0shift sometimes improves results.Like ATAC-seq, DNase-seq peaks can vary significantly in summit height. The same
max_counts/max_jitteradvice from the Recipe: ATAC-seq recipe applies.
Outputs¶
See the “Outputs” table in CLI Pipeline Walkthrough. The
trained model is at dnase_experiment.torch and the TF-MoDISco
report is at dnase_experiment_modisco/.