Introduction

HTAD is an Active Learning Tool for Matrix-based Detection of Chromatin Domain

The instructions below or this video can help you install and run HTAD.

Citation

Shen, W., Zhang, P., Jiang, Y. et al. HTAD: a human-in-the-loop framework for supervised chromatin domain detection. Genome Biol 25, 302 (2024). https://doi.org/10.1186/s13059-024-03445-x

Installing HTAD

We recommend using conda to install HTAD.

$ git clone https://github.com/shenscore/HTAD
$ cd HTAD
$ conda create -n HTAD python=3.11
$ conda activate HTAD
$ pip install -r requirements.txt

Running HTAD

(i) Feature extraction

Calculate the potential TADs and corresponding TAD features.

python /path/to/HTAD/src/calcFeatures.py [cooler] [resolution(s)] [outPrefix]
# example
python /path/to/HTAD/src/calcFeatures.py /path/to/HTAD/test_data/sim1.mcool 10000,20000,40000 Test

parameters:

• cooler: cooler file of Hi-C data

• resolutions: resolutions used to calculate potential TADs and corresponding TAD features, separated by ','.

• outPrefix: prefix of output files

Output:

• outPrefix.di_check_value
• outPrefix.10k.pkl, outPrefix.20k.pkl, outPrefix.40k.pkl

(ii) Manual annotation

Run the HTAD labeler server and train TAD identification model.

(1). Launch TAD labeler

to start TAD labeler, run:

cd dataLabel
python manage.py runserver

Then visit the corresponding port: e.g. 127.0.0.1:8000

(2). Submit input file

Input:

• the file path of your cooler file and feature file of potential TADs

• resolution of the feature file

• tag(prefix) for the output TAD identification model file

Then the server will select samples (50 for each round) for manual labeling.

The webpage will show corresponding heatmap with TAD marked by yellow triangle.

You can quickly judge whether current TAD is positive. We also provide the prediction of current model in the interface.

After each round, the server will generate a XX_roundX.h5 model file.

We suggest that the model should be trained in around 11 rounds with the highest resolution to get the best performance.

Note that we could just use the model trained by the highest resolution to identify TAD from other resolutions.

Output:

• *round*.h5
• *finalmodel.h5

(iii) TAD identification

given the well trained TAD model file model.h5, run:

python predictTAD.py [model] [feature file(s) of potential TADs] [resolution(s)] [n] [outPrefix]
# example
python predictTAD.py model.h5 potentialTAD.10k.pkl,potentialTAD.20k.pkl,potentialTAD.40k.pkl 10000,20000,40000 0 Test

parameters:

• model: TAD model file generated from previous step.
• feature files of potential TADs: feature files of potential TADs, separated by ','.
• resolutions: corresponding resolutions of feature files, separated by ','.
• n: expected TAD number, 0 for None (default).
• outPrefix

Output:

• Test.10k.bedpe Test.20k.bedpe Test.40k.bedpe

(iv) Merge the multi-resolution TAD results

python mergeTAD.py [resolution] [DI check value file] [TAD files of multiple resolutions] [output file]
# example
python mergeTAD.py 10000 di_check_value.10k 10k.bedpe,20k.bedpe,40k.bedpe final.bed

parameters:

• resolution: The highest resolution of TAD files.
• DI check value file: the DI check value file automately generated from the feature extraction step.
• TAD files of multiple resolutions: TAD files of multiple resolutions generated from previous step, separated by ','.
• output file

Output:

• The final TAD identification result in BED format.

For visualization, we recommend utilizing pyGenomeTracks or juicebox.

Contact us

Wei Shen: shenwei4907@foxmail.com