peakachu 2.0

A supervised learning framework for chromatin loop detection in genome-wide contact maps.

NOTE: Peakachu (version>=1.1.2) now supports both .hic and .cool formats.

Introduction

Accurately predicting chromatin loops from genome-wide interaction matrices such as Hi-C data is critical to deepening our understanding of proper gene regulation. Current approaches are mainly focused on searching for statistically enriched dots on a genome-wide map. However, given the availability of orthogonal data types such as ChIA-PET, HiChIP, Capture Hi-C, and high-throughput imaging, a supervised learning approach could facilitate the discovery of a comprehensive set of chromatin interactions. Here, we present Peakachu, a Random Forest classification framework that predicts chromatin loops from genome-wide contact maps. We compare Peakachu with current enrichment-based approaches, and find that Peakachu identifies a unique set of short-range interactions. We show that our models perform well in different platforms, across different sequencing depths, and across different species.

Citation

Salameh, T.J., Wang, X., Song, F. et al. A supervised learning framework for chromatin loop detection in genome-wide contact maps. Nat Commun 11, 3428 (2020). https://doi.org/10.1038/s41467-020-17239-9

Installation

Peakachu requires Python3 and several scientific packages to run. It is best to first set up the environment using conda and then install Peakachu from PyPI::

conda config --add channels defaults
conda config --add channels bioconda
conda config --add channels conda-forge
conda create -n peakachu cooler scikit-learn numba joblib=1.1.0
conda activate peakachu
pip install -U peakachu hic-straw==0.0.6 

Peakachu should now be installed as a command-line tool within the new environment. Options for all peakachu commands and sub-commands can be accessed with the -h option.

peakachu -h

usage: peakachu [-h] {train,score_chromosome,score_genome,depth,pool} ...

Unveil Hi-C Anchors and Peaks.

positional arguments:
  {train,score_chromosome,score_genome,depth,pool}
    train               Train RandomForest model per chromosome
    score_chromosome    Calculate interaction probability per pixel for a chromosome
    score_genome        Calculate interaction probability per pixel for the whole genome
    depth               Calculate the total number of intra-chromosomal chromatin contacts and select the most appropriate pre-trained model
                        for you.
    pool                Print centroid loci from score_genome/score_chromosome output

options:
  -h, --help            show this help message and exit

Example: predicting loops in GM12878 Hi-C

The following example will download an example cooler file containing the GM12878 Hi-C data at the 10kb resolution, train a series of models using H3K27ac HiChIP interactions, and then predict loops using the trained models.

Data preparation

Peakachu requires the contact map to be a .cool file or a .hic file and any training input to be a text file in bedpe format. Example training data can be found at the training-sets subfolder. Cooler files may be found at the 4DN data portal.

wget -O Rao2014-GM12878-MboI-allreps-filtered.10kb.cool -L https://dl.dropboxusercontent.com/s/9d3hqyy0ou0rwsz/Rao2014-GM12878-MboI-allreps-filtered.10kb.cool?dl=0

Train a model and predict loops

It is always a good idea to call the help function immediately before entering a command:

peakachu train -h

usage: peakachu train [-h] [-r RESOLUTION] [-p PATH] [--balance] [-b BEDPE] [-w WIDTH] [--nproc NPROC] [-O OUTPUT]

options:
  -h, --help            show this help message and exit
  -r RESOLUTION, --resolution RESOLUTION
                        Resolution in bp (default 10000)
  -p PATH, --path PATH  Path to a .cool URI string or a .hic file.
  --balance             Whether or not using the ICE/KR-balanced matrix.
  -b BEDPE, --bedpe BEDPE
                        Path to the bedpe file containing positive training set.
  -w WIDTH, --width WIDTH
                        Number of bins added to center of window. default width=5 corresponds to 11x11 windows
  --nproc NPROC         Number of worker processes that will be allocated for training. (default 4)
  -O OUTPUT, --output OUTPUT
                        Folder path to store trained models.

peakachu train -r 10000 -p Rao2014-GM12878-MboI-allreps-filtered.10kb.cool --balance -O models -b gm12878.mumbach.h3k27ac-hichip.hg19.bedpe

This will train 23 random forest models, each labeled by a chromosome. The model for every chromosome was trained using interactions from all the other 22 chromosomes in the provided bedpe file. The purpose of this is to avoid Peakachu to predict loops from the same map it used for training, without overfitting. To use these models, you may either use the score_chromosome function to predict loops in only one chromosome, or the score_genome function to perform a genome-wide prediction.

peakachu score_chromosome -h

usage: peakachu score_chromosome [-h] [-r RESOLUTION] [-p PATH] [--balance] [-C CHROM] [-m MODEL] [-l LOWER] [-u UPPER]
                             [--minimum-prob MINIMUM_PROB] [-O OUTPUT]

options:
  -h, --help            show this help message and exit
  -r RESOLUTION, --resolution RESOLUTION
                        Resolution in bp (default 10000)
  -p PATH, --path PATH  Path to a .cool URI string or a .hic file.
  --balance             Whether or not using the ICE/KR-balanced matrix.
  -C CHROM, --chrom CHROM
                        Chromosome label. Only contact data within the specified chromosome will be considered.
  -m MODEL, --model MODEL
                        Path to pickled model file.
  -l LOWER, --lower LOWER
                        Lower bound of distance between loci in bins (default 6).
  -u UPPER, --upper UPPER
                        Upper bound of distance between loci in bins (default 300).
  --minimum-prob MINIMUM_PROB
                        Only output pixels with probability score greater than this value (default 0.5)
  -O OUTPUT, --output OUTPUT
                        Output file name.

peakachu score_chromosome -r 10000 -p Rao2014-GM12878-MboI-allreps-filtered.10kb.cool --balance -O GM12878-chr2-scores.bedpe -C chr2 -m models/chr2.pkl 
peakachu pool -r 10000 -i GM12878-chr2-scores.bedpe -o GM12878-chr2-loops.bedpe -t .9

The pool function serves to select the most significant non-redundant results from per-pixel probabilities calculated by the score functions. It is recommended to try different probability thresholds to achieve the best sensitivity-specificity tradeoff. The output is a standard bedpe file with the 7th and the final column containing the predicted probability from the random forest model and the interaction frequency extracted from the contact matrix, respectively, to support further filtering. The results can be visualized in juicebox or higlass by loading as 2D annotations. Here is an example screenshot of predicted GM12878 loops in juicer:

Using Peakachu as a standard loop caller

Models for predicting loops in Hi-C have been trained using CTCF ChIA-PET interactions, H3K27ac HiChIP interactions, and a high-confidence loop set (loops that can be detected by at least two orthogonal methods from CTCF ChIA-PET, Pol2 ChIA-PET, Hi-C, CTCF HiChIP, H3K27ac HiChIP, SMC1A HiChIP, H3K4me3 PLAC-Seq, and TrAC-Loop) as positive training samples, at a variety of read depths. Simply download the appropriate model file and directly run the score_genome/score_chromosome function if you want to detect chromatin loops on your own Hi-C or Micro-C maps.

If you are using Peakachu>=2.0, please select a model from the following table:

Total intra reads	high-confidence (5kb)	high-confidence (10kb)	high-confidence (25kb)
2 billion	total 5kb	total 10kb	total 25kb
1.8 billion	90% 5kb	90% 10kb	90% 25kb
1.6 billion	80% 5kb	80% 10kb	80% 25kb
1.4 billion	70% 5kb	70% 10kb	70% 25kb
1.2 billion	60% 5kb	60% 10kb	60% 25kb
1 billion	50% 5kb	50% 10kb	50% 25kb
900 million	45% 5kb	45% 10kb	45% 25kb
850 million	42.5% 5kb	42.5% 10kb	42.5% 25kb
800 million	40% 5kb	40% 10kb	40% 25kb
750 million	37.5% 5kb	37.5% 10kb	37.5% 25kb
700 million	35% 5kb	35% 10kb	35% 25kb
650 million	32.5% 5kb	32.5% 10kb	32.5% 25kb
600 million	30% 5kb	30% 10kb	30% 25kb
550 million	27.5% 5kb	27.5% 10kb	27.5% 25kb
500 million	25% 5kb	25% 10kb	25% 25kb
450 million	22.5% 5kb	22.5% 10kb	22.5% 25kb
400 million	20% 5kb	20% 10kb	20% 25kb
350 million	17.5% 5kb	17.5% 10kb	17.5% 25kb
300 million	15% 5kb	15% 10kb	15% 25kb
250 million	12.5% 5kb	12.5% 10kb	12.5% 25kb
200 million	10% 5kb	10% 10kb	10% 25kb
150 million	7.5% 5kb	7.5% 10kb	7.5% 25kb
100 million	5% 5kb	5% 10kb	5% 25kb
50 million	2.5% 5kb	2.5% 10kb	2.5% 25kb
30 million	1.5% 5kb	1.5% 10kb	1.5% 25kb
10 million	0.5% 5kb	0.5% 10kb	0.5% 25kb
5 million	0.25% 5kb	0.25% 10kb	0.25% 25kb

Instead, if you are using an older Peakachu version (<2.0), please select a model from this table:

Total intra reads	high-confidence (5kb)	high-confidence (10kb)	high-confidence (25kb)	CTCF Models (10kb)	H3K27ac Model (10kb)
2 billion	total 5kb	total 10kb	total 25kb	CTCF total	H3K27ac total
1.8 billion	90% 5kb	90% 10kb	90% 25kb	CTCF 90%	H3K27ac 90%
1.6 billion	80% 5kb	80% 10kb	80% 10kb	CTCF 80%	H3K27ac 80%
1.4 billion	70% 5kb	70% 10kb	70% 25kb	CTCF 70%	H3K27ac 70%
1.2 billion	60% 5kb	60% 10kb	60% 25kb	CTCF 60%	H3K27ac 60%
1 billion	50% 5kb	50% 10kb	50% 25kb	CTCF 50%	H3K27ac 50%
900 million	45% 5kb	45% 10kb	45% 25kb
850 million	42.5% 5kb	42.5% 10kb	42.5% 25kb
800 million	40% 5kb	40% 10kb	40% 25kb	CTCF 40%	H3K27ac 40%
750 million	37.5% 5kb	37.5% 10kb	37.5% 25kb
700 million	35% 5kb	35% 10kb	35% 25kb
650 million	32.5% 5kb	32.5% 10kb	32.5% 25kb
600 million	30% 5kb	30% 10kb	30% 25kb	CTCF 30%	H3K27ac 30%
550 million	27.5% 5kb	27.5% 10kb	27.5% 25kb
500 million	25% 5kb	25% 10kb	25% 25kb
450 million	22.5% 5kb	22.5% 10kb	22.5% 25kb
400 million	20% 5kb	20% 10kb	20% 25kb	CTCF 20%	H3K27ac 20%
350 million	17.5% 5kb	17.5% 10kb	17.5% 25kb
300 million	15% 5kb	15% 10kb	15% 25kb
250 million	12.5% 5kb	12.5% 10kb	12.5% 25kb
200 million	10% 5kb	10% 10kb	10% 25kb	CTCF 10%	H3K27ac 10%
150 million	7.5% 5kb	7.5% 10kb	7.5% 25kb
100 million	5% 5kb	5% 10kb	5% 25kb
50 million	2.5% 5kb	2.5% 10kb	2.5% 25kb
30 million	1.5% 5kb	1.5% 10kb	1.5% 25kb	CTCF 1.5%	H3K27ac 1.5%
10 million			0.5% 25kb
5 million			0.25% 25kb

To make it clear, let's download another Hi-C dataset from 4DN: https://data.4dnucleome.org/files-processed/4DNFI5IHU27G/@@download/4DNFI5IHU27G.mcool. Peakachu provides a handy function peakachu depth to extract the total number of intra-chromosomal pairs in your data and help you select the most appropriate pre-trained model:

peakachu depth -p 4DNFI5IHU27G.mcool::resolutions/1000000

The output of above command will be:

num of intra reads in your data: 592991890
num of intra reads in a human with matched sequencing coverage: 582003409
suggested model: 600 million

Therefore, we recommend using the 30% models (trained with ~600 million intra reads) to predict loops on this data.

peakachu score_genome -r 10000 --balance -p 4DNFI5IHU27G.mcool:resolutions/10000 -O 4DNFI5IHU27G-peakachu-10kb-scores.bedpe -m high-confidence.600million.10kb.w6.pkl
peakachu pool -r 10000 -i 4DNFI5IHU27G-peakachu-10kb-scores.bedpe -o 4DNFI5IHU27G-peakachu-10kb-loops.0.95.bedpe -t 0.95

Not just Hi-C

Peakachu has been tested on Hi-C, Micrco-C, and DNA SPRITE contact maps with good results. For training sets, ChIA-PET, HiChIP, and PLAC-Seq have been tested. The purpose of this software is ultimately to facilitate the interpretation of results from multiple types of experiments, and the user is encouraged to apply Peakachu's training framework to newer approaches as they become available.

Release Notes

Version 2.0 (09/06/2022)

1. Re-trained the models using the latest scikit-learn v1.1.2
2. Used the distance-normalized signals instead of original contact signals
3. Added a 2D Gaussian filter followed by min-max scaling to pre-process each training image
4. Optimized the computation efficiency using numba and matrix operations.