Cluster regions based on similar/dissimilar ChIP-seq signal profile across the whole region and not just overall mean signal.

Question

[rls_08]

When evaluating my ChIP-seq signal across a set of regions, the heatmap metaplot suggests that my signal may have at least 3 distinct profiles of enrichment (a single peak, a bimodal peak, and no enrichment) as indicated by the image below. I would like to derive at least 3 clusters of regions based on the signal profiles, but when using kmean=3 in deeptools plotProfile, I feel like the algorithm does not quite seem to capture the differences in signal profiles but rather the overal mean signal across all bins in a region (not sure if this is how deeptools works, but just a guess based on what I have seen on my data). Do you have any recommendations on what tool(s) to use to better cluster these regions?

[Ming Tommy Tang]

Take a look at https://github.com/jokergoo/ComplexHeatmap/issues/57 I am writing a long blog post for all things related to ChIPseq heatmap.

Answer 1

[LChart]

You can pull the ScaleRegion data directly into scikit learn:

import pandas as pd
import numpy as np
import sklearn

dat = pd.read_csv('output.tsv.gz', sep='\t', comment='@')
dat = dat[dat.iloc[:,6:].isna().sum(axis=1) == 0]
signal = dat.iloc[:,6:].values
sig_nrm = signal / np.linalg.norm(signal, axis=1)[:, None]
train = np.random.default_rng().choice(sig_nrm, 1000)
sklearn.cluster.MeanShift().fit(train).predict(signal)

On my randomly-selected ChIP-seq data I get 3 clusters of very different sizes. You can play around with the clustering algorithm.