guanine
ccre_propensity¶
tags: candidate cis-Regulatory Elements, cCREs, Histone Marks, DHS, functional genomics, sequence function, reference sequences, experimental measurements
tl;dr¶
ccre_propensity (aka ccre_prop) is a multitask dataset of epigenetic markers for functional elements in the human genome. Models fit to one or all subtask(s) in ccre_prop should be able to ‘tag’ sequences containing functional elements, i.e. DNA sequences that do something other than “get turned into proteins.”
overview¶
candidate Cis-Regulatory Elements, or cCREs, are a spectrum of genome annotations that represent genomicists’ mapping of sequence elements, e.g. promoters, or DNA segments that increases downstream transcription. By design, cCREs transcend cell types – a cCRE in one tissue type is still present in another tissue type, even if its regulatory logic is left dusty and unused. The ENCODE registry of cCREs are based on where epigenetic signals – typically histone marks – are found across the human reference genome hg38.
This epigenetic signal allows for functional labels to be assigned to a sequence of DNA, be it promotor, enhancer, structural CTCF site, or even just a DNase Hypersensitive Site (DHS – an ingredient in all cCREs). It was created from the SCREEN v2 database, itself a subset of the ENCODE project. Sequence labels \(y\) correspond to the likelihood an accessible sequence would be annotated as having H3K4me3 or H3K27ac histone marks, CTCF-binding properties, or DNase hypersensitivity (DNase). Unlike typical cCRE tracks, which only measure experimental outcomes in a single cell line or cell type, cCRE is cell-type-agnostic and corresponds to basal tendency, or propensity, to be functional across cell types. ccre_prop, like its unconditional counterpart, dnase_propensity, is cell-type-agnostic, but cell-type-specific models will likely still perform well on this task due to the nature of its construction.
The high-level statistical model is \(y_H \sim p(H \ | \ X, \textrm{DHS})\) for \(H\) a given epigenetic signal and \(X\) a DNA sequence that is presumed accessible (i.e. a DHS). There are four signals used in ccre_prop, one per each subtask, the scores of models are averaged into a single combined score – subscores are often reported in appendices or omnibus tables as ccre-H, e.g. ccre-H3K4me3. Note that the loci in ccre_prop are a strict subset of loci in the dnase_propensity task, since a site must be accessible to be considered a candidate CRE in Encode/SCREEN.
Class labels are ordinal and range from zero (0, no epigenetic signal) to four (4, nearly ubiquitous epigenetic signal across cell types). Intermediate scores of one to three (1-3) represent some degree of cell-type-specific signal, although the type of cells is left unsaid.
Warning
Models trained to predict ccre_prop may not be well-behaved on inaccessible sequences, since only DHS sequences are seen during training. To use the predictions of ccre_prop-trained models, one should first condition their output on the result of dnase_propensity models or similar measures of accessibility.
example models¶
model |
\(\rho\) |
|---|---|
Borzoi |
57.6904 |
Basenji2 |
56.4568 |
SEI |
56.3547 |
Enformer |
55.3816 |
Enformer (Pre-output) |
55.0775 |
5-mer LinSVR (baseline) |
36.3022 |
GC-content (baseline) |
20.5533 |
Note
the scores shown here are averaged across individual subtasks
interpretation¶
ccre_prop is a core indicator of sequence function, and models scoring well on it must at least recognize, if not comprehend, regulatory functional elements in the reference genome. That said, while ccre_propensity corresponds to the degree of cell-type-specific epigenetic signal, it does not indicate which cell types a sequence is functional in, nor can it measure a model’s ability to predict any specific cell type’s H3K4me3, H3K27ac, CTCF, or DHS tracks.
Additionally, while Cis-Regulatory elements are indicative of sequence function, a model that can predict ccre_prop is not guaranteed to understand their cell-type-specific implications, or to understand the impact of sequence or epigenetic perturbations (e.g. variant effects, CRISPR knockout effects).
Supervised models like DeepSEA, Enformer, Borzoi, etc, are prime examples of models built for ccre_prop – their own training data includes the same H3K4me3, H3K27ac, CTCF, and DHS tracks that make up ccre_prop. To restate, ccre_prop is in-distribution for supervised models trained on ENCODE egigenetic tracks, and those that score below 100 should be understood as having underfit their training data.
Tip
The most closely related task is dnase_propensity, which measures the accessibility ingredient of Cis-Regulatory Element functionality (by definition, see ENCODE v3). One can understand the difference between dnase_propensity and ccre-dnase to be the likelihood of a sequence \(X\) being accessible versus its degree of accessibility (conditioned on being accessible), although this is a loose relationship.
example usage¶
first, clone the dataset from huggingface (make sure you have Git LFS installed):
git clone https://huggingface.co/datasets/guanine/ccre_propensity
then, read the file into main memory with your favorite file parser
import pandas as pd
# 1per is the recommended few-shot training split
train_dat = pd.read_csv('ccre_propensity/bed/1per/1per.bed', sep='\t')
train_dat.head()
finally, splice the sequence out with your preferred genome reader, e.g. twobitreader
from twobitreader import TwoBitFile
# download from https://hgdownload.cse.ucsc.edu/goldenpath/hg38/bigZips/hg38.2bit
hg38 = TwoBitFile('hg38.2bit')
CONTEXT_SIZE = 8192 # change this for your model
row = train_dat.iloc[0]
ch = row['#chr']
st = row['center']-CONTEXT_SIZE//2
en = row['center']+CONTEXT_SIZE//2
ta = row['task'] ## split by task, in case you don't wish to multitask
seq = hg38[ch][st:en]
# optionally convert your sequence to uppercase before tokenizing it, etc
seq = seq.upper()
assert len(seq)==CONTEXT_SIZE # we recommend checking for truncation
build details¶
Basal epigenetic signal is approximated by integrating out cell-type-specific signal, i.e. \(\int_{c \in C} \ p(H \ | \ X, \textrm{DHS}, c)\) for \(c\) a given cell-type. Specifically, for the over 1600 cCRE tracks in SCREEN v2, the same tracks used to train DeepSEA, Enformer, etc, we consider the discrete summation \(y_H(X) \propto \sum_{t \in \textrm{H-cCRE tracks}} \ \alpha_t \ \cdot \ \textbf{1}_\textrm{H-cCRE}(X)\) for boolean indicator function \(\textbf{1}_\textrm{H-cCRE}(X)\), which can be decomposed into constituent indicator functions \(\textbf{1}_\textrm{H-cCRE}(X) = \textbf{1}_\textrm{H}(X) \ \cdot \ \textbf{1}_\textrm{DHS}(X)\), where both components represent boolean consensus peak-calling for the underlying tracks. Note this implies that \(\textbf{1}_\textrm{H-cCRE}(X)\) is undefined for cells lacking either signal – far fewer types (still several hundred per task) of cells are included in the construction compared to dnase_propensity.
Also note the weighting \(\alpha_t\), which allows us to downweight cancerous cell lines by half (to help mitigate cancer-specific epigenetic signals).
Because SCREEN v2 relies on consensus peak calling, the peak called loci for reference sequences is consistent across tracks, which allows for this otherwise difficult function to be well-defined.
Finally, for ease of modeling, the raw \(y_H(X)\) scores are binned from one to four (1-4), and a partially G/C-balanced control set of accessible sequences (i.e. from dnase_propensity) are added to the dataset with labels of zero (0).
controlled factors¶
repetitive elements (partial)
G/C content (partial)
immortalized cancer line accessibility (partial)
appears in¶
original citation¶
The ENCODE Project Consortium., Moore, J.E., Purcaro, M.J. et al. Expanded encyclopaedias of DNA elements in the human and mouse genomes. Nature 583, 699–710 (2020). https://doi.org/10.1038/s41586-020-2493-4