ChromatinHD-diff¶

In [ ]:

hide_output

import chromatinhd as chd
import polyptich as pp
import matplotlib.pyplot as plt

import chromatinhd as chd import polyptich as pp import matplotlib.pyplot as plt

ChromatinHD-diff uses accessibility fragments to assess how accessibility changes between (clusters of) cells.

We first load in all the input data which was created in the data preparation tutorial.

In [11]:

import pathlib

dataset_folder = pathlib.Path("example")
fragments = chd.data.Fragments(dataset_folder / "fragments")
transcriptome = chd.data.Transcriptome(dataset_folder / "transcriptome")
folds = chd.data.folds.Folds(dataset_folder / "folds" / "5x1")
clustering = chd.data.Clustering(dataset_folder / "clustering")

import pathlib dataset_folder = pathlib.Path("example") fragments = chd.data.Fragments(dataset_folder / "fragments") transcriptome = chd.data.Transcriptome(dataset_folder / "transcriptome") folds = chd.data.folds.Folds(dataset_folder / "folds" / "5x1") clustering = chd.data.Clustering(dataset_folder / "clustering")

Train the models¶

The basic ChromatinHD-diff model can be trained as follows. Model training and inference will be much faster if a GPU is available and correctly installed:

In [14]:

chd.get_default_device()

chd.get_default_device()

Out[14]:

'cuda:1'

In [12]:

models = chd.models.diff.model.binary.Models(dataset_folder / "models" / "diff", reset=True)

models = chd.models.diff.model.binary.Models(dataset_folder / "models" / "diff", reset=True)

In [ ]:

hide_output

models.train_models(fragments = fragments, clustering = clustering, folds = folds)

models.train_models(fragments = fragments, clustering = clustering, folds = folds)

Interpret positionally¶

Currently, the ChromatinHD-model is purely positional, i.e. it only looks whether Tn5 insertion sites increase or decrease within a region. As such, we can only interpret it positionally:

In [23]:

import chromatinhd.models.diff.interpret.regionpositional

import chromatinhd.models.diff.interpret.regionpositional

In [24]:

clustering.cluster_info.index.name = "cluster"

clustering.cluster_info.index.name = "cluster"

In [25]:

regionpositional = chromatinhd.models.diff.interpret.regionpositional.RegionPositional(
    path=models.path / "interpret" / "regionpositional"
)
regionpositional.score(
    fragments = fragments,
    clustering = clustering,
    models = models,
    force=True,
)

regionpositional = chromatinhd.models.diff.interpret.regionpositional.RegionPositional( path=models.path / "interpret" / "regionpositional" ) regionpositional.score( fragments = fragments, clustering = clustering, models = models, force=True, )

  0%|          | 0/51 [00:00<?, ?it/s]

Out[25]:

example/models/diff/interpret/regionpositional (chromatinhd.models.diff.interpret.regionpositional.RegionPositional)

• regions
• probs ( ENSG00000132465 [13,4001], 367.8Kb, ENSG00000107317 [13,4001], 384.9Kb, ENSG00000134532 [13,4001], 341.9Kb, ...)

Plot differential accessibility at a specific window¶

To avoid overplotting, we will plot only the window close to the TSS.

In [26]:

symbol = "IRF1"
gene_id = transcriptome.gene_id(symbol)

symbol = "IRF1" gene_id = transcriptome.gene_id(symbol)

In [27]:

window = [-10000, 10000]

window = [-10000, 10000]

In [35]:

fig = pp.grid.Figure(pp.grid.Grid(padding_height=0.05, padding_width=0.05))
width = 10

region = fragments.regions.coordinates.loc[transcriptome.gene_id(symbol)]
panel_genes = chd.plot.genome.genes.Genes.from_region(region, width=width, window = window)
fig.main.add_under(panel_genes)

panel_differential = chd.models.diff.plot.Differential.from_regionpositional(
    transcriptome.gene_id(symbol), regionpositional, cluster_info=clustering.cluster_info, panel_height=0.5, width=width, window = window
)
fig.main.add_under(panel_differential)

panel_expression = chd.models.diff.plot.DifferentialExpression.from_transcriptome(
    transcriptome=transcriptome, clustering=clustering, gene=transcriptome.gene_id(symbol), panel_height=0.5
)
fig.main.add_right(panel_expression, row=panel_differential)

fig.display()

fig = pp.grid.Figure(pp.grid.Grid(padding_height=0.05, padding_width=0.05)) width = 10 region = fragments.regions.coordinates.loc[transcriptome.gene_id(symbol)] panel_genes = chd.plot.genome.genes.Genes.from_region(region, width=width, window = window) fig.main.add_under(panel_genes) panel_differential = chd.models.diff.plot.Differential.from_regionpositional( transcriptome.gene_id(symbol), regionpositional, cluster_info=clustering.cluster_info, panel_height=0.5, width=width, window = window ) fig.main.add_under(panel_differential) panel_expression = chd.models.diff.plot.DifferentialExpression.from_transcriptome( transcriptome=transcriptome, clustering=clustering, gene=transcriptome.gene_id(symbol), panel_height=0.5 ) fig.main.add_right(panel_expression, row=panel_differential) fig.display()

Plot differential accessibility for all windows¶

In [48]:

windows = regionpositional.select_windows(gene_id)
breaking = pp.grid.Breaking(windows)

windows = regionpositional.select_windows(gene_id) breaking = pp.grid.Breaking(windows)

In [49]:

fig = pp.grid.Figure(pp.grid.Grid(padding_height=0.0, padding_width=0.05))
width = 10

region = fragments.regions.coordinates.loc[transcriptome.gene_id(symbol)]
panel_genes = chd.plot.genome.genes.GenesExpanding.from_region(region, breaking=breaking)
fig.main.add_under(panel_genes)

panel_differential = chd.models.diff.plot.DifferentialBroken.from_regionpositional(
    transcriptome.gene_id(symbol), regionpositional, cluster_info=clustering.cluster_info, panel_height=0.5, breaking=breaking, window = window
)
fig.main.add_under(panel_differential)

panel_expression = chd.models.diff.plot.DifferentialExpression.from_transcriptome(
    transcriptome=transcriptome, clustering=clustering, gene=transcriptome.gene_id(symbol), panel_height=0.5
)
fig.main.add_right(panel_expression, row=panel_differential)

fig.display()

fig = pp.grid.Figure(pp.grid.Grid(padding_height=0.0, padding_width=0.05)) width = 10 region = fragments.regions.coordinates.loc[transcriptome.gene_id(symbol)] panel_genes = chd.plot.genome.genes.GenesExpanding.from_region(region, breaking=breaking) fig.main.add_under(panel_genes) panel_differential = chd.models.diff.plot.DifferentialBroken.from_regionpositional( transcriptome.gene_id(symbol), regionpositional, cluster_info=clustering.cluster_info, panel_height=0.5, breaking=breaking, window = window ) fig.main.add_under(panel_differential) panel_expression = chd.models.diff.plot.DifferentialExpression.from_transcriptome( transcriptome=transcriptome, clustering=clustering, gene=transcriptome.gene_id(symbol), panel_height=0.5 ) fig.main.add_right(panel_expression, row=panel_differential) fig.display()