Frequency Chord diagram and Heatmap in Python

[maximelepetit]

Hi ,

Thanks a lot for this useful tool !

I'm running Liana and i'm curious that you cannot provide methods to generate Chord diagram or heatmap as in the liana R wrapper.

Have you planned to add these methods ?

Best regards,

Maxime

[dbdimitrov]

Hi @maximelepetit,

Thanks for opening an issue!
I've been considering implementing one of the two to liana-py.

I will do my best to have at least one implemented soon.

In the meantime, ChatGPT does an excellent job when generating network plots with networkx. And a plot with https://github.com/DingWB/PyComplexHeatmap should be relatively quick to make :)

Best wishes,
Daniel

[maximelepetit]

Hi Daniel,

For Chords diagrams i found an easy and quickly way with pyCirclize :
One example :

import pycirclize
from pycirclize import Circos
import pandas as pd
from pycirclize.parser import Matrix


## Extract liana_res
df1 = adata.uns['liana_res'][['source','target']].groupby(['source','target']).size().reset_index(name='counts')
df = df1.pivot(index='source',columns='target',values='counts')

matrix_data=df.values
row_names = list(df.index)
col_names = row_names

matrix_df = pd.DataFrame(matrix_data, index=row_names, columns=col_names)

# Initialize from matrix (Can also directly load tsv matrix file)
circos = Circos.initialize_from_matrix(
    matrix_df,
    space=3,
    cmap="tab10",
    r_lim=(93, 100),
    ticks_interval=500,
    label_kws=dict(r=94, size=12, color="white"),
    link_kws=dict(direction=1, ec="black", lw=0.5),
)

#print(matrix_df)
fig = circos.plotfig()

#print(matrix_df)
fig = circos.plotfig()

Give :

Best !
Maxime

[maximelepetit]

You can highlight specific groups :

import pycirclize
from pycirclize import Circos
import pandas as pd
from pycirclize.parser import Matrix

# Extract liana_res
df1 = adata.uns['liana_res'][['source','target']].groupby(['source','target']).size().reset_index(name='counts')
df = df1.pivot(index='source',columns='target',values='counts')

matrix_data=df.values
row_names = list(df.index)
col_names = row_names
matrix_df = pd.DataFrame(matrix_data, index=row_names, columns=col_names)

# Define link_kws handler function to customize each link property
def link_kws_handler(from_label: str, to_label: str):
    ## Highlith Astro, CR, Microglia source
    if from_label in ('Astrocytes', 'CR', 'Microglia'):
        # Set alpha, zorder values higher than other links for highlighting
        return dict(alpha=0.5, zorder=1.0)
    else:
     return dict(alpha=0.1, zorder=1.0)

# Initialize from matrix (Can also directly load tsv matrix file)
circos = Circos.initialize_from_matrix(
    matrix_df,
    space=3,
    cmap="tab10",
    r_lim=(93, 100),
    ticks_interval=500,
    label_kws=dict(r=94, size=12, color="white"),
    link_kws=dict(direction=1, ec="black", lw=0.5),
    link_kws_handler=link_kws_handler,
)

#print(matrix_df)
fig = circos.plotfig()

Give :

;)

[maximelepetit]

Alternatively you can generate Sankey plot through pySankey :

from pySankey.sankey import sankey
import pandas as pd
df1 = adata.uns['liana_res'][['source','target']].groupby(['source','target']).size().reset_index(name='counts')
sankey(
    left=df1["source"], right=df1["target"], 
    leftWeight= df1["counts"], rightWeight=df1["counts"], fontsize=10
)

GIves :

[dbdimitrov]

Hi @maximelepetit,

Thanks a lot for all your suggestions!
I will pin the issue and will keep it open until I have some of those in place.

I plan to do a rehaul of liana's plotting functions, and I will make sure these are included also.

Daniel

[brainfo]

Hi, I feel like it's also good to have a hypothesis driven layout, showing the specific sender-ligand -- receptor-recepient pairs, with for example senders on one sphere and recepients on another and the color by any score, like the magnitude, or the significance, rank, lrmean, or change between conditions.
I used to use ChordDiagram package in r.
Maybe it's good to have this in the utils in liana+ py?

while so far for convenience with same information we can simply show bar/dot plots

and to compare the same pair in different conditions, which metric is recommended to be tested? combined rank or individual continous magnitude metric?

[dbdimitrov]

Hi @brainfo, thanks for the suggestion. I'm still working on refactoring the algorithmic side of liana, but once I'm done with that. I will refactor the functions and extend them with the suggestions here.

re comparing across conditions: there are a few tutorials, e.g. using differential expression analysis with liana and PyDeSeq2, or the tutorials with MOFA or Tensor-cell2cell.

I would suggest checking those 🙂 though you could compare most scores as long as you normalize such that your conditions are comparable prior to running liana.

Hope this helps.