Decoupler-GPU

Accelerated Transcriptional Regulatory Analysis

Author: Severin Dicks Copyright scverse

Here, we explore the updated Decoupler functions for inferring transcriptional regulatory activity, leveraging the GPU-accelerated implementation from rapids-singlecell.

By running these analyses on GPUs, we achieve faster computation and scalability for large single-cell datasets. We will apply Decoupler methods to estimate transcription factor (TF) activity and pathway activity, utilizing curated regulatory networks such as CollecTRI.

import rapids_singlecell as rsc
import scanpy as sc
import cupy as cp
import pandas as pd
import anndata as ad
import decoupler as dc

import rmm
from rmm.allocators.cupy import rmm_cupy_allocator
rmm.reinitialize(
    managed_memory=False, # Allows oversubscription
    pool_allocator=True, # default is False
)
cp.cuda.set_allocator(rmm_cupy_allocator)

ℹ️ Note: The dataset used in this notebook is generated in 01_basic_workflow.ipynb.

adata = ad.read_h5ad("h5/dli_decoupler.h5ad")

adata = adata.raw.to_adata()

rsc.get.anndata_to_GPU(adata)

CollecTRI network

CollecTRI is a comprehensive resource containing a curated collection of TFs and their transcriptional targets compiled from 12 different resources. This collection provides an increased coverage of transcription factors and a superior performance in identifying perturbed TFs compared to other literature based GRNs such as DoRothEA. Similar to DoRothEA, interactions are weighted by their mode of regulation (activation or inhibition).

collectri = dc.op.collectri(organism="human",license="commercial")

Transcription Factor Activity Inference with ULM

To infer transcription factor (TF) activity, we use the Univariate Linear Model (ULM) from Decoupler.
ULM estimates the regulatory influence of TFs on gene expression by fitting a linear model for each TF-gene interaction.

We apply ULM using the Dorothea regulatory network retrieved earlier.

%%time
rsc.dcg.ulm(data=adata, net=collectri, verbose = True, raw=False, bsize=10000, tmin=3)

CPU times: user 16.4 s, sys: 610 ms, total: 17 s
Wall time: 17 s

score = dc.pp.get_obsm(adata=adata, key="score_ulm")

sc.pl.embedding(score,basis="X_umap_harmony", color=["cell_type", "GATA3"], cmap='RdBu_r',ncols=1)
sc.pl.violin(score, keys=['GATA3'], groupby='cell_type', rotation=90, size = 0)

Transcription Factor Activity Inference with MLM

Next, we infer transcription factor (TF) activity using the Multivariate Linear Model (MLM) from Decoupler.
Unlike ULM, which considers each TF independently, MLM accounts for multiple TFs simultaneously,
providing a more comprehensive estimation of regulatory influences.

%%time
rsc.dcg.mlm(data=adata, net=collectri, verbose = True, raw=False, bsize=10000, tmin=3)

CPU times: user 14.1 s, sys: 1.19 s, total: 15.3 s
Wall time: 15.2 s

Transcription Factor Activity Inference with AUCell

Next, we infer transcription factor (TF) activity using AUCell from Decoupler.
AUCell estimates TF activity based on the enrichment of TF target genes within the top-expressed genes in each cell.
Unlike regression-based methods (ULM, MLM), AUCell is a rank-based approach that does not assume linear relationships.

%%time
rsc.dcg.aucell(data=adata, net=collectri, verbose = True, raw=False, bsize=5000, tmin=3)

CPU times: user 31.2 s, sys: 152 ms, total: 31.4 s
Wall time: 31.3 s

score = dc.pp.get_obsm(adata=adata, key="score_aucell")

sc.pl.embedding(score,basis="X_umap_harmony", color=["cell_type", "GATA3"], cmap='RdBu_r',ncols=1)
sc.pl.violin(score, keys=['GATA3'], groupby='cell_type', rotation=90, size = 0)

model = pd.read_parquet("nets/progeny.parquet")