sctrend 0.1.2

scTREND: An annotation-free single-cell time-resolved and condition-dependent hazard model

scTREND

Description

scTREND (single-cell time-resolved and condition-dependent hazard model) is a novel deep generative framework that integrates single-cell latent representations from a VAE with bulk-level cell-type proportions and hazard coefficients. This enables the computation of patient-level risk scores and the identification of cell populations whose prognostic impact dynamically changes across time and clinical conditions.

Teppei Shimamura's lab, Institute of Science Tokyo, Tokyo, Japan

Model architecture

The model comprises three main components: VAE for latent representation of single cells, bulk deconvolution based on DeepCOLOR, and a conditional piecewise constant hazard model for time- and condition-dependent risk estimation.

Requirements

Python >= 3.8.16

torch >= 1.13.1

lifelines >= 0.27.8

scanpy >= 1.9.5

pandas >= 1.5.3

numpy >= 1.23.5

matplotlib >= 3.7.2

scipy >= 1.10.1

Installation

You can install scTREND via pip:

!pip install sctrend

Application example

Explanation of key functions

• workflow.scTREND_preprocess: Preprocesses single-cell and bulk RNA-seq data to identify highly variable genes and prepare the inputs required for scTREND, with optional incorporation of driver-gene information.

• workflow.run_scTREND: Runs the scTREND workflow, including model training and estimation of time- and condition-dependent hazard coefficients.

Running scTREND

In this tutorial, we present an application of scTREND using a melanoma single-cell RNA-seq dataset (GSE115978) together with a bulk RNA-seq dataset from TCGA-SKCM. BRAF mutation status is incorporated as a driver condition, and the survival time axis is discretized into four time intervals. Under this setting, both the coefficients shared across all patients and the coefficients specific to BRAF-mutant patients can be visualized as shown below.

driver_genes = ["BRAF"]
edges = [...]  # time bin edges used in training
sc_adata, bulk_adata = workflow.scTREND_preprocess(sc_adata, bulk_adata,
     per=0.01, n_top_genes=5000, highly_variable="bulk", driver_genes=driver_genes)
driver_bulk_adata = bulk_adata[:, bulk_adata.var_names.isin(driver_genes)]
driver_bulk_adata.layers["SNV"] = ...  # samples × driver_genes (0:wild-type 1:mutated)
sc_adata, bulk_adata, model_params_dict, spatial_adata, exp = workflow.run_scTREND(
     sc_adata, bulk_adata,
     param_save_path="scTREND.pt",
     epoch=10000,
     batch_key="samples",
     driver_genes=driver_genes,
     driver_bulk_adata=driver_bulk_adata,
     edges=edges
)