DECODEM / DECODEMi: Systematic assessment of the roles of diverse cell types in tumor microenvironment in clinical response from bulk transcriptome

**The relevant manuscript is currently under review:

S. R. Dhruba, S. Sahni, B. Wang, D. Wu, Y. Schmidt, E. Shulman, S. Sinha, S. Sammut, C. Caldas, K. Wang, E. Ruppin. "Predicting breast cancer patient response to neoadjuvant chemotherapy from the deconvolved tumor microenvironment transcriptome", 2023.

We developed a novel computational framework called DECODEM (DEcoupling Cell-type-specific Outcomes using DEconvolution and Machine learning) that can systematically assess the roles of the diverse cell types in the tumor microenvironment (TME) in a given phenotype from bulk transcriptomics. In this work, we investigate the association of the cell types in breast cancer TME (BC-TME) to patient response to neoadjuvant chemotherapy (responder vs. non-responder). The framework is divided into two steps:

1. Deconvolution: we use CODEFACS to deconvolve the bulk gene expression into nine cell-type-specific gene expression profiles encompassing malignant, immune, and stromal cell types.
2. Machine Learning: we use a machine learning (ML) pipeline to build nine cell-type-specific predictors of chemotherapy response using the deconvolved expression profiles.

The output of the framework is the predictive power of each cell type (in terms of AUC and AP) which we use to assess the improvement over the bulk mixture and rank cell types in BC-TME. We further validate these top cell types in multiple independent BC cohorts encompassing both bulk and single-cell (SC) transcriptomics.
_{AUC = Area under the receiver operating characteristics curve
AP = Average precision, equivalent to the area under the precision-recall curve}

^{Figure: The full analysis pipeline for DECODEM and DECODEMi}

Furthermore, we investigate the interactions between different cell types in two ways:

• Multi-cell-ensemble: we incorporate the expression profiles of the top predictive cell types to boost the predictive power even further, yielding the best performance for an ensemble of immune and stromal cell types across two independent cohorts.
• DECODEMi: we extended DECODEM to DECODEMi ('i' stands for interaction) where we use the inferred cell-cell interactions (CCIs) (by using LIRICS) to identify the cellular communications that influence chemotherapy response in BC.

Our findings in breast cancer highlight the considerable predictive powers of the immune and stromal cells in the TME as well as denote key CCIs that are strongly predictive of chemotherapy response.

Dependencies

The deconvolution stage was performed on HPC environment using R and Rslurm (as part of CODEFACS). The CCI inference were performed by using LIRICS on the deconvolved data using R.

The ML predictors were developed on MacOS using python and further tested on linux (on HPC). The ML scripts can be run interactively using a python IDE or on command line as python script_name.py. Complementary analyses i.e., data preprocessing, enrichment analysis, CCI validation in SC, and plot generation were performed locally using R on RStudio.

Dependencies for python scripts:

python >= 3.8  
numpy >= 1.23   
pandas >= 1.4  
scikit-learn >= 1.1  
xgboost 1.6.1
pickle >= 3.0  
matplotlib >= 3.7
seaborn >= 0.12
tqdm >= 4.63  
lifelines >= 0.27  
pickle 4.0  

Dependencies for R scripts:

R >= 3.6  
tidyverse >= 1.3  
plyr >= 1.8
rtracklayer >= 1.57  
GenomicFeatures >= 1.50
clusterProfiler >= 4.6  
biomaRt >= 2.54  
msigdbr >= 7.5  
GSVA >= 1.45  
PRROC >= 1.3  
rstatix >= 0.7  
ggpubr >= 0.6  
glue >= 1.6  
Matrix >= 1.6  

Reproducing the results

All the results presented in the above manuscript can be reproduced by using the scripts provided in analysis. The assumption is that the different bulk expression datasets have already been deconvolved and put in the designated directories within data.

Running deconvolution with CODEFACS and LIRICS

The scripts for CODEFACS and LIRICS should respectively be put in analysis/deconvolution/CODEFACS and analysis/deconvolution/LIRICS. The cell type signature should be in data/celltype_signature.

• Deconvolution using CODEFACS was run by using the slurm scripts in analysis/deconvolution/job_scripts.
• The slurm scripts were run on the NIH HPC system, Biowulf.
• CCI inference using LIRICS was run by using the scripts in analysis/deconvolution/LIRICS.

Data preprocessing

All datasets should be deposited in data using the structure outlined. To process the deconvolved data into the desired formats, use the scripts in analysis/preprocessing.

Examples of some processed datasets are provided in data/TransNEO and data/BrighTNess.

DECODEM: Cell-type-specific prediction

• model_transneo_cv_v1.py: performs the cross-validation analysis using the TransNEO cohort.
• predict_sammut_validation_v2.py: trains the cell-type-specific/multi-cell-ensemble predictors using TransNEO and validates on the ARTemis + PBCP cohort.
  -predict_brightness_validation_v2.py: trains the cell-type-specific/multi-cell-ensemble predictors using TransNEO and validates on the BrighTNess cohort containing triple negative breast cancer (TNBC) patients.
• predict_tnbc_sc_validation_v2.py: trains the cell-type-specific predictors using TransNEO and validates on the Zhang et al. single-cell cohort of TNBC patients (SC-TNBC).
• stratify_tcga_validation_v3.py: trains the cell-type-specific predictors using TransNEO and stratifies survival on the TCGA-BRCA cohort.

If svdat = True in the scripts, the predictions will be saved in data/TransNEO/transneo_analysis/mdl_data (in .pkl format).

DECODEMi: CCI-based prediction

• model_transneo_lirics_cv_v3.py: performs the cross-validation analysis using TransNEO and extracts the corresponding top predictive CCIs.
• predict_sammut_lirics_validation_v2.py: trains the CCI-based predictor using TransNEO, validates on ARTemis + PBCP and extracts the corresponding top predictive CCIs.
• predict_brightness_lirics_validation_v2.py: trains the CCI-based predictor using TransNEO, validates on BrighTNess and extracts the corresponding top predictive CCIs.
• predict_sc_validation_cci_pseudopatients_v1.R: validates the top predictive CCIs in TNBC (using BrighTNess) extracted by DECODEMi with a SC pseudopatient cohort sourced from the Zhang et al. SC-TNBC cohort and generates Figs. S4E-F.

If svdat = True in the scripts, the predictions will be saved in data/TransNEO/transneo_analysis/mdl_data (in .pkl format).

Enrichment & association analyses

The enrichment analyses results and the figures (or panels) in the manuscript can be reproduced using the scripts in analysis/enrichment_and_figures.

• run_enrichment_top_cell_types_v3.R: performs cell-type-specific GSEA analysis and generates Fig. 3E.
• enrichment_cd4_cd8_tcells_v2.R: performs GSVA analysis for CD4⁺/CD8⁺ T-cells, estimates their predictive power and generates Supp. Figs. 3A-D.
• get_abundance_response_corr_v2.py: performs an association analysis between cell type abundance and chemotherapy response, and generates Supp. Figs. 3E-G.

If svdat = True in the scripts, the figure panels will be saved in data/plots (in .pdf format).

Reproducing the figures

Fig. 1 was generated using Biorender. To reproduce the remaining figures, use the following scripts in analysis/enrichment_and_figures:

• generate_plots_ctp_v2.py: generates Figs. 2, 3A-D, Supp. Figs. 1-2.
• generate_plots_cci_v2.py: generates Fig. 4, Supp. Figs. 4A-D.
• generate_plots_sc_surv_v2.py: generates Fig. 5, Supp. Fig. 5.

if svdat = True in the scripts, the figures will be saved in data/plots (in .pdf format).

Examples of the figures generated are provided in figures.

Contact:

Saugato Rahman Dhruba (saugatorahman.dhruba@nih.gov)
Cancer Data Science Lab, National Cancer Institute, National Institutes of Health