Decomprolute: Benchmarking study of proteomic based tumor deconvolution

The goal of this package is to provide various metrics to assess the ability of a deconvolution algorithm to identify specific cell types in bulk proteomics data. The package is fully dockerized and can be run with the installation of Docker and a CWL-compliant tool on your local machine.

We employed a modular architecture to enable 'plug and play' comparisons of different datasets and tools.The modules fall into three categories, each with a data collection and analysis module.

How to use

There are multiple ways to use this package. Given the multiple modules, you can use this package to simply run a deconvolution algorithm a specific dataset, or you can swap out the algorithm for a new one or run on your own data. These various approaches are listed in the workflows directory.

To deconvolve cell types on CPTAC data

The workflow script to run a single algorithm is located in the root of the workflows directory.

cd workflows
cwltool prot-deconv.cwl --cancer hnscc --protAlg mcpcounter --sampleType tumor --signature ../signature_matrices/LM7c.txt

This will run the MCP-counter algorithm on proteomics data from the CPTAC breast HNSCC cohort using our LM7c signature.

To assess performance of existing algorithms

In the absence of a proteomics gold standard, we have implemented three distinct metrics to determine the performance of each algorithm listed below.

Performance on simulated data

We have simulated both mRNA and proteomics data from established experiments as described below. We try to evaluate mRNA data on mRNA-derived simulations, and proteomics data on proteomics-derived simulated data. The datasets themselves are stored in the simulatedData directory.

We have included two YAML files to use as test runs of each simulation.

cd workflows/data-sim/
cwltool simul-data-comparison.cwl rna-sim-test.yml ##evaluate rna-based deconvolution
cwltool simul-data-comparison.cwl prot-sim-test.yml ##evaluate protein based deconvolution

These will produced the necessary summary statistics and figures.

mRNA-Proteomics Comparison

We also wanted to measure how consistent an algorithm was between mRNA and proteomics data. This iterates through all algorithms, data, and matrices to and compares how similar each cell type prediction is across mRNA vs. proteomic samples.

cd workflows/mrna-prot
cwltool mrna-prot-comparison.cwl alg-test.yml

This will run the evaluation in our test YAML file. To update the parameters, create your own YAML file. The algorithm currently has five parameters:

1. mrna-algorithms: List of algorithms to use to deconvolve mRNA data. One of epic, xcell, cibersort, mcpcounter.
2. prot-algorithms: List of algorithms to use to deconvolve protein data. One of epic, xcell, cibersort, mcpcounter.
3. cancerTypes: List of cancer types
4. signatures: List of signature matrices, currently found in the signature matrix directory
5. tissueTypes: list of tissue types: tumor, normal, or all

Pan-Immune clustering annotation

Lastly we can cross-reference known immune types with predicted cell types from the various deconvolution algorithms to ascertain how well

To add your own algorithm or data

To test your own algorithm, or data, we suggest putting it in its own Docker image. For more details please see the Contributing Guide.

Data

For this module we have collected real and simulated data.

CPTAC Data

This algorithm leverages data collected through the clinical proteomic tumor analysis consortium (CPTAC) as the foundation of its benchmarking metrics. This consortium has collected hundreds of patient tumor data, including proteomic and transcriptomic data from the same patients. Given the general confidence in transcriptomic-based tumor convolution, we can use these data to compare transcriptomic and proteomic tumor deconvolution in the same patient samples

We have collect this data via the CPTAC Python API to better match the mRNA data. This CWL tool and Docker image are in the protData and mRNAdata directories.

Below are the available tumor types:

Dataset name	Description	Data reuse status	Publication link
Brca	breast cancer	no restrictions	https://pubmed.ncbi.nlm.nih.gov/33212010/
Ccrcc	clear cell renal cell carcinoma (kidney)	no restrictions	https://pubmed.ncbi.nlm.nih.gov/31675502/
Colon	colorectal cancer	no restrictions	https://pubmed.ncbi.nlm.nih.gov/31031003/
Endometrial	endometrial carcinoma (uterine)	no restrictions	https://pubmed.ncbi.nlm.nih.gov/32059776/
Gbm	glioblastoma	no restrictions	https://pubmed.ncbi.nlm.nih.gov/33577785/
Hnscc	head and neck squamous cell carcinoma	no restrictions	https://pubmed.ncbi.nlm.nih.gov/33417831/
**Lscc	lung squamous cell carcinoma	password access only	unpublished**
Luad	lung adenocarcinoma	no restrictions	https://pubmed.ncbi.nlm.nih.gov/32649874/
Ovarian	high grade serous ovarian cancer	no restrictions	https://pubmed.ncbi.nlm.nih.gov/27372738/
**Pdac	pancreatic ductal adenocarcinoma	password access only	unpublished**

As such, datasets have been updated to following (added hnscc): ['brca', 'ccrcc', 'endometrial', 'colon', 'ovarian', 'hnscc', 'luad']

Simulated data

We also evaluate the algorithms on simulated data.

Algorithms

We have included numerous algorithms in this package. Docker files and requisite data are included in the tumorDeconvAlgs folder.

Algorithm	Source
cibersort
epic
xcell
mcpcounter

Cell type signatures

There are numerous ways to define the individual cell types we are using to run the deconvolution algorithms. We will upload specific lists to compare in our workflow.

List Name	Description	Source
LM7c	Seven cell types (B, CD4 T, CD8 T, dendritic cells, granulocytes, monocytes, NK) collapsed from proteomic data	Rieckmann et al.
3' PBMCs	Seven cell types (B, CD4 T, CD8 T (CD8 T + NK T), dendritic cells, megakaryocytes, monocytes, NK) from scRNA-seq data	Newman et al.
LM10	Ten cell types predicted by MCPCounter signature
LM22	The original matrix from cibersort