In silico Screening of Antigen Combinations

This is a repository for the scripts used to do the analysis in the Dannenfelser et al. paper, Discriminatory power of combinatorial antigen recognition in cancer T cell therapies. Top predicted combinations can be viewed on our antigen explorer webserver.

Citation

Discriminatory power of combinatorial antigen recognition in cancer T cell therapies. Dannenfelser R, Allen G, VanderSluis B, Koegel AK, Levinson S, Stark SR, Yao V, Tadych A, Troyanskaya OG, Lim WA. Cell Systems. 2020. https://doi.org/10.1016/j.cels.2020.08.002

Usage

Due to size contraints, this repository contains expression data for a subset of the genes analyzed in the paper. The instructions below will work off this data. However it is easy to extend these scripts by processing an input gene expression matrix using the same format of the sample matrix data/test-normalized-matrix.txt, with genes as columns and samples as rows. In addition to genes, be sure to have a dataset column with unique identifiers, a type column with values of either cancer or normal, and a tissue.cancer column with a more specific sample label such as the type of cancer (e.g., Glioblastoma Multiforme) or the tissue type (e.g., Adipose Tissue).

To build the version used in the paper, download the 2015 release of RNAseq data from TCGA and version 7 of GTEx. Correct for batch effects using your favorite normalization scheme (we used COMBAT) and make sure the expression values are in log_2 space.

Instructions

1. Build the geometric sketches needed for the training and test sets:
   • setup a virtual environment and install relevant python packages (here we used conda)

   conda create -n antigen python=3.7
   conda activate antigen
   pip install -r requirements.txt

   • run CreateSketches.py

   python src/CreateSketches.py

   • once complete, sketch files should be constructed in results/sketches and the virtual environment is no longer necessary.
2. The remaining steps will be in R. First, load all R functions in the Functions.R file.
3. Calculate clustering-based scores for singles, doubles, and triples with CalculateClusteringScores.R. Depending on your computational resources running on the sample file may take a while to complete. Considering setting the tr variable to True to further reduce the number of double and triple combinations computed for testing purposes.
4. Run the decision tree evaluations in Evaluate.R to calculate performance metrics (F1, precision, recall). Like cluster score calculations, this script may take a long time to run based on your computing resources and the size of the starting dataset.
5. Use Analyze.R to process the top results and conduct the analysis carried out in the paper.

User Agreement

This code is released under a Clear BSD License and is intended for research/academic use only. For commercial use, please contact: Laurie Tzodikov (Assistant Director, Office of Technology Licensing), Princeton University, 609-258-7256.