scCODA reproducibility

This repository contains the code that was used to produce the results and plots for scCODA: A Bayesian model for compositional single-cell data analysis (Büttner and Ostner et al., 2020). The package containing the model can be found here.

Simulated data benchmarks

The directory benchmarking contains code for the three benchmarks (threshold determination, model comparison and overall benchmark). The data generation setup for all benchmarks is in generate_data.py. Each benchmark can then be executed via the according <benchmark_xy>_run.py script. The analysis and plots are produced in a jupyter notebook named <benchmark_xy>_analysis.py for each benchmark.

Note: Since running the benchmarks is very resource intensive, we provided all generated data and benchmark results at TODO!

To ensure compatibility with other versions of some python packages, we provided the data and results as .csv and .h5 files instead of .pkl files, which are produced by the code. Conversion functions are in benchmarking.to_from_pickle.py.

The analysis notebooks can be run with the files benchmark_results and sccoda_effects for each benchmark. For running the benchmarks, please convert the files generated_data and generation_parameters via the function benchmarking.to_from_pickle.benchmark_datasets_to_pickle:

Example workflow for reproducing a benchmark

1. Navigate to the parent directory of scCODA_reproducibility

2. a) Download the benchmark data and results from zenodo and unpack them here

   or

   b) If you want to re-run a benchmark, create the following directory structure:

   <parent_directory>
   |_  <scCODA_reproducibility>
   |_  sccoda_benchmark_data
       |_  overall_benchmark
           |_  data_overall_benchmark
           |_  overall_benchmark_results
           |_  generated_datasets_overall_benchmark
           |_  overall_benchmark_plots
       |_  model_comparison_benchmark
           |_  data_model_comparison
           |_  model_comparison_results
           |_  generated_datasets_model_comparison
           |_  model_comparison_plots
       |_  threshold_determination_benchmark
           |_  data_threshold_determination
           |_  threshold_determination_results
           |_  generated_datasets_threshold_determination
           |_  threshold_determination_plots
   

   Then run benchmarking.generate_data to generate all benchmark data. The generated_datasets_<benchmark_name> directories should now be filled with files.

   Then, execute the according script benchmarks.<benchmark_name>_run. This produces results in the directory <benchmark_name>_results.

   Once you finished re-running all the benchmarks you want, run benchmarking.to_from_pickle to convert the data to standardized file types.

3. To re-do the benchmark analysis, run the according jupyter notebook benchmarking.<benchmark_name>_analysis.

Applications

All applications of scCODA that were shown in the article can be found under applications. For each analysis, a jupyter notebook is provided.