Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/lczech/grenedalf-paper

Code for tests and benchmarks of our paper on grenedalf
https://github.com/lczech/grenedalf-paper

Last synced: 23 days ago
JSON representation

Code for tests and benchmarks of our paper on grenedalf

Awesome Lists containing this project

README

        

# grenedalf-paper

Code for tests and benchmarks of the paper on our tool [grenedalf](https://github.com/lczech/grenedalf):

> grenedalf: population genetic statistics for the next generation of pool sequencing.

> Lucas Czech, Jeffrey P. Spence, Moisés Expósito-Alonso.

> Bioinformatics, 2024. doi:[10.1093/bioinformatics/btae508](https://doi.org/10.1093/bioinformatics/btae508)

We here provide tests scripts to benchmark [grenedalf](https://github.com/lczech/grenedalf) against existing tools:

* [grenedalf](https://github.com/lczech/grenedalf)
* [PoPoolation 1](https://sourceforge.net/projects/popoolation/) (diversity)
* [PoPoolation 2](https://sourceforge.net/projects/popoolation/) (FST)
* [poolfstat](https://cran.r-project.org/web/packages/poolfstat/index.html) (FST)
* [npstat](https://github.com/lucaferretti/npstat) (diversity)

See the `software` directory here for their setup. For the plotting, we furthermore need some python tools, as specified in the `common/conda.yaml` file. As always with these things, versions have to be exact.

We run the following tests here:

* `benchmark-grenenet`: Benchmarks on real-world data from GrENE-net, subsetting one or two files to increasing numbers of positions to show scaling with respect to the genome length.
* `benchmark-random`: Simple benchmarks based on randomly generated files, as a lower boundary of how much faster grenedalf is compared to its competitors.
* `benchmark-samples`: Benchmarks on real-world data from GrENE-net, increasing the number of files to show scaling wrt number of samples.
* `benchmark-scaling`: Benchmarks for strong and weak scaling of grenedalf on multi-core systems, with a small dataset.
* `benchmark-scaling-fst`: Benchmarks for strong and weak scaling of grenedalf on multi-core systems, with a larger dataset that shows better scaling.

Furthermore, we have some auxiliary tests and comparisons:

* `eval-bug-exam`: Examination of the two bugs in PoPoolation Tajima's D implementation.
* `eval-corr-grenenet`: Test how the results from grenedalf correlate with those of other tools.
* `eval-fst-biases`: Evaluation of the biases of different Pool-seq estimators of FST, as shown in our equations document.
* `eval-grenenet`: Quick test to assess the overall gain of grenedalf for our GrENE-net project.
* `eval-independent-test`: An independent bare-bone Python implementation of our equations, to check that the results of grenedalf are exactly as expected.
* `example-cathedral`: A prototype implementation of the cathedral plot for fst.
* `example-fst-ordination`: A simple large-scale example of using grenedalf on thousands of samples.

See the respective subdirectories for details.