This is a reimplementation of lostruct from the original code: lostruct, by Joseph Guhlin with assistance by Peter Ralph.
Please see the Example Notebook
Lostruct-py is available on PyPi
pip install lostruct is the easiest way to get started.
Inputs should be a set of markers in BCF or VCF format. Both should be indexed as appropriate (see: bcftools index). Filtering before running this analysis is strongly suggested (Allele frequency, SNPs only, missingness, etc).
If you use this version, plesae cite it via Zenodo DOI: 10.5281/zenodo.3997106 as well as the original paper describing the method:
Li, Han, and Peter Ralph. "Local PCA shows how the effect of population structure differs along the genome." Genetics 211.1 (2019): 289-304.
This project also uses cyvcf2 for fast VCF processing and should be cited:
Brent S Pedersen, Aaron R Quinlan, cyvcf2: fast, flexible variant analysis with Python, Bioinformatics, Volume 33, Issue 12, 15 June 2017, Pages 1867–1869, https://doi.org/10.1093/bioinformatics/btx057
Please note numpy and R are different when it comes to row-major vs. column-major. Essentially, many things in the python version will be transposed from R.
Python >= 3.6 (may work with older versions). Developed on Python 3.8.5
- numba
- numpy
- cyvcf2
CyVCF2 requires zlib-dev, libbz2-dev, libcurl-dev, liblzma-dev; numa requires libllvm.
These may be installed with conda or pip, e.g. by running pip install -r requirements.txt.
See CHANGES.MD for the full list.
- Smallest circle
- Better in-built JAX support
- Package name changed to lostruct
- Parallelization of get_pc_dists
- Implementation of fastmath parameter for get_pc_dists
Tests were derived from the Medicago HapMap data. The python package and the R package have high correlation of values. If values begin to deviate, theese tests will now fail.
To run tests simply do:
pip install pytest
pytest --benchmark-disable tests/test_fns.py
The tests furthermore require unittest and scikit-bio (and pytest to run them this way).
To run tests with benchmarks, install the following:
pip install pytest-benchmark
Then run ```pytest``
Tox allows you run tests with multiple versions of the python interpreter in venvs. It is best to use pyenv to install multiple versions python to run before submitting pull requests to be certain tests complete successfully across all versions.
To run tests via tox:
pip install tox
To test correlation of results between the R and Python versions we used data from the Medicago HapMap project, specifically SNPs for sister taxa chromsome 1, processed, and run with LoStruct R.
bcftools annotate chr1-filtered-set-2014Apr15.bcf -x INFO,FORMAT | bcftools view -a -i 'F_MISSING<=0.2' | bcftools view -q 0.05 -q 0.95 -m2 -M2 -a -Oz -o chr1-filtered.vcf.gz
Rscript run_lostruct.R -t SNP -s 95 -k 10 -m 10 -i data/
Run 21 Aug 2020, using lostruct R git hash: 444b8c64bebdf7cdd0323e7735ccadddfc1c8989
This generates the mds_coords.tsv that is used in the correlation comparison. Additionally, the existing tests cover correlation.
To test the weight generation, a random sample of weights was created and used. Output was moved to lostruct-results/weights_mds_coords.csv and generated with the random_weights.txt found in test_data/random_weights.txt.
./run_lostruct.R -i data -t snp -s 95 -k 10 -m 10 -w random_weights.txt
Currently the end-user is expected to save the outputs. But would be good to save it in a similar way to lostruct R-code. Please open an issue if you need this.
We are not yet feature complete with the R implementation. If something is needed please check for existing issues and comment about your need.
PCoA returns the same results as lostruct's MDS implementation (cmdscale). In the example Jupyter notebook you can see the correlation is R =~ 0.998. Some examples of other methods of clustering / looking at differences are included in the notebook.
NUMBA and CyVCF2 are used for speeding up processes, and the software becomes multithreaded by default. The Sparse library is used to reduce memory requirements. parse_vcf function is multithreaded. Distance calculation is not.
Below two options are offered, fastmath for get_pc_dists function, and method="fsvd" for pcoa. When using both you will see a performance increase and memory requirement decrease. Accuracy should decrease, but the absolute correlation we see with our test dataset remains ~0.998. Be aware when using fsvd the sign of the correlation may change.
Jax is an open-source library.... optional support in lostruct. Allows for processing on GPU and TPU (untested here, so far). See JAX. You can enable it in the function get_pc_dists by setting jax=True. This results in XX% speedup. Fastmath (below) and JAX are both supported, although JAX outperforms fastmath.
Additionally, a mode implemented Numba's "fastmath" is available. For the function get_pc_dists set fastmath=True. This results in a ~8% speed boost with very little change in the final output (correlation to R code output remains >= 0.995). This was benchmarked on the Medicago data used in the jupyter notebook using timeit, with 100 repeats with fastmath=False and Fastmath=True.
get_pc_dists(result, fastmath=True)
The difference with fastmath=True and leaving it off can be seen here. Note: Downloading the file will allow you to see more detailed information, as some javascript is contained in the SVG but disabled on GitHub.
If you need to limit thread usage, please see Numba's guide
The R implementation handles very large datasets in less memory. The problem arises with the PCoA function. A metric MDS using sklearn may work. Another alternative would be to export the data and run cmdscale in R directly.
The sklearn MDS function differs from the scikit-bio function, here we focus on the scikit-bio version.
There are two options in python for this as well:
pcoa(method="fsvd", ...)
Which reduces memory and increases speed, at the cost of some accuracy.
pcoa(inplace=True, ...)
Centers a distance matrix in-place, further reducing memory requirements.
pcoa(number_of_dimensions=10)
Returns only the first 10 dimensions (configurable) of the scaling. This has no real effect if method is default or manuially set to "eigh" as the eigenvalues and eigenvectors are all calculated, so all are calculated and this becomes a truncation.
You can see the difference between method="fsvd" and method="eigh" (default) here. These are tested with a minimum of 50 rounds. Note: Downloading the file will allow you to see more detailed information, as some javascript is contained in the SVG but disabled on GitHub.
Using all three techniques, correlation is maintained although the sign may change.
mds = pcoa(pc_dists, method="fsvd", inplace=True, number_of_dimensions=10)
np.corrcoef(mds.samples["PC1"], mds_coords['MDS1'].to_numpy())[0][1]
-0.9978147088087447
For more information please see the applicable documentation as well as the relevant changelog. A Zenodo entry is also available on this topic.
Additional citations can be found in CITATIONS (UMAP, PHATE, Medicago HapMap).
We are using the code formatter BLACK. Also, code coverage is actually 100% but numba JIT'd code is not properly counted. As long as all tests complete everything is working.