UniFrac is the de facto repository for high-performance phylogenetic diversity calculations. The methods in this repository are based on an implementation of the Strided State UniFrac algorithm which is faster, and uses less memory than Fast UniFrac. Strided State UniFrac supports Unweighted UniFrac, Weighted UniFrac, Generalized UniFrac, Variance Adjusted UniFrac and meta UniFrac, in both double and single precision (fp32). This repository also includes Stacked Faith (manuscript in preparation), a method for calculating Faith's PD that is faster and uses less memory than the Fast UniFrac-based reference implementation.
This repository produces standalone executables and a C API exposed via a shared library which can be linked against by any programming language.
A detailed description of the Strided State UniFrac algorithm can be found in McDonald et al. 2018 Nature Methods. Please note that this package implements multiple UniFrac variants, which may have their own citation. Details can be found in the help output from the command line interface in the citations section, and is included immediately below:
ssu
For UniFrac, please see:
Sfiligoi et al. mSystems 2022; DOI: 10.1128/msystems.00028-22
McDonald et al. Nature Methods 2018; DOI: 10.1038/s41592-018-0187-8
Lozupone and Knight Appl Environ Microbiol 2005; DOI: 10.1128/AEM.71.12.8228-8235.2005
Lozupone et al. Appl Environ Microbiol 2007; DOI: 10.1128/AEM.01996-06
Hamady et al. ISME 2010; DOI: 10.1038/ismej.2009.97
Lozupone et al. ISME 2011; DOI: 10.1038/ismej.2010.133
For Generalized UniFrac, please see:
Chen et al. Bioinformatics 2012; DOI: 10.1093/bioinformatics/bts342
For Variance Adjusted UniFrac, please see:
Chang et al. BMC Bioinformatics 2011; DOI: 10.1186/1471-2105-12-118
For GPU-accelerated UniFrac, please see:
Sfiligoi et al. PEARC'20; DOI: 10.1145/3311790.3399614
faithpd
For Faith's PD, please see:
Faith Biological Conservation 1992; DOI: 10.1016/0006-3207(92)91201-3
At this time, there is one primary way to install the library, through bioconda. It is also possible to clone the repository and install the C++ API with make.
Compilation has been performed on both LLVM 10.0.0 (OS X >= 10.12) or GCC 9 (Centos >= 7) and HDF5 >= 1.8.17.
Installation time should be a few minutes at most.
An example of installing UniFrac, and using it with CPUs as well as GPUs, can be be found on Google Colabs.
The binaries can be installed through conda
via a combination of conda-forge and bioconda repositories:
conda create --name unifrac -c conda-forge -c bioconda unifrac-binaries
conda activate unifrac
To install, first the binary needs to be compiled. This assumes that the HDF5 toolchain and libraries are available.
Assuming h5c++ is in your path, the following should work:
make clean && make api && make main
#optionally
make install
Note: if you are using conda we recommend installing HDF5 and related compiler using the
conda-forge channel, for example (on Linux):
conda create --strict-channel-priority -n unifrac-binaries -c conda-forge -c bioconda gxx_linux-64 gfortran_linux-64 hdf5 mkl-include lz4 zlib hdf5-static libcblas liblapacke make curl
conda activate unifrac-binaries
For GPU-enabled code, you will need the NVIDIA HPC SDK compiler. This helper script will download it, install it and setup the necessary environment:
scripts/install_hpc_sdk.sh
source setup_nv_h5.sh
If you prefer to build your HDF5 toolchain yourself, more information about how to setup the stack can be found here.
Unifrac uses OpenMP to make use of multiple CPU cores. By default, Unifrac will use all the cores that are available on the system. To restrict the number of cores used, set:
export OMP_NUM_THREADS=nthreads
On Linux platforms, Unifrac will auto-detect the CPU generation, i.e. if it supports avx or avx2 vector instructions. To force the most compatible binary variant, one can set:
export UNIFRAC_MAX_CPU=basic
To check which binary is used (Unifrac will print it to standard output at runtime), set:
export UNIFRAC_CPU_INFO=Y
On Linux platforms, Unifrac will run on a GPU, if one is found. To disable GPU offload, and thus force CPU-only execution, one can set:
export UNIFRAC_USE_GPU=N
To check which code path is used (Unifrac will print it to standard output at runtime), set:
export UNIFRAC_GPU_INFO=Y
Finally, Unifrac will only use one GPU at a time. If more than one GPU is present, one can select the one to use by setting:
export ACC_DEVICE_NUM=gpunum
Note that there is no GPU support for MacOS.
When evaluating the performance of Unifrac it is sometimes necessary to distinguish the time spent interacting with the data from the Unifrac compute proper. Additional informational messages can be enabled by setting:
export UNIFRAC_TIMING_INFO=Y
Below are a few light examples of different ways to use this library.
The methods can be used directly through the command line after install:
$ which ssu
/Users/<username>/miniconda3/envs/unifrac/bin/ssu
$ ssu --help
usage: ssu -i <biom> -o <out.dm> -m [METHOD] -t <newick> [-a alpha] [-f] [--vaw]
[--mode MODE] [--start starting-stripe] [--stop stopping-stripe] [--partial-pattern <glob>]
[--n-partials number_of_partitions] [--report-bare] [--format|-r out-mode]
[--n-substeps n] [--pcoa dims] [--diskbuf path]
-i The input BIOM table.
-t The input phylogeny in newick.
-m The method, [unweighted | weighted_normalized | weighted_unnormalized | unweighted_unnormalized | generalized |
unweighted_fp64 | weighted_normalized_fp64 | weighted_unnormalized_fp64 |
unweighted_unnormalized_fp64 | generalized_fp64 |
unweighted_fp32 | weighted_normalized_fp32 | weighted_unnormalized_fp32 |
unweighted_unnormalized_fp32 | generalized_fp32].
-o The output distance matrix.
-g [OPTIONAL] The input grouping in TSV.
-c [OPTIONAL] The columns(s) to use for grouping, multiple values comma separated.
-a [OPTIONAL] Generalized UniFrac alpha, default is 1.
-f [OPTIONAL] Bypass tips, reduces compute by about 50%.
--vaw [OPTIONAL] Variance adjusted, default is to not adjust for variance.
--mode [OPTIONAL] Mode of operation:
one-off : [DEFAULT] compute UniFrac.
partial : Compute UniFrac over a subset of stripes.
partial-report : Start and stop suggestions for partial compute.
merge-partial : Merge partial UniFrac results.
multi : compute UniFrac multiple times.
--start [OPTIONAL] If mode==partial, the starting stripe.
--stop [OPTIONAL] If mode==partial, the stopping stripe.
--partial-pattern [OPTIONAL] If mode==merge-partial, a glob pattern for partial outputs to merge.
--n-partials [OPTIONAL] If mode==partial-report, the number of partitions to compute.
--report-bare [OPTIONAL] If mode==partial-report, produce barebones output.
--n-substeps [OPTIONAL] Internally split the problem in n substeps for reduced memory footprint, default is 1.
--format|-r [OPTIONAL] Output format:
ascii : Original ASCII format. (default if mode==one-off)
hdf5_nodist : HFD5 format, no distance matrix. (default if mode==multi)
hdf5 : HFD5 format. May be fp32 or fp64, depending on method.
hdf5_fp32 : HFD5 format, using fp32 precision.
hdf5_fp64 : HFD5 format, using fp64 precision.
--subsample-depth Depth of subsampling of the input BIOM before computing unifrac (required for mode==multi, optional for one-off)
--subsample-replacement [OPTIONAL] Subsample with or without replacement (default is with)
--n-subsamples [OPTIONAL] if mode==multi, number of subsampled UniFracs to compute (default: 100)
--permanova [OPTIONAL] Number of PERMANOVA permutations to compute (default: 999 with -g, do not compute if 0)
--pcoa [OPTIONAL] Number of PCoA dimensions to compute (default: 10, do not compute if 0)
--seed [OPTIONAL] Seed to use for initializing the random gnerator
--diskbuf [OPTIONAL] Use a disk buffer to reduce memory footprint. Provide path to a fast partition (ideally NVMe).
-n [OPTIONAL] DEPRECATED, no-op.
Environment variables:
CPU parallelism is controlled by OMP_NUM_THREADS. If not defined, all detected core will be used.
GPU offload can be disabled with UNIFRAC_USE_GPU=N. By default, if a NVIDIA GPU is detected, it will be used.
A specific GPU can be selected with ACC_DEVICE_NUM. If not defined, the first GPU will be used.
Citations:
For UniFrac, please see:
Sfiligoi et al. mSystems 2022; DOI: 10.1128/msystems.00028-22
McDonald et al. Nature Methods 2018; DOI: 10.1038/s41592-018-0187-8
Lozupone and Knight Appl Environ Microbiol 2005; DOI: 10.1128/AEM.71.12.8228-8235.2005
Lozupone et al. Appl Environ Microbiol 2007; DOI: 10.1128/AEM.01996-06
Hamady et al. ISME 2010; DOI: 10.1038/ismej.2009.97
Lozupone et al. ISME 2011; DOI: 10.1038/ismej.2010.133
For Generalized UniFrac, please see:
Chen et al. Bioinformatics 2012; DOI: 10.1093/bioinformatics/bts342
For Variance Adjusted UniFrac, please see:
Chang et al. BMC Bioinformatics 2011; DOI: 10.1186/1471-2105-12-118
$ which faithpd
/Users/<username>/miniconda3/envs/unifrac/bin/faithpd
$ faithpd --help
usage: faithpd -i <biom> -t <newick> -o <out.txt>
-i The input BIOM table.
-t The input phylogeny in newick.
-o The output series.
Citations:
For Faith's PD, please see:
Faith Biological Conservation 1992; DOI: 10.1016/0006-3207(92)91201-3
In addition to the above methods to access UniFrac, it is also possible to link against the shared library. The C API is described in src/api.hpp, and examples of linking against this API can be found in examples/.
A small test .biom and .tre can be found in src/. An example with expected output is below, and should execute in 10s of milliseconds:
$ ssu -i src/test.biom -t src/test.tre -m unweighted -o test.out
$ cat test.out
Sample1 Sample2 Sample3 Sample4 Sample5 Sample6
Sample1 0 0.2 0.5714285714285714 0.6 0.5 0.2
Sample2 0.2 0 0.4285714285714285 0.6666666666666666 0.6 0.3333333333333333
Sample3 0.5714285714285714 0.4285714285714285 0 0.7142857142857143 0.8571428571428571 0.4285714285714285
Sample4 0.6 0.6666666666666666 0.7142857142857143 0 0.3333333333333333 0.4
Sample5 0.5 0.6 0.8571428571428571 0.3333333333333333 0 0.6
Sample6 0.2 0.3333333333333333 0.4285714285714285 0.4 0.6 0