Framework for construction of phylogenetic networks on High Performance Computing (HPC) environment
Introduction
Phylogeny refers to the evolutionary history and relationship between biological lineages related by common descent. Reticulate evolution refers to the origination of lineages through the complete or partial merging of ancestor lineages. Networks may be used to represent lineage independence events in non-treelike phylogenetic processes.
The methodology for reconstructing networks is still in development. Here we explore two methods for reconstructing rooted explicit phylogenetic networks, PhyloNetworks and Phylonet, which employ computationally expensive and time consuming algorithms. The construction of phylogenetic networks follows a coordinated processing flow of data sets analyzed and processed by the coordinated execution of a set of different programs, packages, libraries or pipelines, called workflow activities.
In view of the complexity in modeling network experiments, the present work introduces a workflow for phylogenetic network analyses coupled to be executed in High-Performance Computing (HPC) environments. The workflow aims to integrate well-established software, pipelines and scripts, implementing a challenging task since these tools do not consistently profit from the HPC environment, leading to an increase in the expected makespan and idle computing resources.
Requirements
- Python >= 3.8
- Biopython >= 1.75
- Pandas >= 1.3.2
- Parsl >= 1.0
- Raxml >= 8.2.12
- Astral >= 5.7.1
- SnaQ (PhyloNetworks) >= 0.13.0
- MrBayes >= 3.2.7a
- BUCKy >= 1.4.4
- Quartet MaxCut >= 2.10
- PhyloNet >= 3.8.2
- Julia >= 1.4.1
- IQTREE >= 2.0
How to use
Setting up the framework
The framework uses a file to get all the needed parameters. For default it loads the file default.ini in the config folder, but you can explicitly load other files using the argument -s name_of_the_file
, e.g. -s config/test.ini
.
- Edit parl.env with the environment variables you may need, such as modules loadeds in SLURM
- Edit work.config with the directories of your phylogeny studies (the framework receives as input a set of homologous gene alignments of species in the nexus format).
- Edit default.ini with the path for each of the needed softwares and the parameters of the execution provider.
For default, the execution logs are created in the runinfo
folder. To change it you can use the -r folder_path
parameter.
Contents of the configuration file
- General settings
[GENERAL]
ExecutionProvider = SLURM
ScriptDir = ./scripts
Environ = config/parsl.env
Workload = config/work.config
NetworkMethod = MP
TreeMethod = RAXML
BootStrap = 1000
- The framework can be executed in a HPC environment using the Slurm resource manager using the parameter
ExecutionProvider
equals toSLURM
or locally withLOCAL
. - The path of the scripts folder is assigned in
ScriptDir
. It's recommended to use the absolute path to avoid errors. - The
Environ
parameter contains the path of the file used to set environment variables. More details can be seen below. - In
Workload
is the path of the experiments that will be performed. NetworkMethod
andTreeMethod
are the default network and tree methods that will be used to perform the workloads' studies.Bootstrap
is the parameter used in all the software that use bootstrap (RAxML, IQTREE and ASTRAL)
-
Workflow execution settings
When using SLURM, these are the needed parameters:
[WORKFLOW] Monitor = False PartCore = 24 PartNode = 1 Walltime = 00:20:00
Monitor
is a parameter to use parsl's monitor module in HPC environment. It can be true or false. If you want to use it, it's necessary to set it as true and manually change the address ininfra_manager.py
- If you are using it in a HPC environment (using SLURM), the framework is going to submit in a job.
PartCore
is the number of cores of the node;PartNode
is the number of nodes of the partition; and theWalltime
parameter is the maximum amount of time the job will be able to run.
However, if the the desired execution method is the LocalProvider, i.e. the execution is being performed in your own machine, only these parameters are necessary:
[WORKFLOW] Monitor = False MaxCore = 6 CoresPerWorker = 1
-
RAxML settings
[RAXML] RaxmlExecutable = raxmlHPC-PTHREADS RaxmlThreads = 6 RaxmlEvolutionaryModel = GTRGAMMA --HKY85
-
IQTREE settings
[IQTREE] IqTreeExecutable = iqtree2 IqTreeEvolutionaryModel = TIM2+I+G IqTreeThreads = 6
-
ASTRAL settings
[ASTRAL] AstralExecDir = /opt/astral/5.7.1 AstralJar = astral.jar
-
PhyloNet settings
[PHYLONET] PhyloNetExecDir = /opt/phylonet/3.8.2/ PhyloNetJar = PhyloNet.jar PhyloNetThreads = 6 PhyloNetHMax = 3 PhyloNetRuns = 5
-
SNAQ settings
[SNAQ] SnaqThreads = 6 SnaqHMax = 3 SnaqRuns = 3
-
Mr. Bayes settings
[MRBAYES] MBExecutable = mb MBParameters = set usebeagle=no beagledevice=cpu beagleprecision=double; mcmcp ngen=100000 burninfrac=.25 samplefreq=50 printfreq=10000 diagnfreq=10000 nruns=2 nchains=2 temp=0.40 swapfreq=10
-
Bucky settings
[BUCKY] BuckyExecutable = bucky MbSumExecutable = mbsum
-
Quartet MaxCut
QUARTETMAXCUT] QmcExecDir = /opt/quartet/ QmcExecutable = find-cut-Linux-64
Workload file
For default the workload file is work.config
in the config folder. The file contains the absolute paths of the experiment's folders.
/home/rafael.terra/Biocomp/data/Denv_1
You can comment folders using the # character in the beginning of the path. e. g. #/home/rafael.terra/Biocomp/data/Denv_1
. That way the framework won't read this path.
You can also run a specific flow for a path using @TreeMethod|NetworkMethod
in the end of a path. Where TreeMethod can be RAXML, IQTREE or MRBAYES and NetworkMethod can be MPL or MP (case sensitive). The supported flows are: RAXML|MPL
, RAXML|MP
, IQTREE|MPL
, IQTREE|MP
and MRBAYES|MPL
. For example:
/home/rafael.terra/Biocomp/data/Denv_1@RAXML|MPL
Environment file
The environment file contains all the environment variables (like module files used in SLURM) used during the framework execution. Example:
module load python/3.8.2
module load raxml/8.2_openmpi-2.0_gnu
module load java/jdk-12
module load iqtree/2.1.1
module load bucky/1.4.4
module load mrbayes/3.2.7a-OpenMPI-4.0.4
source /scratch/app/modulos/julia-1.5.1.sh
Experiment folder
Each experiment folder needs to have a input folder containing a .tar.gz compressed file and a .json with the following content. The framework considers that there is only one file of each extension in the input folder.
{
"Mapping":"",
"Outgroup":""
}
Where Mapping
is a direct mapping of the taxon, when there are multiple alleles per species, in the format species1:taxon1,taxon2;species2:taxon3,taxon4
(white spaces are not supported) and Outgroup
is the taxon used to root the network. The Mapping parameter is optional (although it has to be in the json file without value), but the outgroup is obligatory. It's important to say that the flow MRBAYES|MPL doesn't support multiple alleles per species. Example:
{
"Mapping": "dengue_virus_type_2:FJ850082,FJ850088,JX669479,JX669482,JX669488,KP188569;dengue_virus_type_3:FJ850079,FJ850094,JN697379,JX669494;dengue_virus_type_1:FJ850073,FJ850084,FJ850093,JX669465,JX669466,JX669475,KP188545,KP188547;dengue_virus_type_4:JN559740,JQ513337,JQ513341,JQ513343,JQ513344,JQ513345,KP188563,KP188564;Zika_virus:MH882543",
"Outgroup": "MH882543"
}
Running the framework
-
In a local machine:
After setting up the framework, just run
python3 parsl_workflow.py
. -
In a SLURM environment:
Create an submition script that inside contains:
python3 parsl_workflow.py
.#!/bin/bash #SBATCH --time=15:00:00 #SBATCH -e slurm-%j.err #SBATCH -o slurm-%j.out module load python/3.9.6 cd /path/to/biocomp python3 parsl_workflow.py
The framework is under heavy development. If you notice any bug, please create an issue here on GitHub.
Running in a DOCKER container
The framework is also available to be used in Docker. It can be built from source or pushed from DockerHub.
Building it from the source code
Adapt the default settings file config/default.ini
according to your machine, setting the number of threads and bootstrap. After that, run docker build -t hp2net .
in the project's root folder.
Downloading it from Dockerhub
The docker image can also be downloaded from Docker hub. To do that, just run the command docker pull rafaelstjf/hp2net:main
Running
The first step to run the framework is to setup your dataset. To test if the framework is running without problems in your machine, you can use the example datasets.
Extracting the example_data.zip
file, a new folder called with_outgroup
is created. This folder contain four datasets of DENV sequences.
The next step is the creation of the settings and workload files. For the settings file, download the default.ini from this repository and change it to you liking (the path of all software are already configured to run on docker). The workload file is a text file containing the absolute path of the datasets, followed by the desired pipeline, as shown before in this document. Here for example purposes, the input.txt
file was created.
With all the files prepared, the framework can be executed from the example_data
folder as following:
docker run --rm -v $PWD:$PWD rafaelstjf/hp2net:main -s $PWD/default.ini -w $PWD/input.txt
Important: the docker doesn't save your logs, for that add the parameter: -r $PWD/name_of_your_log_folder
.
If you are running it on Santos Dumont Supercomputer, both downloading and execution of the docker container need to be performed from a submission script and executed using sg docker -c "sbatch script.sh"
. The snippet below shows an example of submission script.
#!/bin/bash
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=24
#SBATCH -p cpu_small
#SBATCH -J Hp2NET
#SBATCH --exclusive
#SBATCH --time=02:00:00
#SBATCH -e slurm-%j.err
#SBATCH -o slurm-%j.out
DIR='/scratch/pcmrnbio2/rafael.terra/WF_parsl/example_data'
docker pull rafaelstjf/hp2net:main
docker run --rm -v $DIR:$DIR rafaelstjf/hp2net:main -s ${DIR}/sdumont.ini -w ${DIR}/entrada.txt -r ${DIR}/logs
If you use it, please cite
Terra, R., Coelho, M., Cruz, L., Garcia-Zapata, M., Gadelha, L., Osthoff, C., ... & Ocana, K. (2021, July). Gerência e Análises de Workflows aplicados a Redes Filogenéticas de Genomas de Dengue no Brasil. In Anais do XV Brazilian e-Science Workshop (pp. 49-56). SBC.
Also cite all the coupled software!
Version History
main @ 20ecbe3 (earliest) Created 9th Jan 2024 at 13:04 by Rafael Terra
Merge branch 'main' of https://github.com/rafaelstjf/biocomp into main
Frozen
main
20ecbe3
Creators
Submitter
Views: 955 Downloads: 230
Created: 9th Jan 2024 at 13:04
Last updated: 18th Jan 2024 at 17:50
None