skDER (& CiDDER)

skDER (& CIDDER): efficient & high-resolution dereplication of microbial genomes to select representatives for comparative genomics and metagenomics.

Note: Please make sure to use version 1.0.7 or greater to avoid a bug in previous versions!

Contents

1. Installation
2. Overview
3. Algorithmic Details
4. Application examples & commands
5. Alternative approaches and comparisons
6. Test case
7. Usage
8. Citation notice
9. Representative genomes for select bacterial taxa

Installation

Bioconda

Note, (for some setups at least) it is critical to specify the conda-forge channel before the bioconda channel to properly configure priority and lead to a successful installation.

Recommended: For a significantly faster installation process, use mamba in place of conda in the below commands, by installing mamba in your base conda environment.

conda create -n skder_env -c conda-forge -c bioconda skder
conda activate skder_env

installation with mgecut (for removing MGEs prior to dereplication assessment)

To also use the option to prune out positions corresponding to MGEs using either PhiSpy or geNomad

conda create -n skder_env -c conda-forge -c bioconda skder genomad phispy "keras>=2.7,<3.0" "tensorflow>=2.7,<2.16"
conda activate skder_env

Docker

Download the bash wrapper script to simplify usage for skDER or CiDDER:

# download the skDER wrapper script and make it executable
wget https://raw.githubusercontent.com/raufs/skDER/refs/heads/main/Docker/skDER/run_skder.sh
chmod a+x run_skder.sh

# or download the CiDDER wrapper script and make it executable
wget https://raw.githubusercontent.com/raufs/skDER/refs/heads/main/Docker/CiDDER/run_cidder.sh
chmod a+x run_cidder.sh

# test it out!
./run_skder.sh -h
./run_cidder.sh -h

Optionally, if you are interested in filtering MGEs using geNomad, download the relevant databases:

wget https://zenodo.org/records/8339387/files/genomad_db_v1.5.tar.gz?download=1
mv genomad_db_v1.5* genomad_db_v1.5.tar.gz
tar -zxvf genomad_db_v1.5.tar.gz

Conda Manual

# 1. clone Git repo and change directories into it!
git clone https://github.com/raufs/skDER/
cd skDER/

# 2. create conda environment using yaml file and activate it!
conda env create -f skDER_env.yml -n skDER_env
conda activate skDER_env

# 3. complete python installation with the following command:
pip install -e .

Overview

skDER

skDER will perform dereplication of genomes using skani average nucleotide identity (ANI) and aligned fraction (AF) estimates and either a dynamic programming or greedy-based based approach. It assesses such pairwise ANI & AF estimates to determine whether two genomes are similar to each other and then chooses which genome is better suited to serve as a representative based on assembly N50 (favoring the more contiguous assembly) and connectedness (favoring genomes deemed similar to a greater number of alternate genomes).

Compared to dRep by Olm et al. 2017 and galah, skDER does not use a divide-and-conquer approach based on primary clustering with a less-accurate ANI estimator (e.g. MASH or dashing) followed by greedy clustering/dereplication based on more precise ANI estimates (for instance computed using FastANI) in a secondary round. skDER instead leverages advances in accurate yet speedy ANI calculations by skani by Shaw and Yu to simply take a "one-round" approach (albeit skani triangle itself uses a preliminary 80% ANI cutoff based on k-mer sketches, which we by default increase to 90% in skDER). skDER is also primarily designed for selecting distinct genomes for a taxonomic group for comparative genomics rather than for metagenomic application.

skDER, specifically the "dynamic programming" based approach, can still be used for metagenomic applications if users are cautious and filter out MAGs or individual contigs which have high levels of contamination, which can be assessed using CheckM or charcoal. To support this application with the realization that most MAGs likely suffer from incompleteness, we have introduced a parameter/cutoff for the max alignment fraction difference for each pair of genomes. For example, if the AF for genome 1 to genome 2 is 95% (95% of genome 1 is contained in genome 2) and the AF for genome 2 to genome 1 is 80%, then the difference is 15%. Because the default value for the difference cutoff is 10%, in that example the genome with the larger value will automatically be regarded as redundant and become disqualified as a potential representative genome.

skDER features two distinct algorithms for dereplication (details can be found below):

• dynamic approach: approximates selection of a single representative genome per transitive cluster - results in a concise listing of representative genomes - well suited for metagenomic applications.
• greedy approach: performs selection based on greedy set cover type approach - better suited to more comprehensively select representative genomes and sample more of a taxon's pangenome [current default].

Note

The skDER "greedy" algorithm is just referring to the selection algorithm - all vs all assessment is performed using skani triangle still. This is in contrast to dRep or galah where "greedy" is referring to their iterative process of selecting a representative genome followed by targeted ANI asseessment, to avoid all-vs-all comparison, of the genome against related genomes determined from primary, coarser clustering. The info needed for selecting the next representative genome is then known and the process is repeated as many times as needed. While this stratedgy can speed things up when using fastANI, with skani this does not make much of a difference (in applicaiton we found it can be more memory efficient, but also results in slower speeds than just using skani triangle directly).

CiDDER

Currently only for bacteria - because it uses pyrodigal for gene calling!

In v1.2.0, we also introduced a second program called CiDDER (CD-hit based DEReplication) - which allows for optimizing selection of a minimal number of genomes that achieve some level of saturation of the pan-genome of the full set of genomes (see below for details). Note, CD-HIT determines protein clusters, not proper ortholog groups, and as such an approximation is made of the pan-genome space being sampled by representative genomes.

Details on Dereplication Algorithms

Using Pan-Genome Saturation (CiDDER)

Starting in v1.2.0, CiDDER was introduced to allow representative genome selection based on pan-genome satauration estimates using CD-HIT. After inferring ORFs using pyrodigal, predicted protein sequences are conatenated into a giant FASTA file and clustered using CD-HIT (where parameters are possible to adjust). Each genome is thus treated as a set of distinct protein clusters it features.

Here is an overview of the algorithm:

• Download or process input genomes.
• Predict proteins using pyrodigal.
• Comprehensive clustering of all proteins using CD-HIT
• Select genome with the most number of distinct protein clusters as the initial representative.
• Iteratively add more representative genomes one at a time, selecting the next based on maximized addition of novel protein clusters to the current representative set.
• End addition of representative genomes if one of three criteria are met: (i) Next genome adds less than X number of distinct protein clusters (X is by default 0), (ii) over Y% of the total distinct protein clusters across all genomes are found in the so-far selected reprsentative genomes (Y is by default 90%), or (iii) over Z% of the total distinct multi-genome protein clusters across all genomes are found in the so-far selected representative genomes (Z is by default 100%). Thus, by default, only Y is used for representative genome selection.

Using the Dynamic Programming Dereplication Approach (skDER)

The dynamic dereplication method in skDER approximates selection of a single representative for coarser clusters of geneomes using a dynamic programming approach in which a set of genomes deemed as redundant is kept track of, avoiding the need to actually cluster genomes.

Here is an overview of the algorithm:

• Download or process input genomes.
• Compute and create a tsv linking each genome to their N50 assembly quality metric (N50[g]).
• Compute ANI and AF using skani triangle to get a tsv "edge listing" between pairs of genomes (with filters applied based on ANI and AF cutoffs).
• Run through "edge listing" tsv on first pass and compute connectivity (C[g]) for each genome - how many other genomes it is similar to at a certain threshold.
• Run through "N50" tsv and store information.
• Second pass through "edge listing" tsv and assess each pair one at a time keeping track of a singular set of genomes regarded as redudnant:
  • if (AF[g_1] - AF[g_2]) >= parameter --max_af_distance_cutoff (default of 10%), then automatically regard corresponding genome of max(AF[g_1], AF[g_2]) as redundant.
  • else calculate the following score for each genome: N50[g]*C[g] = S[g] and regard corresponding genome for min(S[g1], S[g2]) as redundant.
• Second pass through "N50" tsv file and record genome identifier if they were never deemed redudant.

Using the Greedy Dereplication Approach (skDER)

Starting from v1.0.2, skDER also allows users to request greedy clustering instead. This generally leads to a larger, more-comprehensive selection of representative genomes that covers more of the pan-genome.

Here is an overview of this alternate approach:

• Download or process input genomes.
• Compute and create a tsv linking each genome to their N50 assembly quality metric (N50[g]).
• Compute ANI and AF using skani triangle to get a tsv "edge listing" between pairs of genomes (with filters applied based on ANI and AF cutoffs).
• Run through "edge listing" tsv on first pass and compute connectivity (C[g]) for each genome - how many other genomes it is similar to at a certain threshold
  • Only consider a genome as connected to a focal genome if they share an ANI greater than the --percent_identity_cutoff (default of 99%) and the comparing genome exhibits an AF greater than the --aligned_fraction_cutoff (default of 90%) to the focal genome (is sufficiently representative of both the core and auxiliary content of the focal genome).
• Run through "N50" tsv and compute the score for each genome: N50[g]*C[g] = S[g] and write to new tsv where each line corresponds to a single genome, the second column corresponds to the S[g] computed, and the third column to connected genomes to the focal genome.
• Sort resulting tsv file based on S[g] in descending order and use a greedy approach to select representative genomes if they have not been accounted for as a connected genome from an already selected representative genome with a higher score.

Test Case

We provide a simple test case to dereplicate the six genomes available for Cutibacterium granulosum in GTDB release 214 using skDER, together with expected results.

To run this test case:

# Download test data
wget https://github.com/raufs/skDER/raw/main/test_case.tar.gz

# Download bash script to run skder
wget https://raw.githubusercontent.com/raufs/skDER/main/run_tests.sh

# Run the wrapper script to perform testing
bash ./run_tests.sh

Usage

If experiencing issues related to "Argument list too long", consider issuing ulimit -S -s unlimited prior to running skDER.

# the skder executable should be in the path after installation and can be reference as such:
skder -h

The help function should return the following:

usage: skder [-h] [-g GENOMES [GENOMES ...]] [-t TAXA_NAME] [-r GTDB_RELEASE] -o OUTPUT_DIRECTORY [-d DEREPLICATION_MODE] [-i PERCENT_IDENTITY_CUTOFF] [-tc] [-f ALIGNED_FRACTION_CUTOFF]
             [-a MAX_AF_DISTANCE_CUTOFF] [-p SKANI_TRIANGLE_PARAMETERS] [-s] [-fm] [-gd GENOMAD_DATABASE] [-n] [-l] [-u] [-c THREADS] [-v]

	Program: skder
	Author: Rauf Salamzade
	Affiliation: Kalan Lab, UW Madison, Department of Medical Microbiology and Immunology

	skDER: efficient & high-resolution dereplication of microbial genomes to select 
		   representative genomes.

	skDER will perform dereplication of genomes using skani average nucleotide identity 
	(ANI) and aligned fraction (AF) estimates and either a dynamic programming or 
	greedy-based based approach. It assesses such pairwise ANI & AF estimates to determine 
	whether two genomes are similar to each other and then chooses which genome is better 
	suited to serve as a representative based on assembly N50 (favoring the more contiguous 
	assembly) and connectedness (favoring genomes deemed similar to a greater number of 
	alternate genomes).
								  
	Note, if --filter-mge is requested, the original paths to genomes are reported but 
	the statistics reported in the clustering reports (e.g. ANI, AF) will all be based 
	on processed (MGE filtered) genomes. Importantly, computation of N50 is performed 
	before MGE filtering to not penalize genomes of high quality that simply have many 
	MGEs and enable them to still be selected as representatives.
	
    If you use skDER for your research, please kindly cite both:

    Fast and robust metagenomic sequence comparison through sparse chaining with skani.
    Nature Methods. Shaw and Yu, 2023.

    and
    
    skDER & CiDDER: microbial genome dereplication approaches for comparative genomic 
    and metagenomic applications. Salamzade, Kottapalli, and Kalan, 2024

    

options:
  -h, --help            show this help message and exit
  -g GENOMES [GENOMES ...], --genomes GENOMES [GENOMES ...]
                        Genome assembly file paths or paths to containing
                        directories. Files should be in FASTA format and can be gzipped
                        (accepted suffices are: *.fasta,
                        *.fa, *.fas, or *.fna) [Optional].
  -t TAXA_NAME, --taxa-name TAXA_NAME
                        Genus or species identifier from GTDB for which to
                        download genomes for and include in
                        dereplication analysis [Optional].
  -r GTDB_RELEASE, --gtdb-release GTDB_RELEASE
                        Which GTDB release to use if -t argument issued [Default is R220].
  -o OUTPUT_DIRECTORY, --output-directory OUTPUT_DIRECTORY
                        Output directory.
  -d DEREPLICATION_MODE, --dereplication-mode DEREPLICATION_MODE
                        Whether to use a "dynamic" (more concise) or "greedy" (more
                        comprehensive) approach to selecting representative genomes.
                        [Default is "greedy"]
  -i PERCENT_IDENTITY_CUTOFF, --percent-identity-cutoff PERCENT_IDENTITY_CUTOFF
                        ANI cutoff for dereplication [Default is 99.0].
  -tc, --test-cutoffs   Assess clustering using various pre-selected cutoffs.
  -f ALIGNED_FRACTION_CUTOFF, --aligned-fraction-cutoff ALIGNED_FRACTION_CUTOFF
                        Aligned cutoff threshold for dereplication - only needed by
                        one genome [Default is 90.0].
  -a MAX_AF_DISTANCE_CUTOFF, --max-af-distance-cutoff MAX_AF_DISTANCE_CUTOFF
                        Maximum difference for aligned fraction between a pair to
                        automatically disqualify the genome with a higher
                        AF from being a representative.
  -p SKANI_TRIANGLE_PARAMETERS, --skani-triangle-parameters SKANI_TRIANGLE_PARAMETERS
                        Options for skani triangle. Note ANI and AF cutoffs
                        are specified separately and the -E parameter is always
                        requested. [Default is "-s 90.0"].
  -s, --sanity-check    Confirm each FASTA file provided or downloaded is actually
                        a FASTA file. Makes it slower, but generally
                        good practice.
  -fm, --filter-mge     Filter predicted MGE coordinates along genomes before
                        dereplication assessment but after N50
                        computation.
  -gd GENOMAD_DATABASE, --genomad-database GENOMAD_DATABASE
                        If filter-mge is specified, it will by default use PhiSpy;
                        however, if a database directory for
                        geNomad is provided - it will use that instead
                        to predict MGEs.
  -n, --determine-clusters
                        Perform secondary clustering to assign non-representative
                        genomes to their closest representative genomes.
  -l, --symlink         Symlink representative genomes in results subdirectory
                        instead of performing a copy of the files.
  -u, --ncbi-nlm-url    Try using the NCBI ftp address with '.nlm' for
                        ncbi-genome-download if there are issues.
  -c THREADS, --threads THREADS
                        Number of threads/processes to use [Default is 1].
  -v, --version         Report version of skDER.

Usage for CiDDER

# the cidder executable should be in the path after installation and can be reference as such:
cidder -h

The help function should return the following:

usage: cidder [-h] [-g GENOMES [GENOMES ...]] [-t TAXA_NAME] [-r GTDB_RELEASE] -o OUTPUT_DIRECTORY [-p CD_HIT_PARAMS] [-mg] [-e] [-a NEW_PROTEINS_NEEDED] [-ts TOTAL_SATURATION]
              [-mgs MULTI_GENOME_SATURATION] [-s] [-fm] [-gd GENOMAD_DATABASE] [-n] [-ns] [-l] [-u] [-c THREADS] [-m MEMORY] [-v]

	Program: cidder
	Author: Rauf Salamzade
	Affiliation: Kalan Lab, UW Madison, Department of Medical Microbiology and Immunology

	CiDDER: Performs genome dereplication based on CD-HIT clustering of proteins to 
			select a representative set of genomes which adequately samples the 
			pangenome space. Because gene prediction is performed using pyrodigal, 
			geneder only works for bacterial genomes at the moment. 
								  
	The general algorithm is to first select the genome with the most number of distinct 
	open-reading-frames (ORFS; predicted genes) and then iteratively add genomes based on 
	which maximizes the number of new ORFs. This iterative addition of selected genomes
	is performed until: (i) the next genome to add does not have a minimum of X new disintct 
	ORFs to add to the set of ORFs belonging, (ii) some percentage Y of the total distinct 
	ORFs are found to have been sampled, or (iii) some percentage Z of the total multi-genome
	distinct ORFs are found to have been sampled. The "added-on" genomes from the iterative 
	procedure are listed as representative genomes.  
								  
	For information on how to alter CD-HIT parameters, please see: 
	https://github.com/weizhongli/cdhit/blob/master/doc/cdhit-user-guide.wiki#cd-hit

	Note, if --filter-mge is requested, the statistics reported in clustering reports (number 
	of proteins overlapping, ANI) in the clustering reports will all be based on processed 
	(MGE filtered) genomes. However, the final representative genomes in the 
	Dereplicated_Representative_Genomes/ folder will be the original unprocesed genomes.

    If you use CiDDER for your research, please kindly cite both:

    CD-HIT: accelerated for clustering the next-generation sequencing data. Bioinformatics.
    Fu et al. 2012

    and

    skDER & CiDDER: microbial genome dereplication approaches for comparative genomic and 
    metagenomic applications. Salamzade, Kottapalli, and Kalan, 2024.
    

options:
  -h, --help            show this help message and exit
  -g GENOMES [GENOMES ...], --genomes GENOMES [GENOMES ...]
                        Genome assembly file paths or paths to containing
                        directories. Files should be in FASTA format and can be gzipped
                        (accepted suffices are: *.fasta,
                        *.fa, *.fas, or *.fna) [Optional].
  -t TAXA_NAME, --taxa-name TAXA_NAME
                        Genus or species identifier from GTDB for which to
                        download genomes for and include in
                        dereplication analysis [Optional].
  -r GTDB_RELEASE, --gtdb-release GTDB_RELEASE
                        Which GTDB release to use if -t argument issued [Default is R220].
  -o OUTPUT_DIRECTORY, --output-directory OUTPUT_DIRECTORY
                        Output directory.
  -p CD_HIT_PARAMS, --cd-hit-params CD_HIT_PARAMS
                        CD-HIT parameters to use for clustering proteins - select carefully
                        (don't set threads or memory - those are done by default in cidder) and
                        surround by quotes [Default is: "-n 5 -c 0.95 -aL 0.75 -aS 0.90"]
  -mg, --metagenome-mode
                        Run pyrodigal using metagenome mode.
  -e, --include-edge-orfs
                        Include proteins from ORFs that hang off the edge of a contig/scaffold.
  -a NEW_PROTEINS_NEEDED, --new-proteins-needed NEW_PROTEINS_NEEDED
                        The number of new protein clusters needed to add [Default is 0].
  -ts TOTAL_SATURATION, --total-saturation TOTAL_SATURATION
                        The percentage of total proteins clusters needed to stop representative
                        genome selection [Default is 90.0].
  -mgs MULTI_GENOME_SATURATION, --multi-genome-saturation MULTI_GENOME_SATURATION
                        The percentage of total multi-genome protein clusters needed to stop
                        representative genome selection [Default is 100.0].
  -s, --sanity-check    Confirm each FASTA file provided or downloaded is actually
                        a FASTA file. Makes it slower, but generally
                        good practice.
  -fm, --filter-mge     Filter predicted MGE coordinates along genomes before
                        dereplication assessment but after N50
                        computation.
  -gd GENOMAD_DATABASE, --genomad-database GENOMAD_DATABASE
                        If filter-mge is specified, it will by default use PhiSpy;
                        however, if a database directory for
                        geNomad is provided - it will use that instead
                        to predict MGEs.
  -n, --determine-clusters
                        Perform secondary clustering to assign non-representative
                        genomes to their closest representative genomes based on shared
                        protein clusters.
  -ns, --determine-clusters-skani
                        Perform secondary clustering to assign non-representative
                        genomes to their closest representative genomes based on skani-computed
                        ANI.
  -l, --symlink         Symlink representative genomes in results subdirectory
                        instead of performing a copy of the files.
  -u, --ncbi-nlm-url    Try using the NCBI ftp address with '.nlm' for
                        ncbi-genome-download if there are issues.
  -c THREADS, --threads THREADS
                        Number of threads/processes to use [Default is 1].
  -m MEMORY, --memory MEMORY
                        The memory limit for CD-HIT in Gigabytes [Default is 0 = unlimited].
  -v, --version         Report version of CiDDER.

Citation notice

skDER relies heavily on advances made by skani for fast ANI estimation while retaining accuracy - thus if you use skDER for your research please cite skani:

Fast and robust metagenomic sequence comparison through sparse chaining with skani

as well as the skDER manuscript:

skDER: microbial genome dereplication approaches for comparative and metagenomic applications

If you use the option to downlod genomes for a taxonomy based on GTDB classifications, please also cite:

GTDB: an ongoing census of bacterial and archaeal diversity through a phylogenetically consistent, rank normalized and complete genome-based taxonomy

If you use CiDDER, please also consider citing pyrodigal (for gene-calling) and CD-HIT (for protein clustering):

CD-HIT: accelerated for clustering the next-generation sequencing data

Pyrodigal: Python bindings and interface to Prodigal, an efficient method for gene prediction in prokaryotes

If you use mgecut (for removal of predicted MGEs) then please cite either PhiSpy (default) or geNomad for their annotation:

Identification of mobile genetic elements with geNomad

PhiSpy: a novel algorithm for finding prophages in bacterial genomes that combines similarity- and composition-based strategies

Acknowledgments

We thank Titus Brown, Tessa Pierce-Ward, and Karthik Anantharaman for helpful discussions on the development of skDER/CiDDER - in particular the idea to directly asses the pan-genome space sampled by representative genomes. We also thank users on GitHub issues for suggesting ideas for new features.

LICENSE

BSD 3-Clause License

Copyright (c) 2023, Rauf Salamzade

Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, this
   list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice,
   this list of conditions and the following disclaimer in the documentation
   and/or other materials provided with the distribution.

3. Neither the name of the copyright holder nor the names of its
   contributors may be used to endorse or promote products derived from
   this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.