Skip to main content

unsupervised classification of polyploid subgenomes

Project description

Quick summary

polyCRACKER is an unsupervised machine learning approach to the classification and extraction of sub-genomes from a set of genomic sequences provided in FASTA format. It currently tailored to the analysis of moderate to recently derived allopolyploid species. It does not require training data or even the number of subgenomes to be known (although this helps). It does require some empirical testing, however, in order to determine the most likely number of subgenomes.

polyCRACKER can be used to:

  1. Identify subgenomes

  2. Extract subgenomes

  3. Validate subgenomes

  4. Explorative analysis of subgenomes relative to genomic features

polyCRACKER works by using repeat kmers (corresponding to viruses, transposons, and other selfish repetitive elements) as molecular barcodes for identifying species of origin. Since such repetitive sequences evolve quickly and copy themselves throughout a genome of a species, but not other closely related species), they can be used to group subsequences based on species of origin.

Given a pool of DNA sequences derived from multiple species, polyCRACKER can be used to identify and separate sequences belonging to one species versus another. In some cases, polyCRACKER performs as well at separating subgenomes of an allopolyploid as the manual extraction of subgenomes by sequence alignment, when the progenitor genome sequences are known and available.

For more information, see the polyCRACKER manuscript preprint. Please cite the following article if you use polyCRACKER in your work.

PolyCRACKER, a robust method for the unsupervised partitioning of polyploid subgenomes by signatures of repetitive DNA evolution. Sean P Gordon, Joshua J Levy, John P Vogel

(First and second authors are co-first authors.)

Getting Started With polyCRACKER

(requires mac or linux OS) (Docker works with Windows)

###Install of polyCRACKER dependencies

  • Docker (recommended)

Install of dependencies can be skipped entirely by using the provided docker image available on github.

Miniconda-based image

The core functionality of polyCRACKER can be accessed by using a miniconda-based image:

docker pull sgordon/polycracker-miniconda:1.0.2

You may need to increase settings in docker to allow additional memory and CPU usage from within docker. Please see this thread: How to assign more memory to docker

We recommend allowing at least 5 Gb of RAM and at least 4 CPU.

Running polyCRACKER on test data

    docker run -it sgordon/polycracker-miniconda:1.0.2

    source activate pCRACKER_p27

    tar -xzvf ./test_data/test_fasta_files/algae.fa.tar.gz && mv algae.fa ./test_data/test_fasta_files/

    polycracker.py test_pipeline -env pCRACKER_p27

Results stored in test_results directory.

To exit the container:

exit

Note that if you want to inspect the results outside of the docker container, you may need to mount a volume.

The details on mounting a volume in the context of docker is outside the scope of this tutorial. Nonetheless, if you have a analysis_results directory on your machine and wish to copy the results from polyCRACKER to that directory, then you may modify the above commands to:

    docker run -v "$(pwd)"/analysis_results:/analysis_results -i -t sgordon/polycracker-miniconda:1.0.2

    source activate pCRACKER_p27

    tar -xzvf ./test_data/test_fasta_files/algae.fa.tar.gz && mv algae.fa ./test_data/test_fasta_files/

    polycracker.py test_pipeline -env pCRACKER_p27

    cp -R test_results /analysis_results/

Then exit the container as above. The results should be persisted within the analysis_results/test_results subdirectory. You may also want to perform this mounting when running on your own data.

You may also build your own docker image by using the Dockerfile at the root of this repository. Details on this are described at the bottom of the page.

  • Manual conda install of the required dependencies and run within a conda environment. See below for more details on manual conda installation of dependencies.

For more testing data:

Running on your own data using docker

  1. Edit config_polyCRACKER.txt (See below)

The flow will be similar to test data, but notably you will minimally need to: 2. Move fasta file in question to ./fasta_files This can be performed by mounting a volume to the docker container as described above, provided that the input FASTA file of interest is in the directory being mounted (for example, "analysis_results"), then copying the FASTA file from the mounted directory into the ./fasta_files directory that already exists within the container.

    docker pull sgordon/polycracker-miniconda:1.0.2

    # assumes we have copied user input FASTA file into analysis_results directory that we will mount
    docker run -v "$(pwd)"/analysis_results:/analysis_results -i -t sgordon/polycracker-miniconda:1.0.2

    source activate pCRACKER_p27

    # copying user input FASTA file into fasta_files directory
    cp /analysis_results/[user FASTA file] ./fasta_files

    polycracker.py run_pipeline -env pCRACKER_p27

    cp -R analysisOutputs /analysis_results/

Results should be in ./analysisOutputs/*/* sub-directories.

  • There's a cluster results directory containing initial clusters of subsequences, and final results directory containing final clusters after signal amplification.

Sometimes signal amplification may fail due to the over agressive iterative recruitment of kmers that are either not subgenome-specific or they are specific to the opposite subgenome and incorrectly recruited. In this case, one can attain intermediate results by going into ./analysisOutputs/*/*/bootstrap_* directories and looking for extractedSubgenomes subdirectory containing fastas.

Note that extracted subgenome fasta files are still "chunked" (split according to the specified subsequence length during normalization), but contain positional information with respect to scaffold of origin.

  • Clustering plots found at in *html files in project directory.

  • Additional plots can be made using polycracker.py plotPositions -h, and there are a few other plotting utilities.

Pro tip: Can rerun/resume pipeline at various parts by setting parts of the config already run to 0 instead of 1.

Pro tip: Use command polycracker.py number_repeatmers_per_subsequence to find a histogram of the number of repeat-mers present in each chunked genome fragment. File saved as kmers_per_subsequence.png

If this histogram is too skewed to low kmer counts in each subsequence, then either:

  • Reduce kmer size
  • Increase chunk size splitFastaLineLength
  • Reduce the low_count threshold
  • Set perfectmode to 1
  • Consider adding the NonChunk = 1 to config
  • And/Or Enforce a higher MinChunkSize.

VERY IMPORTANT!

If there are is not enough repeat content included in the subsequences, they will be hard to bin. When running the pipeline, "kmers_per_subsequence.png" may be run in order to identify the frequency of kmers across subsequences and then tune relevant parameters.

Configuration for Running polyCRACKER Pipeline using Nextflow

polyCRACKER itself is a python module at the root of the repository and contains command line functions that can be individually accessed as noted above.

Because polyCRACKER consists of many individual commandline functions, we provide a pipeline written in nextflow workflow language for the convenience of users. The nextflow implementation allows a single command to then execute all the required steps in serial. This workflow is accessed for test data through:

polycracker.py test_pipeline

or as shown below for use on your own data:

polycracker.py run_pipeline.

The workflow itself is polyCRACKER_pipeline.nf, which is now within the polycracker subdirectory. Currently several resource parameters may need to be edited within the nextflow script itself, namely parameters on the number of CPU to use and memory resoures. These parameters are currently set to conservative values so that it may be run on test data on a modern laptop with 6 cores and at least 5 Gb of memory. When executing on larger datasets you will need to increase these resource settings. In particular, required memory resources scales with the size of the input FASTA sequences being analyzed. The parmeters can be change on these lines:

blastMemStr = "export _JAVA_OPTIONS='-Xms5G -Xmx" + blastMemory + "G'"

CPU requirements are specified in lines prefixed by "cpus" like this:

cpus = { writeKmer == 1 ? 6 : 1 }

polyCRACKER Configuration file setup

The provided config file within the root of the repository is 'config_polyCRACKER.txt'.

Parameters for controlling the amount of resources for individual functions and third party programs are set within the config file. Please modify as described below to suit your FASTA input as described below.

  • File paths: Copy your input FASTA file ( a single FASTA file with all your sequences ) into the fasta_files directory. You may alternatively modify fastaPath to the path for your respective FASTA input file. You may leave the other paths as provided in the example config. FASTA files must end with .fa or .fasta file extensions or they will not be recognized.
    blastPath = ./blast_files/
    kmercountPath = ./kmercount_files/
    fastaPath = ./test_data/test_fasta_files/
    bedPath = ./bed_files/
  • genome: The full filename (not the full path) of your input fasta file.

  • sge interpreter: Unless using an sge or slurm cluster, set local to 1. We do not currently document how to use sge or slurm muti-node clusters, but experienced users may try it on their own.

  • use bbtools: Please leave this set to 1.

  • Settings surrounding the number of anticipated subgenomes: Recommended practice, number of dimensions > number of subgenomes. Modify accordingly. For example, if the number of anticipated subgenomes is 2, then set n_dimensions to 3.

  • FASTA normalization Split FASTA into chunks. This determines the length of subsequences into which the input FASTA is divided into. This is necessary for normalizing subsequences for analysis. This is typically a value between 30000 and 1000000, but depends on the lengths of sequences within the input FASTA file. We recommend these as starting values:

    splitFasta = 1

    preFilter = 0

    splitFastaLineLength = 50000
  • Kmer counting settings 'kmerLength' is an important parameter that may need to be adjusted depending on the analysis. 'kmer_low_count', 'use_high_count', 'kmer_high_count' that control which kmers are used in the anaylsis. 'kmer_low_count' determines what kmers are considered 'repetitive'. 'use_high_count', 'kmer_high_count' limit the use of kmers found in the FASTA at high frequencies. We recommend these initial settings:
    writeKmer = 1

    kmerLength = 26

    kmer2Fasta = 1

    kmer_low_count = 30

    use_high_count = 0

    kmer_high_count = 2000000

    sampling_sensitivity = 1

use original genome for final analysis output Typically this will be set to zero.

  • **re-mapping kmers to the genome and transform of results into clustering matrix. specified memory usage options. 'blastMemory' is an important resource setting. Set this to the amount of RAM that you would like to use. On a laptop we recommend these settings:
    writeBlast = 1
    k_search_length = 13
    runBlastParallel = 0
    blastMemory = 5
    blast2bed = 1
    generateClusteringMatrix = 1
    lowMemory = 0
    minChunkSize = 50000
    removeNonChunk = 1
    minChunkThreshold = 0
    tfidf = 1
    perfect_mode = 0

On a larger single node cluster, you will want to increase the memory setting. 'removeNonChunk' excludes sequences shorter than the specified 'minChunkSize'.

transform and cluster the data: Two critical choices are what dimensionality reduction method to use and which cluster method to employ. 'reduction_techniques' indicates which method to use when performing dimensionality reduction on the sparse repeat-kmer by subsequence matrix. Available dimensionality reducers include:

  • 'kpca': KernelPCA,
  • 'factor': FactorAnalysis,
  • 'feature': FeatureAgglomeration,
  • 'lda': LatentDirichletAllocation, AND 'nmf': NMF.

Description of these methods is beyond the scope of this work.

'clusterMethods' specifies the cluster method that is used.
Supported methods are:

  • 'SpectralClustering': SpectralClustering,
  • 'hdbscan_genetic':GeneticHdbscan,
  • 'KMeans': MiniBatchKMeans,
  • 'GMM':GaussianMixture,
  • 'BGMM':BayesianGaussianMixture.

Example parameters are:

    transformData = 1
    reduction_techniques = tsne
    transformMetric = linear
    ClusterAll = 1
    clusterMethods = SpectralClustering
    grabAllClusters = 1
    n_neighbors = 20
    metric = cosine
    weighted_nn = 0
    mst = 0

extract the subgenomes: Heuristics on subgenome repeat-kmer counts in order to say whether a subsequence belongs to one or another subgenome. Example parameters:

    extract = 0
    diff_kmer_threshold = 20
    default_kmercount_value = 3
    diff_sample_rate = 1
    unionbed_threshold = 10,2
    bootstrap = 1

Using polyCRACKER command line functions outside the automated nextflow pipeline.

polyCRACKER is a python module with command line accessible functions. The nextflow run pipeline scripts allow users to avoid the need to run individual functions in serial for the common purpose of subgenome classification and extraction.

Nonetheless there are instances where execution of individual core and helper functions are useful.

To see the full list of command line available functions:

    docker pull sgordon/polycracker-miniconda:1.0.2

    docker run -it sgordon/polycracker-miniconda:1.0.2

    source activate pCRACKER_p27

    polycracker.py -h

The resulting list:

Usage: polycracker.py [OPTIONS] COMMAND [ARGS]...

Options:
  --version   Show the version and exit.
  -h, --help  Show this message and exit.

Commands:
  TE_Cluster_Analysis             Build clustering matrix (repeat counts vs...
  align                           Align two fasta files.
  anchor2bed                      Convert syntenic blocks of genes to bed...
  avg-distance-between-diff-kmers
                                  Report the average distance between...
  bed2scaffolds-pickle            Convert correspondence bed file,...
  bio-hyp-class                   Generate count matrix of kmers versus...
  blast2bed                       Converts the blast results from blast or...
  blast_kmers                     Maps kmers fasta file to a reference...
  build-pairwise-align-similarity-structure
                                  Take consensus repeats, generate graph...
  categorize-repeats              Take outputs from denovo repeat...
  cluster                         Perform clustering on the dimensionality...
  cluster-exchange                Prior to subgenome Extraction, can choose...
  clusterGraph                    Plots nearest neighbors graph in html...
  color-trees                     Color phylogenetic trees by progenitor of...
  compare-scimm-metabat
  compare-subclasses              In development: Grab abundance of top...
  compareSubgenomes_progenitors_v_extracted
                                  Compares the results found from the...
  convert-mat2r                   Convert any sparse matrix into a format...
  convert_subgenome_output_to_pickle
                                  Find cluster labels for all...
  correct-kappa                   Find corrected cohen's kappa score.
  count-repetitive                Infer percent of repetitive sequence in...
  dash-genetic-algorithm-hdbscan-test
                                  Save labels of all hdbscan runs, generate...
  dash-genome-quality-assessment  Input pre chromosome level scaffolded...
  diff-kmer-analysis              Runs robust differential kmer analysis...
  differential_TE_histogram       Compare the ratio of hits of certain...
  explore-kmers                   Perform dimensionality reduction on...
  extract-scaffolds-fasta         Extract scaffolds from fasta file using...
  extract-sequences               Extract sequences from fasta file and...
  final_stats                     Analyzes the accuracy and agreement...
  find-best-cluster-parameters    In development: Experimenting with...
  find-denovo-repeats             Wrapper for repeat modeler.
  find-rules                      In development: Elucidate underlying...
  find-rules2                     In development: Elucidate underlying...
  generate-genome-signatures      Wrapper for sourmash.
  generate-karyotype              Generate karyotype shinyCircos/omicCircos...
  generate-out-bed                Find cluster labels for all...
  generate-unionbed               Generate a unionbedgraph with intervals...
  generate_Kmer_Matrix            From blasted bed file, where kmers were...
  genomic-label-propagation       Extend polyCRACKER labels up and...
  get-density                     Return gene or repeat density information...
  hipmer-output-to-kcount         Converts hipmer kmer count output into a...
  kcount-hist                     Outputs a histogram plot of a given kmer...
  kcount-hist-old                 Outputs a histogram plot of a given kmer...
  kmer2Fasta                      Converts kmer count file into a fasta...
  kmerratio2scaffasta             Bin genome regions into corresponding...
  link2color                      Add color information to link file for...
  maf2bed                         Convert maf file to bed and perform stats...
  mash-test                       Sourmash integration in development.
  merge-split-kmer-clusters       In development: working on merging and...
  multicol2multifiles             Take matrix of total differential kmer...
  number_repeatmers_per_subsequence
                                  Find histogram depicting number of repeat...
  out-bed-to-circos-csv           Take progenitor mapped, species ground...
  plot-distance-matrix            Perform dimensionality reduction on...
  plot-rules                      Plots normalized frequency distribution...
  plot-rules-chromosomes          Plot distribution of rules/conservation...
  plot-unionbed                   Plot results of union bed file, the...
  plotPositions                   Another plotting function without...
  polyploid-diff-kmer-comparison  Compare highly informative differential...
  progenitorMapping               Takes reference progenitor fasta files,...
  repeat-subclass-analysis        Input repeat_fasta and find phylogenies...
  reset-cluster                   Delete cluster results, subgenome...
  reset-transform                 Remove all html files from main work...
  return-dash-data-structures     Return dash data structures needed to run...
  run-iqtree                      Perform multiple sequence alignment on...
  run-metabat
  run-tests                       Run basic polyCRACKER tests to see if...
  run_pipeline                    Run polyCRACKER pipeline locally or on...
  scaffolds2colors-specified      Attach labels to each scaffold for use of...
  send-repeats                    Use bbSketch to send fasta file...
  shiny2omic                      Convert shinyCircos csv input files to...
  species-comparison-scaffold2colors
                                  Generate color pickle file for...
  spectral-embed-plot             Spectrally embed PCA data of any origin.
  splitFasta                      Split fasta file into chunks of a...
  subgenome-extraction-via-repeats
                                  Extends results of TE_cluster_analysis by...
  subgenomeExtraction             Extract subgenomes from genome, either...
  test_pipeline
  transform_plot                  Perform dimensionality reduction on a...
  txt2fasta                       Extract subgenome fastas from reference...
  unionbed2matrix                 Convert unionbed file into a matrix of...
  update_nextflow_config          Updates nextflow configuration file for...
  writeKmerCount                  Takes list of fasta files and runs...

To obtain information on a specific function, for example, plotPositions:

    docker pull sgordon/polycracker-miniconda:1.0.2

    docker run -it sgordon/polycracker-miniconda:1.0.2

    source activate pCRACKER_p27

    polycracker.py plotPositions -h

result for the above:

    Usage: polycracker.py plotPositions [OPTIONS]

      Another plotting function without emphasis on plotting the spectral graph. Emphasis
      is on plotting positions and clusters.

    Options:
      -npy, --positions_npy PATH      If standard layout, then use these data points to
                                      begin simulation.  [default:
                                      graphInitialPositions.npy]
      -p, --labels_pickle PATH        Pickle file containing scaffolds.  [default:
                                      scaffolds.p]
      -c, --colors_pickle PATH        Pickle file containing the cluster/class each
                                      label/scaffold belongs to.  [default: colors_pickle.p]
      -o, --output_fname PATH         Desired output plot name in html.  [default:
                                      output.html]
      -npz, --graph_file PATH         Sparse nearest neighbors graph npz file. If desired,
                                      try spectralGraph.npz.
      -l, --layout [standard|spectral|random]
                                      Layout from which to plot graph.  [default: standard]
      -i, --iterations INTEGER        Number of iterations you would like to simulate to. No
                                      comma delimit, will only output a single iteration.
                                      [default: 0]
      -s, --graph_sampling_speed INTEGER
                                      When exporting the graph edges to CSV, can choose to
                                      decrease the number of edges for pdf report
                                      generation.  [default: 1]
      -ax, --axes_off                 When enabled, exports graph without the axes.
      -cmap, --new_colors TEXT        Comma delimited list of colors if you want control
                                      over coloring scheme.  [default: ]
      -h, --help                      Show this message and exit.

Re-running subgenome classification and extraction during manual optimization

Two functions of immediate interest in the context of manual optimization and trouble shooting are:

  • polycracker.py reset-cluster -h

result:

Usage: polycracker.py reset-cluster [OPTIONS]

  Delete cluster results, subgenome extraction results and corresponding html files.

and

  • polycracker.py reset-transform -h

result:

Usage: polycracker.py reset-transform [OPTIONS]

  Remove all html files from main work directory. Must do this if hoping to
  retransform sparse data.

The above functions remove some intermediate files as required to be able to successfully re-run the pipeline.

Additional Documentation

Other tips on setting up the config file and running the pipeline are found by running the jupyter notebook ./tutorials/RunningPipeline.ipynb
* Information on what each config parameter means is in this notebook. Highly recommend that you check this out.
* Other examples of old configuration files in ./tutorials/old_configs

Other downstream analyses not included here, but check out the html file described below for more commands.

Accessing additional help docs:
* You can find them here after you download the repository: ./tutorials/help_docs/index.html
* This is an html file that specifies some of the polyCRACKER commands. Still being updated.

Genome Comparison Tool and K-Mer Conservation Rules

  • A separate utility of polyCRACKER that is NOT demonstrated in the paper above is the ability to compare the distribution of k-mers between different genomes/assemblies, and create a plotly/dash app for visualization.
  • To establish a matrix of k-mers versus genomes for downstream analysis, please use bio_hyp_class command (-h)
    * Eg. nohup python polycracker.py bio_hyp_class -f ../../,_,n -dk 5 -w ../../results/ -m 150 -l 23 -min 2 -max 25 > ../../analysis.log &
  • There are then scripts that can be used for downstream analysis (clustering, etc. not detailed here). This aspect will be published in a separate manuscript, in preparation.

Detailed instructions for environment setup

Building your own docker image

(from the provided Dockerfile at the root of this repository.)

The Dockerfile tested should build and run successfully in its current state. To build the image:

    docker build . -t polycracker
    docker run -it polycracker
    source activate pCRACKER_p27

The recipe for conda install of the polyCRACKER environment

(Note that the Docker method is preferred and much easier.)

    conda create -y --name pCRACKER_p27 python=2.7

    conda activate pCRACKER_p27

    conda install -y -c bioconda nextflow scipy pybedtools pyfaidx pandas numpy bbmap

    conda install -y -c anaconda networkx click biopython matplotlib scikit-learn seaborn pyamg

    conda install -y -c plotly plotly

    conda install -y -c conda-forge deap hdbscan multicore-tsne

Test your conda environment by running polyCRACKER to classify algae genomes

  1. Clone the repository to your project directory.
    git clone git@bitbucket.org:berkeleylab/jgi-polycracker.git
  1. change cd [your root of the git project directory containing polycracker.py]
    cd [your project directory containing polycracker.py] 
  1. tar -xzvf ./test_data/test_fasta_files/algae.fa.tar.gz && mv algae.fa ./test_data/test_fasta_files/
  2. Activate your conda environment
    source activate pCRACKER_p27
  1. polycracker.py test_pipeline -env [Your polyCRACKER conda environment]. For example:
    polycracker.py test_pipeline -env pCRACKER_p27
  1. Results stored in test_results directory.

Gallery

Example Plots

Deconvolution of green alga genomes Coccomyxa subellipsoidea and Chlamydomonas reinhardtii

(Plots result of spectral embedding of dimensionality reduced repeat-kmer matrix, Genomes split into 50kb subsequences before classification.)

Assigning sequences in the large tetraploid tobacco genome into two progenitor subgenomes

Classification of sequences in the massive hexaploid bread wheat genome into three ancestral subgenomes

Schematics

Illustrative schematic of polyCRACKER clustering of sequences linked by the repeat-kmers that they contain

fig1

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

polycracker-1.0.3.tar.gz (117.8 kB view hashes)

Uploaded Source

Built Distribution

polycracker-1.0.3-py2-none-any.whl (104.4 kB view hashes)

Uploaded Python 2

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page