This project aims to provide a straightforward end-to-end pipeline that takes as input a HAL-format multiple whole genome alignment as well as a GFF3 file representing annotations on one high quality assembly in the HAL alignment, and produces a output GFF3 annotation on all target genomes chosen.
This pipeline is capable of running both on local cluster hardware as well as on common cloud infrastructure using the toil workflow engine. For full runs on many genomes, a decent amount of computational effort is required. Memory usage is moderate.
Above is a flowchart schematic of the functionality of the CAT pipeline.
The pipeline can be installed by a simple pip install:
pip install git+https://github.com/ComparativeGenomicsToolkit/Comparative-Annotation-Toolkit.git
However, at this time, direct pip installation will mean that the luigi.cfg, logging.cfg, and test files will be buried in your python directory. I am still trying to figure out how to approach this problem. In the meantime, you may be better off instead cloning the directory and installing from your clone:
git clone https://github.com/ComparativeGenomicsToolkit/Comparative-Annotation-Toolkit.git
pip install -e Comparative-Annotation-ToolkitIf you want to do the direct pip installation, you can grab the config files from the repository and place them in whatever directory you want to execute from, or set the LUIGI_CONFIG_PATH environmental variable to point to their location. Or have an ugly log, your choice.
Either form of pip installation will install all of the python dependencies.
Below is a full breakdown of the required dependencies if you are not using Docker. Some of these can be challenging to get to all compile and work properly. In addition to the breakdown below, you may get guidance looking at the Dockerfile or looking at this list of installation commands generated by a helpful user.
By default, you don't need to worry about installing any of these. However, there are also binary dependencies that must be compiled and installed if you are not using the Docker container we provide.
~/bin/$MACHTYPE directory on your path. All of the binaries required by CAT are available pre-compiled on the utilities page. The required tools are faToTwoBit gff3ToGenePred genePredToBed genePredToFakePsl bamToPsl transMapPslToGenePred pslPosTarget axtChain chainMergeSort pslMap pslRecalcMatch pslMapPostChain gtfToGenePred genePredToGtf pslCDnaFilter clusterGenes pslToBigPsl bedSort bedToBigBed wigToBigWig.Augustus. Make sure you are installing augustus >= 3.3.1. If you want to use Augustus CGP, install the development version from the Github repository. You need to follow the instructions to compile augustus in comparative augustus mode. This requires that you modify a few lines in the common.mk file, and also need to have sqlite3, lp-solve, bamtools, and libboost installed. If you are using ubuntu, this should work:
apt-get install libboost-all-dev libboost sqlite3 libsqlite3-0 libsqlite3-dev libgsl0-dev lp-solve liblpsolve55-dev bamtools libbamtools-dev
After you have the primary augustus binaries compiled, add the directory to your path. Note that if you move the augustus binaries from their original location, you will need to set the AUGUSTUS_CONFIG_PATH global variable to point to the species directory.
You will also need to put the contents of the scripts directory on your path. Next, you need to compile the following auxiliary programs from the folder auxprogs:
joingenes. Compiling this program will place it in the augustus binary directory.bam2hints. Compiling this program will place it in the augustus binary directory. Requires bamtools to be installed. If the bamtools headers are not at /usr/include/bamtools, you will need to modify the makefile.filterBam. Also requires the bamtools headers.bam2wig. Compiling this program will NOT place it in the augustus binary directory, you must do so yourself. This program requires you modify the makefile to explicitly point to your installation of htslib, bcftools, samtools, and tabix. Tabix is now packaged with htslib, and both are included in your kent directory at $kent/src/htslib/.homGeneMapping. This program must also have its makefile at $augustus/trunks/auxprogs/homGeneMapping/src/Makefile modified to turn on the BOOST = true and SQLITE = true flags. Then run make clean && make to recompile.$augustus/trunks/scripts directory: wig2hints.pl, exonerate2hints.pl, transMap2hints.pl, and join_mult_hints.pl.Many of the above dependencies are on conda. However, you will still need to install the following things by hand:
clusterGenes: The version on conda is too old.toil: The version on conda is too old (not python3 compatible).cat: You will need to install this repo with pip install after you enter the conda env (conda activate cattest).HAL: Not available on conda.conda create -y -n cattest -c conda-forge -c bioconda -c defaults python=3.7 pyfasta luigi seaborn pandas \
ete3 pysam numpy scipy bx-python bcbio-gff biopython parasail-python configobj sqlalchemy \
samtools bamtools augustus exonerate wiggletools bedtools \
ucsc-fatotwobit ucsc-gff3togenepred ucsc-genepredtobed ucsc-genepredtofakepsl ucsc-bamtopsl ucsc-transmappsltogenepred \
ucsc-pslpostarget ucsc-axtchain ucsc-chainmergesort ucsc-pslmap ucsc-pslrecalcmatch ucsc-pslmappostchain \
ucsc-gtftogenepred ucsc-genepredtogtf ucsc-pslcdnafilter ucsc-psltobigpsl \
ucsc-bedsort ucsc-bedtobigbedIn total, you must have all of the binaries and scripts listed below on your path. The pipeline will check for them before executing steps.
hal2fasta halStats halLiftover exonerate faToTwoBit pyfasta gff3ToGenePred genePredToBed genePredToFakePsl bamToPsl exonerate2hints.pl blat2hints.pl transMapPslToGenePred join_mult_hints.pl pslPosTarget axtChain chainMergeSort pslMap pslRecalcMatch pslMapPostChain augustus transMap2hints.pl joingenes hal2maf gtfToGenePred genePredToGtf bedtools homGeneMapping pslCDnaFilter clusterGenes pslToBigPsl bedSort bedToBigBed sambamba wig2hints.pl pal2nal.pl
This pipeline makes use of Luigi to link the various steps together. First, start the luigid daemon:
luigid --background --logdir luigi_logs
Which provides the central scheduler as well as the web UI, which can be accessed at localhost:8082.
If you don't want to use the daemon, which is highly recommended add the flag --local-scheduler to the invocation.
To run the test data, change directories to the CAT installation folder and do the following:
luigi --module cat RunCat --hal=test_data/vertebrates.hal --ref-genome=mm10 --workers=10 --config=test_data/test.config \ --work-dir test_install --out-dir test_install --local-scheduler --augustus --augustus-cgp --augustus-pb --assembly-hub > log.txt
The test should take around 30 minutes to execute. You can track progress in the log file.
CAT requires valid GFF3 as input. The script programs/validate_gff3 can test that your GFF3 is both valid and meets the requirements of CAT. These requirements include:
The following keys are reserved, and have special meaning:
reserved_keys = ['gene_biotype',
'transcript_biotype',
'gene_name',
'gene_id',
'transcript_id',
'transcript_name',
'ID',
'Name',
'Parent']The keys ID, Name and Parent are required for valid GFF3 and define the hierarchical relationship. The remaining keys, gene_biotype, transcript_biotype, gene_name, gene_id, transcript_id and transcript_name are also all required. In many cases you will not have common names, and so it is fine for transcript_name to equal transcript_id and for gene_name to equal gene_id. The biotypes can be whatever you want, but protein_coding is a special biotype that tells CAT this gene or transcript is coding. You must flag your protein coding genes and transcripts as protein_coding!
You may have any arbitrary set of keys and values in the GFF3 that are not the reserved keys. These keys and values will be propagated on to the resulting transcript in the CAT output GFF3.
If your GFF3 has duplicate transcript names, the pipeline will complain. One common cause of this is PAR locus genes. You will want to remove PAR genes -- If your GFF3 came from GENCODE, you should be able to do this: grep -v PAR_Y $gff > $gff.fixed.
As described above, the primary method to executing the pipeline is to follow the invocation luigi --module cat RunCat --hal=${halfile} --ref-genome=${ref-genome} --config=${config}. Below are the flags that can modify execution and output.
--hal: Input HAL alignment file. (REQUIRED).
--ref-genome: Reference genome sequence name. Must be present in HAL. (REQUIRED).
--config: Path to the config file with annotations and extrinsic hints. See the config section for more information. (REQUIRED).
--binary-mode: How should CAT run its binary dependencies? Valid choices are "docker" (the default, in which case you must have a Docker daemon running) or "local" (in which case ensure your PATH contains the necessary dependencies).
--out-dir: Output directory. Defaults to ./cat_output.
--work-dir: Working directory. Defaults to ./cat_work. Stores all the intermediate files as well as the toil jobStore. Can be removed after completion (but not if you want to re-do any steps).
--target-genomes: List of genomes to use. If not set, all non-reference genomes in the HAL are used. Due to how luigi handles command line tuple parameters, this flag must be formatted as if it was a tuple being passed directly to python, single quoted. So, for example, if your target genomes were Human and Mouse, then you would pass --target-genomes='("Human", "Mouse")'. As always with python tuples, if you have only one member, you must have a trailing comma.
--workers: Number of local cores to use. If running toil in singleMachine mode, care must be taken with the balance of this value and the --maxCores parameter.
--global-near-best: Adjusts the globalNearBest parameter passed to pslCDnaFilter. Defaults to 0.15. The globalNearBest algorithm determines which set of alignments are within a certain distance of the highest scoring alignment for a given source transcript. Making this value smaller will increase the number of alignments filtered out, decreasing the apparent paralogous alignment rate. Alignments which survive this filter are putatively paralogous.
--augustus: Run AugustusTM(R)?
--augustus-species: What Augustus species do we want to use? If your species is not a mammal, please choose one of the species listed here.
--augustus-utr-off: AugustusTMR will crash trying to predict UTRs if your --augustus-species lacks a trained UTR model. You can check if $augustusDir/config/species/$augustusSecies/$augustusSpecies_utr_probs.pbl exists. If it does not, set this flag.
--augustus-cgp: Run AugustusCGP?
--cgp-param: Parameters file after training CGP on the alignment. See the AugustusCGP section.
--maf-chunksize: Size to chunk HAL into. Larger values make the CGP jobs take longer, but reduce problems related to splitting in genic regions. Default is 2500000. If your HAL contains more than 10 or so genomes, reducing this value to 1000000 or so is a good idea to keep job run-times below an hour and avoid going over 8GB of RAM per job. For a 25-way alignment, I set this value to 750000.
--maf-overlap: How much overlap to use in HAL chunks. Larger values increase redundant predictions (which are merged). Default is 500000. For a 25-way alignment, I set this value to 150000.
--cgp-train-num-exons: Number of exons to require in the alignment subset used for training CGP. See the AugustusCGP section. Default is 5000.
--augustus-pb: Run AugustusPB? Will only run on genomes with IsoSeq data in the config file.
--pb-genome-chunksize: Size to chunk genome into. Default is 20000000.
--maf-overlap: How much overlap to use in genome chunks. Default is 500000.
--filter-overlapping-genes: Should genes that get flagged as overlapping be removed? After consensus finding is finished, instances of gene family collapse or paralog mis-assignment may lead to overlapping CDS intervals on different genes. This also in some instances may be a feature of the original annotation set. However, some annotation databases do not like this, so this flag will remove all such instances and resolve them down to one gene.
--overlapping-gene-distance: This controls the amount of exonic overlap two genes must have to be flagged as overlapping. Setting this value higher than 1 is recommended for smaller genomes that likely have real overlapping genes. Default is 1.
--intron-rnaseq-support: Amount of RNA-seq intron support a transcript must have to be considered. Must be a value between 0 and 100. Default is 0.
--exon-rnaseq-support: Amount of RNA-seq exon support a transcript must have to be considered. Must be a value between 0 and 100. Default is 0.
--intron-annot-support: Amount of reference intron annotation support a transcript must have to be considered. Must be a value between 0 and 100. Default is 0.
--exon-annot-support: Amount of reference exon annotation support a transcript must have to be considered. Must be a value between 0 and 100. Default is 0.
--original-intron-support: Amount of original intron support. See transcript evaluation description of original introns a transcript must have to be considered. Must be a value between 0 and 100. Default is 0.
--denovo-num-introns: For de-novo predictions, discard any transcripts with fewer than these number of introns. Important when RNA-seq data are noisy. Default is 0.
--denovo-splice-support: For de-novo predictions, discard any transcripts with less than this percent of RNA-seq intron support. Must be a value between 0 and 100. Default is 0.
--denovo-exon-support: For de-novo predictions, discard any transcripts with less than this percent of RNA-seq exon support. Must be a value between 0 and 100. Default is 0.
--denovo-ignore-novel-genes: For de-novo predictions, discard any transcripts that are predicted to be novel genes. In other words, only retain putative novel isoforms.
--denovo-novel-end-distance: For de-novo predictions, allow transcripts to be included if they provide a novel 5' or 3' end N distance away from any existing ends. Default is 0.
--denovo-allow-unsupported: For de-novo predictions, allow novel isoforms to be called if they contain splices that are not supported by the reference annotation even if they are also not supported by RNA-seq. Without this flag, novel isoforms will only be called if they have one or more splice that has RNA-seq/IsoSeq support and no reference annotation support.
--denovo-allow-bad-annot-or-tm: For de-novo predictions, allow novel isoforms to be called that were flagged as BadAnnotOrTm. These predictions overlap instances where multiple genes transMapped to the same location with significant overlap, and so may be alignment mistakes, collapsed repeats or gene family collapse.
--require-pacbio-support: If set, all isoforms in the final set must be supported by at least one IsoSeq read. This flag is likely to discard a ton of transcripts, so be careful.
--in-species-rna-support-only: If set, all of the above intron/exon support flags will look only at RNA-seq/IsoSeq data from the species in question, and not make use of homGeneMapping to check support in all species. The output plots will always report in-species support.
--rebuild-consensus: A convenience flag to allow you to adjust the flags above. When set, will force the pipeline to re-run consensus finding and will also re-build the downstream plots and assembly hub.
--assembly-hub: Build an assembly hub? Default is false. Assembly hubs allow you to view your alignments and annotation on the UCSC browser.
--hub-email: Optionally, add an email to your assembly hub. Useful if you are planning on publishing the hub.
See below for toil options shared with the hints database pipeline.
The remaining options are passed directly along to toil:
--batchSystem: Batch system to use. Defaults to singleMachine. If running in singleMachine mode, no cluster jobs will be submitted. In addition, care must be taken to balance the --maxCores field with the --workers field with the toil resources in luigi.cfg. Basically, you want to make sure that your # of toil resources multiplied by your --maxCores is fewer than the total number of system cores you want to use. However, I highly recommend using a non-local batch system. See the toil documentation for more.
--maxCores: The number of cores each toil module will use. If submitting to a batch system, this limits the number of concurrent submissions.
To use the autoscale functionality, change the resources to toil = 1 in luigi.cfg
Follow these instructions to setup credentials for AWS and the toil leader node for the cluster.
--zone: AWS region to run on.
--nodeTypes: AWS instance type for the worker nodes.
--provisioner: The provisioner of the cloud system. For AWS it's just aws.
--maxNodes: Maximum number of nodes running at once.
The config file contains the paths to two important pieces of information -- the reference GFF3 and the extrinsic hints (bams).
A major component of producing high quality comparative annotations is making use of RNA-seq and/or IsoSeq information. This information is used as hints to the augustus gene finding tool along with transMap, and is a major component of cleaning up transcript projections. This is also useful if you run the augustusCGP or augustusPB portions of the pipeline.
If the genetic distances in your alignment are high (say maybe an average identity in the 70s-80s), then you may derive great benefit from using a protein reference, if possible. This will be particularly useful for augustusCGP.
A template for the config file is below. At a minimum, your config file must have the annotation section. A example config file is provided in the test_data folder.
BAM files must be indexed!
[ANNOTATION]
Genome = /path/to/reference/gff3
[BAM]
Genome = /path/to/fofn <OR> /path/to/bam1.bam, /path/to/bam2.bam
[INTRONBAM]
Genome = /path/to/fofn/of/noisy/rnaseq
[ISO_SEQ_BAM]
Genome = /path/to/isoseq/bams
[PROTEIN_FASTA]
Genome = /path/to/protein/fastaNote that the BAM/INTRONBAM/ISO_SEQ_BAM fields can be populated either with a comma separated list of BAMs or a single file with a line pointing to each BAM (a FOFN, or file-of-file-names). The reference sequence information will be extracted from the HAL alignment.
For the PROTEIN_FASTA field, every genome you wish to have the protein fasta be aligned to must be on its own separate line. All of these can point to the same FASTA.
It is extremely important that you use high quality RNA-seq. Libraries should be poly-A selected and paired end with a minimum read length of 75bp. If any of these are not true, it is advisable to place these libraries in the INTRONBAM field. Any genome can have a mix of BAM and INTRONBAM hints.
BAM files must be genomic-coordinate sorted and indexed!
If you are using IsoSeq data, it is recommended that you doing your mapping with minimap2. These BAM files must also be genomic coordinate sorted and indexed.
The default mode of this pipeline will perform the following tasks:
These steps will run reasonably fast on one machine without any need for cluster computing. However, to construct a high quality annotation set, it is recommended that the pipeline be run with as many modes of AUGUSTUS as possible.
The primary parameterization of AUGUSTUS for comparative annotation is primarily a method to clean up transMap projections. Due to a combination of assembly error, alignment noise and real biological changes transMap projections have frame shifting indels, missing or incomplete exons, and invalid splice sites. AugustusTM is given every protein coding transMap projection one at a time with some flanking sequence and asked to construct a transcript that closely matches the intron-exon structure that transMap provides. Since AUGUSTUS enforces a standard gene model, frame shifts and invalid splices will be adjusted to a valid form. In some cases this will mangle the transcript, producing either another isoform or something that does not resemble the source transcript. AugustusTMR runs the same genomic interval and transMap derived hints through AUGUSTUS a second time, but with less strict weights on the transMap hints and with the addition of extrinsic hints from RNA-seq and/or IsoSeq. This is particularly useful in regions where an exon was dropped in the Cactus alignment.
AugustusTM and AugustusTMR can be ran by providing the --augustus flag to the pipeline. AugustusTMR will only be ran for genomes with extrinsic information in the hints database. If you are running CAT on a non-mammal, you will want to modify the --augustus-species flag to one of the species listed here. Take care to check if your species has a UTR model, and adjust the --augustus-utr-off flag accordingly.
augustusCGP is the comparative mode of AUGUSTUS recently introduced by Stefanie Nachtweide. This mode of AUGUSTUS takes as input a HAL format multiple whole genome alignment and simultaneously produces ab-initio transcript predictions in all genomes, taking into account conservation as well as any extrinsic information provided. AugustusCGP allows for the introduction of novel isoforms and loci in the final gene sets.
AugustusCGP can be ran by providing the --augustus-cgp flag to the pipeline. If no previously trained model is provided to AugustusCGP via the --cgp-param flag, then the pipeline will automatically train the model using the given alignment. To do so, random subsets of the alignment will be extracted until --cgp-train-num-exons exons are included. In practice, for vertebrate genomes, a few thousand exons corresponding to a few megabases of sequence are sufficient. If your genomes are more dense, this may vary. The trained model will be written to the AugustusCGP working directory, and can be used again on alignments with similar genomes.
AugustusPB is a parameterization of AUGUSTUS to try and predict alternative isoforms using long range data. If any IsoSeq data are provided in the config file, and the --augustus-pb flag is set, the genomes with IsoSeq data will be run through and the results incorporated in the final gene set. AugustusPB runs on single whole genomes.
While the primary mode of operation of the pipeline is to launch the RunCat module, you may want to run specific modules. Any submodule can be ran by changing the luigi invocation to specify a submodule class instead.
This module parses the GFF3 annotation input, creating a genePred format file as well as a sqlite database. In addition, sequence files for all target genomes are extracted and converted to 2bit.
This module will populate the folders --work-dir/reference and --work-dir/genome_files.
This step is the first precursor step to transMap. Pairwise genomic Kent-style chains are produced for each target genome from the designated reference. This step uses Toil and can be parallelized on a cluster.
This module will populate the folder --work-dir/chaining.
This step runs transMap. The chain files are used to project annotations present in the GFF3 from the reference genome to each target genome.
This step performs the preliminary classification of transMap transcripts. This step populates the TransMapEvaluation table in the sqlite database for each target genome with the following classifiers:
This module will populate the folder --work-dir/transMap.
This module relies on the globalNearBest algorithm in pslCDnaFilter to resolve paralogies followed by using clusterGenes to resolve gene family collapse and overlapping loci.
This process has 4 steps:
pslCDnaFilter to work properly, as --minSpan is an effective filter against retroposed
pseudogenes.pslCDnaFilter using the globalNearBest algorithm to identify the best set of alignments. Turning this value
to a smaller number increases the number of alignments filtered out, which decreases the paralogous alignment call rate.-cds flag.GeneAlternateLoci tag.CollapsedGeneIds
and CollapsedGeneNames tags.globalNearBest.After these steps, the transcripts are evaluated for split genes. This process takes the max span filtered set and looks at each transcript separately, seeing if there exists projections on either the same contig or different contigs that are disjoint in original transcript coordinates. This implies that there was a split or a rearrangement.
This module will further populate the folder --work-dir/transMap.
As discussed above, this module runs AugustusTM(R). If the pipeline is ran without a hints database, only the AugustusTM mode will be executed. This process is one of the most computationally intensive steps, and should not be ran without a cluster.
This module will populate the folder --work-dir/augustus.
Running AugustusCGP is trickier than other modes. If your genomes are not closely related to an existing training set, you may need to perform logistic regression to train AugustusCGP before execution. A default parameter set is provided. This mode is also computationally intensive, and requires a cluster.
Each output transcript are assigned a parental gene, if possible. Parental gene assignment is done by looking to see if this transcript has at least 1 exonic base overlap with any filtered TransMap as well as unfiltered transMap. If the transcript overlaps more than one gene, the Jaccard metric is used to try and resolve the ambiguity. If no gene stands out, this transcript is discarded. A sqlite table will record both the filtered and unfiltered overlaps.
Transcripts which are not assigned a parental gene will be considered novel in the consensus finding step. Most often, these are the result of gene family expansion or contraction in the reference. Looking at the raw transMap track in the final assembly hub will help resolve this.
This module will populate the folder --work-dir/augustus_cgp.
Running AugustusPB requires that IsoSeq data be provided. This mode runs on single genomes, and attempts to discover new isoforms. Transcripts predicted in this process undergo the same parental gene assignment described above.
This module will populate the folder --work-dir/augustus_pb.
homGeneMapping is a companion tool of AugustusCGP. This tool uses a HAL alignment to project RNA-seq and annotation information to target genomes. This is used to validate a splice junction in a target genome as being supported in one or more alternative genomes, as well as being supported in the reference annotation. This module populates the *_Hgm database table, where * is one of transMap, augTM, augTMR, augCGP or augPB depending on the transcripts being evaluated. This table has the following comma separated columns:
This module will populate the folder --work-dir/hgm.
The output of the homGeneMapping module has more information embedded in the output files. Each GTF format file in the above folder has a added column on the end with a string like:
"0E-6273,1E-1524,2N:M*-1,3E-742,4E-1912,5E-1208"
Which can be interpreted as 'species 0 had 6273 extrinsic hints (RNA-seq coverage), species 1 has 1524 extrinsic hints, species 2 (the reference) had both a non-coding (N) and coding (M) junction', and so on. The species numeric values are at the top of the file, and correlate to the species ID assigned internally in the hints database. These data can be useful if you want to dig in to a specific annotation.
Transcript alignment allows for AugustusTM(R) transcripts to be compared to their parental transMap. As a result, only protein coding transcripts are aligned. For each transcripts, alignment is performed by parasail two ways -- CDS alignment, and mRNA alignment. The results of these alignments are saved in the folder --work-dir/transcript_alignment. These alignments are used to create functional annotations of transcripts in the EvaluateTranscripts module.
A series of classifiers that evaluate transcript pairwise alignments for transMap and AugustusTM(R) output.
These classifiers are broken down into 2 groups, which will each end up as a table in the database:
\<alnMode>_\<txMode>_Metrics:
These classifiers are per-transcript evaluations based on both the transcript alignment and the genome context.
\<alnMode>_\<txMode>_Evaluation:
These classifiers are per-transcript evaluations based on the transcript alignment. Unlike the other two tables, this table stores the actual location of the problems (in genome coordinates) as a BED-like format. In cases where there are multiple problems, they will be additional rows.
Where txMode is one of transMap, augTM, augTMR and alnMode is one of CDS or mRNA.
The evaluation tables will be loaded as tracks in the final assembly hub.
The consensus finding process takes in transcripts from every mode and attempts to find the highest quality ortholog for a source transcript. The de-novo transcript modes are also evaluated for providing novel isoforms or novel loci. The final gene set is output with a series of features measuring how confident the prediction is.
To evaluate transMap, AugustusTM and AugustusTMR transcripts a consensus score is assigned to each. This score is the sum of the alignment goodness, intron/exon annotation support, original intron support, and intron/exon RNA-seq/IsoSeq support if extrinsic data were provided. The transcript with the highest consensus score is chosen.
If one of the de-novo augustus modes is run, then the those transcripts are evaluated for providing novel information. If a prediction did not overlap any transMap projections, then it is tagged as putative novel and incorporated into the gene set. If a prediction overlaps a transMap projection that was filtered out during paralog resolution, then it is tagged as a possible paralog as well as with the names of overlapping transcripts and incorporated into the gene set. If a prediction overlaps a transMap projection and contains a splice junction not seen in the reference annotation, then it is tagged as a novel isoform and incorporated into the gene set as a member of the gene it overlapped with.
After consensus finding is complete, a final filtering process is performed. This filtering process deduplicates the transcript set. Duplicates most often occur when the augustus execution modes create an identical transcript model from different input isoforms. In this case, the duplicates are removed and the remaining transcript tagged with the names of alternative source transcripts. Strand resolution throws out transcripts that are on opposite strands. The correct strand is chosen by looking at which contains the most high quality transcripts. Finally, the transcripts are again clustered using clusterGenes on CDS intervals to resolve the case where incorporating novel predictions lead to different gene IDs sharing CDS bases.
After consensus finding, a final output gene set is produced in both GFF3 and genePred format. The genePred annotations also have a additional .gp_info file that has the additional fields described below.
The output will appear in --output-dir/consensus.
A large range of plots are produced in --output-dir/plots. These include:
denovo.pdf. If either de-novo mode was ran, this plot will report the results. See the above description of the tags in consensus or GFF3 tags sections.completeness.pdf: The number of genes/transcripts successfully mapped over. The top of the x axis is marked with a red line representing the amount of genes/transcripts the source annotation had.consensus_extrinsic_support: A violin plot of the level of extrinsic support seen across all species for splices and exons, as found by homGeneMapping. Provides a overall plot and a per-biotype plot.consensus_anotation_support: A violin plot of the level of annotation support seen across all species for splices and exons, as found by homGeneMapping. Provides a overall plot and a per-biotype plot.coverage.pdf: A violinplot that shows the overall transcript coverage in the consensus set. Provides a overall plot and a per-biotype plot.identity.pdf: A violinplot that shows the overall transcript identity in the consensus set. Provides a overall plot and a per-biotype plot.transmap_coverage.pdf: A violinplot that shows the overall transcript coverage in the filtered transMap output. Provides a overall plot and a per-biotype plot.transmap_identity.pdf: A violinplot that shows the overall transcript identity in the filtered transMap output. Provides a overall plot and a per-biotype plot.missing_genes_transcripts.pdf: Similar to completeness.pdf, this plot reports the number of genes and transcripts in the original annotation set not found on the target genomes.paralogy.pdf: Stacked bar charts of the number of alignments a given source transcript had in each target after globalNearBest filtering. This represents a best guess of true paralogy.unfiltered_paralogy.pdf: Stacked bar charts of the number of alignments a given source transcript had in each target. This represents all possible alignments for a given source transcript.split_genes.pdf: The number of transMap genes split within and between contigs.transcript_modes.pdf: The number of modes that supported a given comparative annotation. Applies only to protein coding transcripts derived from transMap, because AugustusTMR is not ran on non-coding inputs.augustus_improvement.pdf: A scatterplot + density plot reporting the improvement of primary consensus metrics when an augustus transcript was chosen over a transMap transcript. The density plot may fail in some cases.coding_indels.pdf: The rate of insertions, deletions and indels that are a multiple of 3 are reported from the final consensus set based on the pairwise alignments. Preference is given to the CDS space alignment, if it worked.IsoSeq_isoform_validation.pdf: The number of transcripts in the consensus set whose intron structure is exactly validated by at least one IsoSeq read.gene_family_collapse.pdf: The X-axis for each plot is the number of genes in the source transcript set that were collapsed into one locus, and the Y-axis is the number of this this occurred. So, for example, if X=1 and Y=200 that means there were 200 instances of 2 genes being collapsed into 1.gene_id: Unique gene ID assigned to this gene.transcript_id: Unique transcript ID assigned to this gene.alignment_id: Original alignment ID internally used in the pipeline. Provides a link to the gene sets input to consensus finding.alternative_source_transcripts: If deduplication collapsed transcripts, report the other source_transcript IDs.exon_annotation_support: Was this exon supported by the reference annotation?exon_rna_support: Was this exon supported by the extrinsic database?frameshift: Is this transcript frameshifted relative to source_transcript?gene_biotype: The source_gene biotype. If this is a de-novo prediction, this field will say unknown_likely_coding.intron_annotation_support: Was this intron supported by the reference annotation?intron_rna_support: Was this intron supported by the extrinsic database?source_gene: The gene ID of the source gene, if this is a projection transcript.source_gene_common_name: The common name of the source gene.source_transcript: The ID of the source transcript.transcript_biotype: The biotype of the source transcript, or unknown_likely_coding for de-novo predictions.transcript_class: For projection transcripts, just says ortholog. For de-novo transcripts, will be one of poor_alignment, possible_paralog, putative_novel_isoform, or putative_novel. See the consensus finding section for descriptions.transcript_modes: Comma separated list of transcript modes. The same information as the transcript_modes.pdf plot.pacbio_isoform_supported: Was this isoform supported by at least one IsoSeq read?paralogy: Comma separated list of alignments identified as possible paralogs for this transcript.possible_split_gene_locations: If this gene was split across multiple contigs, this will have a comma separated list of alternative locations.collapsed_gene_names: If this gene was a part of a gene family collapse, this field reports the common names of genes collapsed together here.collapsed_gene_ids: Same as above, but with unique identifiers.gene_alternate_loci: If this gene was identified to have paralogous mappings that were filtered out, these intervals are where the paralogs were found.For GFF3 output, the alignment goodness is in the score field. For .gp_info, it is a column. For .gp_info, the support features are collapsed into comma separated vectors instead of being on their respective features.