Structural annotation with MAKER
We will in this practical session use data from Arabidopsis thaliana, as that is one of the currently best annotated plant organisms and there is plenty of high quality data available. However, working on eukaryotes can be time consuming. Even a small plant genome like Arabidopsis thaliana (135 Mpb) would take too long to run within the time we have for this course. Thus to be sure to perform the practicals in good conditions, we will use the smallest chromosome, (chromosome 4 - 18 Mbp). Nevertheless, this chromosome could take several hours to be proceessed by MAKER on 8 cpu, so in order to work in a managable time, we decided to use only 1 Mbp of the chromosome 4. We will call it
genome all along the excercices.
An annotation project requires numerous tools and dependencies, which can take easily many days to install for a neophyte. For your convenience and in order to focus on the art of the ANNOTATION most of the tools are ready to use.
First of all
Prepare your VM
Before going into the exercises below you need to connect to your Unix virtual machine (VM) following the instruction we provided you (in a terminal: ssh -Y -p
Prepare your aplliance
Once done you will to launch a second virtual machine (Yes twice as much fun !) in a second terminal. This second VM is called
appliance MAKER and will be used to launch the MAKER annotation.
Part 1 - Check of your Assembly (on VM)
Before starting an annotation project, we need to carefully inspect the assembly to identify potential problems before running expensive computes. You can look at i) the Fragmentation (N50, N90, how many short contigs); ii) the Sanity of the fasta file (Presence of Ns, presence of ambiguous nucleotides, presence of lowercase nucleotides, single line sequences vs multiline sequences); iii) completeness using BUSCO; iv) presence of organelles; v) Others (GC content, How distant the investigated species is from the others annotated species available). The two next exercices will perform some of these checks.
1.1 Checking the gene space of your assembly
BUSCO provides measures for quantitative assessment of genome assembly, gene set, and transcriptome completeness. Genes that make up the BUSCO sets for each major lineage are selected from orthologous groups with genes present as single-copy orthologs in at least 90% of the species.
Note: In a real-world scenario, this step come first and foremost. Indeed, if the result is under your expectation you might be required to enhance your assembly before to go further.
Exercise 1 - BUSCO -:
Good news, you don’t have to do anything here. As you already learned how to use busco during the assembly part of the course. Let’s jump directly to the next step.
1.2 Various Check of your Assembly (on VM)
Exercise 2 : Launching the following script will provide you some useful information.
If you don’t see any peculiarities, you can then decide to go forward and start to perform your first wonderful annotation.
Part 2 - Understand MAKER and its control files (on appliance)
MAKER is a computational pipeline to automatically generate annotations from a range of input data - including proteins, ESTs, RNA-seq transcripts and ab-initio gene predictions. During this exercise, you will learn how to use Maker with different forms of input data, and how to judge the quality of the resulting annotations.
The Maker pipeline can work with any combination of the following data sets:
Proteins from the same species or related species
Proteins from more distantly related organisms (e.g. Uniprot/Swissprot)
Transcriptome sequences from the same species or very closely related species
Ab-initio predictions from one or more tools (directly supported are: Augustus, Snap, GeneMark, Fgenesh)
At minimum, most annotation projects will run with a protein data set, possibly complemented by some RNA-seq data. Popular examples of this are most of the traditional model systems, including human. However, a potential shortcoming of such approaches is that the comprehensiveness of the annotation depends directly on the input data. This can become a problem if our genome of interest is taxonomically distant to well-sequenced taxonomic groups so that only few protein matches can be found. Likewise, not all genes will be expressed at all times, making the generation of a comprehensive RNA-seq data set for annotation challenging.
We will therefore first run our annotation project in the traditional way, with proteins and ESTs, and then repeat the process with a well-trained ab-initio gene predictor. You can then compare the output to get an idea of how crucial the use of a gene predictor is. However, before we get our hands dirty, we need to understand Maker a little better…
Maker strings together a range of different tools into a complex pipeline (e.g. blast, exonerate, repeatmasker, augustus…), fortunately all its various dependencies have been already installed for you.
Understanding Makers control files (on appliance)
Check that everything is running smoothly by creating the MAKER config files:
Makers behaviour and information on input data are specified in one of three control files. These are:
What are these files for?
‘maker_exe.ctl’ holds information on the location of the various binaries required by Maker (including Blast, Repeatmasker etc). Normally, all information in this file will be extracted from $PATH, so if everything is set up correctly, you will never have to look into this file.
Next, ‘maker_bopts.ctl’ provides access to a number of settings that control the behaviour of evidence aligners (blast, exonerate). The default settings will usually be fine, but if you want to try to annotate species with greater taxonomic distance to well-sequenced species, it may become necessary to decrease stringency of the e.g. blast alignments.
Finally, ‘maker_opts.ctl’ holds information on the location of input files and some of the parameters controlling the decision making during the gene building.
Part 3 - Create an evidence-based annotation with MAKER (on appliance)
To save you some time, you don’t need to seek for evidence data, we provide you proteins from uniprot as well as EST from NCBI.
The first run of Maker will be done without ab-initio predictions. What are your expectations for the resulting gene build? In essence, we are attempting a purely evidence-based annotation, where the best protein- and EST-alignments are chosen to build the most likely gene models. The purpose of an evidence-based annotation is simple. Basically, you may try to annotate an organism where no usable ab-initio model is available. The evidence-based annotation can then be used to create a set of genes on which a new model could be trained on (using e.g. Snap or Augustus). Selection of genes for training can be based on the annotation edit distance (AED score), which says something about how great the distance between a gene model and the evidence alignments is. A score of 0.0 would essentially say that the final model is in perfect agreement with the evidence.
Let’s do this step-by-step:
Prepare the input data
First you need the genome sequence:
Then the gff file of the pre-computed repeats (coordinates of repeatmasked regions)
You will also need EST previously aligned in gff format:
Finally the protein fasta files:
/!\ Always check that the gff files you provides as protein or EST contains match / match_part (gff alignment type ) feature types rather than genes/transcripts (gff annotation type) otherwise MAKER will not use the contained data properly. Here all the data are fine.
You should now have 1 repeat file, 1 EST file, 2 protein files and the genome sequence in the working directory.
You nust now inform MAKER by editing the following options within the file called maker_opts.ctl (leave the two other files controlling external software behaviors untouched):
- name of the genome sequence (genome=)
- name of the ‘EST’ file in fasta format (est_gff=)
- name of the ‘Protein’ set file(s) (protein=)
- name of the repeatmasker file (rm_gff=)
You can list multiple files in one field by separating their names by a comma ‘,’.
This time, we do not specify a reference species to be used by augustus, which will disable ab-initio gene finding. Instead we set:
This will enable gene building directly from the evidence alignments.
To edit the maker_opts.ctl file you can use the nano text editor:
Before running MAKER you can check you have modified the maker_opts.ctl file properly here
/!\ Be sure to have deactivated the parameters model_org= # and repeat_protein= # to avoid the heavy work of repeatmasker.
If your maker_opts.ctl is configured correctly, you should be able to run maker:
This will start Maker on 8 cores, if everything is configured correctly. This will take a little while and process a lot of output to the screen. Luckily, much of the heavy work - such as repeat masking - are already done, so the total running time is quite manageable, even on a small number of cores.
While Maker is running (1 - On your VM):
If you want to train you eyes to recognize the different gff versions (gff annotation):
*Question: What differences could you see ? Which column change the most between the different version ?
Inspect the output (On the IFB cloud appliance)
Compile the output (On your VM)
/!\ the output have been pre-loaded on you VM machine, the folder name is called genome.maker.output.evidence.
Once Maker is finished, you need gathering all the outputs in some usable form. You have two options: copy select files by hand to wherever you want them; or you can use scripts that do the job for you.
MAKER comes with fasta_merge and gff3_merge scripts but we promote to use the script called ‘maker_merge_outputs_from_datastore’ from the GAAS git repository already include in your VM (not with the appliance). Consequently you need to copy (rsync or scp) the genome.maker.output folder on your VM first. Then, in your VM, if you are in the same folder where genome.maker.output is located you should be able to use this command:
We have specified a name for the output directory since we will be creating more than one annotation and need to be able to tell them apart.
The maker.gff is the result to keep from this analysis.
=> You could sym-link the maker.gff file to another folder called e.g. maker_results, so everything is in the same place in the end. Just make sure to call the link something other than maker.gff (e.g maker_evidence.gff), since any maker output will be called like that.
Inspect the gene models
To get some statistics of your annotation you could launch :
We could now also visualise the annotation in a genome browser (IGV).
Part 4 - Create an abinitio evidence-driven annotation with MAKER (on the appliance)
The recommended way of running Maker is in combination with one or more ab-initio profile models. Maker natively supports input from several tools, including augustus, snap and genemark. The choice of tool depends a bit on the organism that you are annotating - for example, GeneMark-ES is mostly recommended for fungi, whereas augustus and snap have a more general use.
The biggest problem with ab-initio models is the process of training them. It is usually recommended to have somewhere around 500-1000 curated gene models for this purpose. Naturally, this is a bit of a contradiction for a not-yet annotated genome.
However, if one or more good ab-initio profiles are available, they can potentially greatly enhance the quality of an annotation by filling in the blanks left by missing evidence. Interestingly, Maker even works with ab-initio profiles from somewhat distantly related species since it can create so-called hints from the evidence alignments, which the gene predictor can take into account to fine-tune the predictions.
Usually when no close ab-initio profile exists for the investigated species, we use the first round of annotation (evidence based) to create one. We first filter the best gene models from this annotation, which are used then to train the abinitio tools of our choice.
In order to compare the performance of Maker with and without ab-initio predictions in a real-world scenario, we have first run a gene build without ab-initio predictions. Now, we run a similar analysis but enable ab-initio predictions through augustus.
Prepare the input data
No need to re-compute the mapping/alignment of the different lines of evidence (Here the proteins). Indeed, this time consuming task has already been performed during the previous round of annotation (evidence based). While running from the same folder nothing has to be done, MAKER will retrieve automatically those information. In real world you could prefer to launch this second run of annotation in another folder, in such case you have to use protein2genome.gff file from the processed output folder.
This time, we do specify a reference species to be used by augustus, which will enable ab-initio gene finding. To see which species have already a hmm profile included in augustus launch:
Select the appropriate species and inform MAKER in the maker_opt.ctl file (replace my_species by your choice:
augustus_species=my_species #Augustus gene prediction species model (this is where you can call the database you trained for augustus)
If you want to keep abinitio prediction not supported by any evidences you can as well activate the keep_pred option:
/!\ keep_preds is not recommended for plant genomes.
With these settings, Maker will run augustus to predict gene loci, but inform these predictions with information from the protein and est alignments.
Before running MAKER you can check you have modified the maker_opts.ctl file properly here.
Run Maker with ab-initio predictions
With everything configured, run Maker as you did for the previous analysis:
While Maker is running (2- on your VM):
If you want to train to catches problems/differences/flavour in gff3 annotation file:
Look at the following files:
*Question: The 0_test.gff is a reference gff3 annotation file, in its most comprehensive way. What are the features type missing in the other files ? What particular attribute tags could you see in the other files ?
Compile the output (From you VM)
When Maker has finished, copy past the genome.maker.output on your VM and then compile the output (Actualluy it’s already done):
And again, it is probably best to link the resulting output (maker.gff) to a result folder (the same as defined in the previous exercise e.g. maker_results), under a descriptive name (e.g maker_abinitio.gff).
Inspect the gene models
To get some statistics of your annotation you could launch :
Question: Do you have more or less gene models compare to the evidence-based annotation ?
We could now also visualise the annotation in a genome browser.
Part 5 - Complete the annotation
You prabably have now the two maker annotation (maker_evidence.gff and maker_abinitio.gff) in one folder. If not create a folder with a link to those two annotation. Now move into this folder. We will take one of these annotation as reference, and the other as target. We will complement the reference annotation with loci that are annotated in the target but absent in the reference like that:
Get some statistics of your new annotation:
Question: How many genes did you succed to catch from your target annotation ?
Remark: Launchung BUSCO on the proteins resulting of your annotation and comparing the result against the BUSCO you should have launched on the assembly will provide you nice hints about the quality of your annotation. In case of a huge drop of the
complete BUSCOs, something went probably wrong during your annotation (paramters ? Hints provided ? hmm profile use ?).
This concludes the gene building part. We have learned how to use the Maker annotation pipeline and have created gene builds with and without ab-initio predictions. We showed you that the running of an annotation pipeline like Maker is not very hard. Actually the complicated work comes after. How to we best inspect the gene builds? Count features? Visualize it? Most importantly, what steps do we need to take to create a ‘finished’ annotation that we can use for scientific analyses ? In oder to assess the quality of an annotation the use of simple statistics like done here or BUSCO as mentioned several times are good choices. On top of that, an inspection of the annotation in genome browser is really important. We will approach this in another session (Manual curation). But assessing the annotation quality is in any case an easy task. With that being said, the structural annotation was just a step. Now you will have to make an attempt at functional inference for the predicted gene models. Thus will be approach in another session (functional curation).
Keep in mind that the exercices have been modeled to fit in the provided time. Many other important aspects of the structural annotation work haven’t been approached (e.g Gathering evidence data for annotation (Protein, EST and RNA-seq data), Assembling transcripts based on RNA-seq data, how to train ab-initio tools, Comparing and evaluating annotations, etc).