miRNA/isomiR naming

[lpantano]

Hi all again,

cc: @ThomasDesvignes @mhalushka @mlhack @keilbeck @BastianFromm @ivlachos @TJU-CMC

Finally discussing naming proposals. Feel free to adapt the issue to ask more questions or clarify them.

Deadline: May, 16.

Naming goals

In order to set up a functional and coherent annotation system, I think that we should first ask ourselves: what should a microRNA annotation nomenclature accomplish? (mlhack)

Say if you agree or not:

1. Correspondence between mature and pre-microRNA names: Perfect correspondence between mature and hairpin names, i.e. if I have the name of the hairpin --> I can know the name of the mature sequences and vice-versa (this is not possible right now with miRBase because if multiple copies exist of a microRNA, no coherent nomenclature rules exist in miRBase)
2. Definition of the canonical sequence: should define and name the canonical sequence & point out if it is a constitutive canonical (same sequence in all known tissues) or regulated canonical (depends on the tissue)
3. Guide and passenger strand: If a clear distinction between guide and passenger strand can be made at a functional level, this must be reflected in the naming (with the good old ‘*’ for example)
4. Evolutionary information and family naming (I): The naming should include information about paralogues and homologues (like in miRGeneDB): to achieve this, a (evolutionary family) seed definition is needed.
5. Evolutionary information and family naming (II): If the seed region changes --> the function of the microRNA changes: should the microRNAs that are homologous but having different functions (regulate different genes because they have different seeds) receive the same name?

Other technical issues that could be improved:

Say if you agree or not:

A)Length of the hairpin sequences
Right now, in miRBase each hairpin sequence is pre-microRNA + X nt flanking sequences: X can be anything and is not defined by miRBase. This number needs to be fixed (5 nt, 10 nt, 15 nt – what ever).

miRNA

miRNA defined in miRBase: XXXXXXXXXXXXXXXX hsa-miR-X-5p

How to name if:

1. The precursor of the miRNA, and the 3p miRNA
2. Another copy of the precursor (identical)
3. Another copy of the precursor is not identical, only the mature miRs-5/3p
4. Another copy of the precursor is not identical, only miR-5p
5. Another copy of the precursor is not identical, and miR-5p is on the 3p arm this time

In case of novel ‘mirna’ detected by sequencing, how to name if:

i) NGS + structure (‘looks like’ a microRNA due to 5’ homogeneity and Drohsa/Dicer pattern),
ii) phylogenetic footprinting (miRNA is conserved),
iii) Ago loading
iv) Impact of Drosha/Dicer knockout and
v) positive functional assay.

isomiR

Say if you agree or not, or offer other option:

1. Should be annotated all isomiRs?
2. Should be use labels for each type of variants?
3. Other ideas?

Imagine the previous miRNA, how to annotate:

1. InDels XXXXXX-XXXXXXXX
2. template addition at 5’: xxXXXXXXXXXXXX
3. template addition at 3': XXXXXXXXXXXXxx
4. non-template addition at 3’: XXXXXXXXXXXXYY
5. nuncleotide substitution: XXXXXXYXXXXX

[mhalushka]

With the realization that miRNAs by the nature of their genomic loci duplications and mutations, their messy cutting by Dicer, and a host of other variable factors are terribly complex to have a specific naming convention that is 100% accurate, I will take a stab.

Naming goals
In order to set up a functional and coherent annotation system, I think that we should first ask ourselves: what should a microRNA annotation nomenclature accomplish? (mlhack)
Say if you agree or not:
1. Correspondence between mature and pre-microRNA names: Perfect correspondence between mature and hairpin names, i.e. if I have the name of the hairpin --> I can know the name of the mature sequences and vice-versa (this is not possible right now with miRBase because if multiple copies exist of a microRNA, no coherent nomenclature rules exist in miRBase)

Yes, I think a hairpin and the mature miRNAs should have perfect correspondence. Exceptions to this rule may have to exist.

2. Definition of the canonical sequence: should define and name the canonical sequence & point out if it is a constitutive canonical (same sequence in all known tissues) or regulated canonical (depends on the tissue)

At least for humans, I would not peg an miRNA to a tissue. miRNAs should be pegged to specific cell types. Tissues are too complicated. In smaller animals where individual cells are not obtained or obtainable (ex. c. elegans) one could peg a miRNA to the tissue. After that caveat, the most abundant sequence should be the canonical sequence (which is not currently accurate for much of miRBase).

3. Guide and passenger strand: If a clear distinction between guide and passenger strand can be made at a functional level, this must be reflected in the naming (with the good old ‘*’ for example)

It would be helpful for people to inherently know which of the miRNAs is the “passenger” strand. I do like the 5p/3p nomenclature, so I think this mark “*” or other, should be in addition to the 5p/3p.

4. Evolutionary information and family naming (I): The naming should include information about paralogues and homologues (like in miRGeneDB): to achieve this, a (evolutionary family) seed definition is needed.

This would be a good idea, but the devil is in the details. Is the family named just on the seed? A certain percentage of the mature miRNA? The pre-miRNA? What if the mature miRNA is conserved between species but the seed is altered? Certainly some families (let-7 etc) we could all agree on, but there will be slippery slope of evolutionary changes.

5. Evolutionary information and family naming (II): If the seed region changes --> the function of the microRNA changes: should the microRNAs that are homologous but having different functions (regulate different genes because they have different seeds) receive the same name?

I agree this will be challenging. I don’t have any suggestions here.

Other technical issues that could be improved:
Say if you agree or not:
A)Length of the hairpin sequences
Right now, in miRBase each hairpin sequence is pre-microRNA + X nt flanking sequences: X can be anything and is not defined by miRBase. This number needs to be fixed (5 nt, 10 nt, 15 nt – what ever).
For accuracy, this will require experimental data that I don’t believe exists sufficiently across species.
miRNA
miRNA defined in miRBase: XXXXXXXXXXXXXXXX hsa-miR-X-5p
_How to name if:

1. The precursor of the miRNA, and the 3p miRNA
2. Another copy of the precursor (identical)
3. Another copy of the precursor is not identical, only the mature miRs-5/3p
4. Another copy of the precursor is not identical, only miR-5p
5. Another copy of the precursor is not identical, and miR-5p is on the 3p arm this time_

This speaks to the perplexities of naming miRNAs due to their complex evolutionary history. My bias, and it is a bias, is that the mature miRNA is by far the most important part of this. So as long as the mature miRNAs are properly named, I would be happy. Beyond that, linking a miRNA to it’s genomic location(s) is useful. In general, I think the miRBase nomenclature is pretty good. I don’t have specific suggestions.

In case of novel ‘mirna’ detected by sequencing, how to name if:
i) NGS + structure (‘looks like’ a microRNA due to 5’ homogeneity and Drohsa/Dicer pattern),
ii) phylogenetic footprinting (miRNA is conserved),
iii) Ago loading
iv) Impact of Drosha/Dicer knockout and
v) positive functional assay.

I am sure Bastian will have much to say about this. I don’t know the answer here. I thought, at least for humans, that we were near the end of the novel miRNA discovery era. But the Rigoutsis lab’s PNAS paper suggested that may not be true. I think the novel miRNA detection algorithms of miRDeep2 and miRAnalyzer (the only ones I have used) over call miRNAs, thus there is a high false positive rate in any novel miRNA detections made by these methods. I do think NGS + structure is important (particularly the distribution of isomiRs relative to know miRNAs patterns). I do not think conservation is as essential in finding new miRNAs as I think we’ve figured out those major families, but there are certainly lineage specific miRNAs. Ago loading is helpful, but Ago-Seq data is dirty. Perhaps better Ago-Seq data would help. You can probably get function by overexpressing any segment of RNA, so I don’t think that’s the answer either.

_isomiR
Say if you agree or not, or offer other option:

1. Should be annotated all isomiRs?_

I don’t think it’s possible to annotate every isomiR from each miRNA as very deep RNA-seq could have 100+ isomiRs for abundant miRNAs. I think it is important to identify all 5’ changes (that affect the seed sequence), template length variants, and common nontemplated additions on the 3’ end (+A, +T, +G, +C in the +1 position). I don’t know if any other type of isomiR is particularly functional and I have some data that suggests nontemplated additions are somewhat affected by sequencing kit. So that introduces some complexity.
2. Should be use labels for each type of variants?

I think we should cover 5’ changes and identify the nontemplated additions on the 3’ end, such that +1 from the canonical sequence should be defined as +1A vs +1C vs +1T vs +1G and the genomic template nucleotide at this position should be known.
3. Other ideas?
Imagine the previous miRNA, how to annotate:
1. InDels XXXXXX-XXXXXXXX
2. template addition at 5’: xxXXXXXXXXXXXX
3. template addition at 3': XXXXXXXXXXXXxx
4. non-template addition at 3’: XXXXXXXXXXXXYY
5. nuncleotide substitution: XXXXXXYXXXXX

I think most internal changes are sequencing/PCR errors, with the exception of A to I changes. So 1 and 5 are generally an inconsequential percent of the miRNA. I’m not even sure if most programs would identify InDels. I think it is important to document 5’ additions (as they change the seed) and any 3’ additions, where again a distinction could be made between a template and nontemplate addition.

I could be easily swayed on many of these points by convincing data that suggests a better approach.
Sincerely,
Marc

[lpantano]

Hi @mhalushka,

I'll give my two cents.

Say if you agree or not:

1.I agree
2.maybe we don't need the canonical miRNA perse, maybe we need to check the most represented seed region. But I would be happy to name the most expressed sequence as canonical. I don't know if in 10 years, we'll have the same problem though. Because the most expressed sequence based on illumina kits maybe bias (due to ligation bias) and not to be the top one really.
3.I agree
4.I guess this is true if we define a canonical miRNA, otherwise isomiRs themself have this problem as well. Not sure how to solve this specific case.

Say if you agree or not:

A.I agree, longer precursor just to help bioinformatics analysis would be ideal

miRNA

How to name if:

The precursor of the miRNA, and the 3p miRNA
Another copy of the precursor (identical)
Same name, just different locations
Another copy of the precursor is not identical, only the mature miRs-5/3p
I think the idea of same miRNA with an addition letter/number for copy it should be enough
Another copy of the precursor is not identical, only miR-5p
I don't have a super good idea
Another copy of the precursor is not identical, and miR-5p is on the 3p arm this time
Maybe name of the precursor decided to be first annotated with an extra letter

In case of novel ‘mirna’ detected by sequencing, how to name if:

i) NGS + structure (‘looks like’ a microRNA due to 5’ homogeneity and Drohsa/Dicer pattern),
This seems good
ii) phylogenetic footprinting (miRNA is conserved),
I would say yes, but not necessary for all the miRNAs. I would expect some are species specific
iii) Ago loading
Yes
iv) Impact of Drosha/Dicer knockout and
Yes
v) positive functional assay.
Yes, as additional information

isomiR

Say if you agree or not, or offer other option:

Should be annotated all isomiRs?
Maybe this need a meta-analysis of data to make sure is not artifact. I'll say now, naming, like giving a specific name, may not be necessary, only for few of them
Should be use labels for each type of variants?
That, I agree, I think is ok to label the modifications even if they are not in the database, but that would give people a way to share process data to compare results. See https://github.com/miRTop/incubator/issues/10

Imagine the previous miRNA, how to annotate:

InDels XXXXXX-XXXXXXXX: - character is used in CIGAR, I think is ok to use this one for deletion and >NNN for additions
template addition at 5’: xxXXXXXXXXXXXX: here we can go with using position if we don't want to define canonical or sNN for instance.
template addition at 3': XXXXXXXXXXXXxx: same than before but maybe using other letter eNN or enn in case of longer/shorter sequences.
non-template addition at 3’: XXXXXXXXXXXXYY: The can be upper cases at the end of the label of the sequence labelYYY
nuncleotide substitution: XXXXXXYXXXXX: true that many would be error sequencing but anyhow it could be pos and nucelotide on reference.

As I said, I am more convinced that maybe we need a format file just to show what the tool has detected, and then this file can be used to even say what is a PASS/FAIL isomiRs. Again, all this more explain in issue 10.

[gurgese]

Hi all,

@lpantano > Say if you agree or not:

1. I am agree with @mhalushka and @lpantano posts.

2. A new cell type identifyer sould be included in the miRNA name if the mature miRNA sequence will be pegged to specific cell types (as suggested by @mhalushka).
   hsa-miR-23-3p would become hsa-purkinje-miR-23-3p, hsa-pyramidal-miR-23-3p or other. Owever, since this system will grow-up very rapidly and since many cells will share the same mature sequence, maybe is better to store this info in a separate file complementary to the database where the isomiRs of the most abbundant miRNA sequence are annotated as canonical mature forms for specific cell types.

3. The passenger strand is not standardadizable. With more experiments on other tissues or cell types the position of a * passenger can change during the time and the database releases. This can give instability in the notation. I would prefere to have this type of info in an external supplementary file and not embedded in the miRNA name.
   The 5p/3p nomenclature is instead very useful to be found in the miRNA name.

4. Agree to kept the evolutionary information and family naming in the naming system. Issues rised by @mhalushka sounds fine and should be considered during the design of an annotation standard. However, as suggested by @lpantano we should define a canonical miRNA sequence for the alignment algorithms and for the isomiR definition.

5. As we know, the seed region is the most important miRNA-mRNA interaction site. If this six nts sequence is embedded in the miRNA name we can have an implicit and quick label of miRNA family.

I think MirGeneDb (@BastianFromm) have a good miRNA naming system.
If this database is used in the project, I suggest to integrate it with the collection of all the discarded miRNAs that do not satisfy the MirGeneDb selection constraints in a new database of pseudo-miRs. Indeed, even if not classificable as miRNA, these short sequences have been identifyed in published works, thus related to some of the observed biological fenomena.

IsomiR questions:

1. In my opinion, isomiRs should not be annotated since many different variants can be identifyed from a single miRNA sequence.
2. The idea to use a label for each type of variant is the most pawerful. In my paper ( https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-016-0958-0 ) 12 types where used to classify how mapped reads can represent the miRNA for which they are considered expression. These classes can be re-designed to represent the most significant isomiR types.
3. The sites useful for the interaction between miRNA and target mRNA can be annotated using the label assignation to the aligned sequences.

[keilbeck]

I would advise strongly against adding the cell type to the name.
This is an annotation issue rather than a nomenclature issue

"A new cell type identifyer sould be included in the miRNA name if the mature miRNA sequence will be pegged to specific cell types (as suggested by @mhalushka).
hsa-miR-23-3p would become hsa-purkinje-miR-23-3p, hsa-pyramidal-miR-23-3p or other. Owever, since this system will grow-up very rapidly and since many cells will share the same mature sequence, maybe is better to store this info in a separate file complementary to the database where the isomiRs of the most abbundant miRNA sequence are annotated as canonical mature forms for specific cell types."

[ThomasDesvignes]

Hi all,
below I try to give my comments on the above issues, but to summarize my usual point of view/approach, I would say that what we suggest should be somewhat similar to what is done for protein-coding genes when applicable because there are many similarities between prot-coding genes and miRNA genes and instead of re-inventing the wheel using a system that has been proven to be working is best and works towards uniformity and connectivity in systems among different classes of genes.

1. Correspondence between mature and pre-microRNA names:
   I don’t think it’s possible to have perfect correspondence all the time. It should obviously be done when possible but for example when a mature miRNA can originates from two hairpins without us being able to differentiate from which one it won’t be possible or we’ll start making possibly false assumptions.
   A way is to use a common root for the name: e.g. miR-X-5p for example if the mature putatively originates from miR-X-1 or miR-X-2 and there are no other miR-X gene in the family.
   But sometimes it could also be miR-Y-1/2-5p if the 5p mature miRNA could originate from either miR-Y-1 or miR-Y-2 and miR-Y-3 exists. Calling it miR-Y-5p would mislead on the possibility that the mature originates from miR-Y-3.

2. Definition of the canonical sequence:
   The definition of a canonical sequence is very hazardous and to be honest I am not sure how to define it if not somewhat arbitrarily. In our nomenclature paper we suggested it should be of 20-22nt long and should correspond, at least in the seed sequence, to the most expressed isomiRs. But I don’t think a name could safely convey whether it is always this main isomiR or sometime changing because we’ll never have sequenced all tissue in all conditions at all ages etc…
   I don’t think we should use cell type information. This is too biased and risky. Similarly to protein coding genes and splice variants, I would keep a common root but not mention a tissue/cell type specificity.

3. Guide and passenger strand:
   I disagree on the use of the star symbol for the point mentioned above and the fact that we’ll never know for sure if this is always the case. It could be added on a gene web page or something like that but it's not a robust information to include in a name.

4. Evolutionary information and family naming (I):
   I think the actual system with numbers to define paralogs works ok-ish. It is not the best but it works. The best would I think to follow what is done for protein-coding genes (again, I always come back to what is done by other fields if it fits our situation so that we work in a uniformed system) and use suffixes letters (e.g miR-X-a and miR-X-b) to refer to paralogs originating by genome duplications and numbers (e.g. miR-X-1 and miR-X-2) when referring to paralogs that originated by cis-duplications.

5. Evolutionary information and family naming (II): If the seed region changes --> the function of the microRNA changes: should the microRNAs that are homologous but having different functions (regulate different genes because they have different seeds) receive the same name?
   I would think so because even if they potentially have different function they still originate from the same locus/loci and therefore have some similarities in their genomic expression and biogenesis. Also, even if minor or at variable amounts, there are always such isomiRs in sequencing data so it would because a nightmare to have a new name for each 5’isomiR which would actually also complicate the study of the changes in proportion of one 5’isomiR vs another one.

6. Length of the hairpin: I agree there should be a standard. I would say 10nt on each side of the mature pair. But in some cases we clearly see some moRs in sequencing data. Should those be considered? Should we extend to 30nt when moRs exist (20nt fo the moRs and +10nt)… What should we do with moRs...?

7. Novel miRNA: I would Bastian’s criteria set a good base but may be too restrictive (because I consider function above biogenesis => other discussion :) ):

   1. Sequencing data + structure
   2. Proof of 5’ cutting consistency among biological replicates
   3. Consistency of expression among biological replicates
   4. And for sure phylogenetic conservation is a very strong criteria but not always verified if lineage specific miRNA
   5. I would not use Ago loading because it’s not an experiment always performed and even when done we know that we also pull down sequences that are normally not loaded in Ago => it’s not a perfect technique, quite dirty actually
   6. Drosha/dicer knockout: similarly as Ago loading, rarely done…
   7. Functional assay would be great but no one will test all novel miRNA they find…
   8. But I totally agree with Gianvito about conserving at least in some side database (if not in the main) the pseudo-miRNA or miRNA-like short reads because even if they are not canonical they act like miRNAs and in a functional view they matter to me the same way a miRNA does.

8. I would then keep the existing system with 5p and 3p designation and not mention expression level.

9. IsomiR annotation:

   1. I think all isomiRs should be annotated if they have above background level expression.
   2. an accession number is really only what is needed (like the “MR” accession number for each isomiR in miRBAse), and then there could be a tag informing on which type of isomiR it is (like what Gianvito did with isomiR-SEA)
   3. Lorena’s suggestion works with me.

Thomas

[ivlachos]

Hi all,

I'm really troubled with the canonical sequence. It might prove quite challenging for a few (or more?) miRNAs as more data are aggregated and more samples sequenced.
I haven't decided yet on a specific system or approach to propose so this post works as food for thought.

What if the canonical form of the miRNA (which actually is a convention, since miRNAs can have different canonical forms based on cell types, conditions, states, etc or even canonical sequence definition) is not adopted as an approach? Do we have alternatives? I'm not sure that they are viable but I will state some examples as reference.

Example 1: A form similar (actually reverse) to genes and isoforms could be followed (as e.g. in Ensembl): this list of isomiRs has been found to be derived from this (or these) pre-miRNAs. Their order (as in Ensembl) might signify abundance in different tissues or any other loosely or strictly defined metric, while other metadata (as we have now for transcript quality) could be added to help filtering. For instance, evidence level, is an especially common mature (for the most common ones), etc could be added as extra information. Since hairpin(s) of origin for each mature sequence are known (and can also be included in the naming convention as suggested by colleagues above), the interested researcher has all the necessary information to make informed decisions. What is interesting with this approach is that isomiR naming can be decided without having to be compared to an archetypical sequence (e.g. could be based on the pre-miRNA of origin, genomic location and so on).

Example 2: A rigid canonical form could be uniformly defined with the same length for all matures. If this is the case, then isomiR naming could be based on comparisons with this sequence. With this approach a commonly expressed mature (which is now considered as the canonical form) could be actually represented as an isomiR (if it defers to the default). With this approach, depending on the criteria for the uniform canonical form, it would be easy(-ier) to identify a mature sequence just from the naming.

Example 3: [placeholder for another example. Please add]

These approaches have a common aim: so that canonical sequences don't change as we progress and also act as a constant reminder to the community that searching for (or using) a single mature sequence in all in silico/vitro investigations is an oversimplification of the actual biology. It is the same analogy as looking for a single gene isoform in all experiments. For some samples this might work out just fine but there are definitely caveats.

The second aim is to make those of us who don't like naming systems and conventions to change very often more comfortable. My greatest fear of basing the isomiR naming on a canonical mature sequence, is that when (and in many cases will happen based on the definition of the canonical) this sequence changes, then all the isomiRs that rely on it will have to change naming as well. This might create a lot of problems to specialists and non-specialists alike.

[mhalushka]

I wanted to address the canonical form. We have data from primary and cancer cells that looked into the most abundant miRNA form for all miRNAs. Below is some of the data. C represents the miRBase canonical sequence and C+1, C-1 etc represent the location of the most abundant miRNA by 3' length relative to the canonical sequence. Two points - 1) miRBase is not always accurate (we all know that) and 2) the most abundant length can vary by cell type (as noted by non-red colors and by sample size). For example, the most abundant "canonical" form of miR-454-3p can be C, C+1, or C+2 depending on the cell type. We think this is biological (cancer cells are more variable), but also technical. I don't think the potential large role that technical factors (different library prep methods, different miRNA alignment programs, etc) may be playing on both the canonical form and the isomiRs has been part of our conversation, but that should be addressed as well. This figure is from our manuscript on bioRxiv: http://biorxiv.org/content/early/2017/03/24/120394.

[ivlachos]

I completely agree. Based on my experience, alignment can definitely affect isomiR abundance and in some cases significantly. Mapping bias is an actual issue, with some isomiRs being unmappable or heavily handicapped vs others. This occurs in most approaches that we have checked (genome or miRNA alignment-based). Certainly as mhalushka mentioned above, library preps and other technical factors can also affect quantification results.

[mlhack]

Hi all
Sorry for my long silence but by end of this week I will have more time again to dedicate to this project.
For now some opinions on the issues that were discussed recently.
I think that a canonical sequence should be defined although for some microRNAs no clear canonical sequence might exist (canonical sequence == most expressed sequence). For those microRNAs with more than one most expressed sequence (in different tissues), the alternative one should be annotated as isomiR. However, this should be done with caution as the number of isomirs will depend also on the quality of the RNA library - and maybe also the existence of ‘alternative’ canonical sequences might be caused by RNA library quality.
This brings me to the isomiRs itself. I think that those should be only annotated if the concrete isomiR sequence is known to be regulated independently of the parental sequence (the canonical sequence is most cases). I would guess that NTA isomiRs (those with adenylation, uridylation etc events) might be good candidates as other enzymes (not related to the biogenesis) add those nucleotides. Most length variants might go with the parental sequence.
Regarding guide/passenger annotation. I would vote in favour to annotate the passenger strand for the clear cases. For this, in theory apart from expression criterion, we could also use phylogenetic footprinting as the functional guide strand is normally much more conserved than the passenger strand.
Gff3 is a good idea due to the points mentioned by Thomas and others. However, I feel that it should not be the only way to annotate but fasta files should exist as well (the gff3 would be than the annotation of those fasta files in a concrete assembly)