I am pretty sure I gave a lengthy response to this question this morning but may not have submitted it?...

Let's see. First, this is an interesting question and approach that eventually may be the perfect storm for systems biology (MicroRNA-regulated networks: the perfect storm for classical molecular biology, the ideal scenario for systems biology).

Altogether, I would like to completely disagree with my colleagues here. And the simple reason is that neither predictions nor published compilations of "validated" targets are good enough to start a project like yours. The weakest link in your analysis and classification will always be the starting material (kind of arbitrary collections of targets AND kind of arbitrary collections of validated targets).

The reason I say this is that many many publications violate dogma number one in microRNA biology: each microRNA has probably many target genes in each cell and overall can regulate many genes in the entire organisms at different developmental stages. In contrast. most publications focus on one (arbitrary, based on personal bias, fitting the story) microRNA target gene. That is almost unscientific. Therefore, I cannot see where the basis for a new classification system could be derived from. Either you generate such data yourselves or wait a few years. SOmebody has to start it.

Then there is this widespread strong believe that combining prediction data will give you better prediction data. There was once a nice publication that basically showed that this is not the case (Lost in translation: an assessment and perspective for computational microRNA target identification) .But things change. Perhaps by now there are better data out there, but I doubt that you can improve the prediction of targets by using the overlap of predictions. Seems logic, but is probably very silly.

A good example is the conundrum of lys6 and miR-23. Thousands of predictions, hardly any validated targets (Weak seed-pairing stability and high target-site abundance decrease the proficiency of lsy-6 and other microRNAs). Then there is the problem with focusing on "conserved target sites" versus all potential sites and so on and on.

Bottom line, in my opinion, you can calculate the hack out of these data collections but they are so crude that your "results" will be even cruder, probably meaningless.

This is kind of the dark and sad side of microRNA biology. It would be wonderful to do and define these multiple microRNAs targeting one mRNA in the same cell at the same time but we are still far away from good data in this regard. It would be nice to define functional redundancy and cooperation between microRNAs and re-classify them, but that requires better data on the real world microRNA targets, more hard work, less playing with the algorithms.

Try miRDIP, it uses 12 differente databases to precidt it. miRCAT or miRbase can also help.

Hi

I agree with David. It would advice you to use multiple predictors and check what are the applications that predict the same target.

Look at my humble site and see if you can get something interesting.

https://sites.google.com/site/mirnatools2/multiple-predictors

Best. Paco

Take a look at these two sites. Co-miR incorporates multiple pieces of information to look at combinitoral miRNA regulation. The second site is a suite of tools that can help you answer a variety of questions, including pathway associations.

http://www.benoslab.pitt.edu/comir/index2.php

http://diana.imis.athena-innovation.gr/DianaTools/index.php?r=site/index

I am pretty sure I gave a lengthy response to this question this morning but may not have submitted it?...

Let's see. First, this is an interesting question and approach that eventually may be the perfect storm for systems biology (MicroRNA-regulated networks: the perfect storm for classical molecular biology, the ideal scenario for systems biology).

Altogether, I would like to completely disagree with my colleagues here. And the simple reason is that neither predictions nor published compilations of "validated" targets are good enough to start a project like yours. The weakest link in your analysis and classification will always be the starting material (kind of arbitrary collections of targets AND kind of arbitrary collections of validated targets).

The reason I say this is that many many publications violate dogma number one in microRNA biology: each microRNA has probably many target genes in each cell and overall can regulate many genes in the entire organisms at different developmental stages. In contrast. most publications focus on one (arbitrary, based on personal bias, fitting the story) microRNA target gene. That is almost unscientific. Therefore, I cannot see where the basis for a new classification system could be derived from. Either you generate such data yourselves or wait a few years. SOmebody has to start it.

Then there is this widespread strong believe that combining prediction data will give you better prediction data. There was once a nice publication that basically showed that this is not the case (Lost in translation: an assessment and perspective for computational microRNA target identification) .But things change. Perhaps by now there are better data out there, but I doubt that you can improve the prediction of targets by using the overlap of predictions. Seems logic, but is probably very silly.

A good example is the conundrum of lys6 and miR-23. Thousands of predictions, hardly any validated targets (Weak seed-pairing stability and high target-site abundance decrease the proficiency of lsy-6 and other microRNAs). Then there is the problem with focusing on "conserved target sites" versus all potential sites and so on and on.

Bottom line, in my opinion, you can calculate the hack out of these data collections but they are so crude that your "results" will be even cruder, probably meaningless.

This is kind of the dark and sad side of microRNA biology. It would be wonderful to do and define these multiple microRNAs targeting one mRNA in the same cell at the same time but we are still far away from good data in this regard. It would be nice to define functional redundancy and cooperation between microRNAs and re-classify them, but that requires better data on the real world microRNA targets, more hard work, less playing with the algorithms.

Thomas !... My respect to you !... Superb comment !...

Thanks Francisco, after a few years in the microRNA field and seeing so many bioinformatics approaches/publications with both legs in the clouds (meaning "grounded" on hypothetical targets), I appreciate each and every attempt to define the function of an individual microRNA or -if you are a little bit crazy- a group of microRNAs with defining as many good target genes as possible. Very unlikely that there will bit thousands of target genes in one individual cell or even one cell type. So there must be a method to reduce the "predictome" to a real "targetome". And that seems to be hard work, hard work that may not always be appreciated since it is costly, time-consuming and involves a lot of "boring" genes, nobody has ever heard of.

I do not say that bioinformatics is not helpful. I say one has to be careful to put all his/her nuts into that basket since it is a virtual basket. Cells and biology does not care what we calculate, what we suggest as potential targets, does not care what we declare as "validated" targets, does not care about the impact factor of the associated publications. But these cells and this biology belongs to bodies that can get sick, age and need doctors. And these doctors base everything (at least they should) on the work of scientists like us. The closer we come to model cells and biology with hard work, hypotheses, data, the more useful the models will be. Useless models will not result in treatments. Not sure that is true. Even useless models must have a role in scientific progress.

Great Dr Thomas. thought provoking comments. You are right that the available data is mostly crude which will eventually lead to crude results. But presently the datasets of miRNA genes, targets, putative functions are so huge that if we sit, explore and want to come out with some inference, it becomes very difficult. Still, research efforts on discovery of new miRNA genes are increasing exponentially and so are the entries in miRBASE and similar databases. you rightly commented that shift now should be from predictome to targetome.

Hi everyone

Just to introduce some more entropy to the discussion, I would like to ask everybody involved (specially Thomas) about what are the best strategies to understand the entangled miRNA-based regulatory networks. Many people claim that Ago-IP followed by NGS will be the best choice, but of course these results are strictly dependent on the cell type and also on the physiological state of the cells. Also, I agree with Thomas about the idea that we must thing big when we are dealing with miRNAs and we always must look to the overall regulatory networks and not to a particular gene or transcript. But, I feel that we are far to understand how it works.

A few months ago I listen somebody important in the field (not remember who), claiming that the presence of a cognate target for a miRNA must have an influence in the regulatory miRNA levels. This argument was used to defend how the cells select one or other strand from a pre-miRNA to produce the mature molecule (5p or 3p).

So, how we deal with that?. I modestly thing that bioinformatics can help a little bit, at least as a possible orientation strategy. I used it to predict things and then go to the lab, and I can say that at least in 70% of the cases the predictions were right. Of course that the remaining 30% could be very relevant for our particular process, but this is a risk that we must take.

I think that the problem is not in the algorithms, but in the biology of the system. As humans we only can understand if we simplify things. In the genomic output there are so many regulatory layers, that it is almost impossible to take into consideration all the possible factors governing miRNA action.

Do not get me wrong, I always defended a mixed approach combining bioinformatics, genetics, biochemistry and cell biology. As scientists we must deal with all these disciplines and combine them. They are not mutually exclusive.

Thanks to all of you for this fruitful discussion. Best

Paco

Speaking of combining different approaches, I am curious what is other's opinion on tools, such as magia2 (http://gencomp.bio.unipd.it/magia2/start/) or BNSA (http://nugget.unisa.edu.au/bnsa.html). (referred to as MRMs/FMRMs and briefly summarized in http://www.ncbi.nlm.nih.gov/pubmed/23175680).

They aim to provide insight into miRNA functions by combining both target predictions and mRNA/microRNA high-throughput expression data derived from the matched samples. Yes, it's still exploratory and mostly computational, but won't you regard it as a good starting point for choosing the key regulatory nodes and performing downstream analyses?

@Marcin

I have used Magia2 and other approaches like MirTalos and MirconnX, which combine gene expression data with miRNA expression and target prediction. You have more in my site:

https://sites.google.com/site/mirnatools2/integrative-analysis

In my personal opinion these are good tools, however they always depend on the target prediction algorithms that they are using, so be prepared to get different results with different tools.

About your approach, I think that is correct. It would be even better if you can combine also data from Ago2-IP analysis. This will be fenomenal !....

However, do not forget that in some cases the effect of the miRNAs over gene expression is weak. The main advantage of these tools, at least Magia2 and MirconnX is that they usually take into account the combined effect of several miRNAs over a cognate mRNA. So, you can pinpoint good relationships that are not possible to be detected with conventional correlation methods.

Goooood luck. Cheers

Paco

Hello Francisco, thanks for referring to more of these tools. I strongly agree with the additional tips you provide, the influence of different algorithms and cutoffs is still a notable downside, but I hope that one can fell a bit more comfortable after annotating the produced networks with the validated interactions.

Hallo Franscisco, your site is really very helpful. Thanks for the site and your valuable comments

Hi all, wonderful to see this discussion. I agree with most what is said. As I was once told; "it is complicated". Not very helpful. Let's take ceRNAs (competitive endogenous RNAs). A logical and at the same time very hard to digest idea. The weakness of the ceRNA hypothesis is that it goes against some micorRNA dogmas (at least according to my limited understanding of, is it thermodynamics?): microRNAs are often described as buffers of gene expression. Does that fit with the ceRNA hypothesis. another weakness is the dogma that a microRNA has several targets not just one and one pseudo-target, the ceRNA. The strengths are obvious. But I still lack the flexibility in my scheme of things to easily incorporate ceRNAs. They make things terrible complicated. They require the exact knowledge of the concentration of all the targets, their binding affinity, the concentration of the microNA itself.... And if there are co-regulatory microRNAs, then I need to know these data for them as well.

If we see many microRNAs as buffers, does this change the model?

How often does a microRNA really target several more or less equally important targets in a cell? How often does it mainly target one main target and can be influenced by ceRNAs? Here bioinformatics may help.

Dear all,

this discussion is very interesting for me since I worked with miRNAs only in the bioinformatics field and at several levels. Maybe, it is forthcoming to formulate explicite questions that makes "it complicated". Let me start with my questions and maybe I get a hint since related results and insights already exist.

Qs:

1- What will I do if I have different miRNAs for the same target (eg, a TargetScan's family of miRNAs): should I add all RNA-Seq NGS reads (total expression) of all miR-members to a single amount for further interpretation? (Why should a target, a capturer distinguish between the different species of a family?)

2- How are miRNAs regulated? Which influences the levels of the (members of a) mirNAs (family).

3- As Thomas already said, what are the Michaelis-Menton kinetics of an interaction between miR and (real) target?

4- What is the effective concentration (MM-kinetics) if a miR binds to other seeds than the regulatory sites?

5- What about of targets that are normally not used in the reference DB such as pseudogenes (they have often identical 3p ends ! ) or multiply existing splicing isoforms, which both make the picture more complicated?

6- What is the difference of absolute levels (e.g., in NGS reads) even if, in two samples to be compared, the relative fold change is equal, let's say 4? Is there a effective minimum concentration for the targeting effect?

Please, let us all add such general questions, the more fundamental the better, and/or try to formulate simple answers (otherwise links to related literature) ok?

Bests, Thomas-M

@Thomas Meinel

you ask the question that I am struggling with. I believe that I do not have enough biochemistry and math background to address them appropriately but recently there was a major blow to the ceRNA hypothesis (competitive endogenous RNA) that addresses some of your questions and simplifies the situation with microRNAs. A nice summary or "guide" is "Competitive endogenous RNAs cannot alter microRNA function in vivo.

Broderick JA, Zamore PD.

Mol Cell. 2014 Jun 5;54(5):711-3. doi: 10.1016/j.molcel.2014.05.023."

This put a lot of the basic thermodynamic issues on solid ground: there are far too many potential docking sites for microRNA-Riscs to make the concentration of an individual target nor the concentration of the microRNA a relevant factor. A very simple and daring conclusion from this is: don't bother with microRNA expression changes that are in the 2-fold range, don't care about microRNAs with low expression, ignore the ceRNA idea; or put in a positive way: focus on highly expressed microRNAs, if you look for microRNA changes make sure they are really substantial.

Coming back to the bioinformatics approach, we study in the same sample the transcriptome (affy human gene 1.0 array) and the miRNome ( Affy microarray) and built coexpression nets to study the architecture of the nets at different stages (Disease vs control). In our hands this approach seems interesting and gave us valuable information. The idea is that some (of course not all ) of the regulatory effect of the microRNA should be seen at the mRNA level and that could be tested in real data looking at the negative relation (edges) of the coexpression networks, helping us to detect the most relevant nodes of the net.