Target prediction can certainly be useful, but I agree with Raul's original point, that "the use of these data to do globals analysis, like networks approaches, could be a mistake." The problem of target prediction is compounded when you consider context-specific targeting. Certainly whether or not the gene and miRNA are expressed in the same cell will influence the accuracy of a predicted target, but several recent papers also suggest that the expression of other genes targeted by the same miRNA influence how well that miRNA binds to its other targets. This makes it very difficult to predict a miRNA target without out factoring in at least miRNA and mRNA expression data from your cells of interest.

If you are able to culture cells similar to the in vivo cell type you care about, I think the clearest solution to this problem is to begin to move away from target prediction as the methods for experimentally and globally identifying miRNA targets, in the cellular context you care about, have become more refined and more economically feasible to conduct. These techniques include Ago2 pulldowns (Phil Sharp had a nice paper on this recently: http://www.journallab.org/papers/263), biotin-labelled miRNA mimic pulldowns (see Orom's original papers: http://www.journallab.org/papers/8543), if you have access to a good MS lab, SILAC post-miRNA introduction or inhibition is a direct measurement (http://www.journallab.org/papers/614 & http://www.journallab.org/papers/8544), and similarly, ribosomal profiling (http://www.journallab.org/papers/5655).

Though many of techniques can be expensive and time-consuming, the good news is that most of these studies have concluded that the most down-regulated targets of miRNAs can be detected as reduction of the mRNA, not just the protein. As such, an experiment as simple as introducing miRNA mimics or inhibitors and profiling via array or seq can tell you what the top miRNA targets of your favorite miRNA are in the context you care about.

Hi Raúl

Very pertinent and interesting question. In fact, in silico methods for target prediction are quite different, depending on the alrorithms that they used. For example some of them take into account the 2D structure of the target, the compensatory 3' effects besides of the seeding sequences, and many other factors.

It is very difficult to say if a particular target prediction method is credible or not. Some of them are very optimistic (in this group I will put Pita algorithm, for example), and very good if you are depressed because you always will get a target. Others like Diana-microT or Targetscan are more conservative, but in this case you can miss some biologically relevant targets.

My humble and very personal advice is that you should try as many applications as you can and then intersect the results. There are many servers that can help you doing that. For example MirDip server from Igor Jurisica's group in Toronto, is able to scan as much as 14 different target-prediction applications. You can ask the server the minimum number of applications that you will consider to be a realiable target, and intersect the data. If a target is predicted by more than 50% of the applications you must consider it as a potentially interesting one. MirDIp server is located in:

http://ophid.utoronto.ca/mirDIP/

There are also similar servers like MiRecords that also scan several target prediction programs showing a combined representation of the results of them.

I am open for discussion or help if you need it. Best luck.

Cheers. Paco

Hi Paco,

I do just that, with miranda, RNAhybrid, PITA and TargetScan. but the percentage of intersection between pairs are around 50% and among them decrease to 32% (employing a set of concrete parameters in each). In consequence, you are losing a relevant number of targets. A Highly conservative approach.

Other problem is the scoring function. Is it more important the structural accessibility, the binding energy, the canonical seed binding, the dinucleotide frequency around the target site or the relative site position? We must Employ an independent probability function or weighting each algorithms.

I agree Raul.... among the apps that you are using, PITA will be the most optimistic. The other ones are pretty conservative. I think that the parameters are not exactly comparable between apps, but the numbers that you got are pretty standard (around 1/5 of the targets in common). But you should increase the number of apps to get a nice overview of what's happening.

About the scoring function in my opinion the most relevant parameters will be the seed sequencing (try to be conservative and apply a minimum of 7 matches), and in second place the structural accesibility. The binding energy will be integrated in the seed.

The idea of implementing an independent probability function for all the targets sounds great. Let me know. Cheers

I would recommend checking out miRNA targets PCR arrays from QIAGEN where the content development scientist has already put together a properietary algorithm that takes into account the currently available prediction tools. here is the link:

http://sabiosciences.com/miRNATargetsPCRArrays.php

If you do not see the miRNA of your interest, they are able to design target arrays upon request

I think this is an interesting problem, and probably a subset of a more general problem. For instance, in RNA-Seq expression analysis, there are multiple tools for estimating differential expression, and for any given data set each will give a different set of genes called as significantly differently expressed, because each method uses a different underlying model. There is often only a small overlap if you take the intersection.

I suspect that in this class of problems there ought to be a way to take a slightly smarter approach to the overlap - intersect is too strict often, and union is too lax. Maybe weighting each method, and developing a scoring metric taking into account the 'opinion' of each method to differing degrees, is needed for these problems. I think this is what Raul was saying?

In fact I'm using more than five different programs like some of them you had mentioned before. Then I follow them with shuffling analysis, mfe of the secondary structure matching between miRNA and mRNA targets of them. This methodology gave me a good results and a good validation because the prediction with combination of different software only is not significantly but it have to be followe dby some analysis to test if it were a true targets or not. Shuffling analysis is a good one because it confirm that your target is not a chance target but it's almost a true target (calculate the P value and Z score) for the matching results after shuffling (I'm shuffling the sequences 100 times) then run each one again with the targets.

Overall my opinion is that you have to follow the prediction with both in silico validation and in vitro validation if it possible

I Marwa, I did something like that, but I wrote a million of randoms sequences and I tested the predictors against these random sequences. Problems, to produce a acceptable alignment needed to use TargetScan. Then could be better as you said just re-arrange predictions and scores.

I think this is a very good and opt question and very good and insightful answers from Francisco and Marwa. I also get this question always if to and to which algorithm should I belive while predicting miRNa targets. In vitro validation is absolute, but one cant just keep validating 1000s of targets you predict with the present databases and softwares. Therefore, there should be some robust way to short-list your targets from 1000s to 100s and then to very few most promising based on our cell system / work model / pathway or disease of interest and so on. And then it is wise to try validating those selected targets.

PCR arrays suggested by Ravinder also sounds a good idea for fast screening but again the question can be raised if to trust that array or not?

this is critical step in our decision. there are plenty of tools in target prediction. all predicted sites should be validated. just 4 tools have accuracy more than 60%. Diana-microT - PicTar - Targetscan - homotarget. use and trust.

I guys....I think and that was one of my big problem during my Ph.D. training that one of best method is to combine different tools to see the overlap to predicted the most putative targets. Considering this issue recently I developed a new tool called CoMeTa based on co-expression analysis of target genes for each miRNA. In my tool I combined different tools to have the maximum possible targets predicted for each miRNA. Than we rank those targets by co-expression analysis looking at 5000 different microarray analysis. In this way the tool is completely unbiased and based on experimental data. In the same way we was able to predicted the biological function for each human miRNA. We tested this tool and we found 100% of overlap between our predictions and the literature data (looking at 160 different papers). We also tested this tool experimentally and I can say you that it worked very well....And recently using the same tool we was able to identify a human specific miRNA. Why CoMeTa works well? Because that idea was if a target gene A and B are bona fide targets, so they should be co-expressed and maybe regulated by a single miRNA, like a transcription factors. Sure also CoMeTa is a prediction tool but using that rank and fixing a threshold you can reduce the number of false positives and so the number of putative targets to be tested experimentally..... Here the link and I hope that it could be help you..... http://cometa.tigem.it/site/

dears; there is an article entitled "Computational challenges in miRNA target predictions: to be or not to be a true target?" which can help you.

Raul take a look at this: http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0035561

Dear Vincenzo, Your approach is really interesting but, what happen with people like me that work with non model organisms.

Good question....but you need to consider that usually many gene targets are conserved....I agree with you that maybe you cannot manipulate a miRNA in your organism...and maybe you will loose many targets or you cannot be sure about a specific miRNA in your organisms....It's just a prediction but sure you can cut off many false positives...

Is there any way (apart from validating in vitro / in vivo) to decide which targets are false positive and which are not?

Target prediction can certainly be useful, but I agree with Raul's original point, that "the use of these data to do globals analysis, like networks approaches, could be a mistake." The problem of target prediction is compounded when you consider context-specific targeting. Certainly whether or not the gene and miRNA are expressed in the same cell will influence the accuracy of a predicted target, but several recent papers also suggest that the expression of other genes targeted by the same miRNA influence how well that miRNA binds to its other targets. This makes it very difficult to predict a miRNA target without out factoring in at least miRNA and mRNA expression data from your cells of interest.

If you are able to culture cells similar to the in vivo cell type you care about, I think the clearest solution to this problem is to begin to move away from target prediction as the methods for experimentally and globally identifying miRNA targets, in the cellular context you care about, have become more refined and more economically feasible to conduct. These techniques include Ago2 pulldowns (Phil Sharp had a nice paper on this recently: http://www.journallab.org/papers/263), biotin-labelled miRNA mimic pulldowns (see Orom's original papers: http://www.journallab.org/papers/8543), if you have access to a good MS lab, SILAC post-miRNA introduction or inhibition is a direct measurement (http://www.journallab.org/papers/614 & http://www.journallab.org/papers/8544), and similarly, ribosomal profiling (http://www.journallab.org/papers/5655).

Though many of techniques can be expensive and time-consuming, the good news is that most of these studies have concluded that the most down-regulated targets of miRNAs can be detected as reduction of the mRNA, not just the protein. As such, an experiment as simple as introducing miRNA mimics or inhibitors and profiling via array or seq can tell you what the top miRNA targets of your favorite miRNA are in the context you care about.

Most of you provide good tips, pros and cons of using certain algorithms to predict the target of certain miRNAs. I am interested to know what web tools are good and reliable for predicting the miRNAs for a specific target gene.

I agree with Aamir, these are target predictors, thus, there is a chance to get it wrong and a chance to get it right. The ultimate test though is to do functional studies where the predicted targets are assessed.

This is a really interesting point Robert, The activity of a microRNA into the cell are determined not only by the accessibility to its targets but by the relative amount of these targets, the binding energy between them, but also by the accessibility to AGO2 protein against the full set of microRNAS expressed into the cell in a competitive fashion. Is for that that I consider the approach of Vincenzo is really interesting. I am not 100% agree, because not all the microRNAs acts degrading the mRNA and not all of them do that with the same rate. For example If you have two targets in the same mRNA, is the activity increased proportionally or exponentially? Is a target near to the poly-A tail more degradative than other in the middle of the 3'-UTR?

Dear Raul,

Very Very interesting question and there are different opinions on believing upon the target prediction tools. Personally I have used RNAHYBRID, MIRANDA and PITA which are having different algorithms to predict the target site. I agree with Francisco Enguita that it's better to take similar parameters for all the tools and try to find out the common and most representative hits, this will help you out.

There is a good chance that the common targets you may get during experimental validations.

Hi Raul and everybody

Just to introduce some more chaos ... you have also to take into account that two different targets for a single miRNA in the same UTR could have different efficiencies of regulation by themselves. This happens mainly because of the compensatory effects of the 3' region of the miRNA.

As you pointed out, the regulatory meaning of a particular mIRNA within a cell makes no sense if its cognate target is not present. In fact we are dealing with a very complicated subject, specially if we look to the cell as a integrated machine.

I think as more miRNA-mRNA validated results by in vivo assays, much closer we are going to be to find a general rule of this interaction. We are still far from have a universal prediction system. I suspect that it will be a matter of stoichiometry, dose-dependance, and possibility to interact with another miRNAs.

Actually answer of this question is different for different person in my view, it still need alot of work to say more about miRNA-mRNA target prediction, but we need to continue the work and collecting information in the form of manuscript and I believe improvement is underway and after few years it may some other tools which will answer your question but I do not think now it is possible,.......

It's a prediction statistically suggestive

This paper might add to the above comments. http://www.pnas.org/content/early/2011/05/24/1017386108.full.pdf

Hi Raul, This is in response to your post-last week. Thanks so much for the great question and ongoing insightful replies.

First, I completely agree that not all miRNA targets undergo detectable mRNA degradation. However, several of the studies that I cited in my above post that globally identify miRNA targets experimentally (see, especially, Bartel's paper: http://www.journallab.org/papers/5655) find that the targets that are most down-regulated on a protein level, are also measurably down on the mRNA level. It follows that profiling after miRNA over-expression or inhibition, and filtering for genes that have the appropriate miRNA binding sites, is an efficient surrogate for identifying the "strongest" miRNA targets in your cell type of interest - "strongest" here meaning most significantly altered on the protein level.

That being said, there must be functionality behind the evolution of miRNAs to target hundreds of different genes, and thus there is no reason to focus only on the "strongest" most-downregulated targets. Co-suppression of the entire network of targets, both strong and weak, is likely extremely important. Technically, however, these are just much more difficult experiments to conduct, so focusing on a handful of "strongest" targets is just one way to hedge your bets experimentally. As Francisco pointed out, this is a very complicated network.

Concerning your questions on the effect of the number of binding sites, the number of other miRNA targets, and the location of the binding site in the 3'UTR on suppression, I think these are fascinating questions still wide open in the field. This paper begins to address some of them (http://www.journallab.org/papers/562) and, importantly, characterizes a nice system that can be manipulated to directly address your questions.

It seems target prediction is at best complex and uncertain right now. I was wondering whether analysis of the global gene expression in cells treated with an antagonist of a miRNA of interest is a viable strategy to elucidate target genes.

Hi,

Anna, really useful paper and your paper "Extracellular small RNAs: what, where, why?" is also really exiting and expand the network complexity even more. Thanks a lot.

Robert, I will implement a anti-correlation analysis in my RNA-seq datas looking for possibles targets and after that I will compare the candidates with the previously predicted targets. If you are right this analysis should work. Thanks Robert.

If I am using Diana Micro T for predicting targets for a particular miRNA, its give hundreds of predictions with a prediction score, so in this case how many targets(threshold score) can be considered relevant to study.