 The title is a little bit different, but it's the same idea. You'll notice that this talk is a bit different than the other talks we've heard today. We are using TCGA data to infer new knowledge, but not doing the primary analysis. Specifically, I will speak today about our efforts to use TCGA samples to infer post-transcriptional regulation in cancer. I'm mostly interested in regulation of and by microRNAs. MicroRNAs or MIRs post-transcriptionally regulate their targets through multiple mechanisms, including transcribed degradation that is shown in the cartoon above. They can act as tumor suppressors and they can act as oncomears. We and other groups like us are interested in building integrated networks but include both transcriptional interactions and interactions that involve microRNAs. In my opinion, in order to do this, we need to address, at least or at least initially address, the following four aims. We need to be able to predict transcriptional and post-transcriptional regulation of microRNAs, their targets, and to identify regulators of their MIR activity. TCGA samples and their profiles are exactly the kind of data that is useful that we can use to at least make initial attempts to address these four aims. Now today I'm gonna touch on two of these and these two are probably the least, they receive the least amount of attention, both in the systems biology community and in the scientific community at large. These are post-transcriptional regulation of MIRs and regulation of MIR activity. I hope to convince you today that these types of interactions are really worth looking at. Now microRNAs are very tightly post-transcriptionally regulated and changes in this type of regulation can lead to significant swings in the expression of mature microRNAs. MicroRNAs are transcribed, you can see the cartoon on the right, they're transcribed to produce transcriptional units, they're then processed to produce pri, pre, and finally mature microRNAs. The processing machinery includes canonical biogenesis regulators, Adrosha, Dyser, these are non-sequence specific regulators, but they also include sequence specific regulators. We believe that these are in the hundreds, but only about 10 of them are known. The effect of these biogenesis regulators can be seen in the visualization of deep sequencing data of a short RNA library. And what you see there are two tracks on the UCSC genome browser. In the bottom, in red, we visualize fragments, these are RNA fragments, less than 30 bases long, and in the top, in black, fragments between 30 and 150 bases. In the bottom in red, we can estimate the expression of mature microRNAs. Well, in the top in black, we're estimating intermediary products of microRNAs pre and pri. Looking at this cluster, very well studied cluster, of oncomeres, the distance between mirror 17 and mirror 19a on the genome is less than 400 bases. Yet, even though they're co-transcribed, the expression of the mature mirror 18a is almost 20 times less than that of mirror 17. Looking at the black, the top track, we can see evidence for the accumulation of pre and pri product, intermediary products. Intermediary product of the mirror, suggesting that the reason that mirror 18a is poorly expressed is that its biogenesis is not as efficient as the biogenesis of its flanking microRNAs. These are exactly, this is exactly the kind of thing we're looking at when we examine these CGA samples. We're looking for deviations in the expression of co-transcribed mature microRNAs. Genes whose expression correlates with this kind of deviation would be for us candidate biogenesis regulators. We can do the same thing for hosted mirrors. Mirror 218 sits in an intron of slit 3. If slit 3 is expressed, so would be mirror 218. We expect to see an increase in the expression of the mature 218 if slit 3 increases. And we see that in the top right. When the expression of another gene, DDX10 is high, an increase in the expression of slit 3 is followed by an increase in the expression of mirror 218. But this does not happen on the top left where the expression of DDX10 is low. This to us would be an indicator of a DDX10 is somehow regulating the biogenesis of mirror 218. And indeed, upregulation of DDX10 in the glioblastoma cell line SNB19 led to upregulation of both the pre and the mature 218. Now doing this a little bit of a higher throughput, we identified a set of microRNAs of interest, design probes that measure the expression of mature pre and pri products, downregulated silence DDX10 and measured the response of these mirrors to the downregulation of DDX10. And what you see here is a panel of mirrors where the mature microRNA responded to the downregulation of DDX10 while the intermediary products are relatively unchanged. Looking at even a higher throughput way, we measured the expression, the response of about 800 microRNAs to silencing of predicted biogenesis regulators. What you see in the heat map on the right, every row describes the response of a single microRNA to silencing of every one of the genes listed at the top. We selected two positive controls. These were predicted to downregulate microRNA biogenesis, to change microRNA biogenesis in glioblastoma cell lines and are also well known biogenesis regulators. So DDX5, DDX17, positive control. We assembled a panel of predicted biogenesis regulators and a panel of negative controls. These are proteins that are similar to the predicted biogenesis regulators but are not predicted to regulate biogenesis in glioblastoma. There are two things I'd like to point out in this heat map on the right. First, the response of the predicted biogenesis regulators is more similar to that of positive control than negative control. And second, looking at every row in this heat map, you can see that microRNAs are responding differently to the silencing of different regulators, which is in line with the assertion that these are sequence specific. Going back to DDX10, we collected the microRNAs that respond to the DDX10 inhibition and identified an enriched motif in the precursors of these mirrors. Drilling back down again, we could even identify biogenesis regulators that are very different on the protein level but have very similar signatures in terms of microRNA response. And now we are looking to see if these kind of interactions are synergistic. I'm gonna switch to regulation of mere activity. For operational reasons, we distinguish between two types of regulators of mere activity or modulators of mere activity. Oh, what's happening? Sponge regulators compete for microRNA programs but regulate other RNAs. And non-sponge regulators activate or suppress MI risk mediator regulation of target RNAs. I'll give you an example of each one of these. P10-P1 is probably the well-known, the most well-known sponge regulator. P10-P1 is a pseudo gene of P10, a known tumor suppressor. P10-P1 and P10 share homology in their 3-pan UTRs and are targeted by many of the same microRNAs. What Policeno and Salmena argued was that P10-P1 over expression attracts microRNAs that would be otherwise free to target P10 and therefore increase P10 expression. And similarly, downregulation of P10-P1 releases microRNAs that would otherwise target P10-P1. These are now free to target P10, reducing its expression. TNRC6A is an example of a non-sponge regulator. It is required for my risk function at targets and mutations in this protein have been shown to contribute to tumor genesis. So we are interested to identify these things on a genome-wide basis. And this cartoon, I hope, would illustrate the main idea of our method. So looking across TCGA samples, sample each sample in every column, we are looking at genes identifying microRNA programs that target them. And what we would hope to see is that the expression, the total expression of these microRNA programs is inversely correlated with the expression of their target. When we do a genome-wide screen, we are looking for genes that change this correlation. And in this cartoon, you can see the mere program at the top when the modulator is high, our candidate modulator of mere expression is high. The mere program is inversely correlated with the target. In the bottom, when it's low, this correlation is weaker. To us, this would be an indication that the modulator is a modulator of the activity of this mere program. We can look for modulators of mere programs that are target-specific, mere-specific, and oak combinations. When we did a genome-wide scan, we were amazed to see a huge number of interactions. On the left is a visualization of the network that we predicted. But a quarter million interactions involving about 7,000 genes. Not only is this RNA-RNA network, which we call the NPR network, not only is it large, the forces exerted by these interactions are also unusually, well, unexpectedly large. On the right, you can see that these forces depend on the size of the neighborhood. As the number of regulators for a gene increases, so does the correlation between the regulators and the target. And on the right there, we are approaching Pearson correlation of about 0.6. P10, which we mentioned before, is way at the right. Inside this network, we identified a small subnetwork. It's very dense, a lot of edges, and it includes six well-known glioblastoma drivers. To see what response the expression of each of these genes has to changes in 3-prime UTRs, which inspected the cell with P10 3-prime UTR, and so on upregulation, roughly about 50% in every one of these genes. Now, this is not very strong upregulation, but it is consistent and it is significant. Silencing Dicer and Drosha, which are canonical, non-sequence-specific by Genesis regulators of microRNAs, completely abrogated this response. Moreover, as I said in the previous slide, the regulation is combinatorial, and so, transfecting with pairs of UTRs instead of a single UTR, even though we're doing it the same quantity, increase the response dramatically, we went from 0.5 to about two-fold change here. Not only is this affecting gene expression, it's also affecting phenotypical things that we can observe about the cell. So this cartoon here, this diagram here, shows what happens to the cell when we silence P10, that's the purple line at the top, the cell growth rate increases when you silence P10. Transfect with P10 3-prime UTR, and cell growth decreases. Transfect with P10 CDNA, you have forever increase. This is what we expect to see. Now, looking at regulators of P10, and we specifically were interested in regulators that have whose locus is deleted in samples where P10 locus is intact, we compared the effects on cell growth rate of silencing P10 and silencing each of these regulators. I didn't include this, but silencing each of these regulators led to downregulation of P10, and what we see here is the consequence of that. You can see that in blue silencing these regulators leads to a similar, if not the same reaction in cell growth. These genes have never been associated with cancer before, but their effect on P10 is leading to the same effect of silencing of P10 itself. Each one of these genes may be deleted in very few tumors, but there are 500 of these, and combinations of deletions at each of these locus will have the same effect. So with this, I'll end. I'd like to acknowledge Hosheng, who is sitting here and has a poster outside. Wei-jen and Shuri, and this work was done in Andrea Khalifano's lab with help from Jose Siles' lab. Thank you for listening, I'll take some questions. It's... Are there questions for Pavel? Question here? Very interesting approach. So with arachne, you can make pairwise predictions on transcriptional regulatory interactions just through correlations between gene expression patterns and then the sort of sponge MRNA interactions can be predicted from sequence homology. So do you see, when you look at those two independent ways of drawing networks of gene regulatory interactions, do you see that an edge that's predicted by arachne tends to be more, at least between these two genes, is enriched in those predicted sequence homologies that would suggest a sponge regulator? I didn't quite understand, but I'll give you my interpretation. I think you're asking whether the two methods are predicting the same interactions, right? The answer is no, they're mutually exclusive almost entirely. So for example, we use arachne to understand why genes are downregulated when we silence a transcription factor. Looking at data of transcription factor silencing and trying to understand why genes are reacting to the silencing, you'll find that this kind of a method we're talking about mirror activity regulation, these are indirect RNA-RNA interactions. This method explains an entirely different set of genes that are explained by arachne. The size is about the same. There's almost nothing but it's common between the two methods. All right, thanks again.