 Thank you, Neil. So I would like to thank the organizers for giving me the opportunity to speak here. And that's a picture of me jumping from joy because I heard that my abstract would accept it for a talk. And I'll be talking about a pan-cancer proteomic perspective on the cancer genome atlas. We at MD Anderson generate RPPA data, protein expression data for TCGA. So we analyzed, in this particular study, we analyzed 3,467 samples across 11 different tumor types. Those are the tumor types listed there. We did not receive any AML samples, so as opposed to pan-can 12, we had pan-can 11. We had 181 proteins in total, 128 of them were total proteins, one cleaved, one acetylated, and 51 phosphorylated forms. And the data were produced in six reverse-phase protein array or RPPA batches. This is a platform for generating protein expression data. So the first step we did was we developed an algorithm called replicates-based normalization to reduce batch effects between the six batches. After correcting for batch effects, then we performed protein versus other platform comparisons. So the first one was gene versus mRNA versus protein matched comparisons. So what we found that on average, the mRNA versus protein correlations was on the order of 0.3. That sounds a bit on the low side, but this is common for patient samples. In cell lines, we see much higher correlations, 0.7, 0.8. But for patient samples, this is common. But what's interesting is that when we looked at the individual mean correlations within the 11 disease types, we found that the correlation was not uniform across diseases. In colon cancer, for instance, the average correlation between mRNA and protein was on the order of 0.16. However, in ovarian cancer, the correlation was much higher, was 0.33. So we didn't see a uniform distribution of the correlations across disease types. When we compared copy number variations with protein, we found that, on average, amplifications resulted in only a 5% increase in protein levels, whereas deletions resulted in only a 5% decrease in protein expression levels. Now, there is a caveat here that this may be due to the large number of passenger amplifications and deletions that we often see in copy number data. So those might be diluting the results for the copy number data. For mutation versus protein, we saw that, on average, elevating mutations, those that result in higher expression levels, increase protein expression levels by about 20%. Whereas suppressive mutations decrease protein levels by about 10%. So those numbers are more than the copy number numbers. And, again, we think that's because there are fewer passenger mutations than there are passenger copy number variations, so that might possibly explain why we see this discrepancy. Other comparisons we made were between microRNA and protein, so every microRNA versus every protein, mean spearmint correlation was on the order of plus or minus 0.07. And protein versus protein, mean spearmint correlation was on the order of plus or minus 0.15. I'm going to focus on herb B2 as a use case. So what we did was we used breast cancer, HER2 positive versus everything else as our benchmark. And using HER2 positive versus all the remaining breast cancers, we identified thresholds for copy number, mRNA, and RPPA that would separate HER2 positive from HER2 negatives. Then we applied those same thresholds and other disease types to see what comes out. And what's interesting is that, not surprisingly in many diseases, we don't see protein expression level, you know, it's high in, let's say, less than 5% of the samples. But in some of the diseases, what we saw in those diseases where more than 5% of the samples had HER2 positive protein expression levels, we saw that by comparison with copy number and mRNA, proteins predicted a much higher level of samples, a much greater number of samples that had high HER2 levels. So for example, in bladder cancer, we saw that only 7% of the samples were HER2 positive by copy number, 8% by mRNA, but 22% by protein. Similarly, for colorectal cancer, only 7% by copy number, 3% by mRNA, but 37% by protein, and so on. So we saw discrepancies in lung adeno and endometrial as well. So why is this important? This is important because many of the therapies that are currently being used, for example, Herceptin and the new drug that's about to come out, TDM1, they work at a protein level and they actually target the protein on the cell surface. However, many of the tests that we use today to check for HER2 positive versus HER2 negative patients, they rely on copy number versus or mRNA platforms. So this finding is interesting in that we should probably, you know, try to measure a protein directly when trying to assess HER2 positive versus HER2 negative patients and who could possibly benefit from the therapy. Moving on, this is the clustered heat map for the 3467 samples by 181 antibodies. And using unsupervised clustering, we got eight different clusters. We nicely got the women's cancers, that's luminal breast, endometrial and ovarian in one major cluster on the left-hand side. Basil breast and HER2 breast were not in that cluster. They were quite distinct from the remaining samples. And the URL is given there. If anybody wants to look at this in much greater detail, the URL is provided. This is using the tool called Next Generation Clustered Heat Maps that was developed at MD Anderson. It allows you to navigate and zoom and search the entire heat map. So zooming in on this heat map, we see that most of the clusters, they cluster by tumor type. But there are a few exceptions. So for example, cluster F has a mixture of many different tumor types in there. Mainly squamous tumor types, head and neck, lung ad, no lung squamous. Some of the squamous bladders are also in there. The other cluster that didn't have a one-to-one correspondence with tumor type was colorect, was cluster E, in fact. And cluster E had basil breast, HER2 breast, and ironically, a few bladder samples. Those bladder samples were the HER2 high bladder samples. And if you look at bladder, that's a very different animal altogether in this particular picture. What we see is bladder is spread out across four different clusters. This is not in a single cluster. That's in contrary to mRNA clusters where we do see bladder in a single cluster for the most part. But here we have bladder in four different clusters. If you look at outcome differences, so bladder that looks like endometrile, so the ones that clustered with endometrial cancers, they have a much worse prognosis than the remaining bladders. And it's interesting that we found any difference whatsoever because bladder data isn't, you know, there are many examples with sensor data, so it's not the best outcome data that we have. Similarly, in kidney cancer, we found that the squamous-like kidney cancers have worse prognosis than the remaining kidney cancers. And then there's breast cancer. So in breast cancer, for those of you familiar with the marker paper for breast, we had identified a reactive group. This is based on protein expression levels in breast cancer. And that reactive group had good prognosis. Some of the markers of the reactive, breast cancer reactive group were caveolin, collagen, MYH11, and Richtor. Some of the marker proteins that were identified, so what we did was we did a differential analysis to see which proteins were differentially expressed in different clusters. So in women's cancers, we saw ER alpha and AR. Now, AR was a surprise. Now, prostate cancer was not included in Pancana 11, but regardless of that, AR was pretty high in women's cancers. In luminal breast, we found AR, BCL2, FAS, and ACC1 and phosphoACC. They were high. In ovarian cancer, we had CMIC. They were high. So these are potential targets. Not all of them have targeted therapies as of right now, but some of them are under development. So these are potential targets. All of them, except women's cancers, had phosphosarct elevated. And in head and neck, phosphosarct is a downstream target of EGFR. So perhaps we can use EGFR therapies for head and neck. And HER2 was high in a number of cancers including endometrial, bladder, breast, and colorectal. HER3 was high in kidney cancer. And phosphoEGFR, notch 1, HER3, they were all high in GBM. So maybe it could be amenable to combination therapy. The next step we did then was, well, so we find clusters that are mainly thereby tumor type. Those clusters have almost a one-to-one correspondence with tumor type. Can we go ahead and reduce tissue-specific effects so that we can see clusters that span multiple tumors instead of having a one-to-one correspondence for the most part? We did that using a technique called MC that we developed. And I don't have time to go into the details of that, but here's the zoomed-in view of the clusters that come out. So we came up with seven clusters. And as you can see from the top bar, the tumor types are all spread out among the different clusters. There's no single dominant tumor type that's there in each of those clusters. Cluster 1 had PEA and HER2 elevations. Then we divided cluster 2 into two parts, 2A and 2B. 2A had HER2 elevation. 2B had EGFR elevation. Then we had a cluster with AKT pathway high, a cluster with AKT pathway low. Wind signaling high, wind signaling low. And then we had the reactive cluster that I just spoke about. So the reactive cluster had predominantly breast cancers, but also other disease types were there in this cluster. So let me talk a little bit about this. This is the reactive cluster right there. This is the zoomed-in view of the same clustered heat map that I showed earlier. And we can see caviolin 1, MYH11, Richtor, collagen 6, those are the markers for the reactive cluster. So we can clearly see this is a reactive cluster. What is interesting about that cluster is that, like I said, in breast cancer, reactives have good prognosis. And the same is true for colon cancer and kidney cancer. So in colon and kidney, reactive samples have good prognosis. However, if you look at lung squamous, ovarian, and bladder cancers, then in those, the reactives tend to do worse. Now, we don't know why that's the case. All we see is an observation here that the reactives do distinguish from the others, and they can either have good prognosis or bad prognosis. And that's not just true for reactive, but ironically, that's also true for some well-known pathways. So, for example, the AKT pathway, one would expect that when the AKT pathway is suppressed, you would have better prognosis. AKT pathway is responsible for proliferation. And that is indeed what we see in ovarian cancer, that those samples where AKT pathway suppressed do better. But in kidney cancer, that's not the case. It's the other way around. In kidney cancer, when AKT pathway suppressed, those patients do worse. And as a sanity check, we tried the other way around. So when the AKT pathway is activated, we indeed see a trend where the ovarians are doing worse. But kidney is actually doing better. So this is somewhat of an enigma right now that some of the pathways, you know, they don't behave as expected in some of the tumor types. So it looks like pathway activity is dependent on tumor type. This is interesting because there are some studies or some trials that actually try to silo different samples by pathway activity rather than by tumor type. If we're going to, you know, do studies like those, then we should be mindful of such observations. And this is my no means conclusive, but it just raises a possibility that pathways may not be agnostic to tumor type. So to summarize, the take home findings are, when we did cross-platform comparisons, we found that mutations on average have greater meaning, have greater fold changes than copy number variations, possibly due to the larger number of passenger events in copy number domain. mRNA protein correlations can vary widely by disease. They are by no means uniform across disease. Her two protein levels are not predicted well by copy number or mRNA in certain diseases like colorectal, lung adenu and bladder cancer, but they are predicted well in breast cancer. And then several novel markers were identified. There were outcome differences seen across clusters, possibly driven by pathway differences. Pathway effects are not equal by disease. Certain pathway activations may have good or bad prognosis depending on disease. And not shown here, we also looked at protein-protein correlations, and even they vary by disease. So for example, P10 versus phospho-AKT, it was found to have a strong correlation in lung squamous cancers, but not so much in any of the other cancers. Finally, I would like to acknowledge my co-authors. This is the paper that is due to come out in nature communications in about two weeks. The title is Pancancer Proteomic Perspective on Cancer Genome Atlas. I also have poster number one, so anybody who wants to talk to me further is welcome to attend the poster session. And I would like to thank specifically Gordon Mills and the rest of my team. And that's a picture of us celebrating the acceptance of the Pancan paper over dinner. Thank you. Very nice work. Just one comment. The reason you are seeing the AKT pathway in kidney cancer when it's low, the behavior is better, it's because the mTOR pathway is activated, and just a feedback. Because kidney is very, very mTOR-driven disease. You see 6% of mTOR activation mutation, you see TSC mutation. I think that's the reason why you're seeing it. The other question I really have is between F and G, you're talking about kidney cancer, they are squamous type. What are the markers that differentiate them from the most of the G-type kidney cancer that you separate them? So I think you're talking about the very early years, because the main lineage. Right. Let me get back to that slide. Okay. So you're talking about this particular figure, right? So in squamous like, oh, there we go. So in this figure we can see some of the markers in kidney cancer in the squamous type. Those are the markers for the squamous type. So we have phospho-soc, for example, that's elevated in this cluster versus the others. There's also a fast N that is differentially expressed, and also AMPK-alpha. So those are some of the markers that we see between the two clusters. Great. Thank you. Thank you. Thank you very much. Our next speaker is John Martinetti from Mount Sinai, who's going to talk to us about data mining TCGA breast and ovarian exomes for novel susceptibility markers.