William Sellers

Transcription

[applause] William Sellers: So that was a real honor to be introduced by Harold, and I'll take a few minutes to tell you a story. Recently, I built a man cave in the basement. And I've been saving newspapers my entire life from important events in my life, and now I have the man cave, the newspapers are going to go up as posters. And, of course, it's, you know, Red Sox win the World Series, the shuttle launch, things like this. But there was one from my four years in San Francisco, which the headline is "Giants Win Pennant," so, of course, that's why I saved that newspaper. But then on the lower right corner I noticed "Varmus and Bishop win Nobel Prize." [laughter] Of course, quote, the "visionary" that I was, didn't know Harold, didn't know Mike, and didn't know how important the discoveries they made at that time would be for us. And for my own career, really that observation, that cellular oncogenes were the thing that drove cancer, not tumor viruses per se, which was really sort of the thing we were all -- that was being thought of at that time, that, plus the discovery of tumor suppressors, I think, set the trajectory that we've been on for the last 20 years or so with the objective of understanding the combination of oncogenes and tumor suppressors that contribute to cancer, and then making therapeutic advances based on that. So, Harold, you'll be happy to know that we'll be in the man cave since you were lucky enough that you won the Nobel Prize the day the Giants won the pennant [laughs]. So, the other interesting observation that happened during the last 20 years or so was the paradigm shift that was exemplified by the results of imatinib and the disease chronic myelogenous leukemia, and I think this is well known, but still worth reflecting on. Of course, CML is interesting in that it was really the first time a somatic genetic lesion was discovered in 1960 by Nowell and Hungerford. It only took us about 50 years to figure out what DNA was, what a kinase was, what phosphorylation was, that you could make small molecules. But 50 years later, you have a drug, imatinib, that's very effective at inhibiting the output of this oncogene. And this is the efficacy data in the United States, not based on the clinical trial, but rather based on looking at the statistics of mortality and incidents per year in the United States, just going to the Cancer Journal of Statistics and pulling out their estimates every year. So beginning in 1997, the incidence of CML, the number of new cases per year in the United States was around 4,500, and you can see that that's remained constant over the next 14 to 15 years. Imatinib was introduced into the market in 2001, and thereafter mortality dropped from an annual mortality rate of 2,500 deaths per year down to around 440. Nilotinib and dasatinib, second generation ABL inhibitors, were introduced in 2010. Interestingly, the mortality rate in 2011 dropped a little bit. You can sort of speculate based on no data that that may be related to second generation inhibitors coming into the market. So there are a few things worth noting. First, when this really was going, people may forget that there were a bunch of naysayers that thought this was not going to really work. That is, yes, it was working transiently, but ultimately, the dreaded CML stem cell would take off and patients would die because the stem cell population would progress. Cancers were smart. They would find genetic ways around this. And, in fact, if you look at, not the incidence but the prevalence of the disease, this is the number of patients now alive with CML; in 2007 it was around 25,000 patients. By 2010, it was around 37,000 patients. This just illustrates that the mortality per prevalent case is continuing to decline. So there has not been sort of the inexorable progression in this case. Now that being said, there was still a big question as to whether or not this would or would not translate from a relatively simple genetic disease and one that's hematologic in nature, and maybe even represent the benign precursor to advanced leukemia. Would this really be relevant to the more complex, solid tumors that we see at a later stage? And Matthew and I, along with Tom Lynch, and Dan Haber, and Harold, and William Pao, were fortunate to participate in this discovery of EGFR mutations in lung cancer which, you know, illustrated that, yes, you could have fairly dramatic therapeutic results with inhibitors of specifically activated mutant oncogenes even in solid tumors. So I'm not going to go through the other many examples of where this paradigm has translated; just to say that I think this is a translatable paradigm more broadly than just in CML. So now we can sort of look back and think of two different modes of drug development, an era which we still are living in of empiric drug discovery, where therapeutics were applied without knowledge to the underlying pathogenesis, generally to unselected patient populations looking for small benefit. I'm comparing in A the intact trial of Iressa in unselected lung cancer patients to the power of a single patient observation, one patient with a mutation being treated having a dramatic response. So these two modes of drug development coexist today. I think we all hope to have more of B than of A, and the question and conversation for the rest of the talk is how do we make B better. And that, of course, involves sort of identifying the major limitations to making the drug discovery mechanism applicable to the genetics of cancer more robust, more reliable. So I'm going to spend the rest of the talk talking about five key issues for this paradigm. The first is if you want to take on the genetic basis of cancer and this is your job, you have to know the genetic basis of cancer, and for many years, we have not known the genetic basis of cancer. So we need to know this to completion, in my view, and I resonate with Lou's comment that the genetic experiment has been done. It's been done in humans over and over and over again; we just haven't gotten the results yet. So what does it mean to do this to completion? And I know there are people who think we know all the single gene mutations. So what? We need to know this. We need to finish the Atlas. We have a long way to go. We need to know this in all cancer types and all cancer subtypes. That's, in and of itself, a big challenge. We need to know it along the stage of cancer. In prostate cancer, what's the biggest question today? Yes, metastatic, but also the overtreatment of premalignant or benign early stage prostate cancer. What do we know about the genetic differences between that disease and the type of disease that kills patients? Not very much. So we need to know it across the evolution of cancer. And then we need to know it with a robust sample size and a robust depth of analysis that the genetics we understand at the level of functional redundancy, we can understand cooperation, and we can understand antagonistic genetic events. That is, two things are mutually exclusive either because they're redundant or because they're antagonistic. We're nowhere near the sample sizes yet to make those claims. Matthew showed a few examples, but I'm guessing those are still on the verge of statistical significance, and that's when you're only trying to do any two genetic lesions. The second you want to know three genetic lesions or four, we're still very underpowered with respect to the ability to do this. And I believe, firmly, when we have that power, these genetics are going to fall into well-defined pathways. Common nodes of therapeutic intervention could be identified simply by looking at the genetic map. So I think that's the aspiration and the hope. This is your task, I think, and I'm not going to spend any more time talking about it today. So that's the problem number one, and I'm counting on you guys to solve that. Problem number two is where we have to start thinking more deeply, which is, even when we've known the genetic alterations in cancer, some of the absolute best genetic alterations in cancer, we have not made sufficient progress in turning that information into robust drug candidates. And that problem comes in two flavors. The oncogenes; so how many of the oncogenes are actually druggable today? Not very many. And it includes oncogenes like RAS, which are mutant in 90 percent of pancreas cancer; ERG/ETV1 which looks like it's translocated in between 80 to 90 percent of prostate cancer; BCL2 which is a dominant oncogene in lymphoma. These types of oncogenes outside the kinases have really been refractory to drug discovery, and I think they've been refractory for two reasons. One is, they are, in fact, harder, but two is, people may not be trying hard enough, and I think that's where industry also needs to play a more active role. I think a great example of an attack on this problem was the work of Steve Fesik and his colleagues at Abbot Pharmaceuticals who took 10 years, and I think they're still working on this problem, to try to make inhibitors of the interaction between BH3 peptides and the BCL2 family members. And this is difficult because it's a very large surface, a very large protein-protein interaction, it doesn't have the same tight, well-recognized binding structures that kinases do. Nonetheless, they were able, over time, to elaborate ABT-263 which was a lead clinical candidate, and now they have a second molecule going into the clinic attacking this class of oncogenes. So this is an example of taking on the question of, quote, "undruggable," challenging the notion that something's undruggable, and making headway despite the skepticism that one could make a drug against this family. The second class of genetics we talked about earlier is the tumor suppressor pathways and, of course, this is even more difficult because instead of having an activated gene which you could inhibit, you simply have the absence of a gene. Now, I think this is where the concept of synthetic lethality is going to play a big role, and I just wanted to show you a few examples that may not be so obvious to people that suggest this is really working. So the first is the example of the Hedgehog pathway. So we know that mutations in the patched tumor suppressor gene occur in basal cell carcinoma and medulloblastoma, and Phil Beachy's lab discovered the natural product cyclopamine as a natural product antagonist of the receptor smoothened. Now based on the way this pathway works, it was predicted that patched deficiency would lead to constitutive activation of the smoothened receptor, and the work from the Beachy lab suggested that antagonists of the smoothened receptor would reverse the phenotypic consequences of patched deficiency. Now this seems like an obvious one, but synthetic lethality doesn't say it has to be in a parallel pathway or some magically unknown mechanism. This is, or can be, an example of synthetic lethality as well. So I just wanted to show you the example that we've been working on, LDE225, which is a smoothened inhibitor. It has an IC50 for human smoothened and in vitro assays of 11 nanomolar, and then cellular assays of seven nanomolar. So it's a very potent, non-natural product synthetic inhibitor of smoothened. We were very interested in the idea of targeting medulloblastoma, and there were a lot of divergent opinions as to whether you do or don't need blood brain barrier penetrating molecules to treat a CNS or cerebral lesion like medulloblastoma. Nonetheless, we decided to make a CNS penetrating molecule. Here is an example of a pre-clinical study where a patched-deficient orthograft from a mouse was transplanted into a SCID mouse brain and treated with LDE225 versus vehicle, and you can see the two untreated tumors grow very rapidly, while the LDE-treated tumors regressed over time. So we had fairly strong evidence that in a patched-deficient model, in the right location, we were able to effect a therapeutic response in a pre-clinical model. So medulloblastoma is interesting. You would think we would just sequence the patched gene and then take those patients and put them on the drug. It turns out -- this is sort of three to four years ago -- sequencing patched from paraffin-embedded samples wasn't the easiest thing in the world because of the number of exons you have to sequence. Paraffin-embedded sequencing has made a lot of progress since then, but, in the end, we basically made a gene expression signature to capture hedgehog activity as had been defined by some of the key investigators in this field by the transcriptional signature of the pathway rather than by the genetics. So we developed an expression signature using 40 medulloblastomas that were embedded in fresh -- in paraffin-embedded tissue. A multi-gene model was built using the elastic net model. The optimal model selection was validated using an independent data set, and these five genes were selected for evaluation in the clinic using a QRT PCR assay. So four genes that are up in the hedgehog pathway and one gene that's down. So this drug has been through Phase I in both basal cell and medulloblastoma and is now in Phase II studies. This is an example of a pediatric patient who had a complete response to the drug. You can see at baseline, had a tumor near the brain stem and the cerebellum, and then by Cycle 5 had no evidence of disease. The results so far, using unselected medulloblastoma patients and then retrospectively characterizing them for the signature, is that all five patients who have a signature-positive medulloblastoma have responded to the drug, where none out of the patients who are signature-negative have responded. And this number on the right is now zero out of 21 signature-negative patients. So in many ways, this is an interesting control group. We have, you know, a mechanism-based signature, a mechanism-based inhibitor, and activity that seems to be strongly linked to the pathway signature. So this is an example, as I said, of this idea of synthetic lethality, where a mutation in the cancer predisposes -- it enhances the viability of the cancer. Of course, the drug target -- the drug itself has to be viable for the host, but where the cell that bears the mutation plus the drug has a lethal phenotype -- in this case, it would be the patched-deficient medulloblastoma cell being lethal when exposed to LDE225. So I'll just give you one more example of what I consider as synthetic lethal interaction and a successful attack on a tumor suppressor pathway, and that involves the PI3 kinase pathway and, in particular, the tuberous sclerosis gene. So tuberous sclerosis is a fairly rare, hereditary syndrome that's associated with a number of manifestations, but on the cancer side, it's associated with two interesting tumors: angiomyolipomas of the kidney and subependymal giant cell astrocytoma, a tumor in the CNS. Now work in drosophila had shown that -- or suggested that in the absence of TSC, mTOR kinase and S6 kinase would be constitutively disregulated, and, in fact, the first experiments of rapamycin analogs in a TSC-deficient setting were done in drosophila, and you could essentially reverse at least the larval phenotype in drosophila by treatment with rapamycin. So this suggested that rapamycin, or rapamycin analogs like everolimus, would be particularly effective in TSC-deficient settings. So this has been tested in children and adults with tuberous sclerosis, first in the subependymal giant cell astrocytoma setting but also in the angiomyolipomas. This is example of a patient with a fairly large subependymal tumor which has not had a complete response but has had a fairly significant response. We've conducted a Phase III trial led by David Franz and John Bissler at the University of Cincinnati, and the results of the Phase III trial are shown here. In the first year of the trial, no patient with SEGA progressed on therapy. At that time, this was a controlled trial against placebo because there's no approved therapy for SEGA. The overall partial response rate was 35 percent versus zero in the placebo, and in the kidney tumors, the response rate is 53 percent versus zero percent. And based on this data, the FDA approved Afinitor or everolimus in this disease late last year. So again, to me, this exemplifies the notion of synthetic lethality, that a tumor suppressor gene lesion like TSC predisposes preferentially the sensitivity to a TOR inhibitor compared to the normal cells bearing intact copies of TSC, which are, let's say, relatively insensitive or at least give you a therapeutic index when a patient is treated with a rapamycin analog. So everolimus was also treated -- was also studied in breast cancer and had fairly significant advances in progression-free survival in the ER-positive setting of breast cancer. Now, this goes back to model A, because honestly, at the beginning of this trial, I don't think anybody had any idea why a rapamycin analog would work in ER-positive breast cancer. I'm not sure we still understand why it would or wouldn't work, but what's intriguing from the work that's going on in the TCGA project, and I have to admit this was a very difficult figure for me to deconvolute but I've tried to simplify it to these two things, which are in the ER-positive subgroup of breast cancer, there's a fairly high rate of PI3K mutation, and one can hope or speculate that perhaps the effect of TOR inhibitors in ER-positive breast cancer is greatest in the PI3K mutant population. I have no data for that, but, fortunately, we do, in fact, have the samples from this trial, and those samples are being analyzed now in a collaboration with Foundation Medicine, looking at about 500 genes for mutations to see whether or not there are or are not correlates with clinical benefit to TOR inhibitors in this trial. I'll say one thing about this, because some people would think, "Well, now that you have an approved drug, why would you go back and even bother to find out?" I think that's a really good question that is a challenge for the industry. I happen to think there're some really good reasons. One is, if you know you have a mutation, you know the patient is more likely to benefit, you're more likely to optimize the drug for that patient. The second is, you're more likely to work through toxicity, find the dose that the patient can tolerate, find a treatment regimen the patient can tolerate rather than just simply give up because you didn't really understand why the patient would benefit in the first place. And then from an economic point of view, let's face it, if those are all the patients that are benefitting, that's where all the money is being made anyway. So I'm quite hopeful that this is going to make sense not just scientifically but also from a clinical and maybe even economic point of view. So I want to give you one more tumor suppressor example, because I think it's also motivated by the work of the TCGA, which is the common mutation in the PI3K pathways that are found in malignant glioma. The dominant mutation in the PI3K pathway is not TSC1 and 2 in this case but is rather PTEN, but also PIK3CA and the regulatory subunit of PIK3CA, shown further to the right on this side, are also commonly mutated, and I think this was a highlight of the TCGA glioma paper showing the sort of various dominant activation of the PI3K pathway in the glioma. So we've been working on PI3K inhibitors, and one of the central questions is, which PI3K subunit would be synthetic lethal with PTEN deficiency? I wouldn't say this is locked down definitively, but the preponderance of the evidence suggests that in the situation of loss of PTEN, as shown here, where we've compared depletion of PI3Kβ with PI3Kα, PTEN-deficient cells, as shown right here, tend to be pretty dependent on PI3Kβ versus α, and this is a PIK3CA knockdown in a PTEN-deficient cell shown here. Now because PI3Kα can probably take over, we have not taken the strategy of making a beta-selective inhibitor. There are other companies that are doing that. We've taken the strategy of trying to make pan-Type I PI3K inhibitors. People ask me, "Why don't you just make an alpha/beta?" I would if I could. It's not so easy to make an alpha/beta dual-specificity inhibitor because of the structural homology between alpha and delta makes that very difficult. So nonetheless, BKM120 is a PI3K inhibitor that inhibits alpha, beta, gamma, and delta. It has activity against the common alpha mutations. It also has very good blood-brain barrier penetrating properties; in fact, maybe a little too good. It accumulates in the brain over plasma concentrations. So this was also -- the intent of this program was to have a molecule that would work in glioma because of its BBB penetrating properties. This is exemplified in this study where we implanted a PTEN-deficient cell line in the brain. In the top panel, we're comparing BKM120 to GDC-0941 which is a type I PI3K inhibitor that does not have brain penetration, and you can see in blue is the PK of the molecule. So you get very good brain exposure after dosing with BKM120. GDC-0941 does not have brain penetration. Phospho-AKT suppression is shown in yellow, so concordant with high exposure in the brain, we get rapid and profound diminution in phosphor-AKT in the brain, and this is associated with the ability to prevent tumorigenicity in orthotopically-injected, in this case, breast cancer cells that are PI3K dependent. So we are currently studying BKM120 in glioma, and I don't have any results to share with you. I'm hoping it will be successful but this is our attempt to, again, exploit a potential synthetic relationship between -- synthetic lethal relationship between PTEN deficiency and PI3K dependence. So more broadly, there have been a lot of attempts to now discover synthetic lethal interactions, and the dominant way people are trying to do this is by shRNA screening. I would say, like genetic sequencing, this has been noisy so far but it hasn't been done to the extent that we need to do it, at a robust level with deep shRNA libraries across large numbers of cell lines. In many cases, people are using isogenic pairs of cell lines, which we find has a lot of noise to it. Our approach has been to try to use cell line panels that are genetically defined, find shRNAs that are selectively depleting or killing certain mutant cells. An example of this is shown here where β-catenin shRNAs are highly enriched for their -- in depletion experiments in the APC deficient subset of cell lines. So if you didn't know β-catenin was a key player in the APC pathway, you would've discovered it as the very top hit in this particular screen. So I have had a lot of faith still in these types of experiments. I just think that, you know, we're not at the point where they have been great yet, and I think they'll get better and better over time. So that's problem number two, and sort of the idea of taking on the oncogenes and tumor suppressor by working on difficult drug oncogenes, and then trying to exploit systematically this notion of synthetic lethality for discovering druggable genes downstream of either tumor suppressor or, of course, undruggable oncogenes. Now, the third issue, which we're going to face and we face already, is resistance. So there's not going to be a single drug that wipes out cancer like EGFR mutant lung cancer. We know we're going to have to have combinations, and the reason for that is resistance develops to targeted agents. Now, this has caused a lot of hand wringing, too. Right after BCR-ABL success with Gleevec, resistance developed, people said, "Oh, no," but shortly thereafter -- I'm going to just skip to this -- Charles Sawyers and Neil Shah discovered that the resistance was likely mediated by mutations in ABL. Now, this also caused two lines of thought. One was, "Oh, no. The cancers are so smart." Others of us thought, "Wow! That is pretty exciting!" I could've thought of three million other base pairs that might've been mutated, that might've caused resistance but, in fact, these cancers chose or had to mutate ABL in order to survive. To me, this suggested this concept of addiction was very powerful, and also led to the notion that the next best thing you could do in CML was to make a better ABL inhibitor. Now I would contrast that to the situation with Taxol. I'm still not sure we have any idea what Taxol resistance is. Why? Because we don't actually know how Taxol works. We don't know why it works in ovarian cancer. We don't know why it works in lung cancer. The advantage in the targeted therapy paradigm is we generally know why the inhibitor is working, we generally can understand the mechanisms of resistance at a much faster rate, and use that information to leverage further drug discovery. So in the case of Novartis, this led to the generation of a second molecule known as nilotinib. In the case of BMS, BMS developed dasatinib. Both of them are more potent ABL inhibitors. Nilotinib is a very interesting comparison because, structurally, it's very similar to imatinib, as I've shown here. The central difference is that it is 10 times more potent at the cellular level. Gleevec is 220 nanomolar in cellular assays. Nilotinib is 20 nanomolar. For KIT, we know that Gleevec is a KIT inhibitor. For KIT, nilotinib and Gleevec are fairly comparable. So the clinical trial that was done to compare nilotinib and imatinib was a test of whether more potent kinase inhibition matters or not. And I think, you know, that was answered dramatically in the yes, where the more potent kinase inhibitor, nilotinib, doubled the rate of major molecular response and complete molecular response. So that tells us that cells are really addicted to these genes and we really need to inhibit them very well, at least in the case of CML. So, in some cases, improved or enhanced target inhibition is going to be a method for overcoming resistance. We still wonder about this with EGFR, whether we really have the final best EGFR inhibitor yet or not. Such improved inhibitors will not only work in the resistance setting but they will most likely become the better front-line therapy. So the question is, are there other opportunities? I mentioned EGFR. I wanted to share you one new opportunity that we've been working on where we now have clinical data, and that is targeting ALK translocations in lung cancer. So I think many of you know that Pfizer's drug, crizotinib, was approved very rapidly after the discovery of the EML4-ALK translocations in lung cancer. Interestingly, crizotinib is a potent ALK inhibitor but also a potent MET inhibitor. In fact, it's a little better on -- it can be a little better on MET than ALK. We have made a selective ALK inhibitor that is 150 picomolar in in vitro assays and 3.2 micromolar on MET, so very selective for ALK. And 27 nanomolar in the cellular assays versus crizotinib which is 110 nanomolar, so about four to five-fold more potent in cellular assays than crizotinib. In an EML4-ALK driven xenograft single, you know, three to six milligram doses are sufficient to regress the tumors completely. And if we look in the Cell Line Encyclopedia, something I'll describe in a second, across 600 cell lines, you can see for LDK, the three most sensitive cells are all ALK-driven cell lines, and the gap between those and other, let's say, non-genetically ALK-driven cells is quite large. So we're very encouraged by the profile of this drug. We didn't really know at that time what mechanisms of crizotinib resistance would really be evident. It was -- we're following pretty quickly on the heels of crizotinib. But nonetheless, we went into crizotinib refractory patients, and you can see essentially every patient is responding. The response rate at the -- this is the data from the MGH and Alice Shaw -- is 81 percent in crizotinib refractory patients, really simply by making a more potent ALK inhibitor. So we're very excited about this data, but, again, I think it highlights the notion that really targeting the key oncogenes with potent inhibitors is one key mechanism for trying to prevent resistance. Now, that's not the only mechanism of resistance, and I think in the BRAF setting, we're seeing a quite different picture. So in the setting of BRAF, where we have well-defined downstream pathway, almost no BRAF mutations have been found as mediators of resistance to vemurafenib or other BRAF inhibitors. Instead, the work of people like Neal Rosen, and Levi Garraway, and Richard Marais have identified a host of different ways that melanoma cells seem to be able to evolve resistance to the inhibitors. Now that could be also really bad news but, in my view, the good news is still pretty good. Why? Because almost every one of those mechanisms reactivates the MEK/ERK pathway. Again, you know, I could've imagined mutations in the PI3K pathway or a lot of other pathways, but a dominant message that we're getting from studying resistance in the BRAF mutant setting is that pathway reactivation is critical for the development of resistance, and this has led a number of companies -- GSK is sort of out front on this -- to try to develop dual combinations of MEK/RAF inhibitors as a way to create a hurdle over which the cancer cell will not be able to get over. So I'm quite optimistic, in fact, that the study of resistance will ultimately lead us either to the best molecules and/or to the best combinations. But this does lead to problem number four, which is, we know that one drug is never enough. Now, ideally, the sequencing and the genetics of the cancer will tell us what we should be doing. We're, of course, not there yet because we're just starting to understand the single gene mutation frequency. Resistance may be another mechanism by which we get to the right combinations. We're also interested in trying to explore combination space by large-scale systematic screening, and I'm just going to skip this and show you the project that's ongoing. This is a large-scale combination screen that we're doing in collaboration with Zalicus. They're a company that used to be called CombinatoRx, which probably makes why we're collaborating with them more sense. The screen is 70 compounds by 70 compounds over 100 -- it's actually 138 cell lines now -- using this type of combination grid. Just to give you an idea of how hard this is, that's 27 million data points and it's taking us two years to just do that one experiment, and I still think it's underpowered myself. So, some of the things that are emerging are expected. We can see the expected synergies between CHK inhibitors and gemcitabine, or the antagonism between a microtubule stabilizer and topoisomerase inhibitors, but it remains to be seen when the data are complete whether we're really going to get informative stratification of combinations by this sort of large-scale screening. So, just to close, I've mentioned a few of the translational infrastructure model systems we're using, but I think this has been another problem to the advancement of genetic and other forms of cancer therapy, that is, lack of a pre-clinical translational infrastructure. And just to be very simplistic about it, how many papers have you read where the entire paper is about one cell line? Right, okay, that's one patient's cancer. We would never run a clinical trial with one patient's cancer. So we've had this problem that we've had, you know, very limited ability to profile pre-clinically the same number of samples from cancer of patients we're about to treat clinically. So the idea of the Cell Line Encyclopedia was to try to go from one cell line to this encyclopedia of a thousand cell lines. This was a long-term collaboration with the Broad, which is ongoing, where we bought from commercial sources a thousand cancer cell lines. They were bought from commercial sources so that if you want the exact cell line we used, you can order it from the same source. So we bought them, took them out of the vial, grew them in a limited way, and made DNA, RNA as soon after purchase as we could so that the community can hopefully access as close to the same cell line as the data is here. Of course, then we've gone on to do the genetics and expression. When we started this project, there was no next-gen sequencing, so we had sort of aspiration of sequencing, like 50 genes or something like that, and now it's, with the help of TCGA, going well beyond that. So, of course, the key now is to figure out a way to profile the encyclopedia, identify its sensitive cells, and hope that amongst the sensitive cells, there's a biomarker that is enriched in the sensitive cells as compared to the largely insensitive cells. And the trick to this is having a system that allows you to do this. So this is a system that was built first at GNF and then put in place in Novartis. So it's a robotic system with automated cell culture as well as compound handling. So this is the incubator. Those are plates in the incubator. Robot is retrieving the cells from the incubator. The cover of the cells comes off, and then it goes on to the compound dispensing deck. So these are compounds being pipetted into the 15 x 72 well plate. You can imagine trying to do that manually. And then after the compounds are dispensed onto the plate, the plate goes back into the incubator. Three days later, it comes out of the incubator, and then, by CellTiter-Glo, the number of cells on the plate is measured. So with this system, we can profile about 3,000 compounds with triplicate IC50 curves in about a two- to three-month period. In fact, the method of dispensing now has switched from pin dispensing to acoustic dispensing, which turns out to be much more rapid. So just to show you one example from the Cell Line Encyclopedia and how this can motivate the clinical development: This is our PI3Kalpha inhibitor, BYL719. It has a single-digit nanomolar activity against PI3K alpha and reduced activity against beta, delta, and gamma. So, of course, this compound has been run against the Cell Line Encyclopedia, and the types of profiles you're looking for are not the all-green, which are all dead, or the all-red, which are all alive, but compounds that are of interest are the ones that are going to have heterogeneous sensitivities among the encyclopedia. With the help of the Broad, we've built an informatics platform for sifting through 50,000 features that are, in combination, genetics or expression lineage, et cetera. Using compound response measure such as Amax or IC50 or AUC, using those to categorize cells into sensitive, refractory or intermediate, and then throwing out the intermediate and using the sensitive and refractory populations, building models that try to predict compound sensitivity. And Nico [spelled phonetically] is here and others that you can ask about exactly how this works because I really don't know. But what's impressive to me is from 50,000 features, the number one predicted feature for PI3-kinase was PI3-kinase mutation. Now, everybody's saying, "Okay. I already knew that." Still, for all those of you who do a lot of large-scale data analysis, having the right answer, not in the top 20, not in the top 10, at the number one position, I think it's still pretty impressive, and you can see from the list that was true for a number of oncogenic proteins and their cognate therapeutics. The power of that dataset for us is that when we went to do the clinical trial, the data were compelling enough that the clinical trial, from the beginning, was done in PI3-kinase mutation mutant patients. So right away, the Phase I was not an all comers dose escalation; instead, only patients with PI3-kinase mutation were enrolled under the Phase I, and dose escalation was done in those patients. And I can say we know it's well-tolerated. We've seen significant signs of tumor shrinkage as shown here in a patient with PI3-kinase mutated ovarian cancer. So I think this has been really transformative for us. Every one of our project teams now is waiting on an annual basis to see their compounds in this encyclopedia, and to, you know, try to either validate existing therapeutic hypotheses or create new ones. Now, we know that cell lines are deficient for many things. Number one is they grow on a plastic surface, and that can't possibly replicate all of human cancer. We also know that cell lines don't even replicate human cancer because, for example, prostate cancer barely exists in its form as a cell line. So, in parallel to that, we've been trying to establish primary tumor models, and a lot of people are now doing this. We've been doing this since 2007. We've implanted around 2,200 tumors and now have 410 established primary tumors that can be propagated, frozen down as fragments, re-thawed and used as model systems. So we're in the middle of categorizing these since we started in 2007. First, we were on arrays; now, we're doing RNA-seq; then it's whole exome. So it's sort of a mishmash right now, but we expect by the end of this year to finish profiling this, and we're doing ongoing collections to try to fill in the gaps that now exist. So we hope that this will become not as facile as the Cell Line Encyclopedia but still a second source of models that one can use to profile compounds and even mimic mini clinical trials prior to the human clinical trial. So I'm just going to close now by going through these five problems now as more statements of what we need to do. So, first, complete the cancer genome in depth. Work on validated but difficult to drug targets. Discover synthetic lethal drug targets, in particular in the tumor suppressor arena. Study resistance pre-clinically; don't wait until we get to the clinic but try to anticipate resistance and use that to drive either better therapeutic development or novel combinations. We need to discover those novel combinations and start testing them as early as possible in clinical development. And then, finally, we're continuing to build a robust pre-clinical translational infrastructure to allow more of us to sort of explore these questions at a level which will give us confidence and greater direction as we go into the clinic. So I just want to thank the patients who have participated in our clinical trials, and I have the privilege of working, as many of you do, with a great group of scientists, great collaborators at the Broad and elsewhere, and I want to thank them for all their help. So thanks. [applause] Matthew Meyerson: So, Bill, thanks for a really inspirational presentation, really demonstrating some of the ways in which the cancer genomic research that all of us are doing here can start to lead to benefits directly for patients. So we have time for a few questions for Bill to follow up, so please go ahead. Male Speaker: [inaudible] enthusiasm completely on the synthetic lethal approach, and I want to share with you some success we've had with one gene per well siRNA screening approach. Some of the successes we've had are very high level of reproducibility of our hits, about 70 or 80 percent. Secondly, we get a much deeper menu of potential targets. We get the whole iceberg instead of the tip of the iceberg because we can query every gene in the one gene per well approach. And third, it's a new area we're very excited about, is doing screens on primary patient-derived tumor cultures with siRNA, and we can do the assays over a period of a week or two. And these -- a couple of hits we've gone on to validate in pre-clinical models, but we want synthetic lethal with p53 and mixed synthetic lethal in a neuroblastoma model. So I think 2A is -- we agree that's a very good area to focus on. William Sellers: We actually have moved off of well-by-well screening to the pool screens, partly because you have the issue of transfection, and some cells are available to do transfection, others are not. But the heterogeneity of transfection was, at least at the scale we wanted to do this, somewhat difficult. We've done well-by-well lentiviral transduction which was -- as you can imagine, generating each lentivirus one at a time was quite the hurdle. So, anyway, I think there's no right answer but I'm glad to hear it's working. Female Speaker: Greetings. It's a very fascinating talk. Just very naively, I have two comments, and I'm interested in your suggestions. One is that I'm interested to know when is TCGA community going to work on metastatic tumors and so we can compare primary versus metastatic? That's one impediment. And the second thing is that when I worked on discovering subtypes of subtypes in breast cancer, what I did is data integration of different data types, copy number, mRNA expression, so on and so forth. So what I found is that there are many cell lines that are representative of individual genomic alterations, but one of the impediments was that there were limitation of cell lines that show the correlation of events as we see in the TCGA tumors. So what kind of comments or suggestions do you have to overcome those? William Sellers: Well, so maybe I can take the second question first because that's the easy one. Is a thousand cancer cell lines enough to represent human cancer? Not even close, right? So, yeah, the cancer Cell Line Encyclopedia is a limited representation. It represents what it can represent. We're going to try this year starting to convert our primary human tumors, the PDX models, in the cell lines. I would like to see an effort where people who generate cell lines, either through a publication point of view or a grant renewal point of view, are asked to deposit them into ATCC, because I think there are a lot of cell lines out there that are not available or not readily available. So I'm with you that the cell line representation is not great. Accessing the ones that are available, making them readily available would be one thing. Clearly, media growth conditions, different ways to grow cells, that's probably going to have to be important as well. With respect to your first question, like when is TCGA going to do metastasis, since I'm not on the TCGA, I wouldn't know. But probably when they get the metastasis samples, I would guess, right? [laughs] Female Speaker: Thank you. Lou Staudt: So, Bill, could you expand on your -- you said sort of offhand you didn't like the isogenic cell idea, and we have very great abilities now to manipulate cells with exonucleases, and we know that we have very complicated heterogeneity genetically in cancer, and are we going to rely on the luck of whether you get a primary human xenograft to get a model for a particular type, or why can't we engineer it, is the simple way? William Sellers: Well, certainly, if you have no choice, then I wouldn't stay away from that. So a few comments. Isogenic cell lines, in my experience, the ones that are created from cancer are not isogenic, or not necessarily isogenic. So if you take an oncogene and try to create the wild-type version of the cell by knocking out the oncogene, that, to me, is sort of violating the very principle of the idea in the first place. And we've had specific examples where an isogenic pair was provided to us, and the wild-type of the mutant cell line had basically deleted BCL2 during the -- Lou Staudt: Sure. William Sellers: -- isogenic process. The second is that the noise is difficult to overcome, and often what we've seen is you get hits that are differential because of the wild-type cell line not the mutant cell line. And this happened in the setting of VHL deficiency as well as in the PI3-kinase setting. So that's been our experience -- Lou Staudt: I was thinking more if you build up from an immortal but not malignant clone and add things to it that you -- William Sellers: Yeah, I'm not saying don't do it. I happen to like the genetic heterogeneity when you have one consistent lesion and 400 other things -- Lou Staudt: Sure. William Sellers: -- because then if something's consistent, you've already controlled for other genetic events but, yeah, anyway it's certainly harder to do these large-scale panels. And, as you said, if you don't have the -- if there are no cell line models for the disease you want to study then, you know, you can't be a stickler on principle from that point of view. Kenna Shaw: Two comments and a question. So the first comment is about the metastatic cases. We are actively collecting triplets. So this is a source of germline, preferably blood, with the primary tumor, and then if the patient had a metastasis, even if the metastasis was exposed to treatment, we will run those as triplets in TCGA. So if you have them, we will take them. We are actively doing those. We have also a couple dozen recurrences for GBM and ovarian, et cetera, so those data are in the public domain. I did want to also just make a comment since most folks in this audience probably do not know about the TCGA collaboration with Novartis and Broad on the CCLE project. Those cell line exomes will be coming into the public domain through CGHub at UCSC, and David Haussler's group, in the next couple of months, followed in early spring, probably by the end of May, with those same lines in RNA-seq. So those -- there was a decision that was made that those data will be made publicly accessible without, you know, having to go through the DAC process, so you should find those in CGHub in 2013. My question is more, of course, wrought with fear about your statement that we should have to do, you know, all tumor types, all subtypes, et cetera, and whether an endeavor like that would need to, in your opinion, be continued to be led by the federal government as the person who, you know, is responsible for 10,000 sample procurement. Your project scares me. William Sellers: I was thinking much larger than that, actually. [laughs] Kenna Shaw: Yeah, exactly, exactly. That's why I'm scared. So I'm curious whether you think it would be possible for a community-driven effort where the data are generated, you know, in individual sites and then data are deposited centrally? William Sellers: Sure, I mean, it's for you guys to decide on the model. I just don't -- I just want to provide the message that we have a ways to go. And I'm with Lou, the genetic experiment has been done. It's like having taken yeast, you know. The reason we don't do forward genetics in mammalian cells is we couldn't sequence the genome. We did forward genetics in yeast because you could just sequence the genome and find the mutations. The forward genetic experiment has been done. And can we deconvolute it? Can we get enough little yeast tumors and find all the patterns that go together? I think it's very exciting. You know, it may take another log drop in sequencing costs, but every time I see Matthew talk, the chart looks like it's going down, but -- [laughs] Male Speaker: So we sometimes envision a future where cancer is almost chronic, treated as a chronic disease. We do the genomics, treat, cancer recurs, do the genomics again, treat again. What's your view on building the therapeutic armamentarium to make that successful? William Sellers: Yeah, I'm not -- I mean, I'm willing to accept it. I don't want to start at that proposition. The reason I don't want to start there is, if I were a cancer patient, I wouldn't want to have to live with my cancer and the fear of it recurring all the time, and the necessity to go back to the doctor, and to take a treatment, and to get another treatment. I don't think that's the greatest existence if we could actually get rid of the cancer. So why can't we get rid of the cancer? Well, I don't know the answer, but the past history says even diseases like testicular cancer, which people forget looked worse than pancreas cancer. Patients died in weeks of testicular cancer. It's cured today. Now, that may -- yeah, again, that could be an exception, or we may find other ways to engender that kind of curative response. Now, why do I think it's sort of dangerous to go to the chronic therapy model? You are not going to cure patients with homeopathic doses of medications that cause no side effects. In fact, one of the downsides of Gleevec is, in fact, it is so well-tolerated people now think we can treat melanoma and lung cancer with drugs that have zero side effects. Cancers, you need to inhibit the targets in cancers very potently. To do that, you're going to have side effects, but we can manage the side effects. We can find schedules. We can ameliorate side effects with other mechanisms. That's how it works for testicular cancer. That's how it's worked in lymphoma for many years. My fear is if we don't try to cure it, we will stop at doses that are subtherapeutic or subcurative. Now, that being said, if we try all that and it doesn't work, I'm fine -- you know, we're treating to a model where we keep patients alive as long as we can. I think we should do everything we can to do that but, yeah, I'd rather aspire for the bigger cure than to retreat and never have the chance to get there. I think there's a question over here. Male Speaker: So, you know, as a clinician, I'll tell you one thing that today we genotype patients very routinely and our patients get biopsies on a regular basis, particularly at the time of disease progression. I do think there's an opportunity here for industry to collaborate with people like us to actually do sequencing studies not upfront but at the time of disease progression. You know, it used to be a difficult thing before to get repeat biopsies, but today, we do this routinely. But I don't think the industry is there yet, at least in our experience, to support these studies, you know, doing sequencing studies at the time of disease progression, you know. William Sellers: Yeah, so support is, you know, can be a word for "pay for" but -- Male Speaker: [inaudible] [laughter] William Sellers: -- I can tell you that we're very interested in this area. Of course, the first application is we have our own trials, patients relapse. We're actually doing sequencing on biopsies when we get them. I'm in the process of setting up what we call a next-gen diagnostics group at Novartis, that, in fact, wants to collaborate off of Novartis trials and on Novartis trials to answer questions just like the one you proposed. There we imagine we're building the facility and the informatics that a clinician who had interesting samples might want them analyzed, we'd be willing to do it. So we are interested. I think, you know, Mark has had a pretty big investment in an infrastructure in Florida around sort of clinical samples. I don't know if it's focused on resistance or not. But I think people are pretty excited about that. Male Speaker: So I just want to say we'll take the last two questions from Dr. Medico [spelled phonetically] and Dr. Getz, and then there are probably more questions yet to come, but we'll hold off on those for the people who are not yet up at the microphones. Male Speaker: Okay, there's an intriguing problem about acquired resistance and the mutations that drive acquired resistance. Do they preexist in a small fraction of the cells or do they emerge de novo? And if so, how can they come so efficiently during treatment? What's your opinion now on this? William Sellers: I don't know. I think it's a great question, and I was talking earlier about -- or asking earlier about what is our actual sensitivity with NGS right now? And I think NGS is still not actually sensitive enough to answer the question. That is, if it's 1 in 10,000 alleles you can determine, that's probably not -- if you have a mutation in 1 in 10,000, that's a very common population of cells in the human body. So that's -- one issue is the technical limitation of NGS is still a problem. The second is, at some point, it becomes a stochastic problem of biopsying, right? So a human has like 109 cells somewhere in the body, you biopsy one place, as we've seen even in primary tumors, you don't know whether or not you've hit the biopsy point that would have the cell that might be mutated or not. In either case, I still think the answer is the same. I don't know how they generate them so efficiently, presumably mismatch repair and ongoing DNA repair issues, but the answer is the same, that is, to create pressure on the cancer cell from more than one point where no one mutation, and hopefully no two mutations even, is sufficient to overcome the therapeutic pressure. So I think we can extrapolate from lymphoma where at least from one point it was four drugs, now I guess it's down to two, that, you know, combinations at least work in part -- work at least in part by creating this pressure on the genetic evolution that the cancer cells, in fact, cannot overcome. Gaddy. Gaddy Getz: Great talk. I want to ask, what's your take about heterogeneity in cancer and what happens if you find a driver druggable mutation that occurs in 20 percent of the cancer cells. Do you act or not act on that event, and what do you think will happen to the cancer? William Sellers: If it were in 20 percent of one patient's cells, yeah, I wouldn't work on that. I think, for me, you know, I'm, yeah -- the important message to me from the New England Journal heterogeneity paper was not the heterogeneity, it was the non-heterogeneous part. Thirty mutations in that paper were completely conserved in every one of the tumor samples they sequenced. VHL was one of the founder mutations. We know that that's a driver in lung cancer. I want to know the earliest set of mutations that are persistently required for all the clones. That's my own bias. Now, what was interesting in that paper was this idea of sort of mutually functionally redundant mutations, I think it was, what, Satb2 and KDM5 something or other. Okay, so that's sort of an interesting clue that maybe, in fact, during the evolution to the metastatic process, that pathway was, in fact, already turned on. So I'd certainly want to look at that in the earlier stages and see if that pathway was already activated and whether this -- the heterogeneity showing a pathway would be then relevant to the earlier stage of the tumor. Gaddy Getz: Thanks. Matthew Meyerson: Thank you very much. [applause]