Okay. So, I'll be talking about some studies that we've done in trying to identify signatures for genotoxicity. This is sort of a prelude to an ongoing study that's currently ongoing and should be finished later this year with Hesse, where we expanded this to a large--much larger dataset.

What we're talking about--I'll be talking about here is about 29 agents, about 13 genotoxic agents, about 16 non-genotoxic agents. And I'll try to introduce the players as we go along. This has also involved Uri Albrook at Pfizer, who we've been working with for some time.

So, as far as the challenges that we have at this point, there's a need to supplement current genotoxicity battery, which is expensive and has a variety of issues, including lack of mechanistic insight.

And there is precedence for mechanism of action doing cellular assays, whether it's toxicity studies or growth inhibition studies or transcriptomics. We and others have done a lot on the NCI60 cell screen, where agents with similar mechanisms of action would cluster.

There's also many gene expression studies. We had a study with a smaller dataset using cDNA arrays that came out in 2005, showing a separation. There's many other publications out there. I'd draw your attention to the Lai Mateo. They filed a year later, which had a much larger set of pharmaceutical agents.

Now, there are some challenges in doing these types of studies. The first is, are there damage-specific signatures? And I'll use the pointer here. So, can we distinguish DNA damage from non-gene--oops. Now, we have the same problem that Nigel had.

And okay. I won't touch it. Okay. So, another issue is time of exposure. If you leave an agent on long enough, you're going to get cell cycle effects, toxicity effects, cell death effects, and the like.

So, what we've been doing is looking at short exposure times. We know from many studies we've done over many years that transcript induction, stress gene induction for direct-acting agents or agents like heat shock, DNA-damaging agents is relatively rapid, good accumulation within several hours. So, that's when we're doing the study at four hours.

Dose selection is a big issue because, if you use the wrong dose, you're not going to see anything, or you may see some kind of super damage effect, which may not be particularly specific. In the case of the Lamb [sp] paper, they just arbitrarily set a dose I believe of about 10 micromolar, which would work for some agents, but not many others.

The next issue is the signaling complexity. Many of these damaging agents will damage multiple response pathways. For example, alkylating agents like MMS will damage both proteins and DNA. And I'll talk about that a bit and, finally, the sensitivity of the platform.

Okay. Uri Albrook put this slide together. And I'm not going to go through it in detail. But, it sort of highlights the standard genetox battery that's out there. The issues with is that there's a fairly high irrelevant positives in the in vitro damage, chromosome damage assay as well as expense, for example, of animal testing and really all these assays which are used, which I don't have time to go into, are really not providing much mechanistic information. They're just giving you a signal.

Okay. So, as far as dose response issues, one thing that we've tried to push as the idea is, when we're looking at signaling pathways, we want to turn them on robustly. And so, we used ionizing radiation as a model agent and not because of any particular special reason that we're focusing on that but because of the--is that we're free of drug uptake and other types of issues that may occur at low doses.

And what we've seen with this agent and many other agents subsequently is that many of these responses, if we look at changes in RNA expression, are roughly proportional to dose until you get to fairly high doses. So, we're trying to push a dose that gives us a robust response.

And so, down here below, I showed some results for p21 and GAB45, two classic stress response genes. And we're looking at the relative RNA levels in myloid cells, which are treated with ionizing radiation. The RNA is assessed four hours later compared to untreated cells, which are the dotted line.

And we see relatively linear dose response or approximately linear. If we go to higher doses, it plateaus off. With agents like alkylating agents, it'll actually drop as you damage the transcriptional machinery.

So, one point that we've tried to do is use an RTPCR dose-ranging preliminary experiment to optimize a dose, particularly of an unknown agent so that we can see what pathways are getting turned on.

This is an example using RTPCR. Hang Han Lee [sp] in my group and Ren Chin Chin have been doing some of this work. And what we show here is different doses of bleomycin with three stress genes. We had a much larger battery before. But, we found in this particular cell line we're using, which is TK6, which is a lymphoblastoid line commonly used in tox studies, what we found was that these three genes, p21, GAB45A, and ATF, ATF3, gave us good signals for both genotoxic and oftentimes non-genotoxic agents.

On the bottom of the slide, we show for the three genes the doses that we chose for the arrays. The genotoxic agents are on the left. The non-genotoxic are on the right. And as you can see, there's sort of some differences in the profiles, even before you go to the arrays.

GAB45A. Um-hmm. And we can comment on that at the end if there's time.

And as far as the platform, there's--we're using the Agilent two-color approach. Before that, 10 years ago or so, we were using a two-color approach with cDNA arrays.

And for these binary studies, where we're comparing control cells grown at the same time versus cells treated in parallel with the damaging agent, we basically have internal--a normalization since both--we're getting both signals on the same spot on the chip. So, it's really a binary data set.

The stringency of hybridization I think for anything much over 50 is comparable to our old cDNA arrays. So, we get a better signal there. And it gives us the ability to detect low-abundance transcripts. I have some data here, which I updated with some more recent data from Agilent that Mike Buttner gave me.

And the bottom line is that we have a linear dynamic range that is quite broad. And we can see relatively low-abundance transcripts that cannot be seen with some of the other platforms, particularly the ones with the shorter-length arrays and we're not using a two-color approach.

This is some of the Agilent results from their literature, basically showing the yellow and green, which are at the bottom. The lower-abundance transcripts get cut off with some of the other arrays. So, for the higher abundance, they'll work, but not for the lower abundance.

Okay. So, this is an example of our array results. And I don't think that the pointer will work, and I'm afraid to touch the--oh, maybe. Okay. It works a bit.

And what we have here is this is a clustering analysis with--where the computer has basically clustered the responses, red being induced, green being repressed, black being no change.

And what we see here is that there's a cluster here indicated in the purple coloring for a set of genes that are turned on robustly by genotoxic agents and not by other agents.

We have an endoplasmic-reticulum-type response, a heat-shock-type response. Also, we have HDAC inhibitors here, which are giving a classic pattern.

It's hard to distinguish here, but in this heat shock group, there is a chromate down here, also has a genotoxic signature. So, when you do clustering looking at overall responses, what you end up doing is you're forcing these gene patterns to get the best fit. And you're going to--and some are going to drop out of the clusters, for example, the example here, which I pointed out.

So, one of the approaches that we've been trying to do is to address some of these smaller sets of genes, which we referred to as sub-clusters. And it just highlights the fact that damaging agents can turn on a variety of different response pathways. And we won't call them regulons yet, but that's probably what they represent.

And so, what we're going to have is a very rich complex signal. And will we be able to use this to give us a genetox signature and signatures for other types of damage.

So, what Daniel Highduke, who used to be in my lab, and now he has his own setup at UCSD, is he's been using a variety of algorithms and has developed superpower magnetic clustering, SPC, which will look for stable sub-clusters. It was actually used originally by physicists, but it's been applied to biology as well by others previously.

And basically, if we look at our array results here, Daniel has then used an unsupervised SPC type of approach to identify a variety of genes, each row being a particular sub-cluster that shows significant association for various sets of genes.

So, this is just to show that we pull out a lot of sub-clusters. Some of them, like the genotox sub-cluster or the HDAC inhibitor, it's easy for the eye to see. Others are not.

And so, the workflow here is we do the array. And then we use these approaches that are listed here to look at a particular sub-cluster. Here, we're looking at the cadmium heat shock chromate sub-cluster. That sub-cluster is then expanded. And the first three rows on that center heat map on the left show the three genes that are showing the strong association.

That can then be put into ingenuity and other responses--response analysis programs to assess the--what pathways might be associated with this sub-cluster.

Just to give an example of two sub-clusters, we have a genotoxic stress sub-cluster on the left and a heat-shock-related sub-cluster on the right. So, on the left, we have the heat shock--excuse me, on the left, we have the genotoxic responses indicated towards the top of the slide by a large group of genes that are responsive, showing increased induction in red. There's some at the bottom that are green.

And then also, we have on the right-most side the three genes from this heat-shock-cadmium-and-chromate pathway, which we really were not expecting. And interestingly, the heat shock response here does not include the typical heat shock proteins but rather includes, as I listed on the slide, genes involved in amino acid metabolism, ER stress, and the like.

So, we're starting to pull out in these sub-clusters then, which are not obvious until you do the experiments.

Okay. In the case of the 65-gene transcript tomic [sp] biomarker, so this was supervised clustering, where we used the approaches that are listed on the slide to identify 65 genes which would distinguish genotox-type agents from the non-genotox agents. So, the genotox agents are on the left side of the panel, heat map, and the non-genotox are on the right.

We've also done--trying to extend this to the proteomic and metabolomic area, I don't have time to go into it, but just show some limited examples using kinase array type of approach that Hang Hong did in the lab. And what we see, particularly, for the lower-most row, is that, if we look at serine 15, phosphorylation of p53, it correlates completely for this set of agents with genotoxicity.

The other two phosphorylation of gamma-H2X and check two also tend to show responsiveness only in the genotox, but there is an outlier for heat shock, which is known to turn on H2AX.

Okay. So, we're fairly confident of this dataset. But, the question is, what happens if we add additional agents to it. So, Uri Albrook had two Pfizer proprietary agents, which were positive in all the classic assays. He sent them to us as blind agents. We ran them using our dose optimization. And the two spots indicated down there by red indicate that these agents fell in with our genetox cluster. And then he subsequently told us that these are in fact DNA-damaging agents.

Using this 65-gene set, Carol Yok [sp] at Health Canada did somewhat similar experiments but using benzpyrine, which requires metabolic activation. She also included cisplatin at relatively low dose, more than 20-fold less than we use. And she did a four- and 24-hour time point.

And so, what she did then is to send her data, which was done using the Agilent platform but using a PCR-type--or an amplification-type approach for probe preparation is--what she found was that--what Daniel found was that the benzpyrine-type responses, both at four and 24 hours, as well as the cisplatin at the later time, where we had a reasonable signal, all fell nicely into our genotox cluster.

And so, what that highlights is that, once we've identified this--a cluster at an early time point, it can also be applied to later times.

Okay. And finally, Uri and many members of the Hesse Group have been using this as a possible approach to address issues of these positives, particularly in the chromosome aberration, chromosomal damage assay. And so this is a typical scheme that is followed when an agent comes up positive in initial in vitro screens. And I won't go through it in detail. It was published in one of the Nature journals as a review a year or two ago. And I can refer you to that.

So, what we wanted to do was address an agent which we know is positive in chromosomal damage assay, and that's caffeine. So, the key questions are--is relevance of positive findings in this assay, and is it a DNA reactive versus a nonreactive mechanism that's giving this signal?

So, we ran caffeine doing optimization to give us the concentration that gave us the most robust response. And what we found here is that caffeine fell in with the non-genotoxic agents, which now are on the left side of this heat map.

We can do PCA-type analysis. And all the genotox agents will cluster. As you can see, caffeine is quite some distance from them.

So, if we were to use this approach with caffeine then, if we go back to the normal--to the development path that Uri's put together for caffeine, we can remove one arm of this flow chart. And what this will allow us to do is sort of conclude that caffeine does not induce a genotoxic stress response characteristic of typical DNA-damaging agents.

And the positive chromosome aberration findings of the test compound are the non-DNA-reactive mechanism. So, this has been as a voluntary submission to FDA. And that was also discussed in the paper I mentioned earlier.

Okay. So, just to summarize then where we stand, we've--I think in some of our major contributions is a dose optimization, particularly with agents with unknown mechanisms of action. The time of exposure, we can do the analysis when the genes are induced robustly but when the cells are still viable, even with high doses.

This can be applied--this 65-gene classifier can then be applied for a longer time period [unintelligible] to be applied for agents requiring metabolic activation, which is going to be slow to some extent by definition.

The cell selection is probably key. TK6 is a spontaneously [unintelligible] human lymphoblastoid line. It has a relatively robust p53 response. And every one of these cell lines that we choose in vitro is going to have some of its own characteristics, which we can discuss.

The Agilent two-color array platform has the sensitivity that we need. And it also allows us to detect relatively small changes because of this internal standardization.

The SPC approach that I've talked about allows us to--identification of common mechanisms for toxicity presumably, since we're seeing similar responses or subsets of responses. And again, this would be an approach that would drive us towards pathways of toxicity.

So, as far as a proof of principle for application of toxicogenomic analysis and safety risk assessment, I've shown you the example with caffeine. And what we're doing now in cooperation with Hesse and members of the Hesse team is to expand this to a much larger dataset beyond 25, 30 agents and getting up into the hundred range.

And then finally, I'd like to just recognize members of my team. I didn't mention Yuen Wang [sp], who's a lab technician and manager in my group. And I think I've given credit to the others. And I think I'm on time.