Founder Stories, Mark Chee, Founder of Illumina & Encodia and Pioneer in Genomics and Proteomics

This is a very old slide. I just realized I was looking at it. This is the better part of 20 years old, but I think it's still relevant. This was made in the early days of Illumina in the context of genomics., and it was to illustrate the problem we were trying to solve, which was that in order to go from raw sequence data to something that you could apply in the clinic, you really had to do an enormous amount of analysis somewhere in between. And the idea was that we would need to analyze genetic variation at the population level in order to make sense of genetic function. So nowadays I think people would more commonly say precision medicine, but I think this still applies. So this has become, to a first approximation, I would say largely a solved problem. Certainly by no means the end of the road, but now we're many, many orders of magnitude beyond where we were there back then in terms of capacity to analyze sequence., and I used to say that metaphorically, you know, it was like trying to move the Himalayas with a teaspoon, and we needed better technology, we needed better tools. So today I want to look at trying to do the same thing with proteomics, and of course, you know, people have been doing this for a very long time now, and with great success, I would say, but there's still a need for better approaches and complementary approaches. So for example, Mass spec is enormously powerful. People have made great progress with, for example, arrays of peptides and proteins. This is from a paper describing one such fairly recent technology using old methods that were used to synthesize DNA arrays to make peptide arrays, that is photolithography, and this shows an array of, I think, 2.1 million peptides.. But this has some of the issues and limitations of array-based readouts, analog readout. Here's another one where people are expressing proteins on chips. So in this case, the DNA essentially can be immobilized on a chip and expressed in situ using cell-free systems. So these illustrate the more traditional methods of this type of analysis.. And the one I just showed you uses in vitro transcription translation, and it's efficient, it's non-toxic being biological. It has its limitations. Neither is, I would say, you know, obviously hugely better than the other. They both have pros and cons. Chemical synthesis lets you make all kinds of things that are much harder to make using biological systems. You can make things in very large quantity But it's a less efficient process. It's hard to make very long peptides, and you have issues with chemical synthesis, basically. So one way to think about this is, you know, is basically how can you apply some of the paradigms in genomics to proteomics? And conceptually, this So this shows, I think, in basically maybe an oversimplified way, some of the interesting evolution of genomics assays. So if you just take this column at the end, the readout, way back things would be read out individually. And then there was a transition to highly parallel readouts, for example, with microarrays, where you could get a lot of data at one time. People would often say that in one experiment they might replicate the total amount of data or exceed the total amount of data that people had collected in decades previously. And now there's a transition— there's been a transition to highly parallel digital sequence-based readouts, and that's completely transformed genomics. In this dimension, assays were done one at a time. They were done individually and they were read individually. And one of the big advances in the array world was that you did one sample prep and you could read out as much information as you had features on a chip for. And same in this digital sequencing world. But it also applies at the level of making the reagents in the cases where assays depend on reagents. They used to be made individually. At some point, they were made individually and pooled. At some point, they were made as large collections. So, I would say this is to a first approximation where we are today. And we thought that this really is a useful sort of paradigm or perspective for trying to do proteomics assays. So, the need really to transform genomic sequences into information at the proteome level and a very simple rationale is that peptides are known to be useful probes for analyzing the proteome, but a challenge really was how to do this at scale. And so the approach was to try and represent the proteome and variations as large collections of peptides. And as I said, you know, to make use of some of the tools of genomics that are so scalable and so powerful. So in order to do that, we made use of technology whereby you could link the peptide that you made, or protein, with its encoding DNA sequence, and you could use this as a substrate for an assay, and I'll show you some different assays, and then read it out in parallel. So that's the basic paradigm. The process is— it's also, you know, very much templated on the genomics world. You can design sequences in silico. You can start with genomic information. You can translate into peptide sequences. You can apply various algorithms to select the sequences that you want to analyze. And then you can synthesize these sequences as DNA templates in a very parallel efficient way. And there are a number of platforms that let you do this nowadays. You can order pools of oligos, and then one can convert these— this pool of DNA molecules into peptide-DNA conjugates in a very efficient way in a single tube type of reaction and perform the assay and read out by sequencing. So this really is the proteomics part of it. The rest of it is things that typically happen more in a DNA lab. So how do we make these conjugates? This is nothing we invented. This is actually quite old technology. A company called Phylos initially commercialized it for drug screening, screening large combinatorial libraries and doing in vitro evolution. And Jack— Josh Ostek and colleagues developed this method of basically adding a reactive molecule, a puromycin adapter, to the end of the RNA. So this represents a DNA molecule that you want to transcribe into RNA. You modify it with this adapter, and then the ribosome does the work of translation. It pauses and links the growing peptide chain to its encoding DNA sequence, or in this case RNA sequence. You can then convert that RNA to DNA using fairly simple techniques. So I mentioned earlier, so this is just some ways of doing these peptide-based assays, and these two are solid array-based. I just described how we use this methodology of linking peptides to DNA sequence, and very similar to this is really phage display. You can express peptides or proteins on surfaces of phage. You can do this in the same parallel way using the same kinds of techniques, and you can then extract the encoding DNA sequence and read it out on a DNA sequencer. So these two come from the analog world of readout, and these are DNA-based digital readouts. So this is a phage display version. This was done, I think, at a similar time that we were developing the other method.— it's really quite analogous and basically illustrates a similar process whereby DNA arrays are used to make a large collection of oligonucleotides that are then cloned into phage and can then be used in screening assays and read out by sequencing. I'll just show you a couple of examples of assays that we developed just to show the versatility of this process. Excuse me. And this is a protease assay where we attach the conjugate to a solid support. So the library— this can be a large library of conjugates. If you expose this collection to a protease that cleaves some sequences, then you can identify what's been cleaved by collecting these molecules, amplifying this DNA tag, and sequencing it. So again, this can be done on, on very large collections at the same time. And this work was, was published back in 2012 in PLOS One, so I I won't go into any detail on it, but as an illustration of this, we tiled our way through the hepatitis C proteome and showed that we could identify protease sites in the proteome quite efficiently. Similarly, a kinase assay where in this case we phosphorylate the conjugates and anything that's been phosphorylated gets captured and sequenced. Here's some data where we— this is an early experiment where we made a 3,000 or so set of conjugates, and we had one set that was not supposed to be kinase substrates, in this case for ABL kinase. We had a set that should have some substrates in it, and we had a set that was permuted around an actual ABL kinase substrate. So you can see, you know, the negative controls don't give much signal and the positive controls give a lot. And it's quite a nice screening tool. I have to say that although this works at the moment very well with purified kinases, we've yet to get really satisfactory signal-to-noise when doing this with biological samples. So, you know, I think this shows more the principle of how this technology can be used in this type of assay. So it doesn't necessarily work perfectly for any sample type, any assay, but it's a way of, you know, screening a lot of sequences for kinase substrate specificity quite quickly in the way of of testing purified kinases. I'm going to switch gears now because of this strong undercurrent of immunology going on in this meeting, and obviously a very interesting area anyway, and talk a little bit about assays, or at least one or two assays that we've developed in the immunological area. So this is a really interesting model, of course, for this type of proteomics in that with the immune system you are dealing with enormous diversity and complexity of sequences. So it's a good test case, and it's an area where multiplex assays are really needed. So we thought, well, you know, it would be interesting to study this interface between MHC and T cell receptor and peptide and see what we can learn using this type of assay. And so we came up with an assay that, for want of better imagination or time, we call MHC-Pepsi. And basically, you know, the need here is to identify peptide epitopes. That's useful, as I think you all know, for a variety of different applications potentially. And I don't think I need to tell anyone here this. Probably most of you are even more familiar with this than I am. But of course, you know, the MHC complex is the most interesting spot in the human genome in case of, explaining or being associated with various diseases and conditions. So I'm going to start with this assay, this MHC assay. It's again conceptually actually quite simple. We have this library of peptide-DNA conjugates and we expose them to MHC and then we capture this using pretty standard techniques. And anything that's captured basically can be recovered and sequenced. That tells us what peptide's bound to MHC. I'll come back to the other assay later. So in order to test this assay, we chose flu as an example. And that's because there's a fair bit of literature, there's a fair bit of information already out there that we could compare to. To see how well we're doing. So we tiled our way through influenza A virus, I think not everything, but a large chunk of it, in 15 amino acid peptides, stepping through in 1 amino acid increments. And we screened against a set of HLA alleles. I think we had 14 total HLAs. And a little over 3,000 peptides. And we didn't get good data for all the HLA alleles. I want to say that I think 5 alleles not shown gave poor signal background, and we think we probably recover those by optimizing the binding conditions, perhaps. But these are the ones that we got good signals for. Basically, I highlight a position here where I think about 1/3 of the data in public databases is from this region. So we chose this as a region to sort of compare to the publicly available information. But you can see these look like— looks like we have a fairly clean baseline in some— you know, some peptides popping up binding to MHC. And it turns out it's by no means perfect, but it does much, much better than in silico prediction methods. So here's our data assay signal as a function of position in this nuclear protein, and you can see that In colors here are known human T-cell epitopes that we're picking a lot of these up. And this is the equivalent result using, I think, the best in silico prediction methods. And this is one plotted against the other in silico versus experimental. You can see we have much better signal-to-noise. So I think down here, again, we found 6 of 7 experimentally, and SiLKoF finds 4 of 7, and so we have a better positive predictive value. This is in tabular form, and this is sorted into— let's see, these colors are concordant hits. So this is the MHC-PEP-seq assay. And this is the literature. So in green is where we've got general agreement between the two. And this color here is in literature only. So this is something where we don't see it. And then in blue are hits that we see that aren't in the literature. And, you know, so that could be an assay. So these are putative hits. We haven't confirmed them. So these could be assay false positives. They could be real T cell epitopes that we don't— that haven't been identified yet in the literature. Or they could be peptides that bind MHC but nevertheless not immunogenic. It's a complex process to elicit a response. So basically, there are a variety of different methods to do this, and I don't think by any means this is necessarily replacing all other methods. They all have pros and cons, but it's certainly a method where you can do a very high-throughput screening of candidate peptide sequences for MHC binding. And so we're working on putting together related types of assays for the rest of the system as well. But we were pretty happy with this initial result. So, you know, we can get quite highly multiplex competitive. I think this is important because when you look at array-based experiments, you see a lot more signal, but I think you have a lot less specificity because of the nature of the assay. It's not a competitive binding assay. And— it basically, you know, correlates well with what we know, and it's highly scalable. I've shown you experiments where we've only done a few thousand, but basically the nature of the assay is such that it can easily scale to greater than 10 to the 6th. And I think the largest experiments we've done, not of this particular assay, are in the probably 100,000 range at this stage and enables a bunch of interesting applications. So another one, of course, is the antibody response, and this seems like a very natural fit. You've got peptide epitopes, and there's an enormous complexity again in the antibody world. Antibodies are widely used as diagnostic— in diagnostics, and there are pros and cons to them. Again, you know, you basically have some responsiveness to the host environment. You've got amplification by the immune system. It's stable, it's available in blood. It's a very standard type of assay that's— used. It's only an indirect measure, can take time to develop, can persist beyond the pathology, so it may not tell you what's happening in real time, and of course it's subject to variation. Again, there are lots of different ways of doing this, various pros and cons. Together with phage display, this approach has the advantage of very high levels of multiplexing.. And it's actually pretty easy to do, and it has this highly quantitative, specific and sensitive digital DNA readout. So here's an overview of the process. It sort of mimics what I've showed you previously. We can do design in silico representing whatever pathogens, allergens, tumor proteins, and so on that we choose. We make a pool. By methods I've told you about, and then we can do binding to serum samples and collect the antibodies, see what's bound, and do it very quantitatively. As a demonstration of this, we looked at lung cancer early detection. Again, it's something other people have done, so we— and there's something known about it in the literature, so we thought this would be a good model to practice on. And there's a case for early detection. You know, it's a deadly cancer, and there's a huge benefit in detecting early, but it's hard to do. So one needs high specificity, can't tolerate a high false positive rate, and there's a lot of molecular heterogeneity. There are various methods for doing this— direct imaging, of course, detecting circulating tumor cells, proteins or antibodies, and there's also circulating tumor DNA. But as far as I know, none of these tests is particularly great right now. And, you know, we didn't think that we could necessarily do better. It was just a good model for testing. So what we did was we chose proteins that that we thought would be enriched for reactivity, and in total we made about 12,000 or so peptides, including some controls. And then we took a fairly modest number of samples initially, some controls that— described as healthy controls, but of course you never know— non-small cell lung cancer cases, small cell lung cancer cases, smokers, nonsmokers, male, female, and these are the age ranges. So it's about 60 samples, give or take. And the first very striking result I'll show you is these are our positive controls. We put in a bunch of pathogen proteins thinking that everybody just about should react to these, and you get these beautiful patterns. So samples in this dimension, so each column represents a sample, and in the Y dimension is position through the protein. So in this case, this HSV-1 protein, you see red is high reactivity and blue is low. You see particular peptides reacting quite strongly across most of the samples, and this —this largely mimics what's known in immune epitope database. So, you know, we recapitulate the known data that is done by other means. And here is an EBV protein. Again, you see these lovely patterns of reactivity that largely recapitulate—not perfectly, but largely recapitulate what's known. So, you know, we only tiled a few. I can't remember, maybe it was half a dozen or so, but certainly possible to do this on a massive scale. You know, we could take every human pathogen and put down very large numbers of peptides and do high-throughput screening on large numbers of samples. And in fact, another group has done essentially this. So this is a paper published in Science June of last year.. And this group did the phage display version, again very similar. Viral proteins converted into DNA sequences, expressed as peptides on phage, binding assay read out by sequencing after amplification. And some of their data showing— I don't know if you can read this, but some collection of viruses. I think they did a large— I want to say there were a couple hundred-plus viruses they did. This is just a subset. These are subjects that were— they did nearly 600 subjects, samples, older, greater than 10 years in age, and younger than 10 years in age. And you can see obviously older people have more reactivity. And they showed some nice results showing that, you know, this was a high-throughput method that seemed to recapitulate very well the kind of data you'd see in low-throughput methods.. So in our lung cancer samples, we did see hits. We saw, for example, this is for p53 protein. The results, I think, actually seem to be remarkably clean. You know, we— with the pathogens, we saw reactivity across just about all the samples. With the cancer antigens, we saw very very little reactivity, and all the reactivity that we saw recapitulated what was known in literature. For example, this is p53 in our non-small cell lung cancer. We see these patches of reactivity, and these match— when we didn't do this a priori beforehand, when we saw these hits, we then went to the literature and asked what other people had seen, and they saw those same epitopes. So SOX2, the same thing.. And this cancer antigen, NY-ESO, the same thing. These asterisks mark epitopes that had already been seen in cancer patients in the literature. Here we see some reactivity with normal. Again, most of our reactivity is with cancer patients, but we do see some with normal, so it's not perfectly specific, if you will. We don't know what this means. It could be that it's basically just some meaningless cross-reactivity in terms of, you know, cancer. Could be that this patient or this person has an undiagnosed cancer, or it could be that it's a pre-malignancy and it's been rejected in some way. So in summary then, this is an assay that's capable of profiling human serum samples and versus the proteome, human proteome, pathogen proteomes, or any arbitrary set of peptides. And we've shown that can detect biologically compelling patterns of reactivity at very high throughput. We scaled this up. I'll just skip through this very quickly. We were able to screen 384 samples very easily in one experiment. And again, we saw the same general thing where— there's some but very little signal in healthy samples, but a lot of reactivity with these 3 proteins known to be cancer-reactive proteins in lung cancer patients. We took a smaller subset of these things that reacted, and we reassayed using a slight variation of this assay. In this case, we use DNA tags that can be extended and read by sequencing or by qPCR. And in short, basically, I won't go through all of this detail, but in short, you know, we had sort of, I think, respectable sensitivity, maybe 25% range, and good, quite good specificity, 98% range. Not perhaps good enough to use as a clinical diagnostic, but nevertheless I think for a first pass with this, not bad, basically. Because I know there's interest in Lyme here, I just showed the same sort of experiment, just finishing up, that we did in Lyme. Again, there was no optimization here. We just— in this case, we just took a collection known in the literature. And in the x dimension is samples.. In the y dimension is targets. This is the qPCR assay with, I think, a 27-plex. So this is not the big screen. This is— but still quite a highly multiplexed assay by protein standards. And negative controls were all negative. Positive samples, about half were positive in our assay, half were negative, but they were also negative when we tested them with two orthogonal assays. So we don't know quite what to make of that. Either all 3 assays weren't quite up to par in sensitivity, or these— or maybe, you know, something happened to these samples. But regardless, the main point is that we can do this in quite high multiplex. So basically, in conclusion, you know, we and others have tried to tackle one of these challenges in proteomics by developing these multi-analyte protein assays that are much more sample and cost efficient and very, very scalable. And being read out by digital means actually, without a whole lot of optimization, can still give you quite decent sensitivity and specificity. Quite a few people worked on the projects I showed you. I want to acknowledge two people in particular. Have done a lot of the design and analysis and carrying out of the lab work, and we were supported by a number of grants. Thanks. Do you think there's any limit to the size of libraries you could explore? Not really. You know, we've thought about just doing combinatorial libraries of everything, That's kind of what I was thinking, take a whole series of creatures and their genomes. Yes. Yeah. At some point, there's some trade-off in terms of benefit versus sequencing cost. These are actually fairly efficient in terms of sequencing cost, but if you made much bigger libraries of everything, you know, even with a low background, you'd still see You'd still be reading a lot of nonspecific reads. Yeah. Just out of curiosity, from— how long does it take between— let's say I want to pick an organism, Staphylococcus or something— how long between design of the peptide and having it working in assay? Right now it probably takes us about 6 weeks if all goes well, but in practice it actually— in theory it can be a lot shorter than that. It takes us about 3 weeks to get the oligo pool, for example, but the actual time to make that can be more or less overnight. So if this were done in a more production way, it would be very, very short. Yeah.