Scientist Stories: Neil Kelleher, Inventor focusing on developing new proteomic tools

Thank you, Kathy. Thanks to you all for coming. I'm the caboose on this train. We have a session next that will round out a great meeting. It's been great so far, in particular the discussions about the roles of proteins in neurodegeneration and the ECM, extracellular matrix of the aging heart. I want to give you small doses of oncology and immunology.— and as I take up my favorite topic, which is mapping human cells and the proteins within them. And, you know, we've seen it throughout the meeting, you all know the story, you can map in different ways, right? You can map in chemical space, compositional space, actual space, and through time. And in proteomics, that's the central theme of my talk today, is how can we do this and get the sweet spot right there for proteomics and get all of those things to actually happen. And it's that complementarity that I wish to focus on today. And prescient as always, the Allen Brain Atlas years ago, now Brain Institute, spatial mapping is on the lips of everyone, right? That's the normal way in which you think about mapping complex systems. In proteomics, the HPA here took 17,000 antibodies and did a similar sort of thing with a 9-figure investment, stood up in Stockholm some decade or so ago, and that's the largest centralized effort, at least as far as spatial mapping, and mainly tissues and some cell types. But of course, the massive investment from the Human Cell Atlas consortium, working closely with NIH Common Fund initiatives and a lot of other initiatives, NCI, for the spatial mapping piece, right? So that's happening and moving quickly. And we've seen it here today as well, using a mature technology, right? Getting 3,000 sequence gene expression reads from individual cells spatially resolved is happening, right? And it's going to be happening on a billion cells soon. You know, the Human Cell Atlas Consortium just released their white paper last week. Proteomics, the technology is still maturing, so that's what I want to talk about, and that's where my brain is a lot of the time. And it's, you know, near and dear to my heart. And the mature technologies here, you know, I just want to focus also on the compositional aspects, the precise mapping of protein composition. And even that fundamental is in need of further attention, and I'll explain more of that presently. So there is a version 1.0 map of the human proteome. In 2014, there were these back-to-back papers in Nature, and— but by virtue of the technologies used in mass spectrometry-based proteomics, the bottom-up proteomics, trypsin digestion of complex mixtures of proteins, there's knowledge gaps. And the peptides, which is the The unit of currency in bottom-up proteomics are surrogates for the whole proteins. They're not the whole proteins. That's part of the knowledge gap. There's protein variability not captured. They have limited correlation with phenotypes, some, but limited. So let's fill that knowledge gap and map precisely the whole proteins directly, and I'll say within cells of a certain type. So that's the world of proteoforms, and we heard it today, or yesterday, excuse me, And so, what is a proteoform? It has a precise definition, at least as of 2013, courtesy of the Consortium for Top-Down Proteomics. So, when you switch alternative splice variants, that affects the base sequences substantially, and then you multiply that— it's the matrix with all the site-specific features. And the combinations of those events create a proteoform. So, a SNP, the right splice isoform, and with combinations of modifications. So that's a proteoform. And importantly, in top-down proteomics, or use of the whole proteoform as the unit of currency that we use, it always maps back to a single gene. This is getting picked up in the field nicely in the past few years. And so you could think about what the next generation of proteomics might be. What's proteomics 2.0? The next major telescope, if you're to use an astronomy analogy. In 2012, you know, focusing on cells as the— as in the hierarchical organization of biological systems like us, a single cell type and the proteiforms within it would define the cellular proteome. So I tried to frame a project back in 2012 and threw out a number of 4,000 cell types. I actually justified in this paper as well as the 250,000 proteiform estimate. So in each cell type, if you were to determine the composition of 250,000 proteoforms, that would be a billion proteoforms across the human body. And you could throw in some fluids, human fluids, into that as well. And so that would be the frame of a cell-based human proteome project. And in the intervening 5 years, there's been a growing number of labs, still in minority in the field, but trying to assert and assign the functions of these proteoforms. Map them, assign the functions, prove that they have value. And there's some of that laced throughout my talk, as you'll see. So, the next big telescope to map the proteoform universe, Proteomics 2.0, could use proteoforms as the unit of currency. And so, I've been describing these kind of things for some time. Let me try to just show you rather than tell you in the general case. So, I'm just going to show you in the key oncogene, K-RAS, and correlate the proteoforms that are mapped with genotypes and phenotypes in cells and tumors, and do that in just a short vignette by way of closing the introductory part of the talk. So, most everyone here knows about oncogenic K-RAS. Here, I'll be focusing on the G13D mutation, which locks it in the GTP-bound form, the on that stimulates effector pathways, downstream signaling goes crazy, and you get proliferation. So driving cancer mutation, truly the case. Now this gene family is the perfect exemplar to describe how things change when you look at the whole protein. So here we have 3 human genes, HNK-RAS. They are highly identical in this blue region, The sequences are almost all identical, so when you digest those to get little peptides, you blur the isoforms together, and much less the alternative splice variants. So exon 4 is alternatively spliced in 4A and 4B, and much less modifications, right? So we want to be precise about these things because they do clearly matter. And when you get to modifications, there's this interesting discussion in the field. About, well, you get a lot of post-translational modifications have been mapped, a lot of sources of variation, and people explode these large numbers of possible proteoforms. Keyword there is possible. So I'll just show you very quickly that cysteine 118 and the C-terminus is where the action is. Which mods are important? Which mods are functional? Whole protein mass spectrometry, this drops out of doing that process. In this system, how many proteoforms exist out of the possible, the answer is 4. And here's 2 of them. If I'm just growing cancer cells in a dish, they're doubling every 24 hours, I have a 57 Dalton difference because I've got the whole protein and I've switched a G to a D. That's a 57 Dalton shift. That's the raw data. There's just 2 proteoforms and you get the full composition of these proteins. And so it's N-terminally acetylated completely. It's all farnesylated and it's C-terminally methylated, so signaling is good to go in the cell line. And then moving to the— if you knock out the wild-type allele, just leaving the oncogenic one, the cells double the same, they don't change much phenotype. Simple thing, you just drop the red proteoform out and just leaving the oncogenic or oncoproteoform behind.. And the real interesting thing happens here. If you knock out the G13D allele— hopefully the red or the green, you can see one of the two will work— G13D, the cells get 2x slower. They become smaller. They become much more adherent. And their proteoforms completely shift. So now the dominant proteoform isn't the normal wild-type allele. It's shifted 29 Daltons, and it's got a cysteine nitrosylation on it. And, you know, mapping this kind of thing precisely— and it's basically 100% stoichiometry. It's just staring you in the face. And to map these things precisely in combination, or to link the language of genetics and genomics with post-translational modifications, You really have to analyze the whole protein. Now, when you go and try to translate this into primary tumors, lots of different cell types, of course. These data are from where you have 80% cancerous cells in these samples. And the action— where's the action? Where's the post-translational modifications? Well, we map our friend the wild-type proteoform that we've seen before and the oncoproteoform. Even though this is a heterozygous and a homozygous, patient, the het, here you only have 10% dosage of the oncoproteiform, so you can actually determine how much penetrance the genotype has all the way to the protein level. But here, the actions on the methylation, we're seeing highly variable methylation in the first 4 patients we've done this on. And so, there's what you get out of compositional analysis. That methylation, there's just some very recent data from the RAS initiative that shows there's important biology. And there's a lack of signaling if you don't have that methylation. So, that's a quick vignette. And now, you know, to do— you'd like to know the composition, but also spatially resolved, right? That's the critical element. And that's what we've been working on. That's the thing that's been enabled by this ADI award. We've been combining flow cytometry with top-down proteomics., and doing this in immune cells for the last couple years. And we can now, for example, in B cells, look at memory versus naive B cells. And I'll give you the cell counts in a moment. But you can actually now create a quantitative comparison, at least for proteomes, proteoforms 30 kilodaltons and lower. You can actually— each of these circles is a proteoform, and you get the, you know, quantitative information. And you can actually do this experiment now on a couple thousand proteoforms, even from 400,000 cells. So there's been a large improvement in processes from 10 million cells down to 400,000 cells, 25x improvement. We have evaluated two other types of proteomics, or mass spectrometry-based mapping. And so in histone modifications, There's a histone panel we have that monitors quantitatively 80 histone marks on all the core histones, and we've dropped that considerably now. Now we're approaching the 1,000-cell mark, and with technical replicates for that. Cell surface protein capture still needs some work, I think, to get down to where they can be relevant, you know, even down to the million-cell level, and we delineated that in a recent publication. So let's look at the proteoform-resolved information that we've accrued. So, for example, in different cell lineages, CD19 B-cell gating versus CD3+ T-cells, here's a known proteoform. So, the defense in one here, this is a known proteoform, but we can quantify it now. I'll just give you one technical replicate from each of the multiple tech reps that were used in this study. So, we know exactly where the proteoform stops and starts, and we can now quantitatively compare that across different cell lineages. Looking, you can do, we have a 5-color flow sorter, so looking at CD8 versus CD4 T cells, we do find examples of where a proteoform is only present in a given cell type. So here, the acetylated form of cyclophilin A, that's PFR number 16924, the only, we only see that in CD4, not CD8 T cells. The red proteiform over there. So, we're accruing all this data in the Allen Proteiform Atlas, and across 8 B-cell subtypes and 3 T-cell pools and subtypes, along with a slew of cell lines earlier in the project. And this is from bone marrow and blood. We've got the proteiform record numbers, these PFR numbers. Yeah, that's a new thing. And we've gotten UniProt to cross-reference those in a durable way, so this will be the world's proteoform information for the foreseeable future. And we've accrued 8,400 high-quality proteoforms to date from primary sources, and we need to get that by next year to 30,000, which was the original goal of the project. And data are immediately available. In the public domain. And so finally you say, well, what's the— not everybody's cup of tea is cataloging proteoforms. I'm trying to get to catalog a billion of them. That's going to take a lifetime perhaps, or I'm hoping much shorter than that. So what's the value of that? Well, we've been trying to do and address, you know, that as well. And in the context of allograft rejection, so liver and kidney rejection, with the Comprehensive Transplant Center at Northwestern. You may know that the infiltration of immune cells into an organ leads to its— after immune activation, leads to its rejection. So we're comparing AR— sorry, AR versus TX, transplant excellence versus acute rejection. There's a third phenotype kind of in the middle of this that you'll see later too, and that, that'll come up. It's called ADNR. So we now are doing this on PBMCs. So we, from a patient, I'm showing you 10 patients in each phenotype in a moment. We isolate the bulk PBMCs. So this is not cell type resolved. But we're doing our thing here, top-down proteomics. I don't have time to get into all of it. But again, it's the 0 to 30-kilodalton cut and using a whole slew of new technologies that we've got now on greased rails. And looking at 10 transplant excellence, or 10AR, here's the volcano plot from this now mixed— these are 10 different people, right, in each phenotype. Genetic background's quite diverse, but we still get signal, at least, you know, log2-type signals of fourfold. And some of the strongest signals are these interesting CXCR2 signals. CXCL4 is our platelet factor 4, and let me drill down on the top signal there. Let me go back. The CXCL4 here, that's our most— that's our strongest signal. And what it is is a new proteoform this time has been mapped. It's this one, 8140 Dalton variant of this platelet factor 4. Its proteoform record number is 18631. And it's a proteolytic event right at that phenylalanine there. And obviously, if you trypsinize this, you would destroy that information. We've mapped 13 unique proteoforms, and if you look at the correlation to AR of this— just this proteoform here, the 8140 Dalton variant, or the proteoform of AR and TX, clearly separating with some statistical significance. Statistical confidence there, and there's our distraction phenotype, ADNR, and this clearly would be merged had you aggregated all the information from the 13 proteoforms together. You're destroying your statistical power to correlate proteoforms to phenotype. And we haven't done bottom-up proteomics on this, you know, reviewers didn't ask, gratefully. And, you know, you just— the statistical power is higher for the proteiform-resolved information. And 5 years ago, I had to frame that as a hypothesis about the technology, that the proteiform information would be more statistically powerful. Now we're in the process of proving it, thanks to support from the ADI program and Paul Allen, Tom, and Kathy. So, you know, we're building now from discovery mode. You know the drill. You have to go from discovery to validation to then deployment, if you make it through the chasm of death there for biomarker discovery. And, you know, we're developing a panel. We're getting some experience here. The chemokines, cytokines here, that's the one I focused on with you. Lots of different proteoforms now in a targeted panel. We can now shrink the amount of sample required, right, and do this from just a few thousand cells. Because we're just looking for these 30 proteoforms and we know everything about them. We've mapped them already. And then you could say, well, okay, which cell types are they in? These are all in PBMCs. I'm surprised we got signal at all out of this. Now, if you go into the right cell subtypes, your signal should go way up, right? And— or you can just look at the proteoforms and refer to the Allen Proteoform Atlas and you'll know which subtypes these are in. So that's the story, and I'm going to close now with the interaction between the Human Cell Atlas, which is an amazing thing that will— that is moving and happening. You can now think about mapping proteoforms in all those cell types. Maybe 4,000 is not the right number. Happy to have a vigorous argument about what the number might be, particularly in the brain. But once you know the composition, You then can develop technologies to move in the sweet spot, and that would be affinity reagents that then would allow you to map in space, say, for diseased or healthy bone marrow. You could then get back to spatial mapping with proteoform-informed reagents. And that's, that's the story and the plan for the future. If you do this, you know, I think it would create the definitive parts list of proteins from the human body. And doing it at $1 a proteiform for proteomics 2.0, that would be a disruption in the process. And I'm just borrowing that from the Genome Project. You know, a dollar a base in 1994 is about when you hit a dollar a base. 12 years later, amazing disruption in the private sector as well as academics. And so can we domesticate the human proteome in the next 10 or 20 years? I think yes, right? We domesticated plants 10,000 years ago. We domesticated the genome the past couple decades. Can we, in the next few decades, domesticate, get our arms and minds around the human proteome? I think so, at least for the composition. This would provide better biomarkers, stimulate proteoform-informed measurement modalities, enhanced drug development, all of the things, regenerative biology, that you would like to take a whole plateau and ratchet it up from 1.0 to 2.0 and all the attendant benefits of doing so. So with that, I think you can see the Allen Foundation is propulsive in this. And if we look back 10 or 20 years ago, and there's some elements that have moved on, you can see the early investment here. So thanks for your time, attention. And look forward to questions.

Any questions? From a transcriptional standpoint, when you look in some of these specific cell types and have the proteoforms, let's just, you know, without the post-translational modification, just the splices, how much does that match the, you know, what might be predicted from splicing RNA sequencing?

Oh, oh, so that's an interesting open question, right? Rolls right off the tongue. But how do I find, especially when there's low sequence reads, I'm asserting a bunch of alternative splice variants, now I want peptides that map and prove with the exon-exon boundary spanning peptides. They are very hard to find. It's sort of an open question about our low-abundance transcripts that might be alternatively spliced, do they actually make it to the protein level? When things are nicely abundant and you see a couple isoforms in high abundance, you can nail those down by either top-down or bottom-up. [Speaker] I'm wondering if you have a plan to integrate your proteoforms with the structural information in the protein databank or something similar. And would that help in some way to diversify further? Yeah, mapping PTMs onto crystal structures or cryo-EM structures, totally exciting thing. It's, it's just when you have so many PTMs as, as comes up from bottom-up proteomics, phosphoacetylomics, you know, mapping those on and then getting assurance about their stoichiometry. This is what's kind of frustrating that conversation now, but it's a great question.

Neil, very nice talk. I was curious about the idea of making antibodies to the different proteoforms. One of the challenges in that would be to get the antigen itself, the individual proteoforms with post-translational modifications. This has been a real problem in phosphorylated proteins and other PTMs. Do you have any ideas about that, how you could purify the actual antigens that you'd need to make the the antibodies.

Yeah, I think in the case of the nitrosylation, you would try to get the small peptide, get the epitope, and the monoclonal that would do it, or polyclonal. Maybe we could go backwards in time to go forward. The bispecifics are something that I think about too, or nanobodies. You know, there's a lot of already investment in that area to get better.

In terms of the libraries, yeah, but in terms— the key is the antigen. Getting the actual post-translationally modified protein.

Oh, you want the full length? Oh, you're talking about proteiform synthesis.

Well, I think if you wanted to get antibodies to them, you probably would need that, right?

You're thinking of just getting the peptides themselves?

Yeah, I was thinking of—

Those are actually quite challenging. Yeah, so proteiform synthesis, as soon as you catalog them and you know that they're important in biology, you would want to synthesize them. And there are phosphorylations, we're doing this with Mike Jewett. You know, you can put in a stop codon and hijack that to put on a phosphoserine site-specifically. Other, other mods will follow. So I think proteiform synthesis will spin up in the next few years. Okay, thank you so much.