1 00:00:06,000 --> 00:00:09,120 Hey, everyone. On this episode of Translating Proteomics, we're 2 00:00:09,120 --> 00:00:11,920 going to shake things up a bit and have a roundtable discussion 3 00:00:11,920 --> 00:00:16,359 focused on what we're calling proteomics hot takes. For the discussion, 4 00:00:16,359 --> 00:00:19,880 Andreas and I will be joined by Nautilus VP of Scientific Engagement, 5 00:00:19,880 --> 00:00:24,360 Sherry Wilcox. In her day job, Sherry shares how the Nautilus Voyager 6 00:00:24,360 --> 00:00:27,990 platform is advancing biology, helps customers bring the platform 7 00:00:27,990 --> 00:00:32,030 into their labs, and leads our proteomics analysis services team. 8 00:00:32,350 --> 00:00:36,990 She is a proteomics rock star with over two decades of experience 9 00:00:37,070 --> 00:00:40,310 developing new proteomics technologies, and she's going to 10 00:00:40,310 --> 00:00:44,230 make a fantastic addition to this episode. Sheri, I'm so glad 11 00:00:44,230 --> 00:00:45,390 that you can join us today. 12 00:00:46,190 --> 00:00:46,670 Thanks. 13 00:00:47,790 --> 00:00:51,270 Practically the discussion will work as follows. One of us will 14 00:00:51,270 --> 00:00:54,910 present a somewhat contentious or provocative opinion or hot take 15 00:00:54,990 --> 00:00:59,180 on on a topic in proteomics. Then we'll round robin and comment 16 00:00:59,180 --> 00:01:02,820 on and debate that hot take. And then later the next person will 17 00:01:02,820 --> 00:01:05,420 share their hot take. And finally, we'll repeat the process 18 00:01:05,420 --> 00:01:09,420 until each of us has shared a hot take. We hope you enjoy the conversation 19 00:01:09,420 --> 00:01:12,020 and please comment on the episode or shoot us a message or 20 00:01:12,020 --> 00:01:14,900 let us know if you agree with us or disagree with us or have a 21 00:01:14,900 --> 00:01:18,300 hot take of your own. We'd love to hear from you. All right, 22 00:01:18,300 --> 00:01:18,940 ready everyone? 23 00:01:20,540 --> 00:01:21,020 Yep. 24 00:01:21,520 --> 00:01:24,240 All right, let's, let's dive in. Sherry, do you want to kick us 25 00:01:24,240 --> 00:01:24,480 off? 26 00:01:25,040 --> 00:01:29,720 Sure. I'll start out. So my hot take is kind of around what's 27 00:01:29,720 --> 00:01:32,440 the biggest problem in proteomics today. And I think it's 28 00:01:32,440 --> 00:01:38,160 less about any of the technologies and more about the lack 29 00:01:38,160 --> 00:01:41,880 of validation that people tend to want to do once they've discovered 30 00:01:41,880 --> 00:01:45,640 something. I think the biggest problem that we have is that it's 31 00:01:45,640 --> 00:01:49,690 easy to get long lists of potential biomarkers for something, 32 00:01:49,690 --> 00:01:52,450 but a lot of people don't want to do the hard work that comes from 33 00:01:52,450 --> 00:01:57,570 a list of candidate biomarkers to then really understand what's 34 00:01:57,570 --> 00:02:01,610 going on. They just want to take a list of markers, come up with 35 00:02:01,610 --> 00:02:04,770 biological plausibility stories, which I honestly think you 36 00:02:04,770 --> 00:02:09,210 can do for any indication and any list of proteins, and then move 37 00:02:09,210 --> 00:02:13,290 on. But there's a lot of context that matters and people really 38 00:02:13,290 --> 00:02:16,760 need to be doing that type of validation and follow up studies 39 00:02:16,760 --> 00:02:18,200 to get more meaningful data. 40 00:02:20,200 --> 00:02:23,440 Yeah. So Sher, what I take from your question here, from your 41 00:02:23,440 --> 00:02:26,360 hot take is that you primarily talking about validation. When it 42 00:02:26,360 --> 00:02:29,040 comes to functional validation, I think we're really 43 00:02:29,040 --> 00:02:32,640 well positioned when we talk about analytical validation. There 44 00:02:32,640 --> 00:02:37,400 is mass spectrometry techniques for Example, statistical 45 00:02:37,400 --> 00:02:41,320 confidence metrics, spectral libraries, targeted assays, lots 46 00:02:41,320 --> 00:02:46,050 of tools that we have now to sort of identify a protein multiple 47 00:02:46,050 --> 00:02:49,650 ways and get orthogonal evidence. I think where we lack is 48 00:02:49,970 --> 00:02:54,010 the functional validation, in particular the scalability of those 49 00:02:54,010 --> 00:03:00,010 methods. While the discovery tools have relatively high throughput, 50 00:03:00,010 --> 00:03:03,410 the follow up studies remain low throughput. They're very expensive, 51 00:03:03,810 --> 00:03:10,130 time consuming. You have hundreds, thousands of possible large 52 00:03:10,850 --> 00:03:14,090 experiments that give you thousands of candidates, but then 53 00:03:14,090 --> 00:03:17,890 you have, you know, very little tools to follow up. So most 54 00:03:17,890 --> 00:03:22,650 of the studies now are research dependent and the researchers 55 00:03:22,650 --> 00:03:27,970 prioritize that based on biological plausibility, prior literature, 56 00:03:28,370 --> 00:03:31,650 their own intuition, and then most importantly about available 57 00:03:31,730 --> 00:03:35,330 funding. So I think that's where the bottleneck is. I think 58 00:03:35,330 --> 00:03:39,860 the other aspect what you're referring to is should that be followed 59 00:03:39,860 --> 00:03:45,620 up with hypothesis free experiment by a hypothesis driven 60 00:03:45,620 --> 00:03:51,900 experiment? I guess what we can do by hypothesis free approaches 61 00:03:51,900 --> 00:03:55,980 today is that we go through another modality of omics. So we 62 00:03:55,980 --> 00:03:58,980 could follow up a transcriptomic experiment with a 63 00:03:58,980 --> 00:04:02,940 proteomics experiment or a metabolomics experiment and sort 64 00:04:02,940 --> 00:04:07,640 of get more idea about the target. Again, I think those tools 65 00:04:07,640 --> 00:04:11,480 are rather available and so you could do a quick follow up with 66 00:04:11,480 --> 00:04:17,000 a hypothesis free tool. Again, if it's hypothesis driven. I'm not 67 00:04:17,000 --> 00:04:20,840 sure we have the means to do that. And so the real bottleneck 68 00:04:20,840 --> 00:04:26,440 I think is the scalability of functional biological platforms that 69 00:04:26,440 --> 00:04:31,640 allow you to do that. I mean in pharma things are done at a larger 70 00:04:31,640 --> 00:04:35,280 scale. So for example, in genomics we have the CRISPR screens 71 00:04:36,420 --> 00:04:42,540 to study functional genomics, we have perturbed sequencing that 72 00:04:42,540 --> 00:04:46,860 does functional transcriptomics. Obviously there's 73 00:04:46,860 --> 00:04:50,700 lots of drug screening candidates that can give you more 74 00:04:50,700 --> 00:04:55,100 information about function. But I think for the individual investigators 75 00:04:55,100 --> 00:04:58,380 or the individual laboratories, this is around a big 76 00:04:58,380 --> 00:05:02,780 challenge to do functional studies. I think you're right, you're 77 00:05:02,780 --> 00:05:05,540 putting a finger on the right problem. I think we have a tools 78 00:05:05,540 --> 00:05:07,980 problem that we need to be able to overcome. 79 00:05:10,860 --> 00:05:15,020 Yeah, I guess so I'm going to respond a little bit, Andreas, to 80 00:05:16,220 --> 00:05:22,540 your thought because I'd like to. Truthfully, I'm not totally sure 81 00:05:23,180 --> 00:05:29,660 what the right workflow is for validating a finding. So for example, 82 00:05:29,660 --> 00:05:33,530 a very common thing that is asked for by reviewers is, oh, okay, 83 00:05:33,530 --> 00:05:37,930 well you discovered this thing by mass spec, go do a western and 84 00:05:37,930 --> 00:05:41,930 prove to me that you're, you know, you say this protein is overexpressed, 85 00:05:42,410 --> 00:05:47,010 show me by western that it is actually overexpressed. And there's 86 00:05:47,010 --> 00:05:51,890 this whole tree of reasons why that western might not succeed. You 87 00:05:51,890 --> 00:05:54,290 know, the antibody might target a different region of the 88 00:05:54,290 --> 00:06:00,320 protein, the antibody might not work. It's a non trivial exercise 89 00:06:00,640 --> 00:06:03,720 and particularly if you're looking for. So getting to your point 90 00:06:03,720 --> 00:06:07,920 about scale, if my discovery experiment says, oh, these hundred 91 00:06:08,000 --> 00:06:15,840 proteins are differentially abundant, golly, that is a ton to 92 00:06:16,000 --> 00:06:20,000 do an orthogonal experiment and it's just a completely different 93 00:06:20,000 --> 00:06:24,860 universe than a follow up from an RNA SEQ experiment by QPCR where 94 00:06:25,260 --> 00:06:30,060 I can order some primers and stay in relatively the same system. 95 00:06:31,420 --> 00:06:38,740 So there's a question for me about even validating the differential 96 00:06:38,740 --> 00:06:43,260 abundance within the same samples. How should we do this? So 97 00:06:43,260 --> 00:06:48,580 I agree with you that people don't do that a lot. I think the 98 00:06:48,580 --> 00:06:53,090 reason is because we don't have a clear playbook that says here's 99 00:06:53,090 --> 00:06:56,610 how we should do it. And then there's the next step of like, okay, 100 00:06:56,610 --> 00:07:00,130 now I believe you that the protein is overexpressed or knocked 101 00:07:00,130 --> 00:07:07,530 down in some way functionally. What does that do? Is the right answer 102 00:07:07,530 --> 00:07:12,290 to do the knockdown or the overexpression. To your point, that 103 00:07:12,290 --> 00:07:15,610 doesn't scale super well. And then you have all the challenges 104 00:07:15,610 --> 00:07:18,450 of it being a system and all these things working in coordinated 105 00:07:18,450 --> 00:07:23,670 action and so you can do those one by one, but you can't collectively 106 00:07:23,670 --> 00:07:28,950 over express a hundred things yet easily without blowing up your 107 00:07:28,950 --> 00:07:35,590 cell system. So I agree with you that this is a gap. I just have 108 00:07:35,590 --> 00:07:37,910 absolutely no idea how to fill the gap. 109 00:07:38,710 --> 00:07:42,150 Yeah, and like you said, if you want to do knockdowns or something 110 00:07:42,150 --> 00:07:44,750 like that, you can't do that for every protein and have a system 111 00:07:44,750 --> 00:07:49,740 that even works. So I, I kind of think it's just each finding needs 112 00:07:49,740 --> 00:07:52,940 to have its own body of evidence and that body of evidence 113 00:07:52,940 --> 00:07:56,500 may not be the same thing for every single finding. And so there's 114 00:07:56,500 --> 00:07:59,660 got to be like a toolbox and we figure out which tools we get 115 00:07:59,660 --> 00:08:03,420 to use in which circumstance. And depending on what we're trying 116 00:08:03,420 --> 00:08:09,220 to answer, the, the weight of how many tools actually have to be 117 00:08:09,220 --> 00:08:13,950 used can vary. Because if it's something that's really unknown and 118 00:08:13,950 --> 00:08:19,150 really novel, then the weight is probably a little bit lower for 119 00:08:19,150 --> 00:08:21,750 initial understanding, but then it builds over time. Whereas 120 00:08:21,750 --> 00:08:25,510 if it's a protein that there are so many tools out there in the 121 00:08:25,510 --> 00:08:29,310 world that you can, you can leverage, then you would expect a 122 00:08:29,310 --> 00:08:33,110 bit of a higher bar. Right. I also made me a little sad, Andres, 123 00:08:33,110 --> 00:08:40,439 in one of your comments. The funding was one of the reasons people 124 00:08:40,439 --> 00:08:44,159 don't do the right science. And it is unfortunate that the reality 125 00:08:44,159 --> 00:08:49,519 is that that's that's part of the equation as well. That's. 126 00:08:49,839 --> 00:08:54,479 Yeah. I mean these follow ups are extremely generous and it's, 127 00:08:54,479 --> 00:09:01,039 it's pretty quick and easy to do the dis, to do the discovery experiment, 128 00:09:01,199 --> 00:09:04,879 to shoot it through the mass spec and do your DIA and, or run 129 00:09:04,879 --> 00:09:11,490 your SOMA scan, et cetera. But then that one day experiment could 130 00:09:11,490 --> 00:09:14,930 lead to months or years of experimental follow up. 131 00:09:14,930 --> 00:09:16,970 Right? Yeah, somebody's dissertation. 132 00:09:18,250 --> 00:09:21,450 So maybe that's part of the, part of the challenge here. Is that 133 00:09:21,850 --> 00:09:25,250 really what we're trying to. So to your point, what does this 134 00:09:25,250 --> 00:09:30,850 toolbox entail? And I think that depends upon the strength of 135 00:09:30,850 --> 00:09:32,690 the claim, the magnitude of. 136 00:09:32,690 --> 00:09:35,030 The claim and the consequence. 137 00:09:35,110 --> 00:09:38,070 And the consequence, Right, Absolutely. So if for example, what 138 00:09:38,070 --> 00:09:42,350 you're saying is, oh, overexpression of this thing leads 139 00:09:42,350 --> 00:09:47,510 to cancer initiation, well then the functional follow up is 140 00:09:47,910 --> 00:09:53,870 a classic set of cancer initiation assays. And if it is, 141 00:09:53,870 --> 00:09:59,390 oh, this affects motility, well then that's a functional phenotypic 142 00:09:59,390 --> 00:10:06,350 screen. And so. But all of those tend to be one by one perturbation 143 00:10:06,590 --> 00:10:10,510 phenotyping. They don't tend to be, hey look, I'm going to move 144 00:10:11,070 --> 00:10:15,910 a hundred things a little bit and see how that broadly affects 145 00:10:15,910 --> 00:10:25,390 the phenotype. So is this a particular problem in all of biology 146 00:10:25,550 --> 00:10:29,080 or is this more of a problem in proteomics? 147 00:10:30,120 --> 00:10:34,680 To your point you made earlier, it may be a problem everywhere, 148 00:10:34,680 --> 00:10:39,920 but I think proteomics maybe has some extra high hurdles compared 149 00:10:39,920 --> 00:10:46,880 to other things like genomics, where you can make the tools that 150 00:10:46,880 --> 00:10:49,360 you need to do the right experiments. Every protein is so 151 00:10:49,360 --> 00:10:55,800 different that it's harder to have just a unique set of go to. 152 00:10:55,800 --> 00:11:00,200 This is what I do every time. But that's just one thought. I don't 153 00:11:00,200 --> 00:11:03,560 know what I'm going to put the question back to you or to Andreas. 154 00:11:04,280 --> 00:11:08,160 So yeah, I mean the biomarkers we get from fluids probably have 155 00:11:08,160 --> 00:11:12,360 a different follow up going to the original tissue, maybe doing 156 00:11:12,360 --> 00:11:16,040 bigger studies. But I think fundamentally it's something that's 157 00:11:16,040 --> 00:11:20,160 related with proteins themselves because they're diverse 158 00:11:20,160 --> 00:11:24,960 functional set within the context of a cell or with the context 159 00:11:24,960 --> 00:11:30,170 of the tissue. And so you better have a good question in the 160 00:11:30,170 --> 00:11:34,010 first place how you want to validate the protein or you're not 161 00:11:34,010 --> 00:11:37,050 even doing the right experiment. So I think it's a little 162 00:11:37,050 --> 00:11:40,650 bit of the complexity of biology that stands in the way as 163 00:11:40,650 --> 00:11:45,370 well, besides the fact that there's no really good high throughput 164 00:11:45,370 --> 00:11:49,010 tools that are available. And then I think I like your comment, 165 00:11:49,490 --> 00:11:54,130 Parag. What do we mean by validation? Right. I mean, you could 166 00:11:54,130 --> 00:11:58,230 probably take the same cell line and I've proliferated it over 167 00:11:58,230 --> 00:12:01,030 several years in different laboratories and then do the exact 168 00:12:01,030 --> 00:12:03,950 same study and come up with a very different answer because you're 169 00:12:03,950 --> 00:12:07,350 talking about a different cell line. Right. And so I think that 170 00:12:07,350 --> 00:12:10,350 is a fundamental problem we have to deal with in biology, that 171 00:12:10,350 --> 00:12:15,470 it is not as reproducible, at least on the protein level, as you 172 00:12:15,470 --> 00:12:17,630 might be able to do that on a genomics level. 173 00:12:18,990 --> 00:12:22,910 Yeah. So I think that question, and if we want to be, if 174 00:12:22,910 --> 00:12:28,830 we want to be able to provide some practical guidance, maybe part 175 00:12:28,830 --> 00:12:33,790 of it is in the validation defining upfront, what aspect of 176 00:12:33,790 --> 00:12:37,390 the experiment are you trying to validate? Are you trying to validate 177 00:12:37,390 --> 00:12:41,430 that in these specific samples that the measurement was correct? 178 00:12:42,390 --> 00:12:46,990 And maybe part of that comes from in experiment metadata, spike 179 00:12:46,990 --> 00:12:50,810 in controls, positive controls, negative controls, et cetera, 180 00:12:50,810 --> 00:12:55,170 that further build confidence that the measurement itself was correct, 181 00:12:57,490 --> 00:13:01,530 which again gets to sort of the QPCR equivalent for transcriptomic 182 00:13:01,530 --> 00:13:05,370 data, which really is just asking, do I believe when I say that 183 00:13:05,370 --> 00:13:09,290 I measured this quantity, changing that it actually did, then 184 00:13:09,290 --> 00:13:14,260 there's a next layer of biological heterogeneity and was 185 00:13:14,260 --> 00:13:23,260 there a bias somehow in the samples? And that is addressed by 186 00:13:23,820 --> 00:13:29,580 replication more biological samples of the same sort, running 187 00:13:29,580 --> 00:13:33,820 cell lines today and running them a year from now. And so there's 188 00:13:33,820 --> 00:13:39,140 sort of a chain of custody for biological sample diversity. And 189 00:13:39,140 --> 00:13:42,060 to what extent is the statistical significance preserved 190 00:13:42,960 --> 00:13:50,200 in a similar but wider biological set? And then to the other 191 00:13:50,200 --> 00:13:55,040 point is the next level beyond that is the functional impact. And 192 00:13:55,040 --> 00:13:58,720 that gets really deep and really hard. But I haven't really 193 00:13:58,720 --> 00:14:03,080 seen a good rubric out there that says, okay, well, to validate 194 00:14:03,080 --> 00:14:07,840 at these three levels, here's the playbook. Plenty of statistical 195 00:14:08,510 --> 00:14:13,470 literature that says, okay, well, I'm doing my significance test 196 00:14:13,470 --> 00:14:18,230 in this way over this support set, but that's different than the 197 00:14:18,230 --> 00:14:22,910 post facto analysis. So I think, unfortunately for this hot 198 00:14:22,910 --> 00:14:32,990 take, we all agree, but this is maybe an opportunity for creating 199 00:14:32,990 --> 00:14:39,700 some structure around what is the standard for expectation of validation, 200 00:14:39,940 --> 00:14:40,740 at what level. 201 00:14:42,420 --> 00:14:47,220 Great. So to summarize, I think we all agree that analytical 202 00:14:47,220 --> 00:14:51,860 validation has some tools that we can employ in order to understand 203 00:14:52,020 --> 00:14:55,540 the measurements that we're actually making. The functional validation, 204 00:14:55,540 --> 00:15:00,580 however, is a bit harder and we don't really have a playbook defined 205 00:15:00,580 --> 00:15:04,830 and don't really have the tools that today that we need in 206 00:15:04,830 --> 00:15:08,310 order to make that happen. But I look forward to this community 207 00:15:08,550 --> 00:15:13,590 being able to step up and come up with Great ideas so that we can 208 00:15:13,590 --> 00:15:17,950 build this playbook and understand how we can take all of 209 00:15:17,950 --> 00:15:21,430 these findings that we have and turn them into the biological 210 00:15:21,430 --> 00:15:23,190 meaning that will advance the science. 211 00:15:25,030 --> 00:15:28,670 Thanks, Sherry, that was a great discussion. Andreas, I'll turn 212 00:15:28,670 --> 00:15:30,070 it over to you. What's your hot take? 213 00:15:31,280 --> 00:15:34,320 Well, we stay a little bit with protein and protein detection 214 00:15:34,320 --> 00:15:38,720 here, sort of building on the story we had around Western blots. 215 00:15:38,720 --> 00:15:42,840 But recently we have seen a number of scientific papers questioned 216 00:15:42,840 --> 00:15:48,000 and even retracted because manipulated fabricated Western blots 217 00:15:49,520 --> 00:15:53,680 and that extended even to a major vendor in the very recent weeks. 218 00:15:54,320 --> 00:15:58,480 In some cases, images were duplicated, altered, assembled in 219 00:15:58,480 --> 00:16:03,680 different ways. In any case, the data that were present misrepresent 220 00:16:03,680 --> 00:16:07,640 the underlying data. At the same time, you actually have a cottage 221 00:16:07,640 --> 00:16:12,120 industry evolving for investigators that call themselves 222 00:16:12,920 --> 00:16:19,680 in scientific integrity experts that use AI tools to go through 223 00:16:19,680 --> 00:16:24,840 existing publications and find disconnects between the actual claims 224 00:16:24,840 --> 00:16:29,120 and the the proof points. But that's not really what I want to 225 00:16:29,120 --> 00:16:32,680 talk about. I think what I want to talk about is the challenge 226 00:16:32,680 --> 00:16:39,480 of actually measuring proteins and the nutritical variability that 227 00:16:39,480 --> 00:16:43,680 is often associated with those tools, such as Western blots. What's 228 00:16:43,680 --> 00:16:47,040 interesting is despite the fact that we have tremendous advances 229 00:16:47,040 --> 00:16:52,320 in biology, many of the core protein detection tools are fundamentally 230 00:16:52,320 --> 00:16:57,910 not different. The last 60 years, I mean 1997, you know, best 231 00:16:57,910 --> 00:17:01,510 and blood essentially was created out of a laboratory at Stanford 232 00:17:01,510 --> 00:17:05,990 and further improved through hairy topians using electroplotic 233 00:17:05,990 --> 00:17:09,150 approach. But basically that's what we use today. So it's a tool 234 00:17:09,550 --> 00:17:12,269 that has been going on for a lot, I've been living on for a long 235 00:17:12,269 --> 00:17:16,470 time in our laboratories and we all know what the best practices 236 00:17:16,470 --> 00:17:20,470 are of using such a tool, which is to spend significant time 237 00:17:20,470 --> 00:17:27,180 and resources to create a process control. Then you test your 238 00:17:27,820 --> 00:17:33,420 antibody specificity, you look for cross reactivity, you test lot 239 00:17:33,420 --> 00:17:38,020 to lot consistency and you make sure that the conditions you're 240 00:17:38,020 --> 00:17:41,060 measuring the protein in are represented in your controls. And 241 00:17:41,060 --> 00:17:44,380 so there's a lot of work, a lot of resources involved to really 242 00:17:44,380 --> 00:17:52,190 validate antibodies. The question for you guys is if antibody 243 00:17:52,190 --> 00:17:57,270 validation remains the main responsibilities of individual laboratories, 244 00:17:57,270 --> 00:18:02,430 is that a model that can sustain ultimately modern biochemical 245 00:18:02,430 --> 00:18:06,630 research potentially? What could a better future look like for 246 00:18:06,630 --> 00:18:09,310 protein detection? Is there new tools on the horizon that could 247 00:18:09,310 --> 00:18:14,990 take the role of a Western blot? I kick it over to Sherry. 248 00:18:15,960 --> 00:18:21,160 Great, thanks. It's a very interesting question. I think that 249 00:18:21,240 --> 00:18:28,040 the burden can't be on every lab to do all of the analytical validation 250 00:18:28,040 --> 00:18:32,240 again and again and again for all of these materials. That's just 251 00:18:32,240 --> 00:18:38,520 too burdensome. I am hopeful and encouraged by some of the initiatives 252 00:18:38,840 --> 00:18:43,640 to try to standardize at least the commercially available antibodies. 253 00:18:44,790 --> 00:18:49,030 You know, there are a lot of antibodies that get resold under 254 00:18:49,510 --> 00:18:52,430 different companies and it's the exact same material, but the 255 00:18:52,430 --> 00:18:56,470 general user doesn't tend to know that. And so I do think that 256 00:18:57,270 --> 00:19:03,110 things like antibodypedia and some other types of databases that 257 00:19:03,110 --> 00:19:08,270 are compiling both the validation data and the essentially 258 00:19:08,270 --> 00:19:12,520 the primary source of the antibody so that people can know, 259 00:19:12,520 --> 00:19:15,960 oh, actually this, this antibody is the same as that antibody 260 00:19:16,520 --> 00:19:18,920 is going to be really important for people going forward. 261 00:19:20,680 --> 00:19:25,800 As far as additional tools, I think again, it comes back to it's 262 00:19:26,760 --> 00:19:33,480 the same tool isn't always the right answer for every single question, 263 00:19:33,480 --> 00:19:38,430 which makes it really, really hard. There were some talks at an 264 00:19:38,510 --> 00:19:42,670 affinity proteomics workshop I went to a couple of years ago where 265 00:19:42,990 --> 00:19:47,190 a lot of there are some initiatives have an exact same panel 266 00:19:47,190 --> 00:19:52,630 of tests that are always done on every antibody in exactly the 267 00:19:52,630 --> 00:19:56,750 same way. And while at face value that sounds great, we can get 268 00:19:56,990 --> 00:20:01,830 much more comparable data when you dig into it, given this application 269 00:20:01,830 --> 00:20:04,990 versus that application, the concentration might not be the same 270 00:20:05,470 --> 00:20:08,970 because the affinity is not the same. Exact way you run those 271 00:20:08,970 --> 00:20:11,850 tests does still have to be nuanced. So I'm not sure I have a 272 00:20:11,850 --> 00:20:18,650 great answer as to what we can do to kind of make it better, aside 273 00:20:18,650 --> 00:20:24,129 from cataloging as much data as we possibly can and making as 274 00:20:24,129 --> 00:20:31,890 much data available to people as possible and trying to be as transparent 275 00:20:31,890 --> 00:20:35,730 as people can for some of these tools and the level of validation 276 00:20:35,730 --> 00:20:38,810 that has been generated thus far. 277 00:20:40,730 --> 00:20:42,050 So I'm, I'm just, I'm. 278 00:20:42,050 --> 00:20:42,410 You've. 279 00:20:42,410 --> 00:20:45,530 You've got me thinking and there are many things spiraling in 280 00:20:45,530 --> 00:20:51,490 my head at the moment. So I think part of why what I'm thinking 281 00:20:51,490 --> 00:20:57,930 about are again a couple different levels of. One is polyclonal 282 00:20:58,570 --> 00:21:04,890 antibodies, sera that are not even purified. There are many form 283 00:21:04,890 --> 00:21:12,550 factors for affinity reagents that are not just recombinant expressed 284 00:21:12,550 --> 00:21:19,030 monoclonals. They're also with different, lots of even monoclonals. 285 00:21:19,350 --> 00:21:24,829 Some proteins might decide to misbehave and aggregate or misfold 286 00:21:24,829 --> 00:21:28,310 in some way. The percentage of active antibody might be different. 287 00:21:29,590 --> 00:21:38,030 And so the challenge is, is I think at two levels, one is on the 288 00:21:38,030 --> 00:21:43,030 part of the vendor to know that they have built a good reagent 289 00:21:43,030 --> 00:21:48,030 in the first place. They have a toolkit of why you should trust 290 00:21:48,030 --> 00:21:53,230 this reagent versus that reagent, that finding things out 291 00:21:53,230 --> 00:21:59,150 like what epitope was used to immunize for this particular reagent 292 00:21:59,150 --> 00:22:03,150 is actually quite hard. It's quite scarce information. And then, 293 00:22:04,110 --> 00:22:09,630 and so somebody might run the same antibody on two different systems, 294 00:22:10,110 --> 00:22:13,470 one of which is the dominantly, the N terminal form of 295 00:22:13,470 --> 00:22:16,590 the protein is present and the other the dominantly, the C terminal 296 00:22:16,590 --> 00:22:20,430 form of the protein is present. And so the antibody itself 297 00:22:20,430 --> 00:22:23,030 might be great. It might just be targeting a completely different 298 00:22:23,030 --> 00:22:26,670 part of the protein. And so you might get disparate results for 299 00:22:26,830 --> 00:22:32,220 biologically reasonable reasons. And so that aspect of the 300 00:22:32,220 --> 00:22:36,900 complexity and having more information available, not just the, 301 00:22:36,900 --> 00:22:43,340 hey, look, I tried it on the protein and I got a band. The burden 302 00:22:43,340 --> 00:22:49,820 of proof, I think is higher than what we have demanded in the 303 00:22:49,820 --> 00:22:57,300 past. To your point, Sherry, about Antibodypedia and the similar 304 00:22:57,300 --> 00:23:01,820 efforts of Protein Atlas has put out the lists of antibodies that 305 00:23:01,820 --> 00:23:05,260 they've used and validated their sets of gold standard antibodies 306 00:23:05,260 --> 00:23:08,580 that have been validated by multiple people. I think these resources 307 00:23:08,580 --> 00:23:11,579 are invaluable. What's interesting is we have these really 308 00:23:11,579 --> 00:23:17,740 great databases for every RNA SEQ experiment deposits their data 309 00:23:17,740 --> 00:23:22,740 in geo. I promise you, everybody who does an immunoassay, 310 00:23:22,740 --> 00:23:29,440 a Western, does not deposit their gel in some central resource. 311 00:23:29,760 --> 00:23:33,040 So we could go back and say, okay, well, this antibody from this 312 00:23:33,040 --> 00:23:37,400 vendor number, from this lot number, here are the 12 gels, here's 313 00:23:37,400 --> 00:23:40,720 the positive control on the gel. And so maybe one of the things 314 00:23:40,720 --> 00:23:44,960 that's necessary going forwards is actually expanding the 315 00:23:44,960 --> 00:23:50,920 scope of these publicly available deposited databases. I 316 00:23:50,920 --> 00:23:57,410 think now many journals are requesting the unedited raw images 317 00:23:57,730 --> 00:24:03,170 of immunoassays in part to help mitigate the fraud concerns. 318 00:24:03,250 --> 00:24:07,690 But beyond the fraud concerns, there are just the practical realities 319 00:24:07,690 --> 00:24:11,170 that that is very valuable data to be able to connect a lot 320 00:24:11,170 --> 00:24:15,770 number and a vendor and an antibody target to a series of data. 321 00:24:15,770 --> 00:24:19,010 And if we started having thousands or hundreds of thousands 322 00:24:19,010 --> 00:24:24,000 of images from a given antibody from a given vendor, I think 323 00:24:24,880 --> 00:24:27,680 that would be an incredible resource that would allow us to go 324 00:24:27,680 --> 00:24:31,040 back and troubleshoot and understand the performance of these 325 00:24:31,040 --> 00:24:35,600 reagents and performance over time and in different contexts. So 326 00:24:37,120 --> 00:24:42,560 to answer your question, I do believe it's a shared responsibility 327 00:24:43,680 --> 00:24:48,640 between the vendors and the users across the scientific community. 328 00:24:49,390 --> 00:24:54,470 It shouldn't be necessary for every lab to fully revalidate every 329 00:24:54,470 --> 00:25:00,350 antibody, but it should be that there's a minimum set of things 330 00:25:00,350 --> 00:25:03,870 that are in every experiment, the positive control, the negative 331 00:25:03,870 --> 00:25:12,190 control, et cetera, before we fully trust an immunoassay output. 332 00:25:12,190 --> 00:25:15,440 And I think there are Many experiments that, and many westerns 333 00:25:15,440 --> 00:25:19,040 that you see that don't necessarily have the titration series 334 00:25:19,040 --> 00:25:24,840 of the positive control, for example, or multiple forms of a protein 335 00:25:24,840 --> 00:25:29,040 or a background lysate negative control. And those reagents 336 00:25:29,040 --> 00:25:33,920 themselves can be hard to find. But if we can agree as a scientific 337 00:25:33,920 --> 00:25:39,240 community to start upping our expectation, we have these things 338 00:25:39,240 --> 00:25:43,680 like the Miami standard for MRNA expression, the MyAPI standard 339 00:25:43,680 --> 00:25:46,960 for protein expression, and these are mostly applied to high 340 00:25:46,960 --> 00:25:49,720 throughput discovery technologies. Maybe there's a new 341 00:25:49,720 --> 00:25:55,880 standard that is wanting to be made around minimum information about 342 00:25:56,120 --> 00:25:57,960 an immunoassay experiment. 343 00:25:58,360 --> 00:26:01,320 Right. And the International Working Group on Antibody Validation 344 00:26:01,320 --> 00:26:06,840 several years ago put out essentially the five pillars of validation 345 00:26:06,840 --> 00:26:11,720 that all antibodies should at least try to achieve. And even if 346 00:26:11,720 --> 00:26:16,020 it's hard to get data for all of those pillars because some aren't 347 00:26:16,020 --> 00:26:20,260 relevant to a particular type of target that an antibody might 348 00:26:20,260 --> 00:26:26,140 be targeting, having again a body of evidence that is as much 349 00:26:26,140 --> 00:26:33,340 as one can do, I think is useful to the community. Andres, 350 00:26:33,340 --> 00:26:38,340 you also asked about alternate ways to do validation and I do think 351 00:26:38,910 --> 00:26:44,030 tech, you know, some of the, the protein arrays are certainly 352 00:26:44,750 --> 00:26:48,670 another way to, to get some of this validation beyond a western 353 00:26:48,670 --> 00:26:54,430 blot. Additionally, the FIP SEQ phage immunoprecipitation sequencing 354 00:26:54,430 --> 00:26:58,990 is a tool that can be used to again, look at specificity. And then, 355 00:26:58,990 --> 00:27:03,550 and then those pillars that are in the International Working 356 00:27:03,550 --> 00:27:08,470 Group on Antibody Validation paper are really good, good ways 357 00:27:08,470 --> 00:27:11,230 to look at this validation beyond just a western blot. 358 00:27:12,270 --> 00:27:15,590 Yeah. So here you're both making an argument that, you know, 359 00:27:15,590 --> 00:27:18,190 there needs to be more standardization, there needs to be 360 00:27:18,350 --> 00:27:24,190 more sharing of resources, some way of really documenting, you 361 00:27:24,190 --> 00:27:26,990 know, the experiments and validation experiments are being 362 00:27:26,990 --> 00:27:31,590 done. And I think I agree because most of the antibody market 363 00:27:31,590 --> 00:27:36,070 is basically, you know, wild, wild west. You produce an image that 364 00:27:36,070 --> 00:27:38,510 shows that it migrates somewhere where you would expect 365 00:27:38,510 --> 00:27:41,650 it, and then you can sell it, put a label on and sell. And then 366 00:27:41,650 --> 00:27:44,930 there's obviously, you know, as you mentioned, Sherry, many, many 367 00:27:44,930 --> 00:27:47,250 things going on in the background in terms of, you know, 368 00:27:47,250 --> 00:27:51,170 good antibodies, you know, available commercially through many 369 00:27:51,170 --> 00:27:55,890 channels. I think it is, you know, probably a good idea to look 370 00:27:55,890 --> 00:27:59,010 at what other analytical technologies have done or techniques 371 00:27:59,010 --> 00:28:02,050 have, you know, sort of done. You can buy, I mean, you can buy 372 00:28:02,050 --> 00:28:05,730 standards for any analytical technique, whether it's a chromatography 373 00:28:05,730 --> 00:28:08,340 system, a mass spectrometry, spectrometry system, you know, a 374 00:28:08,340 --> 00:28:11,900 spectroscopy system. They all have standards. You buy and you basically 375 00:28:11,900 --> 00:28:15,980 validate the instrument before every single experiment. You do through 376 00:28:16,140 --> 00:28:20,260 a calibration procedure. And I think that it's probably worthwhile 377 00:28:20,260 --> 00:28:26,700 to think through why not making the actual, you know, epitope 378 00:28:26,700 --> 00:28:29,940 or the protein itself that, you know, you raised the antibet 379 00:28:29,940 --> 00:28:34,510 against it, available, maybe even the media, and then have the 380 00:28:34,510 --> 00:28:37,670 community have access to the tools and say, yes, I can reproduce 381 00:28:37,670 --> 00:28:41,670 it consistently. My lab, particularly if the lab is very much 382 00:28:41,670 --> 00:28:44,390 focused on that protein because they're doing the validation 383 00:28:44,390 --> 00:28:47,270 studies. We talked in the first question, in the first part 384 00:28:47,270 --> 00:28:51,630 of the hot takes here. I think it makes a lot of sense to really 385 00:28:51,630 --> 00:28:56,030 put the pressure now on the providers, but also on laboratories 386 00:28:56,030 --> 00:28:59,350 in the academic space that produce antibodies to come together 387 00:28:59,350 --> 00:29:04,250 and really form that resource that we need that ultimately helps 388 00:29:04,250 --> 00:29:11,930 us avoid these very frustrating exercises to validate 389 00:29:11,930 --> 00:29:16,890 the data in retrospect, retrace some of these papers. In 390 00:29:16,890 --> 00:29:21,330 some cases, careers have been damaged or delayed. And so it's all 391 00:29:21,490 --> 00:29:25,970 pretty much unnecessary if we work together. I haven't heard any, 392 00:29:26,320 --> 00:29:30,200 you know, you guys talk about a new technology or other ways to 393 00:29:30,200 --> 00:29:33,600 maybe detect proteins. I would put forward that, you know, there 394 00:29:33,600 --> 00:29:39,120 is tools like modern affinity reagents that use two antibodies. 395 00:29:39,360 --> 00:29:42,480 There's obviously mass spectrometry is more compositional, 396 00:29:43,360 --> 00:29:48,680 you know, focused analysis, but maybe there's also things we 397 00:29:48,680 --> 00:29:53,330 can do to combine methods. What do you think? And I've got, 398 00:29:53,330 --> 00:29:55,290 you know, handing this off to you, Parag. 399 00:29:55,690 --> 00:29:59,210 Yeah, well, I, I think about, there are these alternative platforms, 400 00:29:59,370 --> 00:30:05,610 things like the Licor and the Wes and Jess and similar non ex. 401 00:30:05,850 --> 00:30:10,010 They're not exactly Westerns, but they are immunoassays. I think 402 00:30:10,090 --> 00:30:13,810 isoelectric focusing. Immunoisoelectric focusing has been 403 00:30:13,810 --> 00:30:18,490 a very interesting technique. Ironically, it dates, it sort of 404 00:30:18,490 --> 00:30:24,370 reminds you of the earlier, the very, very old immuno electrophoresis 405 00:30:24,370 --> 00:30:30,530 methods that you mentioned earlier, from the late 70s. And so 406 00:30:30,530 --> 00:30:34,970 I think there are, there certainly could be newer methods 407 00:30:34,970 --> 00:30:39,290 that are brought in, but all of those methods essentially still 408 00:30:39,290 --> 00:30:44,080 depend upon your confidence that your antibody is binding to 409 00:30:44,080 --> 00:30:49,560 your protein in the state that that protein is. And for Westerns, 410 00:30:49,560 --> 00:30:52,760 things are a little bit easier in many cases because the proteins 411 00:30:52,760 --> 00:30:59,080 are denatured unless you're running a native Western. But for 412 00:30:59,080 --> 00:31:05,600 ELISA reagents or reagents that are being used in ihc, where 413 00:31:06,320 --> 00:31:10,420 the antibody is targeting a structural motif, you can imagine 414 00:31:10,420 --> 00:31:13,980 that as the protein breathes or gets modified and changes shape, 415 00:31:14,380 --> 00:31:19,340 that the antibody might bind in a different way. And so I think 416 00:31:19,340 --> 00:31:23,500 the part of what needs to be answered is, all right, what do we 417 00:31:23,500 --> 00:31:31,180 think we are measuring with this particular probe and is it purely 418 00:31:31,260 --> 00:31:34,780 protein abundance or is it something more complicated? 419 00:31:36,200 --> 00:31:38,800 And that gets back to the transparency of the epitope, like 420 00:31:38,800 --> 00:31:40,520 you said earlier. Yeah, yeah. 421 00:31:41,960 --> 00:31:46,520 So maybe the takeaway here from the conversation is we definitely 422 00:31:46,520 --> 00:31:50,080 want to make sure that the community stays on the current path 423 00:31:50,080 --> 00:31:56,200 and really invests in common tools, common platforms that actually 424 00:31:56,360 --> 00:32:00,320 allow us to validate the antibody that we use. And I think 425 00:32:00,320 --> 00:32:04,810 what you're saying, Parag, is that we also need to keep the context 426 00:32:04,810 --> 00:32:08,770 in mind of how we validate these antibodies because sometimes 427 00:32:08,850 --> 00:32:15,170 it is not as simple as running Western blot. It might involve additional 428 00:32:15,570 --> 00:32:22,610 effort spent. And so I think that probably provides a good path 429 00:32:22,690 --> 00:32:28,610 forward for the current conversation around antibodies in 430 00:32:28,610 --> 00:32:32,460 the context of the fraud that, that I started introducing the topic 431 00:32:32,460 --> 00:32:36,580 with. And hopefully we see some, we see some rallying around 432 00:32:36,580 --> 00:32:37,780 this particular cause. 433 00:32:39,460 --> 00:32:39,820 Great. 434 00:32:39,820 --> 00:32:42,740 Thank you, Andreas, for leading us through that hot take. 435 00:32:43,300 --> 00:32:47,300 Just as a reminder to the viewers, listeners, we'd love to 436 00:32:47,300 --> 00:32:52,500 hear your thoughts on these topics, send us comments, argue with 437 00:32:52,500 --> 00:32:58,510 us. These are important topics and we'd love to hear your feedback. 438 00:32:59,790 --> 00:33:02,990 So I'm going to turn it over to myself now for our last hot take. 439 00:33:03,550 --> 00:33:08,670 And I think I feel like our current hot takes, we can be even 440 00:33:08,750 --> 00:33:15,710 more dramatic. And so I'm going to go for the hottest take 441 00:33:15,710 --> 00:33:23,350 that I can possibly go, which is that the recent wave of AI virtual 442 00:33:23,350 --> 00:33:30,470 cell models are grossly overhyped and unlikely to be as performant 443 00:33:30,550 --> 00:33:36,350 as they are believed to be. And so I'll explain what I mean by 444 00:33:36,350 --> 00:33:40,710 this and why. It has certainly been the case that over the last 445 00:33:40,710 --> 00:33:44,510 two years there have been a number of both frontier models and 446 00:33:44,510 --> 00:33:48,470 foundation models that have come out that have been trained over 447 00:33:48,470 --> 00:33:53,340 large corpuses of data, typically large amounts of transcriptomic 448 00:33:53,340 --> 00:33:58,940 data, and then in some cases cell imaging data on top of it to 449 00:33:59,340 --> 00:34:06,980 try to get something related to phenotype. The stated claims of 450 00:34:06,980 --> 00:34:12,460 these models or genomic data, DNA models, is that they can perfectly 451 00:34:12,460 --> 00:34:17,380 predict phenotypes that you can put in some mutation, and out 452 00:34:17,380 --> 00:34:22,120 the other end comes a perfect prediction of what the cell behavior 453 00:34:22,120 --> 00:34:30,840 or the human physiologic behavior is. Now, I the the effort 454 00:34:30,840 --> 00:34:36,440 of building one of these kinds of models is gargantuan. And absolutely 455 00:34:36,440 --> 00:34:41,640 I, I'm not in any way diminishing the effort that the researchers 456 00:34:41,640 --> 00:34:45,310 went through to build these sorts of models and even just the 457 00:34:45,310 --> 00:34:50,190 data curation, organization, model structure, the actual mechanics 458 00:34:50,190 --> 00:34:54,910 of training are huge and complicated. But the question at 459 00:34:54,910 --> 00:34:59,430 the end of the day is how much of biology is being captured by these 460 00:34:59,430 --> 00:35:07,070 models, particularly disease biology. The other question is, when 461 00:35:07,070 --> 00:35:10,150 we think about some of these models, they're there to say, oh, 462 00:35:10,150 --> 00:35:17,320 okay, well I can take it a DNA measurement and I can predict protein 463 00:35:17,320 --> 00:35:23,240 abundance from my DNA measurement. I'm going to come out 464 00:35:23,240 --> 00:35:26,240 and declare that that is crazy town. 465 00:35:27,200 --> 00:35:31,760 Yeah, so I think you're right in the aspect that, you know, AI 466 00:35:31,760 --> 00:35:34,600 models are typically overhyped. And it's not just in the 467 00:35:34,600 --> 00:35:39,080 biology space. I mean, it's, it's across the entire AI space. 468 00:35:39,320 --> 00:35:45,080 But I would submit to you that I think it's a really important first 469 00:35:45,080 --> 00:35:52,840 step for us in biology to go beyond the typical phenotypical observations 470 00:35:54,520 --> 00:36:00,600 and low throughput biology I would call it. I think what is probably 471 00:36:01,320 --> 00:36:05,520 more honestly claimable in that conversation is that we actually 472 00:36:05,520 --> 00:36:09,600 have a fundamentally new way how to learn biology rather than 473 00:36:10,720 --> 00:36:15,360 try to actually predict biology. At this point, the arguments 474 00:36:15,360 --> 00:36:20,719 I would put forward is the following. The AI models can learn 475 00:36:20,719 --> 00:36:26,280 beyond a single cell. And so you can take millions of cells, not 476 00:36:26,280 --> 00:36:29,360 just one experiment, but millions of cells and learn from 477 00:36:29,360 --> 00:36:39,290 that. The other argument is that the models can capture biological 478 00:36:39,290 --> 00:36:43,690 context rather than the individual gene that we typically 479 00:36:44,490 --> 00:36:49,210 study in particular cell models, for example. And so because 480 00:36:49,210 --> 00:36:52,610 the gene doesn't have a fixed biological role, the same gene can 481 00:36:52,610 --> 00:36:58,240 have a different context in different cell types. AI models are 482 00:36:58,240 --> 00:37:03,360 much better tools that actually capture the context. And 483 00:37:03,360 --> 00:37:08,280 transformer models are built to actually extend those functions 484 00:37:08,280 --> 00:37:12,440 amongst different, for example, cell types. I would also 485 00:37:12,520 --> 00:37:17,080 say, for example, that AI models are much better in integrating 486 00:37:17,080 --> 00:37:23,830 multi omics aspect of it. We struggle a lot to sort of even combine 487 00:37:23,830 --> 00:37:28,350 these multi omics data set on a, you know, on a single protein 488 00:37:28,350 --> 00:37:32,630 or single gene level. And AI machines or computers in general, 489 00:37:32,870 --> 00:37:36,470 you know, because of their scalability are much better in sort 490 00:37:36,470 --> 00:37:40,190 of capturing the modality of many modalities. And so I feel that 491 00:37:40,190 --> 00:37:44,310 future biology really requires the integration of these transcriptomics, 492 00:37:44,310 --> 00:37:48,230 epigenomics, proteomics, spatial information metabolomics 493 00:37:48,230 --> 00:37:52,230 data. And the transformer architecture is inherently flexible 494 00:37:52,230 --> 00:37:57,270 enough to incorporate these multiple data types. I think what 495 00:37:57,270 --> 00:38:01,870 you will end up seeing, or hopefully what we end up seeing, 496 00:38:02,270 --> 00:38:08,870 that these models create transferable representations of cells 497 00:38:08,870 --> 00:38:13,390 rather than the prediction of a particular cell. And that ultimately 498 00:38:13,390 --> 00:38:22,280 those cell models are much, much more practical in now testing 499 00:38:22,600 --> 00:38:26,720 biological hypotheses because they can access a scale that ultimately 500 00:38:26,720 --> 00:38:31,000 we cannot assume. And so you might be able to take this rather 501 00:38:31,480 --> 00:38:35,480 virtual cell and ask many, many questions and say, hey, which 502 00:38:35,480 --> 00:38:39,240 one of them is even a good question to ask in the first place 503 00:38:39,320 --> 00:38:42,740 based on the information we have gotten from many different multi 504 00:38:42,740 --> 00:38:48,980 omics experiments. And so I think the, the cell models are not 505 00:38:48,980 --> 00:38:53,100 there really to predict biology, but essentially guide us. 506 00:38:53,260 --> 00:38:56,780 How do we actually study biology in a much, you know, on a 507 00:38:56,780 --> 00:39:01,900 much, on a much bigger scale. So I think that's my current thinking 508 00:39:01,900 --> 00:39:05,220 about it. AI develops so quickly these days. So, you know, 509 00:39:05,220 --> 00:39:07,860 we may have another podcast where we have changed our whole minds, 510 00:39:07,860 --> 00:39:08,860 but. Sherry, what do you think? 511 00:39:10,200 --> 00:39:15,880 I have a few thoughts. One is as soon as someone uses the word 512 00:39:16,120 --> 00:39:21,840 perfect, as a scientist, I have a pretty visceral response to 513 00:39:21,840 --> 00:39:25,000 that that I'm going to not believe anything else. That said, 514 00:39:26,840 --> 00:39:33,520 biology is way more complex. I think there's an element of hubris 515 00:39:33,520 --> 00:39:36,520 for us to think we're going to develop something that is going to 516 00:39:36,520 --> 00:39:41,260 be able to give a perfect prediction of anything. And so it's 517 00:39:41,260 --> 00:39:46,100 kind of, it's. It's not. My concern is that it's not unlike what 518 00:39:46,100 --> 00:39:48,820 my hot take was to a certain extent that people want to get a 519 00:39:48,820 --> 00:39:54,020 list and not think and just go with it. I'm concerned that that's 520 00:39:54,020 --> 00:39:57,180 what can happen here too is that people are going to believe 521 00:39:57,820 --> 00:40:01,660 this is perfect. They're not going to ask hard questions of it 522 00:40:01,660 --> 00:40:05,770 because AI generated it. And then we're going to get some flawed 523 00:40:05,770 --> 00:40:12,650 understandings, particularly knowing that the model is going to 524 00:40:12,650 --> 00:40:15,770 only be as good as the input that went into it. Of course, and 525 00:40:15,770 --> 00:40:19,210 we know that in scientific studies. And I'm not this is not 526 00:40:19,210 --> 00:40:25,330 necessarily just about a cell model or in particular, but diversity 527 00:40:26,130 --> 00:40:32,570 of participation in generating the data to develop any of these 528 00:40:32,570 --> 00:40:36,650 kinds of any models has not been great historically. And it's 529 00:40:36,650 --> 00:40:40,050 something that a lot of people are certainly working on. And yet 530 00:40:40,050 --> 00:40:44,690 we know that the data that is fed to date is probably going to 531 00:40:44,690 --> 00:40:50,890 have some biases. And so for us to just take anything as at face 532 00:40:50,890 --> 00:40:56,210 value as this is, this is the absolute and this is the correct 533 00:40:56,290 --> 00:41:00,640 answer without asking those other questions of how might something 534 00:41:00,640 --> 00:41:05,680 else that is different in one context than it was in the context 535 00:41:05,680 --> 00:41:09,680 in which this was built. If we can't ask that question, then we're 536 00:41:09,680 --> 00:41:12,680 going to miss a lot of nuance and we're going to do a disservice 537 00:41:12,680 --> 00:41:17,720 to the greater community. I think AI has, has a lot of good uses. 538 00:41:17,720 --> 00:41:21,480 I agree that it is so overhyped. Anytime people want to 539 00:41:21,480 --> 00:41:26,080 use AI, even if they just mean normal machine learning, you know, 540 00:41:26,080 --> 00:41:30,880 they, it's just, it's great to plug in the word AI, but we do have 541 00:41:30,880 --> 00:41:35,160 to be careful that we're really, really being cautious of 542 00:41:35,160 --> 00:41:40,080 knowing exactly what went into the model and, and what the implications 543 00:41:40,160 --> 00:41:44,080 of that might be on, on the decisions that we make based on, 544 00:41:44,080 --> 00:41:47,280 on what we're, we're trying to study. Yeah. 545 00:41:47,520 --> 00:41:51,320 I'll point out a few specific things that lead to my declaration 546 00:41:51,320 --> 00:41:59,180 of overhyped. One is the nature of feedback loops, that biological 547 00:41:59,180 --> 00:42:06,380 regulation is all about these sense and response and across multiple 548 00:42:06,460 --> 00:42:10,900 length and time scales. So things that occur at the second time 549 00:42:10,900 --> 00:42:16,660 scale, things that occur at the decade timescale and many of 550 00:42:16,660 --> 00:42:21,280 the models are feed forward models and so they don't have an 551 00:42:21,280 --> 00:42:26,160 explicit notion of a feedback loop. Now there's a thought that, 552 00:42:26,160 --> 00:42:30,240 okay, well if you upscale enough, maybe you abstract away the 553 00:42:30,240 --> 00:42:36,800 feedback loops. But I would argue that the genome to phenotype 554 00:42:36,800 --> 00:42:42,520 is such a big jump and it's a jump over so many timescales that 555 00:42:42,920 --> 00:42:48,210 having some concept of being able to play this forward, see what 556 00:42:48,210 --> 00:42:51,890 happens at multiple timescales and then coarse grain from there 557 00:42:52,690 --> 00:42:57,210 is an architectural thing that is currently quite hard to build 558 00:42:57,210 --> 00:43:01,610 into most AI models and hasn't really been anticipated. I'm not 559 00:43:01,610 --> 00:43:04,370 saying it can't be. I'm just saying that the current generation 560 00:43:04,370 --> 00:43:08,850 of cell models haven't really accounted for space and time very 561 00:43:08,850 --> 00:43:13,010 well. And so we've talked about that on other podcast episodes 562 00:43:13,010 --> 00:43:16,370 about just in general how important space and time are to modeling 563 00:43:16,370 --> 00:43:20,950 biolog. I think the other thing that in our conversation with 564 00:43:20,950 --> 00:43:25,710 Olga Witek, we were talking about causal models as a branch of 565 00:43:25,710 --> 00:43:31,310 AI models. And that notion of I poke here and this thing happens 566 00:43:31,710 --> 00:43:37,710 is again, it's a directional, it's a magnitude component that isn't 567 00:43:37,710 --> 00:43:42,710 necessarily well captured by simple associations or even complex 568 00:43:42,710 --> 00:43:48,980 associations. And so that next generation of AI tools that explicitly 569 00:43:48,980 --> 00:43:56,700 have causality baked in, or has physics baked in, or has multiscale 570 00:43:56,940 --> 00:44:01,780 multi omics baked in at the appropriate scale, I think that's 571 00:44:01,780 --> 00:44:06,300 the gap. So when I see these models making predictions I like, 572 00:44:06,300 --> 00:44:10,020 okay, it is very easy to make lots of predictions, particularly 573 00:44:10,020 --> 00:44:13,020 when you have large, huge parameter systems. How sensitive 574 00:44:13,020 --> 00:44:16,150 are they? And then when we talk about disease processes, my 575 00:44:16,150 --> 00:44:19,990 sort of third argument is that disease processes typically are rare 576 00:44:19,990 --> 00:44:25,550 events. And so they're at the edges of distributions. And when 577 00:44:25,550 --> 00:44:29,110 you're trying to estimate any parameter, you need a ton of data. 578 00:44:29,430 --> 00:44:32,230 And when you're trying to estimate parameters that are at the 579 00:44:32,230 --> 00:44:35,270 edge of distributions, you definitionally don't have as much 580 00:44:35,270 --> 00:44:39,150 data because they're rare events. And so it's actually quite 581 00:44:39,150 --> 00:44:43,880 hard to Figure out how to learn these extremely rare low frequency 582 00:44:43,880 --> 00:44:48,680 events, and of course working on transfer learning methods, etc. 583 00:44:48,680 --> 00:44:52,840 Is meant to address these. But I think these three fundamental challenges 584 00:44:52,840 --> 00:44:56,800 are why I say that the current models, though very impressive in 585 00:44:56,800 --> 00:45:03,080 scale, are still not quite as performant as they may be hyped in 586 00:45:03,080 --> 00:45:03,920 the media to be. 587 00:45:05,120 --> 00:45:09,390 Yeah, I would just again state, as you say, we are at the 588 00:45:09,390 --> 00:45:17,390 very beginning of learning how to construct these models to more 589 00:45:18,350 --> 00:45:23,630 reflect actual the phenotypes we can observe over space and time. 590 00:45:24,510 --> 00:45:27,230 I'd share with you the concern that most of these models are built 591 00:45:27,230 --> 00:45:31,150 on single cells and single cells are not tissue. Single cells 592 00:45:31,390 --> 00:45:38,870 put together the pipes would create a different response. So it 593 00:45:38,870 --> 00:45:44,310 goes back to what you just stated, that it's very hard to predict 594 00:45:44,630 --> 00:45:50,510 forward. However, I think the fact that we have a number of people 595 00:45:50,510 --> 00:45:54,350 working on this using computational tools maybe also brings 596 00:45:54,350 --> 00:45:58,950 it back to what Jerry said. We have to have a fair amount of hubris 597 00:46:00,150 --> 00:46:04,230 in mind when we do biological studies because we simply cannot, 598 00:46:04,830 --> 00:46:09,870 you know, even comprehend the scale of information that's happening 599 00:46:09,870 --> 00:46:13,830 in single cell process processes or, you know, multiple 600 00:46:13,830 --> 00:46:17,750 cells, not to speak of, you know, trillions of cells in a human 601 00:46:17,750 --> 00:46:23,110 body. And so I think those models will help us maybe keep, you 602 00:46:23,110 --> 00:46:26,830 know, keep grounded and continue to, you know, do the, the 603 00:46:26,830 --> 00:46:29,790 studies we can do that we have confidence in. 604 00:46:30,920 --> 00:46:32,760 Cherry, any last thoughts on this topic? 605 00:46:33,720 --> 00:46:38,800 I just look forward to seeing how it changes and people doing the 606 00:46:38,800 --> 00:46:43,880 right kind of thinking about what they're seeing and not just 607 00:46:44,120 --> 00:46:47,400 taking everything at face value. Great. 608 00:46:47,880 --> 00:46:52,680 So just to summarize, I think the hot take is the current cell 609 00:46:52,680 --> 00:46:59,170 models are overhyped, but I think the consensus is from the crew 610 00:46:59,170 --> 00:47:03,530 here is that we're just at the beginning of a journey and that there 611 00:47:03,530 --> 00:47:06,730 is strong belief amongst all of us that we're going to be able 612 00:47:06,730 --> 00:47:11,650 to improve them, that they are already a valuable tool, even if 613 00:47:11,650 --> 00:47:16,690 they are not quite the panacea that they are sometimes being promoted 614 00:47:16,690 --> 00:47:23,490 as. Many thanks to both Andreas and Sherry for your hot takes 615 00:47:23,490 --> 00:47:28,410 today and thank you all for watching and listening and in the 616 00:47:28,410 --> 00:47:31,570 comments. Please, please, please, please let us know what you 617 00:47:31,570 --> 00:47:33,970 think about our takes and share your own. We look forward to 618 00:47:33,970 --> 00:47:37,570 hearing from you and and as always, please do, please do like 619 00:47:37,570 --> 00:47:41,290 and subscribe. Thanks so much to my guest today and I'll see you 620 00:47:41,290 --> 00:47:42,730 all soon on the next podcast. 621 00:47:43,690 --> 00:47:45,250 Bye bye, bye. 622 00:47:45,250 --> 00:47:46,330 Thanks for having me. 623 00:47:49,930 --> 00:47:53,770 We hope you enjoyed the Translating Proteomics podcast brought 624 00:47:53,770 --> 00:47:57,730 to you by Nautilus Biotechnology. To contact us or for 625 00:47:57,730 --> 00:48:02,130 further information please. Email translating ProteomicsNautilus 626 00:48:02,130 --> 00:48:02,610 bio.