On the latest episode of the Translating Proteomics podcast, Nautilus VP of Scientific Engagement and proteomics industry expert Sheri Wilcox joins hosts Parag Mallick and Andreas Huhmer for a thought-provoking discussion of their proteomics "hot takes.”

Each “hot take” is a provocative opinion on a topic in proteomics. Specifically, they cover:

• Insufficient biological validation is one of the biggest problems in proteomics.
• Antibody validation remains a distributed, inconsistent, lab-by-lab responsibility, and this model for validation may not be sustainable going forward.
• AI cell models are overhyped.

Check out the episode for deep explorations of each of these hot takes and their potential remedies.

After listening, be sure to let us know whether you agree, disagree, or have your own hot take to share!

Hey, everyone.

On this episode of Translating Proteomics, we're going to shake things up a bit and have a roundtable discussion focused on what we're calling proteomics hot takes. For the discussion, Andreas and I will be joined by Nautilus VP of Scientific Engagement, Sherry Wilcox.

In her day job, Sherry shares how the Nautilus Voyager platform is advancing biology, helps customers bring the platform into their labs, and leads our proteomics analysis services team.

She is a proteomics rock star with over two decades of experience developing new proteomics technologies, and she's going to make a fantastic addition to this episode. Sheri, I'm so glad that you can join us today.

Thanks.

Practically the discussion will work as follows. One of us will present a somewhat contentious or provocative opinion or hot take on on a topic in proteomics.

Then we'll round robin and comment on and debate that hot take. And then later the next person will share their hot take. And finally, we'll repeat the process until each of us has shared a hot take.

We hope you enjoy the conversation and please comment on the episode or shoot us a message or let us know if you agree with us or disagree with us or have a hot take of your own. We'd love to hear from you. All right, ready everyone?

Yep.

All right, let's, let's dive in. Sherry, do you want to kick us off?

Sure. I'll start out. So my hot take is kind of around what's the biggest problem in proteomics today.

And I think it's less about any of the technologies and more about the lack of validation that people tend to want to do once they've discovered something.

I think the biggest problem that we have is that it's easy to get long lists of potential biomarkers for something, but a lot of people don't want to do the hard work that comes from a list of candidate biomarkers to then really understand what's going on.

They just want to take a list of markers, come up with biological plausibility stories, which I honestly think you can do for any indication and any list of proteins, and then move on.

But there's a lot of context that matters and people really need to be doing that type of validation and follow up studies to get more meaningful data.

Yeah. So Sher, what I take from your question here, from your hot take is that you primarily talking about validation.

When it comes to functional validation, I think we're really well positioned when we talk about analytical validation.

There is mass spectrometry techniques for Example, statistical confidence metrics, spectral libraries, targeted assays, lots of tools that we have now to sort of identify a protein multiple ways and get orthogonal evidence. I think where we lack is the functional validation, in particular the scalability of those methods.

While the discovery tools have relatively high throughput, the follow up studies remain low throughput. They're very expensive, time consuming.

You have hundreds, thousands of possible large experiments that give you thousands of candidates, but then you have, you know, very little tools to follow up.

So most of the studies now are research dependent and the researchers prioritize that based on biological plausibility, prior literature, their own intuition, and then most importantly about available funding. So I think that's where the bottleneck is.

I think the other aspect what you're referring to is should that be followed up with hypothesis free experiment by a hypothesis driven experiment? I guess what we can do by hypothesis free approaches today is that we go through another modality of omics.

So we could follow up a transcriptomic experiment with a proteomics experiment or a metabolomics experiment and sort of get more idea about the target. Again, I think those tools are rather available and so you could do a quick follow up with a hypothesis free tool. Again, if it's hypothesis driven.

I'm not sure we have the means to do that. And so the real bottleneck I think is the scalability of functional biological platforms that allow you to do that.

I mean in pharma things are done at a larger scale.

So for example, in genomics we have the CRISPR screens to study functional genomics, we have perturbed sequencing that does functional transcriptomics. Obviously there's lots of drug screening candidates that can give you more information about function.

But I think for the individual investigators or the individual laboratories, this is around a big challenge to do functional studies. I think you're right, you're putting a finger on the right problem. I think we have a tools problem that we need to be able to overcome.

Yeah, I guess so I'm going to respond a little bit, Andreas, to your thought because I'd like to. Truthfully, I'm not totally sure what the right workflow is for validating a finding.

So for example, a very common thing that is asked for by reviewers is, oh, okay, well you discovered this thing by mass spec, go do a western and prove to me that you're, you know, you say this protein is overexpressed, show me by western that it is actually overexpressed. And there's this whole tree of reasons why that western might not succeed.

You know, the antibody might target a different region of the protein, the antibody might not work. It's a non trivial exercise and particularly if you're looking for.

So getting to your point about scale, if my discovery experiment says, oh, these hundred proteins are differentially abundant, golly, that is a ton to do an orthogonal experiment and it's just a completely different universe than a follow up from an RNA SEQ experiment by QPCR where I can order some primers and stay in relatively the same system. So there's a question for me about even validating the differential abundance within the same samples. How should we do this?

So I agree with you that people don't do that a lot. I think the reason is because we don't have a clear playbook that says here's how we should do it.

And then there's the next step of like, okay, now I believe you that the protein is overexpressed or knocked down in some way functionally. What does that do? Is the right answer to do the knockdown or the overexpression. To your point, that doesn't scale super well.

And then you have all the challenges of it being a system and all these things working in coordinated action and so you can do those one by one, but you can't collectively over express a hundred things yet easily without blowing up your cell system. So I agree with you that this is a gap. I just have absolutely no idea how to fill the gap.

Yeah, and like you said, if you want to do knockdowns or something like that, you can't do that for every protein and have a system that even works.

So I, I kind of think it's just each finding needs to have its own body of evidence and that body of evidence may not be the same thing for every single finding. And so there's got to be like a toolbox and we figure out which tools we get to use in which circumstance.

And depending on what we're trying to answer, the, the weight of how many tools actually have to be used can vary.

Because if it's something that's really unknown and really novel, then the weight is probably a little bit lower for initial understanding, but then it builds over time.

Whereas if it's a protein that there are so many tools out there in the world that you can, you can leverage, then you would expect a bit of a higher bar. Right. I also made me a little sad, Andres, in one of your comments. The funding was one of the reasons people don't do the right science.

And it is unfortunate that the reality is that that's that's part of the equation as well. That's.

Yeah.

I mean these follow ups are extremely generous and it's, it's pretty quick and easy to do the dis, to do the discovery experiment, to shoot it through the mass spec and do your DIA and, or run your SOMA scan, et cetera. But then that one day experiment could lead to months or years of experimental follow up.

Right? Yeah, somebody's dissertation.

So maybe that's part of the, part of the challenge here. Is that really what we're trying to. So to your point, what does this toolbox entail?

And I think that depends upon the strength of the claim, the magnitude of.

The claim and the consequence.

And the consequence, Right, Absolutely.

So if for example, what you're saying is, oh, overexpression of this thing leads to cancer initiation, well then the functional follow up is a classic set of cancer initiation assays. And if it is, oh, this affects motility, well then that's a functional phenotypic screen. And so.

But all of those tend to be one by one perturbation phenotyping. They don't tend to be, hey look, I'm going to move a hundred things a little bit and see how that broadly affects the phenotype.

So is this a particular problem in all of biology or is this more of a problem in proteomics?

To your point you made earlier, it may be a problem everywhere, but I think proteomics maybe has some extra high hurdles compared to other things like genomics, where you can make the tools that you need to do the right experiments. Every protein is so different that it's harder to have just a unique set of go to. This is what I do every time. But that's just one thought.

I don't know what I'm going to put the question back to you or to Andreas.

So yeah, I mean the biomarkers we get from fluids probably have a different follow up going to the original tissue, maybe doing bigger studies.

But I think fundamentally it's something that's related with proteins themselves because they're diverse functional set within the context of a cell or with the context of the tissue. And so you better have a good question in the first place how you want to validate the protein or you're not even doing the right experiment.

So I think it's a little bit of the complexity of biology that stands in the way as well, besides the fact that there's no really good high throughput tools that are available. And then I think I like your comment, Parag. What do we mean by validation? Right.

I mean, you could probably take the same cell line and I've proliferated it over several years in different laboratories and then do the exact same study and come up with a very different answer because you're talking about a different cell line. Right.

And so I think that is a fundamental problem we have to deal with in biology, that it is not as reproducible, at least on the protein level, as you might be able to do that on a genomics level.

Yeah.

So I think that question, and if we want to be, if we want to be able to provide some practical guidance, maybe part of it is in the validation defining upfront, what aspect of the experiment are you trying to validate? Are you trying to validate that in these specific samples that the measurement was correct?

And maybe part of that comes from in experiment metadata, spike in controls, positive controls, negative controls, et cetera, that further build confidence that the measurement itself was correct, which again gets to sort of the QPCR equivalent for transcriptomic data, which really is just asking, do I believe when I say that I measured this quantity, changing that it actually did, then there's a next layer of biological heterogeneity and was there a bias somehow in the samples? And that is addressed by replication more biological samples of the same sort, running cell lines today and running them a year from now.

And so there's sort of a chain of custody for biological sample diversity. And to what extent is the statistical significance preserved in a similar but wider biological set?

And then to the other point is the next level beyond that is the functional impact. And that gets really deep and really hard.

But I haven't really seen a good rubric out there that says, okay, well, to validate at these three levels, here's the playbook.

Plenty of statistical literature that says, okay, well, I'm doing my significance test in this way over this support set, but that's different than the post facto analysis.

So I think, unfortunately for this hot take, we all agree, but this is maybe an opportunity for creating some structure around what is the standard for expectation of validation, at what level.

Great.

So to summarize, I think we all agree that analytical validation has some tools that we can employ in order to understand the measurements that we're actually making.

The functional validation, however, is a bit harder and we don't really have a playbook defined and don't really have the tools that today that we need in order to make that happen.

But I look forward to this community being able to step up and come up with Great ideas so that we can build this playbook and understand how we can take all of these findings that we have and turn them into the biological meaning that will advance the science.

Thanks, Sherry, that was a great discussion. Andreas, I'll turn it over to you. What's your hot take?

Well, we stay a little bit with protein and protein detection here, sort of building on the story we had around Western blots.

But recently we have seen a number of scientific papers questioned and even retracted because manipulated fabricated Western blots and that extended even to a major vendor in the very recent weeks. In some cases, images were duplicated, altered, assembled in different ways. In any case, the data that were present misrepresent the underlying data.

At the same time, you actually have a cottage industry evolving for investigators that call themselves in scientific integrity experts that use AI tools to go through existing publications and find disconnects between the actual claims and the the proof points. But that's not really what I want to talk about.

I think what I want to talk about is the challenge of actually measuring proteins and the nutritical variability that is often associated with those tools, such as Western blots.

What's interesting is despite the fact that we have tremendous advances in biology, many of the core protein detection tools are fundamentally not different.

So it's a tool that has been going on for a lot, I've been living on for a long time in our laboratories and we all know what the best practices are of using such a tool, which is to spend significant time and resources to create a process control.

Then you test your antibody specificity, you look for cross reactivity, you test lot to lot consistency and you make sure that the conditions you're measuring the protein in are represented in your controls. And so there's a lot of work, a lot of resources involved to really validate antibodies.

The question for you guys is if antibody validation remains the main responsibilities of individual laboratories, is that a model that can sustain ultimately modern biochemical research potentially? What could a better future look like for protein detection? Is there new tools on the horizon that could take the role of a Western blot?

I kick it over to Sherry.

Great, thanks. It's a very interesting question.

I think that the burden can't be on every lab to do all of the analytical validation again and again and again for all of these materials. That's just too burdensome.

I am hopeful and encouraged by some of the initiatives to try to standardize at least the commercially available antibodies.

You know, there are a lot of antibodies that get resold under different companies and it's the exact same material, but the general user doesn't tend to know that.

And so I do think that things like antibodypedia and some other types of databases that are compiling both the validation data and the essentially the primary source of the antibody so that people can know, oh, actually this, this antibody is the same as that antibody is going to be really important for people going forward.

As far as additional tools, I think again, it comes back to it's the same tool isn't always the right answer for every single question, which makes it really, really hard.

There were some talks at an affinity proteomics workshop I went to a couple of years ago where a lot of there are some initiatives have an exact same panel of tests that are always done on every antibody in exactly the same way.

And while at face value that sounds great, we can get much more comparable data when you dig into it, given this application versus that application, the concentration might not be the same because the affinity is not the same. Exact way you run those tests does still have to be nuanced.

So I'm not sure I have a great answer as to what we can do to kind of make it better, aside from cataloging as much data as we possibly can and making as much data available to people as possible and trying to be as transparent as people can for some of these tools and the level of validation that has been generated thus far.

So I'm, I'm just, I'm.

You've.

You've got me thinking and there are many things spiraling in my head at the moment. So I think part of why what I'm thinking about are again a couple different levels of. One is polyclonal antibodies, sera that are not even purified.

There are many form factors for affinity reagents that are not just recombinant expressed monoclonals. They're also with different, lots of even monoclonals. Some proteins might decide to misbehave and aggregate or misfold in some way.

The percentage of active antibody might be different. And so the challenge is, is I think at two levels, one is on the part of the vendor to know that they have built a good reagent in the first place.

They have a toolkit of why you should trust this reagent versus that reagent, that finding things out like what epitope was used to immunize for this particular reagent is actually quite hard. It's quite scarce information.

And then, and so somebody might run the same antibody on two different systems, one of which is the dominantly, the N terminal form of the protein is present and the other the dominantly, the C terminal form of the protein is present. And so the antibody itself might be great. It might just be targeting a completely different part of the protein.

And so you might get disparate results for biologically reasonable reasons. And so that aspect of the complexity and having more information available, not just the, hey, look, I tried it on the protein and I got a band.

The burden of proof, I think is higher than what we have demanded in the past.

To your point, Sherry, about Antibodypedia and the similar efforts of Protein Atlas has put out the lists of antibodies that they've used and validated their sets of gold standard antibodies that have been validated by multiple people. I think these resources are invaluable.

What's interesting is we have these really great databases for every RNA SEQ experiment deposits their data in geo. I promise you, everybody who does an immunoassay, a Western, does not deposit their gel in some central resource.

So we could go back and say, okay, well, this antibody from this vendor number, from this lot number, here are the 12 gels, here's the positive control on the gel. And so maybe one of the things that's necessary going forwards is actually expanding the scope of these publicly available deposited databases.

I think now many journals are requesting the unedited raw images of immunoassays in part to help mitigate the fraud concerns.

But beyond the fraud concerns, there are just the practical realities that that is very valuable data to be able to connect a lot number and a vendor and an antibody target to a series of data.

And if we started having thousands or hundreds of thousands of images from a given antibody from a given vendor, I think that would be an incredible resource that would allow us to go back and troubleshoot and understand the performance of these reagents and performance over time and in different contexts. So to answer your question, I do believe it's a shared responsibility between the vendors and the users across the scientific community.

It shouldn't be necessary for every lab to fully revalidate every antibody, but it should be that there's a minimum set of things that are in every experiment, the positive control, the negative control, et cetera, before we fully trust an immunoassay output.

And I think there are Many experiments that, and many westerns that you see that don't necessarily have the titration series of the positive control, for example, or multiple forms of a protein or a background lysate negative control. And those reagents themselves can be hard to find.

But if we can agree as a scientific community to start upping our expectation, we have these things like the Miami standard for MRNA expression, the MyAPI standard for protein expression, and these are mostly applied to high throughput discovery technologies. Maybe there's a new standard that is wanting to be made around minimum information about an immunoassay experiment.

Right.

And the International Working Group on Antibody Validation several years ago put out essentially the five pillars of validation that all antibodies should at least try to achieve.

And even if it's hard to get data for all of those pillars because some aren't relevant to a particular type of target that an antibody might be targeting, having again a body of evidence that is as much as one can do, I think is useful to the community.

Andres, you also asked about alternate ways to do validation and I do think tech, you know, some of the, the protein arrays are certainly another way to, to get some of this validation beyond a western blot. Additionally, the FIP SEQ phage immunoprecipitation sequencing is a tool that can be used to again, look at specificity.

And then, and then those pillars that are in the International Working Group on Antibody Validation paper are really good, good ways to look at this validation beyond just a western blot.

Yeah.

So here you're both making an argument that, you know, there needs to be more standardization, there needs to be more sharing of resources, some way of really documenting, you know, the experiments and validation experiments are being done. And I think I agree because most of the antibody market is basically, you know, wild, wild west.

You produce an image that shows that it migrates somewhere where you would expect it, and then you can sell it, put a label on and sell.

And then there's obviously, you know, as you mentioned, Sherry, many, many things going on in the background in terms of, you know, good antibodies, you know, available commercially through many channels. I think it is, you know, probably a good idea to look at what other analytical technologies have done or techniques have, you know, sort of done.

You can buy, I mean, you can buy standards for any analytical technique, whether it's a chromatography system, a mass spectrometry, spectrometry system, you know, a spectroscopy system. They all have standards. You buy and you basically validate the instrument before every single experiment. You do through a calibration procedure.

And I think that it's probably worthwhile to think through why not making the actual, you know, epitope or the protein itself that, you know, you raised the antibet against it, available, maybe even the media, and then have the community have access to the tools and say, yes, I can reproduce it consistently. My lab, particularly if the lab is very much focused on that protein because they're doing the validation studies.

We talked in the first question, in the first part of the hot takes here.

I think it makes a lot of sense to really put the pressure now on the providers, but also on laboratories in the academic space that produce antibodies to come together and really form that resource that we need that ultimately helps us avoid these very frustrating exercises to validate the data in retrospect, retrace some of these papers. In some cases, careers have been damaged or delayed. And so it's all pretty much unnecessary if we work together.

I haven't heard any, you know, you guys talk about a new technology or other ways to maybe detect proteins. I would put forward that, you know, there is tools like modern affinity reagents that use two antibodies.

There's obviously mass spectrometry is more compositional, you know, focused analysis, but maybe there's also things we can do to combine methods. What do you think? And I've got, you know, handing this off to you, Parag.

Yeah, well, I, I think about, there are these alternative platforms, things like the Licor and the Wes and Jess and similar non ex. They're not exactly Westerns, but they are immunoassays. I think isoelectric focusing.

Immunoisoelectric focusing has been a very interesting technique.

Ironically, it dates, it sort of reminds you of the earlier, the very, very old immuno electrophoresis methods that you mentioned earlier, from the late 70s.

And so I think there are, there certainly could be newer methods that are brought in, but all of those methods essentially still depend upon your confidence that your antibody is binding to your protein in the state that that protein is. And for Westerns, things are a little bit easier in many cases because the proteins are denatured unless you're running a native Western.

But for ELISA reagents or reagents that are being used in ihc, where the antibody is targeting a structural motif, you can imagine that as the protein breathes or gets modified and changes shape, that the antibody might bind in a different way.

And so I think the part of what needs to be answered is, all right, what do we think we are measuring with this particular probe and is it purely protein abundance or is it something more complicated?

And that gets back to the transparency of the epitope, like you said earlier. Yeah, yeah.

So maybe the takeaway here from the conversation is we definitely want to make sure that the community stays on the current path and really invests in common tools, common platforms that actually allow us to validate the antibody that we use.

And I think what you're saying, Parag, is that we also need to keep the context in mind of how we validate these antibodies because sometimes it is not as simple as running Western blot. It might involve additional effort spent.

And so I think that probably provides a good path forward for the current conversation around antibodies in the context of the fraud that, that I started introducing the topic with. And hopefully we see some, we see some rallying around this particular cause.

Great.

Thank you, Andreas, for leading us through that hot take. Just as a reminder to the viewers, listeners, we'd love to hear your thoughts on these topics, send us comments, argue with us.

These are important topics and we'd love to hear your feedback. So I'm going to turn it over to myself now for our last hot take. And I think I feel like our current hot takes, we can be even more dramatic.

And so I'm going to go for the hottest take that I can possibly go, which is that the recent wave of AI virtual cell models are grossly overhyped and unlikely to be as performant as they are believed to be. And so I'll explain what I mean by this and why.

It has certainly been the case that over the last two years there have been a number of both frontier models and foundation models that have come out that have been trained over large corpuses of data, typically large amounts of transcriptomic data, and then in some cases cell imaging data on top of it to try to get something related to phenotype.

The stated claims of these models or genomic data, DNA models, is that they can perfectly predict phenotypes that you can put in some mutation, and out the other end comes a perfect prediction of what the cell behavior or the human physiologic behavior is. Now, I the the effort of building one of these kinds of models is gargantuan.

And absolutely I, I'm not in any way diminishing the effort that the researchers went through to build these sorts of models and even just the data curation, organization, model structure, the actual mechanics of training are huge and complicated. But the question at the end of the day is how much of biology is being captured by these models, particularly disease biology.

The other question is, when we think about some of these models, they're there to say, oh, okay, well I can take it a DNA measurement and I can predict protein abundance from my DNA measurement. I'm going to come out and declare that that is crazy town.

Yeah, so I think you're right in the aspect that, you know, AI models are typically overhyped. And it's not just in the biology space. I mean, it's, it's across the entire AI space.

But I would submit to you that I think it's a really important first step for us in biology to go beyond the typical phenotypical observations and low throughput biology I would call it.

I think what is probably more honestly claimable in that conversation is that we actually have a fundamentally new way how to learn biology rather than try to actually predict biology. At this point, the arguments I would put forward is the following. The AI models can learn beyond a single cell.

And so you can take millions of cells, not just one experiment, but millions of cells and learn from that.

The other argument is that the models can capture biological context rather than the individual gene that we typically study in particular cell models, for example. And so because the gene doesn't have a fixed biological role, the same gene can have a different context in different cell types.

AI models are much better tools that actually capture the context. And transformer models are built to actually extend those functions amongst different, for example, cell types.

I would also say, for example, that AI models are much better in integrating multi omics aspect of it. We struggle a lot to sort of even combine these multi omics data set on a, you know, on a single protein or single gene level.

And AI machines or computers in general, you know, because of their scalability are much better in sort of capturing the modality of many modalities.

And so I feel that future biology really requires the integration of these transcriptomics, epigenomics, proteomics, spatial information metabolomics data. And the transformer architecture is inherently flexible enough to incorporate these multiple data types.

I think what you will end up seeing, or hopefully what we end up seeing, that these models create transferable representations of cells rather than the prediction of a particular cell.

And that ultimately those cell models are much, much more practical in now testing biological hypotheses because they can access a scale that ultimately we cannot assume.

And so you might be able to take this rather virtual cell and ask many, many questions and say, hey, which one of them is even a good question to ask in the first place based on the information we have gotten from many different multi omics experiments. And so I think the, the cell models are not there really to predict biology, but essentially guide us.

How do we actually study biology in a much, you know, on a much, on a much bigger scale. So I think that's my current thinking about it. AI develops so quickly these days.

So, you know, we may have another podcast where we have changed our whole minds, but. Sherry, what do you think?

I have a few thoughts.

One is as soon as someone uses the word perfect, as a scientist, I have a pretty visceral response to that that I'm going to not believe anything else. That said, biology is way more complex.

I think there's an element of hubris for us to think we're going to develop something that is going to be able to give a perfect prediction of anything. And so it's kind of, it's. It's not.

My concern is that it's not unlike what my hot take was to a certain extent that people want to get a list and not think and just go with it. I'm concerned that that's what can happen here too is that people are going to believe this is perfect.

They're not going to ask hard questions of it because AI generated it.

And then we're going to get some flawed understandings, particularly knowing that the model is going to only be as good as the input that went into it. Of course, and we know that in scientific studies.

And I'm not this is not necessarily just about a cell model or in particular, but diversity of participation in generating the data to develop any of these kinds of any models has not been great historically. And it's something that a lot of people are certainly working on.

And yet we know that the data that is fed to date is probably going to have some biases.

And so for us to just take anything as at face value as this is, this is the absolute and this is the correct answer without asking those other questions of how might something else that is different in one context than it was in the context in which this was built. If we can't ask that question, then we're going to miss a lot of nuance and we're going to do a disservice to the greater community.

I think AI has, has a lot of good uses. I agree that it is so overhyped.

Anytime people want to use AI, even if they just mean normal machine learning, you know, they, it's just, it's great to plug in the word AI, but we do have to be careful that we're really, really being cautious of knowing exactly what went into the model and, and what the implications of that might be on, on the decisions that we make based on, on what we're, we're trying to study. Yeah.

I'll point out a few specific things that lead to my declaration of overhyped. One is the nature of feedback loops, that biological regulation is all about these sense and response and across multiple length and time scales.

So things that occur at the second time scale, things that occur at the decade timescale and many of the models are feed forward models and so they don't have an explicit notion of a feedback loop. Now there's a thought that, okay, well if you upscale enough, maybe you abstract away the feedback loops.

But I would argue that the genome to phenotype is such a big jump and it's a jump over so many timescales that having some concept of being able to play this forward, see what happens at multiple timescales and then coarse grain from there is an architectural thing that is currently quite hard to build into most AI models and hasn't really been anticipated. I'm not saying it can't be. I'm just saying that the current generation of cell models haven't really accounted for space and time very well.

And so we've talked about that on other podcast episodes about just in general how important space and time are to modeling biolog. I think the other thing that in our conversation with Olga Witek, we were talking about causal models as a branch of AI models.

And that notion of I poke here and this thing happens is again, it's a directional, it's a magnitude component that isn't necessarily well captured by simple associations or even complex associations.

And so that next generation of AI tools that explicitly have causality baked in, or has physics baked in, or has multiscale multi omics baked in at the appropriate scale, I think that's the gap.

So when I see these models making predictions I like, okay, it is very easy to make lots of predictions, particularly when you have large, huge parameter systems. How sensitive are they? And then when we talk about disease processes, my sort of third argument is that disease processes typically are rare events.

And so they're at the edges of distributions. And when you're trying to estimate any parameter, you need a ton of data.

And when you're trying to estimate parameters that are at the edge of distributions, you definitionally don't have as much data because they're rare events.

And so it's actually quite hard to Figure out how to learn these extremely rare low frequency events, and of course working on transfer learning methods, etc. Is meant to address these.

But I think these three fundamental challenges are why I say that the current models, though very impressive in scale, are still not quite as performant as they may be hyped in the media to be.

Yeah, I would just again state, as you say, we are at the very beginning of learning how to construct these models to more reflect actual the phenotypes we can observe over space and time. I'd share with you the concern that most of these models are built on single cells and single cells are not tissue.

Single cells put together the pipes would create a different response. So it goes back to what you just stated, that it's very hard to predict forward.

However, I think the fact that we have a number of people working on this using computational tools maybe also brings it back to what Jerry said.

We have to have a fair amount of hubris in mind when we do biological studies because we simply cannot, you know, even comprehend the scale of information that's happening in single cell process processes or, you know, multiple cells, not to speak of, you know, trillions of cells in a human body.

And so I think those models will help us maybe keep, you know, keep grounded and continue to, you know, do the, the studies we can do that we have confidence in.

Cherry, any last thoughts on this topic?

I just look forward to seeing how it changes and people doing the right kind of thinking about what they're seeing and not just taking everything at face value. Great.

So just to summarize, I think the hot take is the current cell models are overhyped, but I think the consensus is from the crew here is that we're just at the beginning of a journey and that there is strong belief amongst all of us that we're going to be able to improve them, that they are already a valuable tool, even if they are not quite the panacea that they are sometimes being promoted as. Many thanks to both Andreas and Sherry for your hot takes today and thank you all for watching and listening and in the comments.

Please, please, please, please let us know what you think about our takes and share your own. We look forward to hearing from you and and as always, please do, please do like and subscribe.

Thanks so much to my guest today and I'll see you all soon on the next podcast.

Bye bye, bye.

Thanks for having me.

We hope you enjoyed the Translating Proteomics podcast brought to you by Nautilus Biotechnology. To contact us or for further information please. Email translating ProteomicsNautilus bio.