AI-generated phages that work: ChatGPT for phage biologists? — Podovirus podcast
1 hr 04 min

AI-generated phages that work: ChatGPT for phage biologists?

The conversation

Episode notes

Sam King, a Stanford bioengineering PhD candidate, and Claudia Driscoll, a postdoc in the same lab at the Arc Institute, explain how they turned AI-generated DNA into working phages. Listen to understand the biological judgment behind designing an entire genome and deciding what counts as success.

The model was not explicitly prompted to infect E. coli C, so the researchers used biological knowledge to filter generated genomes for spike proteins resembling those of PhiX174.

Sixteen phages worked from approximately 300 tested genomes, but the researchers prioritized novelty, meaning that result does not establish the success rate for designs selected primarily for confidence.

After passages against resistant bacteria, sequencing suggested that successful cocktail-derived phages combined segments from multiple generated genomes, although the researchers could not confirm that recombination had occurred.

Episode timeline

  • 0:00 Introduction to AI-generated phages
  • 7:14 How Evo learns from genome sequences
  • 18:27 Why PhiX174, what worked, and what remains
  • 32:57 What prompting supplies and filtering decides
  • 49:18 Measuring phage behavior and resistance
  • 56:43 DNA synthesis costs and access to Evo

Learn more

Studies

Tools and resources

Transcript

AI-transcribed and edited for readability. May contain errors or omissions; check the recording for exact wording.

Select a word or timestamp to listen. Use arrow keys to move between words.

  1. Jessica Sacher

    Okay, hello everyone, welcome back to pot of virus. We have been on hiatus for a few months, but we have been drawn back out of vacation mode. Because we saw a really awesome paper come out, or preprint that you may have all seen. I think everyone has seen this in the phage world because I just went to a conference. And there was not a single person who hadn't heard about the AI generated phages, so that's really exciting.

    And I happen to work right next door or down the road from these guys at the Arc Institute, so we have Samuel King and Claudia Driscoll here from Brian He's lab at Arc Institute. And so they were, I met them. I feel like it was several months ago, I don't know how many. Where they were trying to get like these AI generated phages to actually reboot and be real phages.

    So they've built a system based on, as I understand it, it's as everyone talks about all the time. An AI model, or a machine learning model, similar to, basically, if you think about Chatgpt, but for DNA. So they built this. And then they tried to get it to make phage genomes from scratch, and they tried to get it to make ones that would actually live and work and kill.

    E. coli. So they came, they we. We got connected somehow. And they were like, still at the point at that time where they were trying to get the first phages to work from the system. And that's where I remember chatting phages, talking about propagation and like agar percentages and just how do you get plaques and things. And then, you know, we didn't talk for a few months.

    And then all of a sudden they're like, Okay, we're done, we did it, we made all the synthetic phages, it's working and the preprint is out. So now this whole conversation is partly a mission to just be like, whoa, like, how did you finally get it to work? And also, what does it mean to make a model like this? Like, how do you prompt a model like this to make a biological thing?

    How can we use it as a phage field, as phage researchers? Cause I know you guys put out your models publicly and everything is all open, so we should be able to use this eventually. Hm. And also just some of our questions on behalf of the phage field of what, like, what this really means for, or the phage research field, like, can we really use it to make therapeutic phages that are better?

    And like, we'll maybe address some of the pushback we've heard, too, which is, you know, biology could have already engineered Phi X 174 phage and variants of it, it already has people already have, so this making a new set of variants of it might not be like. It's not like, shockingly novel on a biological scale to some biologists. But I think once we dig in, we'll find why I am excited about it and why Joe's excited about it.

    That this is like a really cool S proof of concept, first step, and like, kind of what the constraints are with it that have us where we are right now. So with that, I would love to hear just some intros from you guys, from Samuel and Claudia. About who are you before you came to this role at Arc, and what are you doing there? We'll start there.

  2. Sam King

    Awesome, yeah, thanks for having us, it's really great to be here. So I'm Samuel, I'm now a third-year bioengineering PhD candidate at Stanford. I work in Brian Hie’s lab, we have our lab space at the Arc Institute, but we're also part of Stanford. And before coming here, I did my undergraduate degree at UBC in Vancouver, British Columbia. I studied mostly molecular biology there. I did a lot of cool cell biology research as well as synthetic biology at the time.

    And came into my PhD wanting to do something really out of my comfort zone. And that's when I started getting increasingly involved in computational work and especially seeing, the implications for machine learning in biology. Everything just fell into place when I saw Brian's lab and I was really excited to be here.

  3. Claudia Driscoll

    Yeah, again, thank you so much for inviting us to be here. I'm Claudia, I'm a postdoc in Brian's lab.

    I did my PhD work in Oxford in the UK. My background is very much protein biochemistry, so I was kind of working in protein engineering for designing by specific antibodies. And I came to Brian's lab. As I think a lot of people do, the hope that I would be able to kind of apply myself, to AI for bio. Using pro both protein language models and then, as the lab has evolved, also genome language models.

    And I was very lucky to join the lab at a time when Samuel needed some help with the experimental side of things. And so, yeah, I contributed to this project on the kind of experimental side. And I think we can probably get into it a little bit in the podcast. But coming from, I guess, a non phage background was quite challenging. Cause we had to kind of establish a lot of, I guess, basic protocols, you know, that you would consider basic, but we kind of had to establish those from scratch. But I think it was almost helpful to not have like prior biases. You know that meant, oh, no, it has to be done this way, or we have to, you know, do these phage protocols in one specific way.

    Whereas we were able to kind of come at it with a very fresh mind and figure things out, you know, from first principles, almost, which I think helped us develop quite exciting protocols in this paper. But yeah, we can chat about that. Yeah,

  4. Jessica Sacher

    Yes, I loved seeing, by the way, I loved seeing your protocols. Because, yeah, they were like, written completely differently than I'm used to. But there were like several things about them were like, so much more efficient. And I was like, Wait, oh, you just, oh, I wish I'd you. Just like, have some from the freezer, put it in and see if it kills. And of course, why not? You don't have to do a full plaque assay evolution series every time if that's not what you're checking for. And anyway, so many versions like that. And I was like, Kate, please send me these, because these are just, I don't know. They're so crisp and to the point, and they just work. So,

  5. Claudia Driscoll

    kind of blocky, I guess, as a yeah, space,

  6. Jessica Sacher

    Yeah, love it. And I also wanna make sure we let the audience know Joe is here, so you can't see him on the camera if you're not watching. But sup Joe, how are you?

  7. Joe Campbell

    I'm doing well. I was at the same meeting that Jessica was. And yeah, I enjoyed reading your paper and curious about a few things, but we can get into them.

  8. Jessica Sacher

    So can we just walk through the little bit of the nuts and bolts of how? What you've done in the paper? Let's just say, you know, you're talking to a bunch of phage researchers, they know all about phages, so you and to set the scene a little bit. And some people have read these other papers of yours or your group's from Evo, the Evo model and then Evo 2 model.

    So we understand that there's like millions of genomes, and then a trillion genomes, and the other one, something like that. The first one had prokaryotic and phage genomes, as I understand it, and then the second one had eukaryotic genomes as well. And you kind of, can we talk about, like, what those models were? And how that sort of forms the basis for this, as I understand it, a third model that's like, very specific on making Phi X 174 phage genomes.

    Is that the right way to think of it, and how does that all work?

  9. Sam King

    Yeah, I can get into that, our group, we put out a model called Evo One last year, I think it was published, fully published November 2024. And this project, the phage project, actually started before the Evo paper was completely done.

  10. Jessica Sacher

    Cool,

  11. Sam King

    but at the time, we. We really wanted to leverage advances in machine learning architectures, which enabled, especially long sequence modeling. These models like Chat, it, as you said earlier, it works very similarly to ChatGPT. You put in a prompt and the model spits out what it thinks should follow after that prompt. In this case, the prompts or the sequences the model was trained on is not natural language or English language, it's DNA sequences.

    And we specifically were able to leverage an architecture called Striped Hyena, which was a mixture of these Hyena modules and Transformer modules. Transformers are typically the architecture used for these other natural language models. In this case, the Hyena modules and the architecture were able to improve the long range processing of the sequences that it was trained on. I won't get too much into the technical details of that.

    It was. It was amazing to see that, you know, after we had set up the training process with it was around, I think, two, seven million phage genomes, and then around 80 zero bacterial and archaeal genomes. So many, many more phages actually were seen by the model, but in the Evo One paper, most of the validation was done with bacterial systems and bacterial modeling. We wanted to see in the 1st part of the phage work if the model understood phage genomes quite well.

    And so we just did some pretty general sequence generation. So in the training data for the model, every genome was prepended with these phylogenetic tags, essentially describing the entire taxonomy of that genome. So since that's in the training data, we can actually use those phylogenetic tags to prompt the model to generate sequences that it thinks should match that phylogenetic tag. So in the first part of the phage work, we use those tags to broadly generate phage realms, so double stranded DNA phages, single stranded DNA phages and RNA phages. And it really did look like the models understood phage sequences quite well. They could generate very reasonable looking genes that annotated as phage genes, but also were very novel when blasted against other natural phages. And so that was like, quite encouraging in the direction that we could then go on and actually maybe generate.

  12. Joe Campbell

    What's the criteria? I mean, I remember reading your paper, you said phage like, and you just used it. Can you define how are you classifying something as phage like or not phage like?

  13. Sam King

    Yeah, so we were comparing these generated sequences specifically against a database called the Frogs database, which is a curated collection of these different clusters of phage genes. And we were seeing how similar the sequences were to these natural phage genes. And then we were also just looking at them in terms of general metrics, so structural conference confidence. So we were folding them with ESM fold and seeing if they had reasonable confidence, similarly to how ESM fold might have confidence around natural structures.

  14. Joe Campbell

    And you had pre set like cutoffs that were gonna say if it's above this cutoff, we're gonna call it phage like, and if it's below it, we're not. Or how exactly did you make those decisions as to whether it was phage like?

  15. Sam King

    Yeah, it wasn't as strict as that, it was mostly just: we wanted to see the distribution of these metrics and how it compares to the distribution of natural phage sequences. So we also when we looked at the generated sequences, we compared them against a small data set of natural phages from these three different viral realms. And then we plotted the same distributions. And then we also used scrambled, what we called scrambled controls, which were scrambled versions of the natural sequence sequences.

    To see if these sequence or structure based statistics would also work on just fully S randomly scrambled sequences. And that wasn't the case. So we do see that a very coherent, actual architecture of these genes.

  16. Joe Campbell

    So you have a lot of phage genomes and you're training your model on this. Are you training it? As are you just telling that this is a phage genome, or are you telling it it's a phage genome and it infects these bacterial genomes? Or are you just telling it that it's a phage? What are you telling it about the phage genome? Be beyond the fact that it's a phage genome?

  17. Sam King

    Yeah, that's a great question, so, and that also leads, well, I think, into further explanation on Evo one and Evo two. So it's called unsuper supervised training, so these both Evo one and Evo two were just apart from these phylogenetic tags, which are essentially just given with each genome. To hopefully tell the model that that's the taxonomy related to that genome. But apart from that, there is no kind of guided information on what the sequences actually are that the model is looking at.

    And so this goes for all of the phage genomes a as well. So in the stage where the model is trained on this really large corpus of DNA, this is called pre training, and it's essentially to hopefully train the model well enough. So that you have this very general view of all of this DNA. And in a way, it's kind of this one single AI model that has now seen a very, very great representation of the evolution that we as humans have seen and sequenced.

    And all in a single model, where it's learned probabilistically, which nucleotides come after which in these sequences. So from there, we could use some these pre trained models to generate the phage sequences that we want. And that was the case for when we were generating broadly these phages that match these different viral realms. But when we actually wanted to generate specifically phages that were very close to PhiX174, we needed to do what's called fine tuning.

    And this is where it's essentially just more training. So we curated another data set of around 15 zero microbiome day sequences. Microbiome day is the family that PhiX comes from, and we just further trained the model on these sequences. Essentially to kind of sharpen or hone its understanding of microbiome day. And from there, we're able to use much smaller prompts to point the model in the right direction towards what we want.

    And so this is what we did for both Evo one and Evo two, so Evo two was not even a thing when we first started this project. And throughout the project, it quickly rose to be a much larger and better version than Evo one. And the preprint came out in February, it's not fully published yet, but we were then partway through the project. Able to also fine tune Evo two and see if we could get similar or better performance on generating phage genomes,

  18. Jessica Sacher

    And did it help? cause it didn't have more phages, did it, or did it also have more phage genomes?

  19. Sam King

    It was roughly the same number of phage genomes in this case. So the main difference between Evo one and Evo two in terms of the training data is that Evo two was additionally pre trained on eukaryotic genomes. So this includes plants, animals, fungi, humans. And what we tend to see when we're generating sequences is actually since Evo two has seen a much more diverse set of genomes in its training.

    That it seems to maybe actually require slightly longer prompts to get it generating sequences. I think since it just has so much more understanding across all of life, I think maybe it just needs a little bit more information to tell it, like, which sequence you're actually trying to generate. And so that was one thing we pretended to see when generating sequences. But also, I think because of that, it seems like Evo two is able to generate more diverse sequences.

    So one way we measured this was through this simple formula called Shannon diversity, and this is often used to measure, like species diversity in ecosystems. In this case, we were measuring the diversity of the genomes that were generated using different combinations of parameters to for the model. And we saw that Evo two tended to have higher Shannon diversity in different batches of sequences we were generating. So it seemed like, with the same set of constraints, it could generate more diversity, which was cool.

  20. Jessica Sacher

    Yeah, it's wild. How it could learn from eukaryotes how to design phages more diversely, and more well, I guess, stands to see if they're more realistic or with, or more powerful. But I feel like the concept of thinking about the models as just like learning, like broadly, like what matters in DNA, and like in regulatory elements, and just like order of genes. And I feel like you got into this in your paper and it was.

    It kind of expanded my mind, cause I was always previously thinking that, like biology, AI models are like whatever training data you put in, like, it can generate something like that. And I guess that I at the si same time play around with chatgpt all the time. And it's like, clearly it just has a body of knowledge

  21. Joe Campbell

    No,

  22. Jessica Sacher

    that it can, like, rest on. And it's not requiring

  23. Joe Campbell

    definitely

  24. Jessica Sacher

    me to give it anything to have a conversation about something. And I guess it's like, is it kind of the same? Where it just kind of learn? Has learned all these kind of subtle rules about what's allowed in biology? And like, there's always a stop codon, kind of a little ways away from a start codon, and like every other complex like rule that we've managed to find, plus all the ones we haven't. Is that kind of how you think of it?

  25. Sam King

    Yeah, I guess that's really the goal with these deep learning frameworks is to learn this extremely general representation of the data. And hopefully, if the model is trained well enough, then the goal is that it could maybe actually reason out of distribution from the data that it's seen. And based on, you know it, I mean, this is, I guess, a big critique in deep learning is that these models are what people call black boxes, where you can't really peer into them and actually fully understand what's going on because they're so ginormous and complex, and there are some really, really exciting methods now, where we're able to interpret a bit better how the models are understanding different sequences.

    So in the Evo Two paper, we actually show this, where we. We can essentially see, like, which parts of the model are activating on different features within DNA sequences. And how this kind of informs, different aspects of these sequences that it's come to learn, like codons, or stop codons, or, you know, even concepts as general as a phage or a prophage. It seems like specific neurons within the model are associated, with these things. And I guess that that's a very simplified way of saying it, but we're able to sort of tease that out of the

    model, through cool machine learning methods. And but I think maybe I just wanna take a step back and actually come back to, like, the big picture o of this project. And it really was about that concept, like, what does it mean to design a genome and navigate around this evolutionary design space? This was just even philosophically something at the beginning of the project where we were really creatively challenged.

    As you know, what kinds of, like, metrics do we want to implement to evaluate generated genomes, like what does a generated genome even? And what should that look like? What kind of architecture should it have, which genes should it have, or, you know, regulatory elements, and how should they be arranged? How long should the genome be? And that led us to having to come up with this framework for designing the genomes, specifically using a natural sequence as a template or a source of inspiration.

    And then, in this case, we chose PhiX 174 as that source of inspiration. Because of its really long history in molecular biology, being such a prominent phage, and also because it was small enough for us to feasibly synthesize DNA at that scale. And because also the decades of experimental characterization it has as a very safe, tractable model for phage biology. It just seemed like the one of the most logical starting points for us.

    And so we've kind of just worked around it.

  26. Claudia Driscoll

    But also just to add to that, I think Phi X is kind of like the perfect challenge for these genome language models because of how complex. You know, it's like an incredibly highly evolved sequence, highly entangled with, you know, regulatory elements. And not only does you know the model have to output a set number of proteins, you know, that can perform different functions in the phage life cycle. But it also has to be structured in the right way. The genes have to come in the right order and the regulatory elements have to fall in the right place. And I think that, yeah, beyond, I mean, it's hard enough, you know, just to think of how you would design a protein from scratch, you know?

    But being able to do this on the order of complexity of an entire genome really is like the ultimate challenge. And I think Phi, due to its small size, actually, you know, lends itself more of a challenge than something as simple as you know, like a cluster of genes or something like this.

  27. Sam King

    That's, like, the way we see it is. It's almost like you have a smaller canvas to work on. And you have to fit all of these different things with so many constraints, like internally, but also like these constraints. With the host cell and everything interacting with each other. And phi X has evolved such this, this extremely careful stoichiometry of the way its all of its genes are expressed. and so that really, really small working space provided a lot of extra challenge for us. And we think maybe if we were to move up to a larger genome that has a bit more wiggle room in things that are essential or non essential, we might actually have a bit more success.

    So it was nice having this really hard challenge, and it was almost like the perfect challenge for us.

  28. Joe Campbell

    Related to that, I’m curious in terms of experiments that you've done, I believe in your paper you isolated 16 phages that will infect the E. coli that phi X does. And I wonder, Claudia mentioned regulatory factors. Do you have you looked at whether there's sort of the classic early, middle and late expression of genes that sort of are the hallmark of, I'll say, natural phage phages infecting? Do you see that also with your 16 phages that your model generated?

  29. Sam King

    Yeah, we didn't, that's a good question. We didn't measure gene expression at that resolution.

  30. Joe Campbell

    I see,

  31. Sam King

    Our readouts were mostly like, yes or no, does this work? And then, generally, and it outcompete phi X and these kinds of assays. But what? Something I will note there is that many of the mutations that we observed in these genomes, especially the completely novel segments of DNA that we couldn't find mapping to anything in nature. These segments were in some of the non coding regions directly on terminators. And maybe it's.

    We're not sure, but it's probably affecting the expression of those genes in interesting ways that could affect, like early, late, or or late expression.

  32. Joe Campbell

    I mean, I think that, I mean, I don't know who you could if you wanna collaborate or you wanna do that yourselves. But I mean, I think certainly I would be. And I think a lot of people in the phage field would be really curious as to when you do that, do you get that? And I guess another question that would be, so if am I understanding it right?

    That you designed something like 300 phages, and did you make the DNA for all of them? And then you transformed them into bacteria and see saw which ones made phages and which ones didn't. Is that basically what you did?

  33. Sam King

    Yeah, that, that's exactly right. So we, well, we generated many, many more sequences than that. The approximately 300 phages that we actually tested were a very highly curated set of phages that we believe just. S. M. Many of which were phages we were. We had a relatively high degree of confidence that they might work. But then the most of them, actually, probably about two thirds of the 300 genomes that we tested were just really wacky or interesting looking. Microbiome day phages, where, you know, the architectures were really different, or there were complete deletions of large genes or small genes, and these kinds of features.

    And so, we were with the 300 phages that we actually ordered. We were really going for novelty, not confidence. So I

  34. Joe Campbell

    I

  35. Sam King

    think, the 16

  36. Joe Campbell

    see.

  37. Sam King

    phages that we that ended up ultimately working. It's possible we would've actually gone in many more than that if we had only tested phages that were all within a very high confidence range. But we were really looking at those strange looking ones.

  38. Joe Campbell

    So, do you think would it be possible? And what would be the utility of taking some of the phages that weren't in the 300 that you tested? And asking, you know, training the model and saying, Okay, here's the 300 we tested, here are the 16 that were. Can it pick out? Among the sequences you didn't test, would it get the right answer if you asked it to guess, like which one of the sequences that weren't tested were most likely to work?

  39. Sam King

    Yeah, it's so the model can. What it can do is calculate a likelihood of a sequence that you give it, and generally the likelihoods correlate with, I guess, model confidence and also just the naturalness of the sequence. So we do see that the phages that worked were all quite high likelihood from the model, but also the ones that worked were quite high similarity to PhiX174. And so it's hard to pick apart that. But a good thing to make is that actually there were around, I think, 50 to 60 of the phages that didn't work were in the same range of sequence similarity to PhiX174. And so

  40. Joe Campbell

    Right,

  41. Sam King

    it says something about actually how non trivial it is to design at this scale, across a whole genome, and how even maybe single point mutations might have made all the difference in what worked or didn't work.

  42. Joe Campbell

    Right, I mean, I guess what I'm getting at is maybe you would learn some rules, some further rules. If you took the things that were you had, maybe it sounds like you had something like 80 or 90 that you thought were the most likely to work, and 16 of them worked. Now, if you train the model that these are the ones that worked and these are the ones that don't.

    And add it to that training and ask it to once again sort of do the experiment again. Of picking out things that are likely to not only be phage like, but actually work, would the model be able to learn that and learn it? Would you be training it on some with new information? I guess is what I'm

  43. Sam King

    Yeah, that, that's a great idea. There are really exciting reinforcement learning methods in the field where you could do that type of thing, where you have this experimental feedback. And it would be amazing to get this kind of pipeline going. Where you generate, test, improve the model, generate more, and you kind of get to a point where you're able to, like, really confidently and with a high success rate, generate desirable qualities.

    This is a pretty, I think, new area, particularly in generative biology, but the methods from other areas of deep learning, I think, could be ported over. And really help this kind of space. Yeah,

  44. Jessica Sacher

    cool, yeah, I wanna. Yeah, I really think that getting my head around practically, like, what does it take to start doing that? If someone did want to expand on your work and use the Evo model, and can they like? Is it even in the realm of possibility that biologists can start asking these questions and like building on the models? Or is it still kind of in the research area? Like the reinforcement learning people, the deep people in the machine learning side of things, have to, like, get that working with biology. Like some, like, figure out, how do you speak, you know, results to the DNA model, cause, as I imagine it, it only speaks DNA. But like to give it words that aren't DNA and get it to learn, like, is that part figured out?

    Or are we still in the area where that needs to get figured out before, like, biologists can go? Okay, I? I tested a bunch of sequences that Evo said to test and I have the results. They kill these strains. How long till I can, like, put that information into the model's brain?

  45. Sam King

    Yeah, it's. It's definitely really early days. This is very much still a research field and there are, I think so. A. A big contribution of this that that we feel is important about the work is the showcasing that the computational methods are capable of producing novel phage genomes. And the strategies that had to go into that, so fine tuning and these other, post training strategies.

    Where we had to come up with many different ways to evaluate the phage sequences and and then all of the experimental strategies on how to actually go from generated DNA to viable phage in the lab. This whole framework, I think, is just the start, and it's really part of an existing toolbox of ways to generate novel DNA sequences and also testing them in the lab. And I think we would like to see from now, I think, methods that especially help with evaluating the sequences.

    So the model, you know, in a single day, maybe generate a million sequences. But as the scientists, how do we actually tell which ones, which of those sequences are worth spending money on synthesizing? And the heuristics that we came up with, I think helped get us there a little bit. But it's very, very primitive. And hopefully it can become more automated and require less intervention and time from the scientists in terms of figuring out what looks good and what doesn't look good. And so there's cool work happening in the field around this concept of agents, where these much more automated models can go about.

    And essentially do everything that you would do, but they just do it on their own and navigate all of the data themselves. And so I think with the development of agentic frameworks like that, and also maybe better ways to better interfaces, especially on how to, like, look at sequence data. And

    use all of the bioinformatic tools that we had to use to, like, quickly evaluate those sequences. There, and there's, I think, also very, very cool work very recently coming out on that type of thing as well.

  46. Claudia Driscoll

    I think just to add to that and to, like, give credit, you know, to the paper, what you know, in its current state. Samuel did, you know, an excellent job at really outlining, you know, a protocol and kind of a set of design, you know, intuitive design steps. That allows you to guide these genome generations towards, like, a very targeted goal. So in this case, you know, we were trying, or Samuel was trying to generate phi X like phages that can infect E.

    coli C specifically, you know, like one host that phi X can infect. And I think Samuel, like in the paper, outlined a very logical, you know, set of filters. That would allow you to both screen the generated genomes for things that look like genomes, things that can replicate, things that are, like, you know, would would be realistic in biology, but also things that look like the phage that you're trying to to generate.

    And I think using that logic, and this is kind of like, actually quite a key innovation in the paper itself. To be able to use that logic to, for example, apply that same method towards a different, like type of phage, or a different, designing phages towards a different host. And I think that this wasn't kind of like, you know, a very open ended design task.

    It was quite focused. And the current state that the, you know, the field is in, could allow you to apply that same logic towards a different kind of design task. And so I think, yeah, I mean, Samuel's obviously correct, this is just the start. But I think I'm, you know, trying to give him credit where it's due. And, there's actually a great innovation here. And being able to, even in the current state, you know, being able to design things or understand how to design things. That look like real genomes and also look like real genomes that will infect a particular host, I think that's, yeah, important. And good kind of stepping stone for people to kind of build off for when they wanna apply this to other kind of phage design tasks.

  47. Joe Campbell

    Just to make sure I completely understand, because what the prompting was to make something that looked like Phi X, but there was no part of the prompt that was make something that looks like Phi X. That will infect coli. It was just asked to make something that looks like Phi X. And the presumption was that if it succeeded in making something look like Phi X, there would be a reasonable chance that it would infect coli, right?

  48. Sam King

    Yeah, yeah.

    So that's exactly correct. The prompt that we ended up, or we used several prompts, but the strategy there was: we got the consensus genome of natural Phi X variants, and then we used the just anywhere between like four to ten nucleotides of the beginning of those cons. That consensus genome as a prompt. We wanted the smallest prompt possible, so that most of the genome generated was completely from the model and not from our prompt.

    And so four to ten nucleotides was about as small as we could get it. And then the rest of the over five zero base pairs was all just completely generated by the model. And also by using a small prompt, we kind of were able to find a sweet spot in the generations where they looked like Phi X, but also they were quite diverse. And you could imagine if you used a really long prompt, then you.

    You're just giving the model too much context to the point where it's just going to completely memorize or recall the natural Phi X genome. So that kind of careful prompt engineering allowed us to generate natural looking yet diverse sequences. And then after generating sequences, we were able to then determine which ones are sequences that might have a higher chance at infecting our target host E. coli C.

    And so the way we did that was we decided because the spike protein of PhiX174 is so important for actually binding to E. coli C cells. We filtered out genomes that had spike proteins that weren't very similar to PhiX, and we only kept the ones that had very high sequence similarity spike proteins to natural PhiX. And so I think by setting that kind of constraint on the generated genomes after they were generated, we were able to pretty quickly filter down.

    Which ones, were more confident? And then also the rest of the context that comes with that. So if you, if you filter for sequences that have a high similarity spike protein, you're also going to just consequentially with that, get other parts of the genomes that are probably more similar to PhiX. And so it was a pretty strict constraint. And I think could serve as a strategy for other phages by filtering on those those more tropism, post tropism specific proteins.

  49. Jessica Sacher

    is really cool to think of it. I'm glad you drilled into this, Claudia, about thinking about like the whole work of filtering as like the next step after the prompting step. And it's not just like you made an AI model and then you clicked go and, like, these ones are aligned with Phi X, and then you made them. It's like, they're, you are giving it like such a small prompt, like four nucleotides, like a t, c, g, and then it's going, okay, no worries. I got it for the entire genome. And then you now have, probably, I don't know how many sequences you had at this point. But like, you have to come up with what you're gonna filter by.

    Like, it's not like you could say make Phi X, cause you can't, you can only say a t c g or a t t t. And somehow, like, how did you even approach

  50. Jessica Sacher

    what you would filter on. And then like, was that? Did you always know that was gonna be kind of? It sounds like that might have been the biggest like lift, but did you know that that was gonna be going in? Like, is that known to machine learning people that like, oh, that part's gonna be the thing? Or cause? I don't think we have that concept from the outside of, like, what that work is?

  51. Sam King

    Yeah, that, that's such a great point that it's so. In other areas of machine learning, generally, the outputs are much more intuitive or interpretable for humans.

  52. Jessica Sacher

    Hm,

  53. Sam King

    So if you're generating an image, you can look at it and pretty easily tell, Okay, is this an image of a cat or a dog? For us, since we're generating DNA sequences from the model, we don't know right away, just by looking at them, if the model is generating things that we want. And so we have to implement all of these different bioinformatic metrics in order to tell, is this a good DNA sequence or is this a bad DNA sequence?

    And so it was a very big lift, just trying so many different ways to evaluate the sequences. And then coming up with a specific combination of those methods to choose. Sequences that were very diverse yet realistic. And we ultimately landed on the set of constraints that we described in the paper. But there are many, many more ways that we could have filtered these genomes or looked at them, maybe with even fewer filters or more filters.

    And I think that's where the really fun part, as a biologist, comes into play, is you get to choose these ways. You want to evaluate the sequences and implement your understanding of DNA to get to the final product. But the earlier part is just trusting that the model is good enough to generate desirable DNA, and then what will you actually do with that DNA once you get it from the model?

  54. Joe Campbell

    I have a naive question, so you took, if I understand it correctly, you took roughly 300 sequences, you made them in a test tube and then you introduced them into the bacteria. Now, I could imagine at least three possible scenarios. One, it just goes in and doesn't do anything. One, it goes in and, like the 16 of them, it makes a phage, and when you put that phage, it can now infect the strain that you put the DNA into. I could also imagine a scenario where if the DNA is introduced by transformation, it can make a phage particle.

    But that phage particle doesn't recognize that bacteria and therefore isn't a phage. So were there, beyond the 16, were there pieces of DNA that made phage particles that could not in infect the coli strain that you put them into?

  55. Sam King

    Yeah, the it's. It's hard to know, because our initial screen of all 300 genomes was mostly this binary readout of does the E. coli population die off or does

  56. Joe Campbell

    I

  57. Sam King

    it?

  58. Joe Campbell

    see, right,

  59. Sam King

    It? It is conceivable, though, that if we were to, if we had the capacity to test these 300 genomes in many, many different bacterial strains, we would have gotten other hits that we didn't get. Just in E. coli C. We did also screen all of the genomes in E. coli K12. But beyond that, the initial screen of all of the genomes was just those two strains. And then we tested the 16 hits against eight different strains of E.

    coli.

  60. Joe Campbell

    right? But I guess it's possible that if you went back to some of the phages, some of the sequences that looked most phage like that. If you tested like ten of them, you might find some that, like, made a bunch of phage particles. Because once they got in, they could make a phage, but that phage couldn't couldn't go back and recognize it.

  61. Sam King

    yeah,

  62. Joe Campbell

    Ah, and I wonder

  63. Sam King

    I, I feel like,

  64. Joe Campbell

    if those would be interesting in their own way. Cause that might get sort of at the issue of like, well, I guess. With Phi X, it's probably already known what the receptor is and what the s, what the receptor on the bacteria is, and what the what, what the phage is recognizing,

  65. Sam King

    yeah. I think that is probably almost certainly the case that there were generated phage genomes that could produce something in the cell, almost definitely, like many, if not all of them, were expressing something in the cells. Whether or not you can get this extremely complex assembly of all of these hundreds of subunits together in perfect geometry and stoichiometry. And then interacting with different elements in the host cell, and then ultimately bursting from the host cells and infecting others.

    It's just like there's so many steps in that lifecycle

  66. Joe Campbell

    right?

  67. Sam King

    that something could go wrong.

  68. Claudia Driscoll

    yeah, I actually we had, you know, when we actually last met with you, Jessica, we had this case where we had a promising looking phage candidate. That had, you know, shown some signs of life, both in plaque and in liquid culture. But we were not able to propagate that phage, and that's when we were troubleshooting with you, Jessica. We were thinking of all sorts of different ways for, you know, really trying to.

    At different temperatures, different, you know, humidities, different agar percentages and things that we could try and encourage this phage to, if it was having difficulty, you know, entering it, entering its host. As you say, it could have been creating viable particles, but then those were just not able to infect host cells. And eventually, you know, we had to kind of cut our losses and move on.

  69. Jessica Sacher

    had to give up on that one?

  70. Claudia Driscoll

    That's still one that's close to our heart.

  71. Sam King

    That that's a good insider perspective that that hasn't been shared very publicly is, yeah, there were. There were multiple candidates throughout the course of this paper where, but we spent probably months of our time pursuing them and seeing. Like, it just seems like it's almost working, but then

  72. Claudia Driscoll

    but I do,

  73. Sam King

    it just couldn't be propagated. Or,

  74. Claudia Driscoll

    I. I definitely think like, if we were to go back and look at those candidates. And, you know, assess whether potentially they are making particles that are just not able to infect host cells. Or something like, there could be some really interesting mechanis mechanistic insight, in why those were not able to be propagated. But it might also be something as simple as different salt concentrations when he did it or something.

    But, you know, within the realm of time, one has in a scientific…

  75. Jessica Sacher

    Yeah,

  76. Joe Campbell

    I mean, I wonder how hard would it be to do something like, take all 300 of those and infect them and have a, you know, a qPCR assay? Just is the DNA replicating? Just a really simple like, you know, can it make? Can it replicate its DNA? Because a lot of pieces of DNA would would fail that test. And I just wonder if you would learn something like, I mean, obviously it's not as complicated as making a phage, but just reprogramming the cellular machinery to replicate the DNA is a lesser ask than making a phage particle, which can reinfect.

    But it's sort of like helping you, sort of like, okay, these pieces of DNA can do that, but they can't do the whole shebang. I mean, I guess it's just, you know, trying to understand what steps things can and can't do. Certainly has been like a tried and true thing that people do when they do sort of classical genetics. And trying to understand, okay, I have this piece of DNA that, you know, it can't do it, it can't do everything, but it can replicate.

  77. Sam King

    Yeah, I really like what you're saying, like this is. This gets at one of the things that are closest to our hearts about this. This work is that it's. It's an entirely new way to study genomes and evolutionarily understand what's going on and how different parts of these genomes work or don't work. And traditionally, it's through what you're saying. We're. We're much more used to taking a natural genome and then perturbing it or mutating it in different ways and seeing what happens from there. But in this case, we have all of these very evolutionarily distant genomes generated by a model that's seen most of observed evolution. And giving you candidates that way. And then going back and like, reverse engineering some of these changes, or and just probing, what is even going on?

    I think it's such a new way to study deep evolution.

  78. Jessica Sacher

    Yeah.

  79. Sam King

    Yeah,

  80. Jessica Sacher

    And like the idea of, instead of being like, Oh, I wanna make a mutant, and which gene do I wanna make a mutant in? It's like, I get to design the filters, what should I filter for, or what should I select for? But not like you're putting it on a petri dish with an antibiotic and that kind of selection. Like, what should I choose as, like the creator of all these genomes to let let get to the next step?

    And just like the idea of seeing this, like filtering, I think it was called something in the paper, like inference, time filtering. Is that right? Like, I was like, skipped that part, like my brain just died at that part. Cause I, you know, just looking for biology, I guess, but like, I don't know. I'm now eye open to the fact that that's like, like, as biologists, like, that's your selection and you get to design and be creative.

    That's the chunk of where you can have a lot of. That's where you ask your questions. Maybe it's like, Yeah, you showed us, like you guys showed us how you filter to get functional phi X by like, yeah, the RBP being like. So he decided that maybe that's critical, and like, you know, that to simplify it all. But we could decide, oh, I think this gene of this phage is critical. I'm gonna only filter for this and then A/B test that against Oh, I'm not gonna filter for that.

    And I don't know. Like, feels like a really something that didn't come through when I read the paper. Cause I just assumed it was some machine learning stuff that didn't apply to me as a biologist. But that seems like a very cool area and like a totally different way of doing science. and just like, Yeah, this is all just. I see it as a whole new tool

  81. Jessica Sacher

    for us, but it's like a tool on, like the meta level. Yeah,

  82. Jessica Sacher

    yeah,

  83. Claudia Driscoll

    also kind of love that. You don't need to have a complete understanding of how the entire phage life cycle works. Or, just, you know, phage biology. But you have a good understanding that, you know, spike protein is necessary for host binding. And hm, and you know, infection. And that, you know, in a very simplified form, is like, could be all you need, you know, to produce things, you know, produce viable phages that are going to infect the thing you want it to infect.

    I also think, like, I just wanna add that having this very neatly outlined filtering pipeline is really important. Also from like a safety perspective, because you're able to, you know, know with a pretty high confidence. Be ordering genomes and testing them in a lab, that are going to be very constrained to, like, one host that you want it to infect. And I think that was important from our perspective, you know, conducting this project.

    Is, you know, these were very carefully filtered and curated, and it wasn't just ordering, a bunch of random sequences. And I think that's actually important for future applications of similar, you know, design tasks. Is that you, yeah, you know, you have these safety filters, I guess, in place to ensure that you're designing curated things.

  84. Jessica Sacher

    yeah,

  85. Sam King

    I agree with everything that's been said, and yeah, I want to add that the filtering isn't the. It is definitely one of the main parts that a biologist would will get to implement their knowledge here, but it's—we tried to set it up so the entire genome design framework was modular. So going from choosing the host to which phage you want to be inspired from, and the phylogenetic data you might use to fine tune the model, and then the steering strategies. So how you're pointing the model in different directions to generate what you want. So that was for us, like using different prompts and then all of these filtering metrics and constraints. And then, yeah, and then also, like, you know how you will conduct the experiments.

    Like at every part of the pipeline, there's decisions, design decisions to be made by a biologist. And hopefully it will become clear in the future that it is quite modular. That w. That was really what, what we wanted with this. And yeah, it would be super cool to see, another group sometime soon, maybe doing a different phage in this way.

  86. Jessica Sacher

    yeah, that's so cool. And I feel like the other aspect seems like, and maybe this is part of the modular like pipeline that you're laying out is like the whole lab. Part. What you choose to use as fitness tests for your phages like that seems like another huge block of creativity possibilities. Cause you guys also came up with a bunch of methods we don't usually see in phage papers, but like, make a lot of sense. Like, how many generations did it take for the phage, for the strain to escape the phage. Like, and counting that and having that as your metric, like, I just, I don't think that that's usually what I see, but why not?

    And

  87. Jessica Sacher

    yeah, just like the single phage, like, oh, like how you kind of for each of the different phages you generated. Like the difference in just like very fine little metrics on it, like time to burst or whatever, or like rate of OD per second drop, like, I don't know.

  88. Jessica Sacher

    I don't know. I love that.

  89. Claudia Driscoll

    I think

  90. Jessica Sacher

    You kind of could get a. You could design whatever fitness test you want, and if you had a whole battery of them and they're seem very high resolution. Like they could detect differences that are pretty close. Whereas we might just do a host range test and be like, does it kill E. coli or not? Like, or like, pretty much quickly or not, really like. But getting much more, granular with your metrics that you design in the lab seems like you could also start to pick up small changes. And like, that's just a whole other creativity block for the biologists, too.

  91. Sam King

    Yeah, yeah, we—once we had our 16 viable phages, we wanted to like we. We technically could have just stopped there, like we had functional phages from an AI model, and that's that. But we really wanted to see what kind of phenotypic diversity was in these genomes, beyond the genomic diversity. And okay, yeah, and I mean, before those assays, we did characterize the genomic and structural diversity quite deeply. We saw the structure of one of the phages that used a DNA packaging protein in an interesting way.

  92. Claudia Driscoll

    I think, though, you know, both of us as we were working more and more with these 16 phages and propagating them and performing different assays and things. We really started to, you know, just observe that they really did have very different characteristics. And we kind of, like, called them personalities, you know, of the different phages. And so I think that as well encouraged us to do kind of more further phenotypic, you know, analyses on the different phages.

    Cause it was very clear to us, just in propagating them over and over again, that they really did have clear, different differences in how they were killing the host. Like the speed and you know, the density that the bacteria would get to before you start observing killing. And I think, it was actually really fun, it was like a really good, fun, you know, platform for us to just be creative and investigate phage biology, you know, I think it was like, kind of just,

  93. Jessica Sacher

    yeah,

  94. Claudia Driscoll

    yeah,

  95. Sam King

    Yeah, there's a fun graphic in one of Claudia's lab meeting presentations where she generated,

  96. Claudia Driscoll

    it's

  97. Sam King

    it was kind of like a Pokemon card for

  98. Jessica Sacher

    Love it, put that in the paper, throw it in the supplementary,

  99. Claudia Driscoll

    we'll start selling them.

    But I think that was like, because we did see these clear differences to like, how. What wild type Phi X was behaving, despite, I guess them being like, on a sequence level, SIM similar to wild type Phi X, they clearly did. They were behaving differently.

    And so I think, like, that's why, it was important for us to do that further investi investigation into how they were differing. Because small changes in the genome really do make can make a big, a big difference in how they behave.

  100. Joe Campbell

    One thing that you might find interesting to do is if you take phi X and the 16 phages that you generate, so maybe you've done this. But if you just do the old fashioned geneticist experiment of taking the bacteria and selecting for resistance, like you select resistance against phi X. How many of the of these do you get cross resistance to? And if you go through and do that with each one of the 16 and select resistance, you know, this is sort of like, I mean, I'm just the old genetics playbooks. But I think people did learn a lot from those, and there's been a lot of understanding of those. And it kind of gets to possibly like what they're recognizing on the cell surface or something like that, you know? So do you have any idea about if you select for resistance against one of your phages?

    Does it tend to be? Do the other 15 tend to be? Does it tend to be resistant to the other 15? Or do you get cross resistance? I guess, yeah,

  101. Sam King

    Yeah, we haven't done a cross resistance screen at that scale. I guess that would require 16 phages, plus wild type Phi X times at least 16 or 17 resistant strains, and then doing the entire matrix. What we did do was evolve, many resistant strains against wild type Phi X, and then this data didn't make it into the paper. We, but we individually tested each phage against those strains and on just the initial test.

    They didn't weren't able to infect the Phi X resistant strain. And that's when we had the idea of what might it look like creating a cocktail of these phages and then seeing if the cocktail has some kind of evolutionary resilience over time against the resistant strains.

  102. Joe Campbell

    and did. Did you sequence the phages from the cocktail? So you saw it, the ability to overcome it. So I presume that there was at least one phage that was generated.

  103. Jessica Sacher

    yeah,

  104. Sam King

    Yeah,

  105. Joe Campbell

    Did it look like one of the 16 that just had a mutation? what, what? what were those resistant phages? Did you sequence them?

  106. Sam King

    so what we ended up doing was over the multiple passages against the resistant E. coli strains, we had this final phage harvest of the resistant phages. And we took that tube and we streaked it out on a plate and got individual plaques, and then we sequenced many plaques from those plates to see which genomes tended to be the most predominant from the resistant culture, and for generally, it was like one main genome for the different resistant strains. Where it was, it looked like a phage that was kind of a re recombination of multiple of the generated phages, two or three of them into a single, I guess like super phage genome or something where it was like harnessing the power of each of these different phages in the cocktail into one recombination.

    We can't say for sure that a recombination event happened, but based on the sequencing data, we were able to match different segments of the genome perfectly to each different.

  107. Joe Campbell

    would those phages able to? Maybe this was in the paper, but still able to kill the original strain, or would have they lost the ability to kill the originals?

  108. Sam King

    Yeah, they could still kill the original strain too. So we were always doing the assay where we were at the same time challenging them against the original susceptible strain and then against the resistant strain.

  109. Jessica Sacher

    Cool, yeah, that's so cool. I think we probably should get towards wrapping up, but I feel like one thing I don't wanna leave people on a scary note. But well, it's Halloween week. But you know, how much does it cost to do all the synthesis that you were doing to do this experiment? Because I think that's a core one of the first. I wanna mention it as like a relevant reason why to stay in a very small phage, like a 5KB phage, people go and say, Well, why didn't they do a bigger phage? How much does it cost? Every time you wanna print one of these candidates, like the 300 that you printed? And

  110. Jessica Sacher

    then I think we were talking about, like, cause I'm one of those people being like, let's do Pseudomonas phage. I have a 260 KB one and, you're like, how much do you wanna invest? Cause it's like many thousands per candidate once you start getting... So where is it, ballpark, right now, and how do you get those sequences? I think through Twist right now.

    Yeah,

  111. Sam King

    Yeah, we worked with Twist to synthesize these genomes, and, they were really gracious with us. But generally, I th. I think it was on the order of, like, hundreds of dollars per genome, and all of the genome it was. It was tens of thousands of dollars. Hopefully, DNA synthesis will only get cheaper and cheaper from here, and this will become like, much more feasible and accessible.

    Definitely, like something like a jumbo phage is a big endeavor, especially if you wanna test many different jumbo phages. That's the same problem I think that the synthetic genomics community faces with synthesizing their large scale DNA. And so,

  112. Claudia Driscoll

    I will say we did think about and plan out more kind of cost efficient ways of synthesizing genomes. On the scale of, you know, Phi X, where you can, you know, fragment your genome into smaller chunks and order those chunks as like an oligo pool or something, for example, which is definitely much, much cheaper. And if you can think of like clever ways of assembling those via, like Golden Gate strategies.

    And that's like, you know, a very feasible way of cutting costs and also being able to assemble genomes on the same scale.

  113. Sam King

    yeah, there are good protocols out there already to do like combinatorial Golden Gate assembly from very cost effective oligo pools. So, but we would definitely recommend that people think about those

  114. Jessica Sacher

    Okay, yes, that is good to know that there's like, some other avenues. But I feel like if we have pressure on this part of the I don't know funnel, like that's good, because then it'll incentivize people to work on this. And I'm sure there's companies trying to make synthesis cheaper. Cause obviously phage world is not the only one who needs it, which is always helpful. So okay, but yeah, it sounds like we can do a ton. Even before we start printing genomes on the on, like just the generating sequences. And like looking through them and filtering them down, and like getting sort of a score around how, or just like looking at them to see how much are they like a phage? And what is weird? What's the pattern that they're coming up like? It sounds like we can do a lot for free.

  115. Sam King

    Absolutely, yeah.

  116. Claudia Driscoll

    Yeah,

  117. Sam King

    And if you're willing to like, you know, test fewer, fewer designs or fewer mutations, like you could use some phage DNA you already have in the lab as a starting point for your cloning process and go. So there, I think there are many, many different ways that, many smart researchers will be able to come up with to make this more accessible.

  118. Jessica Sacher

    Okay, awesome. And I just also wanted to point people if they really wanna get hands on to try the Evo tools. Or, like, what would they see? I should have done this before, but I've been to your website where you have the actual like, what do you have online that people can use? Is there like a browser based tool that you can already kind of play with, or what?

    What's available?

  119. Sam King

    Yeah, there's a browser, so the Arc website, I think Evo like in a very nice user interface where you can just put in your prompt and like, get generations, or you can score sequences. I think there is a limit on the length of sequences. And but if you want to do more hardcore generative biology with these models, everything is completely open source. The models are uploaded on Hugging Face. And then they're, all of the code is on GitHub as well, so you can install Evo in your environment and go from there. So I, it's all very, very accessible, and if people need help, there's also great documentation out there, or they're definitely free to email us too.

  120. Jessica Sacher

    awesome. I hope that we can get a small army of phage people that are curious about the AI biology interface like me. And try to, yeah, like, get the all the tools are there. So you can start playing with it and wielding it in your own hands. And then asking your PI to get some grant money to pay for all the synthesis. But I think they sometimes get excited when there's like a big, expensive thing to put on a grant.

    I don't know, it's kind of like something to shoot for.

  121. Sam King

    Yeah, and this is, it's, yeah, again, like this is just the start. We're really excited about this method for, you know, as a computational strategy. And developing all of the concepts and ways to go about. Phage genomes and other genomes broadly, and for studying evolution, and hopefully this can be implemented in phage biotechnologies or phage therapies. And yeah, it is part of a larger toolbox that's available to phage researchers and other genomics researchers. And I think just being aware of what these AI models are now capable of, and thinking about how you can use them in innovative ways for your tasks.

  122. Jessica Sacher

    Yeah, yeah, I think that's a whole new realm. It's not just like, Oh, you make a model, like, it's like, Okay, there's a model, use it. And I think that's a whole new, brain space. And just like new set of skills for us biologists. But thank you for what you have done and what you've put out in the world. And like, for this framework of how to think about even starting, like, I think it's a huge service to just everyone in phage field, but the entire world.

    Or this is like, very pioneering, like, legitimately pioneering, not just like Chatgpt, fluffy pioneering. Like. I don't say this very often, but I think it's really, really amazing what you've done, so thank you, Joe. Do you have any parting remarks?

  123. Joe Campbell

    No, thanks for talking to us and entertaining all of our off the wall questions.

  124. Jessica Sacher

    Angry phage community doesn't think this phage is big enough. Prove that it's cool. You have, I think, convinced at least a whole bunch of them with this. Cause. Yeah, cool, awesome. Well, I will see you around. And if you ever wanna resurrect more half dead phages, I'm all ears. And yeah, looking forward to see what you guys do next.

  125. Claudia Driscoll

    you so much for having us on. Yeah,

  126. Sam King

    this was a ton of fun, thank you.

  127. Claudia Driscoll

    hi,

  128. Jessica Sacher

    Yay, okay.

  129. Joe Campbell

    Bye, bye, nice to meet you.

  130. Claudia Driscoll

    Great to meet you.