2026-05-01-nextflow-summit-day-2

Date: 2026-05-01 Source: nextflow-2026/transcripts/sessions/2026-05-01-nextflow-summit-day-2.md Word count: ~25,571

Auto-generated captions

This transcript is from auto-generated English captions on the live event stream. Expect misheard names and technical terms (e.g., "Vicksville" for "Nextflow"). Use search for exact quotes.

Meeting Title: Nextflow community training and ambassador program overview Date: May 1

Transcript:

Them: Sharing with your colleagues there, really going deeper into And when she joined Securis, she took one look at our training material and said, we can do better. So we're gonna hear about some latest news from the NextFo community with Geraldine and Martha. Thank you, Phil, for that very kind introduction. Phil knows about wearing the many hats of the community team because he used to run the community team I got his job by fighting him gladiator style. That would have been fun, but he's so tall. Anyway, so no, I did not fight him, but he very kindly let me take over his job and has been fantastic journey. So yes. And so Marcel and I are going to talk about some of the activities of the community team. I'm not going to give you a laundry list of all the things that we do. But today we are going to focus on, two things specifically. Hello? Yes. Nope. There's next one. Ah, there we go. We're going to talk about training, and Marcella is going to talk about the ambassador. Program, which is a really cool program. So this is kind of split half and half between for the community and by the community, to some extent. So let's talk about training. As Phil mentioned, I'm quite passionate about training. I do believe that is one of the best ways we can make people successful. With NextFlow is offer quality training that helps you get started, that helps you learn what you need to be successful with the tools. And so over the past couple of years, we've been putting a lot of effort. And it is very much team effort. So it's collaboration between our little community team and the in house bioinformatics team, that you might have heard, referred to as scientific development team, You saw several of them give demos yesterday. You saw Florian's amazing talk about his work with, using coscientists for the protein prediction. Those folks are all, on that team. And several of their members have been especially helpful in, helping us update the training materials. So what I'm going to do is kind of give you an overview of the resources. That we provide, and I will talk a little bit about how we maintain them up to date and so on. There's some lessons learned here that might be of interest for some of you. Who yourselves might be involved in maintaining training materials, in developing training materials, and training users of maybe software you're involved in developing or maintaining. Okay. So this is our training portal. It's everything I'm going to show you is freely accessible. All the time. It's open source under a Creative Commons license. So anybody can use that for self training, but also to teach others and to use as a starting point if you want to develop your own materials. Now, the goal is, in the spirit, of not reinventing the wheel as much as possible to provide a resource that hopefully caters to the needs of many. That's if you need to train users at your institution or if you want to help a colleague learn Nextflow or something like that, you don't have to develop your own materials. You can, if you have your way of teaching it, and that's totally fine. That's great. But these resources exist so that you don't have to if you don't want to. So, that's our training portal in general. I'll walk briefly through the main courses that we have, that we provide. They're primarily developed to be used as a self -service thing, but they can be used for classroom teaching. We do that regularly. Those of you who took the trainings earlier this week will have, worked through some of that. They use GitHub code spaces. So we provide a training environment that's fully loaded, that's got everything installed. You just open it's freely accessible. It's run by GitHub. You have some quota that is typically sufficient for working through the trainings. And you can spin that up. Or you can spin it up locally by yourself. We have some of the instructions if you want to do it locally. But the idea is that we provide this training environment so that you don't need to worry about installing anything. You can just go ahead, get started, focus on learning NextFlow. And that's really the point. And I'll show you what that looks like in a minute. I always also want to mention that all of this is available in 11 different languages, so English plus 10 others. And there were supposed to be flags on my slide, but they've all been replaced by the code except for the Catalan flag, because the Catalan flag does not exist as a preload in emoji. You can't get that one. So that one I had to go and copy paste the picture. And so the Catalans have the privilege of getting their flag when everybody else just has their country So there you go. Congratulations. As a reminder, Siqueira is based in Barcelona, which is in Catalonia. So I guess it's fitting. All right. Yes, that's our training portal. What's in it? Here. Do I need to do some kind of invocation to make it progress? This is very bioinformatics. Do I need to turn it off and on again? Oh, no, no, far too far, far too far. All right. Oh, look at it. He has a sneak preview. All right. All right. No hiccup there. Sorry. It's Friday morning. All right, so we have a number of courses. We have two main tracks. Like, we loosely organize these things as intro and advanced track. But really, we have the resources that are the most mature at this point NextFlow for newcomers, that is we have kind of our flagship intro course, Hello Next Flow, six parts. It takes about a day to teach in a class room. You can go through it faster if you're doing it on your own, of course. But that walks you through all the basics of the key things you need to know about NextFlow, but not more than you need to know. So it's very focused on having a gentle introduction to all the pieces, but in a way that's very goal oriented. You learn to develop a workflow step by step. Every step builds on the next one the previous one. Like a workflow. So that is really our flagship course for beginners. Please recommend it to your friends and colleagues if you have people who are interested in NextFlow. We have Nextflow Run, which is a short version, kind of a bridge, less focus on code, more focus on running things, but with still some explanation of how it works under the hood. Because I believe in looking under the hood, even if you're not ever going to develop code yourself, it's useful to understand how it works, what are the main because when something goes wrong, it helps for debugging if you understand what a process is in the first place. So it's kind of based on that kind of philosophy. We have a few courses, a small growing portfolio of courses in the Nextflow for Science category. Where the idea is that all the newcomer materials are domain agnostic. It's all designed to be usable for someone who doesn't have any domain specific background. So it's not specific genomics or NANC or anything like that. We have people using Nextflow for material science, for astronomy, So this is appropriate for anyone from any backgrounds. But then Nextflow for Science shows, okay, this is how you apply the principles you learned in the basic, inter course to a specific use case, like genomics, like RNA seq, We have a bioimaging one. We have a metagenomics one that's in development. Thanks to some contributions from the community. So that's something that we're going to develop further. On the advanced side, many of you already know NextFlow or already have some familiarity. You're already developing with the tools. So there we have some resources that go a little deeper. Hello, Next Flow is the introduction to NFCore. We know that getting to grips with NFCore can be challenging. There's a lot going on in there. So this is a course that in the same spirit as Hello Nextflow walks you through gradually how to learn how to work with that. And we've had people who have been using Nextflow for years come to us and say, Okay, this the NFCore course really helps because now I understand how NFCOR pipelines are structured and so on. So even if you're already very familiar with Nextflow itself, you might be interested in checking that out. We also have more advanced training that goes deep into specific topics like metadata handling, like composing workflows, debugging. There's a ton of topics. They're called side quests. And you can take them in no particular order. If there's a particular topic that you need to level up on, to progress in your project, you can browse that. And hopefully that will help you. Anyway, a bit long winded, sorry, but the this is the catalog of training that we offer at this time, and we are very actively continuing to develop that further. And we're very happy to hear feedback from, like, what's missing? What's what are the topics that you feel like we need to cover? Okay. In practice, what it looks like. This is just these are two browser tabs shown as one within Chrome. But basically, we have the instructions. This is to show you kind of how you would work through it. You have the instructions on the web page that detail what to run, that show you expected outputs, etcetera. It's very step by step. And then on the right side, you have the that's the GitHub code space. That's loaded by default. It gives you a Versus code. Interface, and you have everything that you need You have example code. You have starter scripts and so on. And you just basically run through the instructions in the code space, and everything just is designed so that you can work through the exercises as you go. All right. So that's what it looks like in addition for Hello Nextflow specifically, Phil has recorded an amazing series of walkthroughs, basically showing in practice working through all of these courses and it's it's an amazingly valuable resource so thank you, Phil, very much for putting that effort in. It takes a huge amount of time to do that. So it's very useful. And there's a lot of additional little commentary and tips in addition to what's written in the instructions. People seem to be pretty happy with the course. This is the kind of survey that we take for Hello Nextflow. This was last year. This year is shaping up to be even better with the latest changes we've made People are generally really thrilled. I'm not saying that to brag. I'm saying that's just say, like, please do use these resources. Please do recommend them. To your friends and colleagues, because we do think that this is a way to get people leveled up, with a minimum amount of pain and maximum amount of results. Okay, I'm So really important part of this is, as you can imagine, fairly sizable amount of material. It's a lot of material. And as you might have noticed yesterday, the pace of development in next flow is pretty unrelenting, right? Paulo and Ben keep coming up with new features every six months. It's quite fast. You feel like it's fast for you? It's fast for us too. So we have to update the documentation. And because of how not just documentation, but training materials, because of how they're built and they're step by step and everything builds on the previous step, you change one thing. You add, like, for example, workflow level outputs. You add that to the starter script. Now boom, domino, you have to update every single snippet of code across dozens and dozens of pages. And have I mentioned it's available in 11 languages? It's a lot. So we have this year started using Claude for maintaining and updating the material. And it has been transformative in terms of enabling us to very quickly adapt the materials. There's still some work that we need to do in terms of designing, deciding, for example, at what point do we introduce a new feature? But there's a lot of the work of just going and checking all the updating code snippets, checking that everything still works, regenerating expected results in the solutions All of that now we can handle with cloud skills. And I want to shout out Jonathan Manning, from the side of team who has done a really cool amount of work to set us up with this. Because now I can go in I can make some changes to their tutorial and just ask Claude to rerun everything, update all the solutions, It is phenomenal. I cannot overstate how useful this is. And if you are maintaining any kind of training materials users or something like that, I really recommend you look into that. You check that. It turns things that would normally take hours into a task that takes minutes. It's days to hours. It's really transformative. So I'm not going to go too far into detail. Because I'm eating up all the time, but I know we have a little bit of buffer time. So I'm taking advantage of that. I hope you don't mind that I'm holding you all hostage, but I really want to get this across. The other parts of our use of AI for the training material updates is for the translations. And that is Phil's work, when I said he was irrepressible. While he was doing all the rest stuff, he still found the time to set us up with an AI power translation harness so that now okay, we work on some development branches, We make a bunch of updates to the courses. Now we're ready to release. All the updates. Do we have to go and update all of the translations? No, we do not. Because we just run a GitHub action. There's literally a button that I go click that says run workflow. And it'll work for, I know, it's a few dozen minutes, I feel like. It doesn't, it takes less than an hour. It goes through, it looks at what's at the diffs for what has changed. And then it will go update translations across all, 10 other languages. It's fantastic. The key principle on that it's really clever in terms of saving on tokens and saving on processing time. There's two things, right? One is it just retranslates the areas with the diffs. Don't retranslate everything. And part of it is also you don't fix the prompt you don't sorry. It's the opposite. You don't fix the file itself. So if you notice something is a little off in a translation somewhere, you don't make a fix to the translation itself. You fix the prompt so that next time, every time, it'll be automatically correct. And by that, mean, the kind of things that you can do are improve the grammar, improve the style. For example, in so my native language is French, so I was looking at the French translations, and I realized we did still have a lot of cases of, you know, we're saying, the developer can do this, the developer can do that. In French, it'll be automatically translated to male pronouns. Just because that's the default in the language. But French has a way of using inclusive notation so that you can have, like, a gender inclusive form for that. And so we just had to add that to the prompt. Rerun the translation, and boom, all of our training materials now are gender inclusive, which is really cool. And you can do that in any language. That supports that. So we do this with the help of people in the community who are native speakers of those languages. The biggest part is adapting the per language glossary so that we can adapt technical vocabulary to have the correct terms for like, for a channel or something like that. Anyway, so that is very powerful. Huge thanks to Phil. For doing a fantastic job on that and enabling us. With that, I invite you, if you would like to, level up your skill Nextflow skills some more, please join us in two weeks for Nextflow training week. It's online. It's self directed. You choose when you work on courses. You can get the certificates. You can get all sorts of cool stuff. But please sign up. Check it out if interested. And with that, having consumed plenty of time, I am going to hand the mic over to Marcel, who is going to tell you about the ambassador program. Thank you very much. Everyone, it's great to be here again. I was just watching Gigi's talk and looking at the audience here and thinking about what an amazing community we have. Right? This month is a very special month for me because it's it marks twenty years of running open source communities. Like the first time I ran an open source community, was in April. 2006. And I can confidently say that in the past twenty years, the NextLow community is the best community I've ever worked with. We had the hackathon before, the training, the summit here, Nobody comes for an award at the hackathon or certificate of hours in the training. Or something very like industry focus in a summit. Is a lot of things about learning more talking about Nextflow and get to know more people who use Nextflow and all that. Right? I'm talking about these things because that's what we are always thinking in the community team. What an honor we have of being here in the Nextflow community. And about two years and a half ago, we were discussing this thing. Look, this amazing community we have. So many people who are eager to talk about Nextflow. Over the world, in their fields of expertise, with their friends, with their work colleagues, They talk about Next World because they're not paid to do so, because they like it. They use Nextflow. They are happy. It helps them. They're really happy about it. And very quickly, we identified what I call the two profiles of very active community people. One of them we just called them ambassadors. Like, they already talk about Nextflow whenever they can. They go to conferences, they give talks on Nextflow. Train their friends. They use it at work. They use it at university. Right? So they have ambassadors already. They just don't don't have the title, but they're already working as ambassadors. They already even call themselves sometimes ambassadors. And some people they're a bit shy, maybe they don't know. They can talk about next, so they don't have the confidence, right? But it could be a great job. They could do a great job. And that's what I call the potential ambassadors. Right? So we were looking at the community, looking at these two profiles and thinking like, we can help them. I mean, maybe they need their help. We'd like to help them go further. Right? So for the ambassadors, if they are doing this great job, let's make sure they don't stop. They are happy doing that. Let's help them in any way we can. And for those who have the potential to do so, let's make sure they feel confident. Let's help them spread the word about next slow enjoy more next slow, and all these things. So we're thinking about programs, right? I ran ambassador programs before. And when I joined the community team in Sequeira, was like, you know, this is the project we need to do. And since the first day I was in Sequeira, I was always talking about this. Let's create an ambassador program. Was like, talking to a few, we have to create an ambassador program. This is gonna be amazing. But I mean, company was too small. We were too organizing a few things. Lots of work to do. But eventually, my persistence won, and we decided to ambass the program. So this happened at late twenty twenty three. We already knew this ambassador that already behaved like one, so it was just a matter of inviting them. And talking to some people if they were interested. So we started the first cohort of the ambassador program in 2024. We had 45 people and it quickly grew. Currently, we have 143 ambassadors Some of them are with us from the very beginning, since the first cohort. Some of them joined then they leave for some reason. They change jobs. They change countries or some or they get too busy. And some joined for the first time in continue in the program. Right? At the beginning we had 16 countries of residents right? Some ambassadors they were the five ambassadors we had, they were in 16 countries. But it quickly grew also to the 40 countries we have today. In all these two years and a few months, the ambassadors have conducted almost 700 activities. This is almost like 25 activities per month. There's a lot of work they've been doing We don't count activities more than one. So if three ambassadors worked in one activity, here I'm counting them only as one. So for the past two years and a half, it's a bunch of work. They've been doing. And this is the map of where we currently have living. This changed throughout the cohort. Of course, we had people in different countries, now we have in different countries. But the thing that I want to highlight is that this is where they live. Not where they act. So let's take an example of CUBRA. She's in Germany, but she's originally from Turkey. She has been doing lots of stuff in both countries. We have people who are in UK but are from Iraq or in Brazil but are from Japan or vice versa. So the the countries where we act as ambassadors are actually much more many more than what we have in the map. But still is a very nice geographical representation. Even though it's important to highlight that ambassadors, even though this name has a very is very meaningful related to geographical geographical oh, regardless, the word I've been there. Location, It can also be a dispute of expertise. Right? It doesn't mean that you have to be the ambassador in Brazil or in China. It could be the the ambassador in spatial omics, for example. It's not only about the geographical location, also your field of expertise. So some numbers about the program. So this is the categories of the activities we have had by the ambassadors. As you can see, lots of blog posts, trainings, events that are organized, like hackathons, talks they've given, and so on. And one question we usually have is, Okay, but what's the portion of that that is in person Because we do value this in person interactions, like here we have the hackathon that's training the summit, right? And when it comes to in person interactions, in many of these categories, a bunch of them are in person, like events, training, talks. So this is very nice. You see our ambassador really like, acting in the real world, not only online. Right? So in 2025, which is the last year we completed the program, we got some numbers about how active our were And we usually ask for a commitment of two activities per semester. So we have to do two things. Two talks, two trainings, two posters, two blog posts, choose something related to outreach and next flow. Right? And it's not always easy, like sometimes their troubles happen during the semester, you change jobs or something like this. So it's understandable not all of them can do that. But we had over half of our ambassadors doing at least two activities. Then 32 of them doing five or more, up to four of them doing 10 or more activities in the year. So five times or more of what we asked. And then I would like to take this opportunity to thank our top 10 ambassadors that got Certificate of Excellence. We have James, Julia, Matías, Jose, Firas, Evangelos, Maxim, Cubra, and Antoine. They're mostly in Europe, except for Fides, who is in Tunisia. But they've done amazing work in multiple different places. Of the world, in multiple different fields, different types of contributions. So we are really thankful to the whole community, of course, all the ambassadors. But this top 10 they are not only the most active, but if you want to join the program, if you feel a bit lost, if you know how to contribute, if you like to ask some suggestions, these guys, they know a lot about the program, a lot about the community, a lot of next flow. So they're the best persons mean, can talk to me, of course, also. But they are great people in the community. That you can talk and learn more. About how is it to be an ambassador. So James and Julia combined had 62 activities 2024 and 2025. That's almost 10% of all the ambassador activities. So they are really active. And again, they do a lot of different stuff. It's not like the same type of activity every time. They do a bunch of stuff. They're great ambassadors. So usually what we do is that if you do something that sometimes doesn't fit in an outreach category, like you wrote a This doesn't count as an ambassador activity. But if you talk about it, it does, right? If you do a blog post about it, if you do a talk about your pipeline, if you record a video on YouTube teaching how to use it, So all these things are outreach and they count as ambassador activities. So usually, we ask people to write blog posts about these things, but not only that. Like sometime you organize a training like Latter did, is an outreach activity, which she also wrote the blog post about the experience. So that other people who want to run a similar thing, they can just read and learn from her experience. So here we have a few blog posts we had last year. About ambassador activities. We had ambassadors bringing information from COP sixteen that was related to workflows, computational workflows. We had people who wrote pipelines like Zohibe, on the virology field. Went to events to give talks about it. Wrote materials about it. We have who was traveling over Africa to give talks about next level trainings, writing content online, recording videos, So here are just some examples of the activities we had. In 2025. And I would like to talk a bit about hackathons. Because the big thing we do, like the one we had a few days ago, and we also have a hackathon at the summit in October or November, depending on the month we have every year. And also one in March, is the Habit Hackathon. I like to say it's the largest one we had, every year because it's Hibrid. So we have people all over the world participating. We have recently like over a thousand people participating in this hackathon. And what's the connection between the hackathons and ambassadors? So very often ambassadors like spearheading lots of hackathon sites and stuff like this. Like for the hackathon we had here, some group leaders, they were an excellent ambassadors. When we have the hybrid hackathons with local sites around the world, we have this 70% of local sites that are always hosted by ambassadors. It's not a aesthetic number, but for some reason, every year, 70% of the number of local sites we have around the world, they're hosted by next to ambassadors. And a bunch of them who are not hosted by ambassadors, they are supported by ambassadors. So ambassadors are helping in almost any activity we can think of when it comes to community. So I think after saying all these amazing things about the program, what to do, and so on, why not join the program, right? Like maybe you think you're not ready. Maybe you think you know next to an off. Maybe you don't think you speak well or write well. Well, if you really want to help and you are already part of the community, have contributed to the community, you know Nextflow, you're more than welcome. A lot of people, we help them how to better ride. We review their taxes. We help them with slide and all these things. So if you want to help, and you like the NextLoc community, welcome. I'm going to show the instructions to apply, but basically you're more than welcome to try. To apply, right? So why joining the program? Like if you like the next tool and you like the impact of the task, why not help increase the impact with your contributions? Again, maybe in your geographical location, Nextflow is not well known or in your field of expertise. So why not bring a bit of Nextflow there? And it's like a two way thing. We can also help increase your impact. Like if you're sharing something on social media, we can repost we can mention you in our post, so we can increase your reach You get more followers, increase your reach and your impact in social media and all these things, right? You also get some behind the scenes about how to run a community, but not only that. Very when we have a new thing in secure, before we release to the public, we actually release to our ambassadors. So it was the same thing with SecureAAI, which so many projects we have had so far. For the VSQL extension. We first share with them back ambassador so they can try it. They can give their feedback to us. We can improve technology before it's released. To the public. So you get this behind the scenes information on not only how to run a community, community, but also the technologies we have. We are always looking for nice events where it would be good to have We can let you know. We can help you go there. We can keep you posted about news in the topics associated with Nextflow and then of course. You get training. So if you feel confident enough, we can help you get more confident with Nextflow. We can help you do a training. We can help you understand the training materials. Even though anyone can ask questions about Nextflow and F4 and so on, there's a private ambassador channel where we talk, and you can get some special support, let's say, over there. It's by me, but I mean, is this an extra level of support, I say, we can do? In terms of recognition, you get a certain at the end. If you don't your two activities, you get a certificate. A recognition for participating in the program. And if you get to, like, an excellent level of contribution, get an excellence a certificate of excellence just like a second certificate some people like. And all these things that you are doing or you want to do, we can support you in doing so. Again, we have this profile that I'd like to call like the original ambassadors. They're already doing great work, but maybe they can do more. And the reason they can't is because maybe they have to travel, they don't have the money, or they don't lose want to spend money to another event in that year. We can help you with that. So we have a budget for supporting with snacks for the pants, a traveling accommodation sometimes for some strategic events and all these things. Have a budget for that. Oops. And so far, funding went mostly to this discount in official events, swag is next for hackathons or trainees if you want to give to the audience. And travel and accommodation for some specific events. So we have had multiple ambassadors traveling, like Zohaib had mentioned with the virology pipeline. He went to Portugal Everything for him to go there and talk about his pipeline, the Nexlo and the NFC and so on. So if you want to join the program, you can just go to the NextLow ambassador Nextflow website, nextdoor.io, get the resources. There's a Nextflow Ambassador button, description, and a summary of the program. At the bottom, you're going to have a link to the next real handbook lots of information about what it means to be an ambassador, what you should know, what you should do, how we can help you and so on. And also the apply here button that you can apply to the next cohort that's going to start in July. So our cohorts, they're a six month cohort. People apply during the one semester to join the next one. We have a kickoff call. Follow ambassador at the beginning and so on. So thank you very much. I'd be more than happy to answer questions if you have. And thanks again for making this community amazing. You're a part of that. You very much both of you. You can probably tell that we're all pretty excited about the community. So if we get the opportunity to talk for a really long time, we'll take it. Yes, any questions in the audience? Before I kick off on a couple of my own? Hi. Sorry. You said that ambassadors are on a six month rotation, but are people admitted on a rolling basis? Or is that the beginning of each six month period? So you start after you apply, right? You can say, So when it gets to the end of the six months, I ask, like, you know, which of you guys want to stay? Most people say, yeah, they wanna stay. Some people are like, you know, I change jobs, I'm in a different country, I'm too busy, wanna leave. Wanna come back later. And that's totally fine. So we have people who stayed for six months and left. And have some people who joined in twenty four point one and are still here today. So it's up to you. As long as you want to contribute and you're still there, I can stay. I think you might not have mentioned that we have a call in January and in June. So we have two fixed time points of entry in the year, if that's Yeah. The January or July? Oh, that's I think about the rolling. Sorry, sorry. Yeah. I think so. Maybe. So if you apply in June or January, it's the same thing. You're gonna start in July. Thank you. There was one thing, you were talking about the translation work on the training and I had a I just wanted to point out that I wanted to give credit to the community of a fast API docs because I basically stole the idea from them. But yeah. Been fascinating seeing people come together with these these translations there. Especially languages which are so completely kind of foreign to me like Korean and stuff. It's I love seeing all these Nextflow terms pop out amongst these wall of characters. And And and definitely, if there's if there's a a language in there that you are competent in and you'd be interested in seeing that represented, If you know that it would make a difference to a community that's speaks that language? Because we know that it can be challenging if you have the double wall of cognitive you know, challenge of you have to learn something technical that's very difficult, and it's in a language that's not your native language. I'm very familiar with that problem. It is it's like a double problem. So if we can take away half of that, by making things available in your language, or in the language of the people that you're in contact with. We're more than happy to. We just we don't want to just make a dump of all possible languages. We we want to have at least one native speaker or somebody who is very fluent who can help us gauge that, yes, this is correct. This is not outrageously, you know, we don't sound like gangsters. Or something. You hear about translations based on not quite the right style. Anyway, yeah, we'd love to get help with that. So thank you. So Alexia, a few words about this. In 2020 I think, there was a roundtable and a summit. And I talked about basically localization of training materials. I was just remembering that. And the thing is, people think like, okay, everyone speaks English. Right? Well, in some places that's not the case. Even people like doing master degrees and PhD, even though they can understand English they don't feel comfortable. It's not so easy for them to read content in English or to listen to content in English. It really helps to have content in different languages. And in 2022, we already started trying to fix this. And it's such a huge amount of work because first, translate everything manually right and when you're done something changed. Right? And then speak Portuguese. Jordan speaks French. I speak a bit of French. And then we have English and Spanish. But what about language we don't master? So we rely on someone else to volunteer, proofread that. Not only proofread, to translate that. And at some point we had, I think, the basic training material in Italian, in French, in Spanish, in Portuguese, in English, And for a few days it was so nice until everything got outdated. Right? So at some point, it wasn't worth having it in different languages because the quality was not good. So we would rather not have anything in a different language than have broken code, broken instructions, and all that. So I was particularly very sad. Because that's what we had to do. We couldn't have so many languages because it was just broken in different languages. When Phil came with the AI thing, I was so happy. And he showed you now all the skills we have with Claude and all these things. And the smart only fixing the translation for the dev in the recent and so on. This makes everything so good and easy because we can fulfill this goal of having it in different languages, but also with quality, correctness, and being to date. Just amazing. I I know if I can transmit to you guys how happy I am about this thing, but it's just amazing. It's a dream come true. And I will say those skills are in the training repo too. So if you wanna do that for your project, you're more than welcome to just copy it. Yeah, it's Creative Commons. You just give credits, and and boom, you're you're done, and you've got you can set that up. And if you need help setting up to actually work with the GitHub actions and so on, Ask us, and we will point you to Phil, because he's the only one who understands how it works. But anyway. Oh, I have an NFCOR duck for the winner of our training quiz. From earlier this week. So if you were the winner of the quiz, come and talk to me and claim your duck. Otherwise, we'll contact you. Yes. Sorry. Massively. Thank you. You, guys. Next up on the agenda, we were hoping to have Vipava from Grail, but unfortunately she's had to duck out. Due to personal circumstances. So she has recorded to very graciously her talk for us and we will put that up on YouTube along with all the other after the event. So I do encourage you to check back in a few days when the talks are online. And listen to her talk. It's a shame because the abstract was sounded really interesting. So we all push on and next up, I'm very I'm delighted to have Reema come up and speak to us about her pipeline for doing downstream analysis of RNA seq data. Reem is based at Northeastern University and formerly was at a stint at Novo. Is that right? Novo Nordisk. And we had a great conversation earlier. And she's also presenting here in Boston at BioIT. I'm sure probably many people in this crowd will be there as well. Yes, welcome. Thank you so much, everyone, and good morning. Today, I'm excited to present my work. Which I had developed and initiated during my time at No More No Risk. For the NextFlow downstream analysis pipeline. So So I will start with the briefly introducing the motivation behind this pipeline. And the upstream foundation for the RNA sequencing analysis. Then the main scientific questions we want to solve with the count matrices and downstream bottlenecks for the analysis. Then I will cover the DSL two architecture, for the custom pipeline development and key takeaways from my workflow. So basic questions of the downstream analysis is like, this question motivates me to build this workflow. Like, well, many workflow and methods for the stream RNA sequencing processing has become available and standardized. But still, there is the bottlenecks for downstream analysis, for differential expression and pathway analysis, which which introduced the result variability, result reproducibility, and, makes the scaling difficult. So I started with the solving the question after the count matrices, like this is the NF score up analysis pipeline, which is everyone is familiar with this pipeline and many of have used this pipeline for the pre processing steps and after the pre processing step, the count matrices are mainly used for the differential expression and pathway analysis. So once the upstream, analysis is done, the focus is mainly towards the differential expression. So, for multifactor analysis and different different workflow designs sometimes required sometimes Sometimes it requires the downstream or different strategies. So the question is about what are the main biological effect we want to solve with this, and does the experiment have a sufficient power for the analysis and the same time we want to understand that the experimental factors such as treatment, donor variability, cell type contribute to the gene expression changes, particularly in the multifactor study design. And in more complex experiments, repeated measure or correlated sample introduce additional modeling challenges that require appropriate statistical approach, So these questions are critical because they determine reliability and biological relevance of the results. And in terms of the downstream RNA sequencing bottleneck, these factors are mainly critical a critical role in the analysis. First of all, the study design, hybrid and multiplex design can help to balance signals across the batches. Repeated measure designs improve the power and detecting temporal changes. And group comparison designs captures the difference between the subjects alongside the study design, the power analysis is become the important portion to that require that estimate the sample size estimation and evaluate the power of this study. The second major issue is the biological and technical variability So biological variation refers refers to the natural fluctuation in the analysis comparing the within and between subject variability while the technical variability is the technical artifacts that are introduced during the library preparation and sequencing methods. Along with this, along with this proper metadata harmonization and standardization of results, in the downstream analysis is also, bottleneck in the analysis. So to solve this problem, like, I started with introducing the combat seek in my workflow. So, combat seek is helpful for batch correction effect. When it is based on the negative binomial framework and one of the advantage, it's that preserved account matrices by generating the integer like count making it compatible with the DC2, Lima, and other Azure method, which required count matrices as integer values and combat seek is useful for specific scenarios. Because other models are also included, the batch variation in there. Models. But in some of the situation when a study is different, have a different batch if effect than including combat seek as a separate method. Is useful for capturing the true biological effect. And for another question is RNA seq. Power. Sometimes like after the study is designed, analysis is run perfectly. But if we want to validate our study with the power analysis, is also a good approach. And for the RNA seq, power, particularly the RNA seq power package is available So user can use that to estimate the study design in the workflow. It helps to determine the effect size, sample size, biological variability, sequencing depth. And significance level of the analysis. And other question is about the tool types, like the dSeq2 and Lima is useful for different approaches. DSeq to model a raw count data using the negative minomial distribution which directly captured mean variance relationship in RNA data. In contrast, the Lima transforms count into log two count per million. They seek to require raw unnormalized integer count and internally corrects for the sequencing depth using the size factor while Lima, on other hands, requires preprocessed data typically using TMM normalization followed by WHIM transformation to model variants. DSeq two applies this, shrinkage on lock form log four changes will help to reduce noise in log count, log count genes. Lima applies the empirical B base, inference which estimate the variance improve the stability, and especially for for smaller sample size. Both methods relies on replicates for the reliable inference. DCC2 strictly requires biological replicates while Remy is more flexible in certain design as a time course data and where trends can be modeled. These seek to often prefer for smaller sample size when working directly with the count data. And LimaWome is more powerful for larger data set for when library size are very significantly. So I built this modular DSL two multifactor RNA sequencing workflow with the NF core standard and this is based on the tools I have earlier, which I have included in my workflow. So main workflow start with the configuration and input layer, where the user defined parameter through the analysis, through the configuration file. Along with the Nextflow configuration file for the resource allocation and execution profile. This enables flexible execution across environments such as HPC, or singularity. This pipeline orchestrate through the main entry point of main NF file, which load the parameter, build channels, and connect different sub workflows. These sub workflows represent key analytical steps, including batch correction, power analysis, differential expression, and pathway analysis. Each sub workflow calls the modular processes, that execute specific task using the R script, where the code statistical analysis is performed. Data is passed between using the Nextflow channel along with the clear separation between data handling and computations, There is also the API config. So if user want to load the data set with the other database or other platforms, So API config can help to load metadata and count data without the manual download. And other option is for manual download, like if a user want to upload locally their account data and metadata to the next flow. So two types of option is available for the data input. And during the execution, next Flow manager managed the resources of the CPU and memory that generate structured output, including reserve tables, logs, files, and for the traceability and debugging. Overall, this architecture, is based on an, of course, standard. It is produced it is for reproducibility, flexibility and create organizations of the result across the workflow. And if a user want extend the analysis, so it generates plot and figures for the differential expression and pathway analysis. In the PDF form. So yeah, moving to the next slide. Like, this is the metro metro map based on the workflow I have designed. The green line is showing the core workflow steps. Which is it start with the metadata configuration parameter and count matrices. From there, it moves to the moves to the gene filtering, which is the first and mandatory step for RNA sequencing analysis. After that, batch correction is optional steps. If there are significant library differences or batch effects, user can apply. Combat seek after the filtering, which generates corrected count. File, which you can use in the further step of power analysis. Or differential expression analysis. In the differential expression analysis is divided into two branch, DC, two and Lima. This, in this pipeline, Lima is a main core method. While DSeq2 is optional alternative depending on the data set. The analysis uses the corrected count matrices if batch correction is applied. Otherwise, use the filter count in the differential It perform normalization, applied contrast definition from the configuration file, and generates the eg list to which are saved for each contrast. Finally, in the last step, pathway analysis and enrichment analysis. It is also, optional step, but it it generate the rank gene list file from the enrichment of using the FGACA and MCPB pathway collections. And these are the output of generates through the differential expression and pathway analysis. So moving to the next slide, The strength of this workflow is about, to it is based on the modular DSL two pipeline. Which have reusable components. Parameterized VML driven execution, standardized output, log files, and for the traceability, containerized reproducible runs across the environment, And for the testing and validation, the core module is functionally validated while cross design data set delivery and testing is underway. Runtime and resource profiling is also ongoing. And in the next step, it the next step is to optimize scalability and resources. And for the portable Docker and Updater profiles. Overall, the architecture is stable, and based on the NFCore standard and the next step is more focused on validation and optimization. And in the my final key takeaways is that metadata harmonization and quality control are the most critical steps. Because error at this stage directly impact the downstream analysis. Strong study design matter more than tool choices alone, and statistical rigor is what transform broad count data into reliable biological questions. And there is no perfect tool section, no one fit size, all method. Means tool selection always depends on the experiment design. And sample size. Replicates are essential for accurate and variance estimation in method like DSeq2 Lima and Azure. And finally, the modular containerized workflow like Nextflow are ensures the reproducibilities scalability in the RNA seq analysis, Overall, the goal is to not just run the analysis, but generate the results that are reliable, reproduce and biological meaningful. I would like to acknowledge my mentor for the guidance and support. From the Novo Nordisk and the Nextflow orp community who provides the trainings and, advanced training for and resources for the for the community. Thank you. You so much. That's a great talk. The topic is very close to my heart, this one. Like I remember before we started NFCORVA, RNA Seq pipeline, we had endless discussions about whether we should sort of ship scripts for differential gene expression analysis and it's a really difficult thing to do. Any questions from the cloud? Yeah, there's a couple over here. Hi, Reema. Thank you for your talk. Really enjoyed learning about this pipeline. This it also takes me back like Phil because this is like the one thing that I did gene expression analysis related to bioinformatics. Back in grad school. And we tried we also tried lots of different methods And I'm curious forgive me if I misunderstood. It looks like your pipeline you have the two methods and you you choose one or the other for a given run. Have you looked at all in trying to perform use multiple methods and try to fuse them try to come up with some sort of composite gene ranking. Or does that not make sense? So for combined above methods, Lima and they seek to combine gene list right? For example, yes. Yeah, no, it's generated separately, not combined results, but it runs one time, one method, not combined method. And generate a result based on select method in the initial configuration of file. Presumably, you choose one or the other depending on your experimental setup as well. Yeah. So based on the sample size. Basically. Yeah. Another question there? I have seen the batch correction. How about like the batch sample filtering? How would you make that kind of decision? When you do it? Normally, iterative process from sample to sample, it varies. There is no streamlined way of conducting because human has to make a decision how you handle those kind of quality filter, a decision to exclude certain sample, a decision to make sometimes certain COVID For example, in mouse, if you combine sex, you know, you get different things. You have to make a decision which line you have to go depending on that specific data set. How you make those kind of decisions when you just use a streamlined workflow? Yes. So it is basically in the starting point when so I also tried the user defined approach in with the Python script when users start the analysis. There is multiple options pop up at the, like, window. What user want to perform, or what sample or user want to select for the analysis. So still, I'm working on that portion to, like, user defined script, but otherwise manually a user can select the parameter contrast type and exclude what the sample, which have a low gene and other specific sample user want to perform for the analysis. So it is during the script in the YML. It's generate Yeah. That's something we've discussed before is, like, having a a config generation tool kind of thing. Yeah. Question here. Hi. So I noticed that you in for your pathway enrichment step, that you maybe baked in m sig d b as the reference for GSE And so I'm wondering, does that mean that this pipeline is limited to human and mouse references? And is there any way you would see in the future for someone to be able to sort of bring in their own reference for GSEA and the whole pipeline if they had non human and non mouse models they wanted to use this pipeline for? So currently, I have tested with the human sample only. Yeah. So that's why I've used this M6DB pathway collection for it. For future, like, we can elaborate and extend the pipeline once it properly execute with many complex data set of the human Yeah. Just call out there. There's an NF core pipeline called differential abundance. Which is also slightly more generic and I think that should work on any species. And we were talking yesterday about whether we can kind of try and come together and and you know, take take take some of the excellent work you've done and and see see how we can fit this ecosystem together. I'd add one more question related to the the topic about customization. Differential abundance isn't like an our shiny interactive app, which kind of gets bundled with the results. So my question was going to be, have you thought at all about interactivity? It sounds like you have a little bit with like a Python wrapper. Is that something that we're kind of kind of be bundled with the pipeline somehow or do you have any thoughts about that? About interactivity? Or the visualizations? Yeah. Visualization and and configuration of the the experimental setup. Yes. So can you will you Maybe a refrain the question. Could you tell us a little bit more about how your wrapper works to help people configure their experimental setup? Oh, yes. So it is based on the user defined setup, which I've was trying to implement in the Python. So yeah, it is, so user can't so in that way, like, it generate first user defined parameters. So according to the pipeline, it shows which contrast or which experiment setup do a user want to perform and which model they want to use Otherwise, like, pipeline execute model by itself with the conditions of the parameter. Okay. Thank you, right? Thank you very much, Amy. Okay. Yeah. We can discuss that further. Okay. Next up, we have a familiar face. I've asked someone will join us again and tell us a little bit about what's happened been happening this week with the hackathon? Long time no So as you're probably aware, we have had the the next little hackathon a few days ago here before the summit. We have the hackathon every March. Which is Hibri briefly talked about it. We have in the Boston Summit, we have in the Barcelona Summit, And every once in while, people do their hackathons also. In different locations, right? Sometimes quite often, I would say, I hear people saying like, yeah, know I didn't participate in the hackathon because I'm not good enough or I don't know next went off, I don't have a project. I think people won't like me because I am not expert enough for Nextflow. And it hurts me because this makes no sense at all for any of the hackathons. Everyone should be able to come. Of course, if we have the training opportunity, anyone to learn, if you want to learn, great, you have this opportunity here. But the hackathons themselves are also a great opportunity to learn more about Nextflow. Of course, and related topics. Right? The thing is, I don't know about your background, but at least for me, whenever I think of a hackathon, the first thing that comes to my mind is like competition. I know why. Maybe it's because in the past that's how most were. People are fighting for an award. Or a title or something. There are teams battling each other around the same topic. So, well, I would feel threatened by joining this thing too because I'm just learning next slow and I have to fight these people. Have to battle them. I'm gonna get the last position in the ranking. I'm not going. Thing is, the NFC or hackathons here are nothing like this. We have changed in multiple different ways throughout the years, but never in this direction It's not a competition. So oh, sorry. So the first thing I would like to say is, well, there is one competition, which is the quiz, right? So there's a quiz when when we're in Boston. There's some questions about and some questions about NextLow. And then, of course, but it's a game. It's just a game. You get to learn more about next low in Boston or Barcelona depending on where you are. And that's it. It's part of the thing. So it's the complete opposite of a competition. The NFCore, the Nexo hackathons, in general, they are for fun. They're a way for you to get to know more people, to have fun, to learn next flow, to do something that you already wanted or something someone is working on. That's it. Right? So it's a collaboration opportunity. And it's a very flexible one too. Maybe you wanna just work on your project Or maybe you don't have anything to work on and you want see who has a project that interested in collaborating? Or like Harrison, you make this huge AI agent that's doing a bunch of stuff, and it's not even him anymore. It's just like the AI. So I don't know. Maybe we should count the AI as participants in the Haggadah, which was bigger. So many agents doing lots of interesting stuff and amazing project by Harrison. So I would say these are the three things that should come to your mind when you think of an of an NFCore hackathon. It's fun, It's collaborative. And it's flexible. You have a personal project you want to work on, Great. You want to work something in an NFCorp project? Great. You want to work on someone else's project? Whatever. It's just about getting together, getting to know more people. Having fun, learning more in Xflow in the minute of core. So we had some awards, not like prize awards. Money awards, but we have this NFCOR DUCs, which are extremely valuable. In the near future, they're gonna be worth billions of dollars on Amazon or something like this. How do you call the the not Amazon, eBay. They're gonna pay billions on eBay for these things. Believe me, these they are like, I bought I got one for my daughter. She's gonna pay her university in the future with this duck. So believe me when I say people would kill for that. We probably we have had so far, don't know, 40 maybe, I don't know. But like, it's really valuable. So congrats to people who got like, Benji got for the one with the quickest and most correct questions to the quiz. And Joe was randomly selected to get. So I mean, you know, you even have to win the quiz. You can be randomly selected to get the lottery. So it's good, right? So if you guys get rich with the don't forget me. So a brief So a brief recap of the March hackathon, not the one we had a few days ago. The March hackathon which was hybrid, has some interesting new things that I want to highlight here. So it was Hibbett, so we had people online and people in person like we had also last year. We had 41 projects registered in five different categories. 29 local sites around the world, 600 participants in 19 time zones, lots of PR and so on, blah, blah. Great. The thing is, last year we had a bit of a bad experience, let's say, that some people they were kind of lost. Like, okay, I want to work on this project, but it's not very active online. Apparently, it's people in Australia working on that and feel a bit disconnected. So we heard that. One thing we did in the March hackathon is all the projects they have very clear if they're going to be mostly online, mostly in person. And well, if you really wanna work on something which is mostly in person, great. You're more than welcome. But you know that maybe there won't be a lot of activity online. So we gave this heads up to people so that they know what they were working on getting involved. And so far, the feedback has been great. People loved it. We had people working multiple different projects, just like they won a few days ago. You don't have to work the whole hackathon on the same project. On the first day, can work on something. Maybe the second day, you want to try something new, right? You go to another project. There's no problem at all. Actually, that's quite amusing because you have new ideas come to different projects every day. So it was an amazing hackathon. We have the NFCo Statistics page that I forgot to put the image here. But you can see the bumps in the charts. On the days during the hackathon. So lots of activities being done lots of people happy with their personal projects and projects. Work projects, whatever, whatever they bring to work on the hackathon. People got together to help. So now let's go to the Boston hackathon. So I'm going to talk a bit about some of the projects that we worked on during the hackathon. It's in person in Boston. Now it's not an easy time to come to The US. So we don't have a bunch of people, but we still have about 30 people working on 30 non secure ones plus secure ones. We'd say maybe 40. Something people working on these projects. One of the projects that we worked on was the Typex migration project. It's something that we've been working on since three hackathons ago, like the one in Barcelona last year. Basically, it's to adapt the NFCore modules to this new functionality we have in XLO, which are the topics channels. So it's much cleaner code, easier to read, easy to understand. It makes much more sense to have the versions of the tools in the pipeline being shared through this topic channel. It's like instead of passing information between processes, you just say, you know, this information go to topics channel. This one, go to this topic. It's an easier way for different processes to share information in a single place, which is a topic channel. So we have the versions topic channel, right? I'm actually quite happy because, as you can see, we merged I think, 30 or 40 PRs. This was a very productive project. And the best thing is almost everyone was doing that for the first time. So the next NFCore hackathon, if you feel you are not if you don't feel confident enough, you don't feel that you're an expert on Nextflow, or NFCore, just remember we have this first good first good issues Good first issues, which basically means that if you're the first time you're there, you don't feel very confident, you can work on that because there are issues that we have organized that are great for people starting. Not only because they are not hard, but because they're quick to merge so that at the end of the hackathon, you can have your peer merged. So we have people who have been here for the first hackathon ever And during the hackathon, in the first day sometimes, they got their PR merged. I don't know how you feel about it, but if it was me, I mean, I had this experience in the past, it's amazing. Know you were able to contribute something. There's something now out there being used by people all over the world And there's a piece of you that contributed to that. I think that's amazing. Some projects, it takes a bit longer to get nurse. I mean, the topics migration one the first time we did last year, was actually only merged the days after the hackathon. It's fine. But I mean, when you are starting and you see something you did, merge during the hackathon, that's amazing. Like people are using something you did now. It's a contribution You're officially part of the community as an active contributor. I think that is great. So this is a top immigration. We merged a bunch of PRs converting NFCore modules to use the versions topic. And every project this is a law, is a very important law. Projects only exist if they have a meme. So if your project doesn't have a meme, you're out. You'd need a meme. In a good meme, you're evaluated. Okay. This is a competition. This is the competition part of the hackathon. So this is the in every day you have a new meme. So this is the best meme I found for the topics one. Because you know, when you you test your modules locally, with all the test files, the version, everything else, it takes a while. Then when it's done, you push it to the repository. And then the GitHub is actually going to trigger and do also tests. And they do, I don't know, 30 tests or something like this. Because sometimes you're chatting like you're not working. You know, it's testing. I have to wait until the testing is done. So I can go back to work. So this means we had lots of moments during the Haggai Doll chat about live. And random things. For the second project, we had this massive beast, which is Sarek. Having something else added to it. So Gary and Oh, actually, sorry. So we had a bunch of people in the topics project. We had me, Eamon, Chris, Duane, Harrison, Anabella, Louis, and all the people who joined. Throughout the days. Right? For the GPU variant, colors being added to Sarek, we had Gary and Joe. So is getting bigger and better. Right? So you guys probably saw, like, in talk, and discussions recently how we are speeding up some tools. With GPUs and other stuff. So this was an attempt to add some GPU variant colors to SREC, which is this pipeline you have in NFCore. One of the most famous ones, right? That means RNA Seq, Maybe it's Rx second or third. And again, it's a big chunk of contribution. So it's amazing. We are really happy about it. But not everything got merged yet. And that's how hackathons go. Right? So if you are resilient, you don't want your peer merged the first day, you can go to a more challenging project like this one. And here we have the meme. Right? And, oh my god, I think probably everyone who was in the hackathon kind of felt the same thing. Like, you have your task, you're failing. And when you finally make it work and all the tasks pass, you realize the output is not there. And that's why it's having a successful test. And then you go back, you work again. Eventually, you manage to make it work. So here, no was being generated, but they managed to fix it. For the self painting one, so the self painting is an NFCore pipeline. Cell painting is a high throughput image based assay using drug discovery and biology to analyze cellular changes by staining cells with first and lies, I'm reading yes. So that's same cell painting, and there's an NFCore pipeline for that. Ken has been spearheading this for a long time. His personal project, I would say, very nice pipeline, lots of progress. Every hackathon we have new things. Bunch of progress happened here in this hackathon as well, and many people got involved. Ali Khan, as I said, I had Rachel, Riley, Ronnelly, Jake, Faye. Also all the people that here and there did some contributions. So lots of progress. Lots of PRs be merged. And I personally love this meme. The pipeline too, this is amazing, right? The interpolating thing can break some hearts. Then we have the chin, final share her But again, amazing pipeline. And again, as you've seen some of these stories, right, here they're mostly end of core projects, end of core assets. But you can work also on on personal In in the next other projects. So we had this work by Michael, Edmond, and Ben on bringing some dev d b integration to Nextflow. So this discussion started in '25. So very often you have this discussion happening, and then you do something and you stop, you know to do, how to progress, then come back to that. It'll Eventually, Michael created her two for DAS, and now they're trying to make an pronounce that. E I d e r, I So it's a way to bring some integration with duct d. To I think a lot of people agree with that. And this is like the project by Harrison. Right? So it was Harrison, a lot you, Harrison and the AIs. And this what he did was using anti gravity in a bunch of different agents. He went to find some cleaning tasks cleaning tasks. Right? Something that you have in your repository it's a bunch of work with manual work. Don't wanna do that all manually. So we thought of asking these agents to do for that and find way to fix all these things automatically. It's a bit not to write anything. What's the word? It's a bit like a you can teach your AI too better. Right? Of course. So if you do, like, a massive PR and that's not nice, why don't you play that in more PR? And then he's agency that. And then sometimes when someone's reviewing, he saw he did something wrong. So you fix that in this PR, but why not make the agents fix that in all PRs? Not only And, they're so manual. Right? because then we have more free time. Work on other things. And eventually, when these things learn and they get better, the review is gonna be so simple. So this manual labor this manual task they're annoying to, but they have almost 2,000 modules. 40. Then we have the last project by Lucas, Chris, Chase, and there was this a meme that accidentally didn't change for the next project. But I love that. This year this year's call AI. Ask questions, it's confidently why I see you. So I love this new name. So I actually forgot to replace for the name of of Lucas I'm sorry for that. But then we have this project on writing and and, of course, compliance, let's say. Single cell prescript pipeline. The visible group perturbation And I think that after all that, maybe some of you are your class, in your lab, in your company, in your city? Why not hosting Harkon? But myself, because it's way more and just So the thing is we created a how how they are? What they can assist of? What things happen during What can I pick up? Maybe they want to do all these things, all these Maybe they socks they're socks Right? And all these things. So you have all these options, how to plan, how much advance, some documents to make it easy for you to manage your hackathon if you want to have a of site or not. Not, how to organize a project. It's a it's a almost it's almost 30 pages is involving almost 30 pages document with lots of information. You don't have to use everything. Again, no problem with spikes So thank you all for your attention, and I hope you join our next hakakong in hotel room. Thank you. Next Some some keyword. Last talks before the coffee break. We have the the lightning talks for the remaining lightning talks that have posted session. So first up is Marcus. Good morning, everyone. My name is Marcus Sujanski, I'm a biotechnician at the MDI biological laboratory in Bar Harbor, Maine. For those of that are not familiar with MDIBL, sensors don't feel bad. We are a very small research institution. But the main focus of our research is actually regeneration and aging, which we study using a variety of modernisms, including the axolotl, the African turquoisee, kellyfish, the zebrafish. However, with such a wide variety of species all across the phylogenetic tree, all working towards asking and answering the same questions, it's important to have a robust and reliable way to be able to compare cells and genes across these organisms. Which unfortunately brings me to our problem. Especially in evolutionarily distant organisms, the ability to identify the same cell type or genes across these species in a comparative study which we will often wanna do is heavily reliant on mammalian cell type annotation and gene which when you're primarily dealing with non mammals like we do MDIL adds a ton of unneeded bias into the data and fails to capture enough of the biological variation purposes. So to help solve this problem, I designed and developed SC Samant, which is an excellent pipeline wrapping the existing SAMAP algorithm written by Alyssa Terschansky and Boeing Boeing from Stanford. I don't wanna spend too much time on the math behind the algorithm, but very briefly, the algorithm constructs a shared feature space between two species and iterates between using last words and the single cell expression data in order to statistically identify similar populations between the two data sets. However, the pipeline adds crucial pre and processing steps that format the data directly beforehand and perform novel downstream analyses including home log detection, cross species, cell type annotation excuse me, not annotation. Cross species differential expression a more robust cell type to cell type mapping. All of these additions transform the already powerful same into a ready to use plug and play single cell comparison tool. All in all, SC SAMM app capitalizes upon industry leading single cell prosthesis comparison tool of SAMM in order to extract live cleaning for results Hi. My name is and I'm a final year of teaching student at the City University of New York. So my last study and the outputs are not usually very easily shareable or accessible to everyone that would need access to it. So to address this, developed two pipelines. One is called is based on a previous pipeline that was published on our lab that was a galaxy web web server based pipeline, so it's a modified and updated version of that. The second one is an analysis pipeline that takes integrate all the outputs of the first pipeline, including data if it's available, and it gives the user a list of transcripts that can then be used for phylogenetic analysis or for downstream functional tests. There's also interactive components to help people to make it more enjoyable to, like, do the data and every day day. The little QR code shows you an example of one of these samples. Thank you. Coffee, I guess. But before I start, just want to ask Some kind of R and C data sitting around in our computer from a collaborative from a project from and we never thought about looking From लिए never certain Raise your hands and put this in Even if you have anything to say So essentially, what were how do we think this? Essentially what I'm showing you today is a pipeline which helps in identification of our RNC business and looking at the right of the I feel that my job is saving. So I'll just stop Non canonical splicing happens, we found this closed covalent loop that we just made in our come to next finger. Very cool, the physical model of the But, essentially, what it is is Oh, I have to go here also. Sorry. So I think what happens is that it's any other quantitation pipeline, gets a fast new file as input. And it forms a count matrix where the certain identifiers are on the rows, and the columns are the samples. Now what makes it special is the speed of it, which basically what I show you over here. Is the speed of certainty, which I call it 10 out of 10 because it's very efficient, I think. And compared to the benchmark of what is a fastest part fast is useless, unless it's accurate. Right? If you haven't seen it, you go back and see the recording of very interesting talk from yesterday. With the QPCR cycle that is mixed to make the RNA reach the 25 level. So a good correlation means the molecular variation matches with the computational quantification. And what I show you over here is that it matches very well with the published benchmark of CD three, and also overdoes many of the other pipeline that the field has. So it's not just fast, but also is very accurate. Which, you know, of course, comes at a cost of very low CPU and low run time also, makes it very like, powerful and useful, think. And if you want to know more about it, you should come to my poster in the next room. And, I have a good search surprise for all of you who will visit me at the poster. So see you in the next room. Thank you. Thanks very much. I I missed your your bio, He needs wanted to ask you about, getting coffee with a slight side of Alright. You very much, everyone. You're all fine in there. Take your seats. That'd be great. We're gonna kick off the final session of the summit. So it's a sweet moment. We've got some fantastic talks, lined up for you in in the next session. We try and kind of roughly group the talks then And I remember previously I I mean, I'm I'm not very familiar with the Zero myself, I I I'm Hello, hello. Tal vez rappeen la persona So two years ago, I gave a talk on the stage about, scaling cloud workflows workflows the cloud using high performance speeding technologies like Azure And in the last two years, quite a bit has changed. With AI now. And so today, idea is that we're talking about building multi agent workflows for our programmatic processes. Using Microsoft foundry Nextflow. So a little bit what we'll go through is, I'll give you a bit about me just to so you can see where I'm coming from. Give a bit of an introduction on where we stand today with, Bio AI, and then every tier on the same page of, like, what's model versus an agent? Then I'll get an introduction to Microsoft Foundry, and then I'll give a demo. Assuming the video works, we'll have a demo. So two seconds about me. A computational biologist by training best in most of my career in building cloud architecture for mathematics. I am Two So some questions and goals. What are agents, and what can they do in the science world today? What is Microsoft Foundry, and how does that enable, enterprise grade AI? And then how do we implement those workflows for Mathematics? Plus using our favorite tools, course, in Nestle? So what is a model? I think we all understand what a model is individually. This is a large image model. Can accept users input, text to input, give them text to output. We're all aware of that. And you can provide us some prompt that sort of guides it into a particular type of behavior. But by itself, a model doesn't actually connect anything isn't connected to anything, it's limited in its knowledge to its training data and whatever information that you've given it as part of context when has no memory between different conversations and really connected. It's pretty same alone. So some examples of models are these foundation models like GPT 5.4 or Qualia OS 4.7, also open source models like what you see with Microsoft three four models and the millions of others that are on Hugging base. But when we start connecting those models to other sources and other capabilities, this what we call an agent. So agents utilize that model, they perform a particular task as defined by its instructions. So the agent can use tools, and connect to other other platforms through what sources like NTP. Can search the web, or they can also query data from a RAG based architecture. They can have they can detect you to memory store that allows them to remember things in between chat sessions. Think about you use Cloud or ChatGPT, it it remembers a bit about you between the sessions, even say, ask for your agent to make political content, or you could also prevent the output. For example, you could limit agent from writing code. So examples of these instructions or the purposes for agents are, you are a research agent, you're an extra Python programmer, or you're a famous Italian chef in Palo, Is Palo a good good? Yeah. Of course. So if we zoom out a little bit, and we if we attach multiple agents together, this is what we call a workflow. So workflows work really well whenever there's a more complex task that may require multiple steps to get to the solution. This contains multiple different agents that have different purposes. And then also we make some logical steps that help to either throw out an error or re loop back, to to try again. And so the very basic diagram that I've been showing on screen here is we've got two agents in a sequential flow where I've the first agent that maybe can reason on a problem, break that problem up into, into a more manageable chunks in a more consistent way, and then hand that task over to a coding agent that is more specialized for a coding task. I'll show the the next move all is this. There are other architecture options including concurrent architecture where multiple agents can work at the same time on different tasks. More group chat where multiple agents work the same time on the same task, and then another architecture that's being that's pretty popular is the human in the loop where agents can work but then if there's some sort of action where we need a human to say, to approve it or to approve it or to start a task or something like that, then there's a human in loop's a pause where the the human can say yes or no. There's a million of our architectures, please, which the examples. So now we're on the same page of that. So where are we with bio agent agents in the field today? You may have seen with OpenCall, some people have formed some wrappers around that. Call Bio and BioClaw, of course, two different platforms that almost the same name. And there's some of is an open source one that can obviously and then also the big players are trying to pay out their own life science specific things like GPT or call for life sciences. But all in all, these have some very common utilities, analysis, astrobiology, visualization, which research, And recently, some of them are coming out more quality control automation and regulatory prep stuff, which is great because that's part of most of us are not doing. If you wanted to run some of these tools locally, it's cool, right, that you can run this on your laptop. It works. But it's it's great until it isn't, until you hit some sort of limits. So some of limits are what I show here. One is scalability. So if I wanted to run a very, large model, can you run that on your laptop? Or can run that if you have Mac Studio or a a pretty GPU at home? Maybe. But can you run multiple of them? Can you run a big workflow with multiples? Right? And if you have things running and you but now you want to connect other data sources, do you not configure that? If you can't configure that, are you doing it well, are you doing it in a in a way that that works more optimally for for your agent workflow? And then let's say you get all that set up. Now we've got the security question. So how do control who can use your AI agent? How do control what they can do it? And this is where it's really difficult to sort of string together, to take these things together when you bring this locally. This is the only marketing slide in the whole poll deck here, this is where Microsoft Foundry comes in. Microsoft Foundry is was announced in November at night. Is the generative AI platform in the Microsoft Azure stack. And it includes various capabilities that I think are really cool that really easy to use. The first thing The first is the model section, you can internally add your organization. You can deploy your own version of, say, cloud or a GPT model. You can also deploy a series of open source models and over 30, like, 14,000 different models from Hugging Face. When you deploy these models, they're internal to your network, to your security, to your everything in your in your environment, which means that you can achieve the security and compliance requirements that you need for the privacy of your your data. Once you deploy these models, you can wrap them up as agents, and those agents can then connect to other services and foundry. Versus every IQ, which allows you to build knowledge bases on your own data. It's super easy. If you have it in a daily, for example, or the databases Foundry IQ can do the the inbound process, and you can connect to the the knowledge bases there. And then also Foundry tools And then also Foundry tools includes tons of tools that are already pre configured for other cloud services that could be, like, Azure SQL DB or Databricks or, also things are not in Azure Stack Superbase, for example, is a full. And there's also a generic NTP tool connector, which why I use a demo that you'll see in a minute, it where you can connect your agents to these external services. There's also a local foundry. It's an open source project, foundry local that sort of Ooma, where it allows you to download and run models locally on your laptop. The difference with Foundry Local versus other tools like Ooma is that it will re reoptimize the models for your specific hardware setup. You don't have to worry about, like, picking the right model in CUDA or they can run, like, quantize it for your specific memory restrictions. The foundry local does that automatically. And all of this is wrapped up in a control plane that allows you to achieve the security governance and compliance that you need so you can keep an eye on who's doing what. Talking to you, is costing, those and places, etcetera. When you spin up a foundry project, you get a home screen that looks like this where you get a project endpoint. This is an open AI compatible endpoint If you use OpenAI Python library, this endpoint works for that. It's kinda funny because this endpoint also works regardless of what kind of model you're using. You can deploy at call And but also, you'd be the similar endpoint for all your agents and end models. So I'll show a little bit of demo. And in Microsoft Foundry, workflow builder tool is this nice dagger, nice GUI based interface. Abstracts out to a YAML file, which I'll show you at the end. What we'll build in this workflow is two agents. The first agent will be live from the active evaluator agent. It is prompted to analyze a specific scientific question and break it down into consistent steps. For example, what data sets are needed? What type of analysis is this? And it will give a more structured way of handling the for the cooling agent next to handle that that process. So the next agent is a, is a cooling agent that is connected to my secure NCP endpoint. And it also can write Python code and NextLoc code and actually perform the analysis. And we wrap this up in workflow that has some logical steps. To tell us we need loop this backtrack in or fail until user. So, again, there's two agents for two purposes. So one of left here is Bioinformatics eVyware agent. Been prompted to say, you're super great scientist. Break break up the problem into this consistent format. And you notice that it's connected to a tool. The tool is the web search capability this agent actually has the ability to search the Internet. And then the second one is a secure agent. This one is your coding agent that's an next for Nextflow, and the tools that it has connected is a code interpreter, which can write and execute code on your behalf, it also connects to my secure account through a token with through the endpoint. And the way this works infinitely. Start with loop count one, and then we store the initial question as a variable. Then that first agent takes that problem and breaks it down. If it's not successful and break it down, takes this left path where it will say, complete, but sorry. I'm not able to actually help. Please ask me another question. If it's successful, it takes that right path and goes and hands that problem over to this agent. That agent can then take one of three paths. If it's successful, it will complete and give the output. If it's unsuccessful, it will loop back a couple times or a few times to the the FiveMac's ValueAgent. Then if it hits a limit count, we'll just set the five here. Say, sorry. Maxed my And, hopefully, this video works. Bye. Okay. So in this, mentoring I'll zoom in a little on this. I've asked this agent to help me, perform analysis on a bio project from This is a from some plasmadium buyback samples I produced, like, six, seven years ago. I gave it no other information other than helping do this. And the evaporator agent is able to break down that that problem. Actually went out and found the and then headed to the Square Agent, which was then able to go out, grab the individual samples. I'll show that. It's able to grab the individual samples and then execute or sorry. Go to the next little core, pick a perfect pipeline, and then execute the code. So, again, here's where we see how it designs our product. And then I didn't tell anything about plasmonium. I didn't tell anything about RAC. It actually found out It found it from the 10 of them here. Notice that it what kind of data it is. And then it goes and the pipeline. Then it starts actually writing the next code. And then if I allow it, then it will it will create a file execution plan and execute this in the in the platform. Then it'll give me the results and outcomes and also give me any initial output information, like, many tries it took, etcetera. So all in all, this is where tools like Microsoft Foundry allow you to achieve the enterprise sale AI, so it gives you the security clients, etcetera, but also gives you the flexibility to build a network flows that work really well for next flow just out of the box because security also gives build files into your project and and recreate this whole thing, as I mean, you get your own into key type. And then the director at the bottom here is the O'Reilly Amazon page book just came online this morning. Not quite available reorder, favorite that. Then it also started writing some extra code. Do you know what was the extra code it was writing in one? Yeah. So there's there because I didn't actually tell it to execute in my script account for demo purposes, the it it sort has both That's something where the prompt had changed a little bit to say, if you found one, don't write a new one. If you didn't, then then help. So that's what's going there. Okay. It's writing a new by play based on I found it. Yeah. It's more of demo workflows that's sort showing how the that language should be used is sort what's What's your experience with, or what's your feedback on using the big models like GPT or Opus on NextRoll? You think they are there yet in the right next to like a good NextLow partner or the or they are things that are getting better? The short answer is I'm not really sure. I think the challenge is cost trade off right now. So, for example, the book I just finished the the rest of book and sent it off to for peer review. And to get through the example in the book, it's, $600 just to get through the examples. And I'm not running anything huge. You may see that I running GPT-four O's. Not the latest and greatest models. So in terms of, like, is this better or as good as a certain level of next level program? Because at some point, should just pay for the next level program, to be honest, I do think that some of the open source models are are doing really well, especially for coding particular purposes. I have a conversation with someone about, like, client tuning in some of these models for, like, Nexlopecific or MacroMettisific use cases. And you can just swap those into the models to swap out GD to whatever for that fine tune model, that's Awesome. I can I no? No. No. Not thinking. Wanna do also on the on the data perspective, when you get the code, does it or The only weird thing that it does is it has no concept of just the endpoint itself have any concept of where the data is stored. And so some of pathing is a little off. So that's where, again, making the prompt a little bit better or being a little bit more verbose in in what I'm asking it to do or in the in the instructions would improve that. Otherwise, could look straight to me, but the is a little weird for some examples. You took my question about this. Was gonna ask it. Final one for me. Mean, I really enjoy seeing, real life examples of people using those to go MTP in different environments. I mean, I think it's a common theme for we try not to have any kind of old gardens when you lock in. And and I love that people can now kind of take our tools to where they're already working. Is there any kind of anything what's your experience like with the security being? Is it new? We're kind of still feeling a little bit about how we should work on what best interaction with this. Did you find anything was missing with it? Or The let me on the spot here. Not say bad things. No. The it was actually really easy. So with the foundry generic MCP tool that's there, all I had to do was just figure out how authentication took and worked. Which wasn't in here that obviously got working. Otherwise, the rest of the tools are working well. One one thing I found it that's really nice is when you're playing with this, can get, like, a demo chat window up on the side, and so you can continue to play while you're building to see, like, oh, I changed the prompt. I changed the model where I'm testing this tool. And it was it was really easy. One thing with NTP is that and when with calls NTP is they authenticate permission to each individual action. So that's a little bit going. That's not that's not a super problem. That's a, like, a security thing where it's saying, security thing where it's saying, oh, I need you I need you to grant permissions to list the datasets. I need you grant permissions to actually go clear your data. I need you to grant permissions to actually execute. And so gotta do multiple times before the thing would finally actually work for the first time. Otherwise, was Sorry. Said fine one before I have one more. In terms of, like, as we move more to this model using agents, how do you find the determinism, like, aspect of things? I mean, agents are inherently not deterministic by is there a way we can kind of make them more reliable? Yeah. Think using using agents to execute is a lot better. You think about it as you're in your design process. Right? This is the conversation I had last night of when an agent is having to, like, rewrite this encode every time, excuse me, that's a poor use of tokens and also just gives you variability in what's being run. Whereas if you just have, like, a run script that's sitting there and all the agent has to do is recall that same run every single time, then that's way more consistent. I think that the design is don't have your agents redo work just like you would have a human employee redo work, have the have the agent or or whatever. A standard operating procedure that they're following, and then that that gives a a higher level deterministic feeling to the to the process. Okay. Thanks very Thank you. Fantastic. Next up, we have Joe from Pfizer Fantastic. From So yes, So I'm at Pfizer. I'm in the machine learning and computational science group who started the medicine design research unit. So our group sits in between the research units and our digital infrastructure team. So our team is a mixture of bioinformaticians, confirmations, and data engineers. Today, I'm gonna talk a bit about how we're updating our competing infrastructure that supports our research at Pfizer. So first, just a little bit about how we use Women's Data at As it seems like, use it to inform our decision making. So research, like you can see on slide, have this nice linear path. For how research is done, yeah, it's always a of jumble. But we can't impose some on her on it. Looks like this. So, typically, it starts with using to identify new targets. And once those target targets are, you validated in some way, moves into the preclinical space. So those preclinical systems typically generate lots of own data. And we use that OMM's data then to push the programs into clinic. And so if our therapeutic then makes it into clinical studies, then biomarkers and homics data are used to enhance our understanding of patient populations and the disease itself. And what we learned in the clinic, you know, ultimately can come back to the top here and and inform our selection of new targets again. If we then look at who is doing this work, advisory, we have a group called integrated biology which actually spans all of our research units. So each research unit has embedded computational team, So I'm showing the the the research units are part of integrated biology here. It includes psychology, information, immunology, vaccines, integrative drug safety. Biostats, and then our group, gene learning, and then comes So, you know, this is a pretty big topic, and it's one that we thought a lot about at which is the promising reality of, artificial intelligence in And so a few couple of years ago. Now, all these groups in IB got together and started thinking about, what's gonna be our path forward to to benefit from from these new large link models and the agent capabilities that that come with them now. So I think everyone has a similar concept of of what this could be, this sort of desired state that we're gonna get to. Where a researcher can ask a question, and that question can trigger an LN based workflow that has access to data, access to tools, and it can return synthesized results that can lead to to new insights. And then as that keeps moving forward, you get to this this next stage where you can actually use these models to basically go straight from your experiment to to new insights. All that is, you know, a company that's advised, much more easily said and done. So before we we're gonna get to the state, we needed to to, like, review the ways in which we're currently registering our experiments and tracking them Where's our data? How are we accessing it? Are our store practices? Where is all our compute running and what methods are we using? For that, we began this initiative, and to make this into reality, we started to invest in forty years, and that includes, generating end to end data at sale, data and creating a nice data and analysis environments. Providing sufficient computing power, and then attaching to it all these paper purpose AI models and methods. So call this project iFox. Stands for infrastructure for all mixed cross functions. And what this group is, it's actually a very tightly net knit group of both our R and D colleagues and our information technology colleagues in digital. It only works that we have buy in from both sides. Because we need this infrastructure to work for for this type of type of project. It is helpful, though, to give you picture where we started with all this. Had a very hergenous landscape across all the different research units in terms of where data was stored. Where, like, the the actual storage solutions for each of these. We had lots of different ways and platforms for people to access compute and multiple HPCs. Cloud environments, people running work on Windows versus Mac. Pipelines, we had very few joint efforts across the research units. We had no joint standards. Lots and lots of shiny apps that probably are not really working anymore. So this is gonna be very difficult to bring together to answer the current framework we're gonna start making our data available for for these types of agentic tools. And so we also realized, like, you know, what what do wanna get to? What's the future state that we're trying to get to? Essentially, it's just making time for science. So getting rid of all of these, you know, bloggers that are people from getting their data quickly in access to the people they need the allergic sources, and really just make time for science. So of the first components of this is organizing your data and pipelines. So making your data easy to access and share. Having that data be well For acute and AI, we really need to get this set up for scale. We also needed to make it easy for people to access to different machine learning and AI models. And make our interactive workspace very depicurable, and then I'll set up a solution for fashion queue. But we also want it turns into, like, a community effort where we're actually trying to make our IB communities stronger together by learning with from each other. So training development by our colleagues for the latest and greatest tools, providing documentation support, then also standing on some self-service infrastructure that they can easily go in and get things started. And so this is sort of our vision of what that looks like. So we think that in terms of going from seeking sort insight. And so all of our all of our own data, you know, we're starting from some type of sequencer. And now as we move into cloud, need to make it much very, very simple for the data end of where we're gonna be accessing from. So all of our sequencers now are connected to their gateways, they all end up in s three. And that can get registered inside our daily house, which is built on this platform called these different types of datasets, sequencing, genetics, proteomics. It's also there in in the cloud. And then to to attach and get that data, we need our compute solutions of BatonCube, Pine W US, of And someone starts by going over with this daily concept because it really does underlie a lot of the work we're doing. So this is an open day framework for for biology from layman And way we think about it, it is, Lamin is the data as GitHub is to your code. It's a way to basically track and trace the full image of your data as it works GitHub It's to its capabilities. You can directly attach the inputs and outputs of any transform, which is the Python script and markdown file and Slow pipeline, etcetera. It's a very generic abstract way of viewing code that you're running. And any any data set that's coming into the transform that coming out of it is gonna have full image of who did it when and then all those artifacts can be versioned as well. And it's something you can actually work with directly. So inside, like, a computer node, printing load, the the language package along with the BioNT oncology and actually starts connecting to instance. So this case, you are no longer blocked by S3 policies or different file systems with different access control. All you need is your API key, and then you have access to anything that's in that that it leak us. And when you connected, can also search. So in this case, you can search for a game DigiTag and find any RTAC that's associated with that And immediately, actually, instead of just, like, downloading file and trying to load it, you can load that directly into a 100. It'll start start working. This is what the front end of it looks like. It's it makes it easy to attach in a controlled way. For all of our relevant entities and ontologies. And then you can search It's got a nice fast bit search. So if wanna search for, here's our file, that's a given tissue type or cell type. You can you can find it here. And then easily, get the unique idea of that artifact to pull into your your analysis. So, yeah, gonna move on to then something that we've assume is a common problem because it's very common for us. Is getting people to attach the data to their experiment. Generally seems to as easy as possible. And so Rebecca White is actually here today. She's been working on this tool. I think we're still deciding what we're calling. But this is an easy way to basically create such temas that that people can then use NLMs to correct and quickly get example sheets set up for one of their pipelines. The the underlying back end of this is actually flexible enough that it could work with an interactive analysis too, and so you just use your Schema, and then you can curate metadata that's going back into Lyman. And then we have a more harmonized set of of metadata that you can use. Pipeline strategy looks a bit like this. We do leverage, as much as can from the open source community, The reason we do that is because there's a lot good standards and best practices going on, especially at end of court, and there's always new innovation happening within these It also lets us dedicate more time to working on pipelines we have to make from give us a bigger breadth of the pipelines that we can supply to the R and D groups. I said, is a partnership between R and and digital. And so way we set up our relationship is R and D is responsible for figuring out what pipelines we need, and how there's how they need to run. And digital is in charge of setting up infrastructure that allows us to run them. So this is all running on AWS Match, and, yeah, it's been running on the Scare platform. And to kinda bring it back this pipeline into Lamin, we can actually use Lamin as a way to set our pipelines, register our data, and then the pipeline becomes transformed, which we track in post several works together pretty well. So that's pretty cool. And this is the object from GD's pipelines. He's also here today. For interact computing, we've been using the DATE studios from Zecara, which have been very great. We we did a comparison among a number of different interacting platforms, and the thing I really drove with the studios is how configurable they are. So this lets us do one beyond letting the the researcher take the resources and attach data that they want. We can actually provide our containers And within those containers, we can provide access to all the different fixed environments for topics that the researchers are working on. Each of them play a result in it. We also kinship these, period set of skills for the co pilot. So they're a client seal that anyone can use in TalkTwo and learn how to quickly use this new tool. And then eliminate their their data. So, yeah, it's been a very popular platform for us. And then we also went in nice, creative set of reference data. For that, we have this IFOS database. So it runs on a, airflow deck. The reason we we set this up is everyone had their own client set up for connecting to these these public data sources. And this provides a single entry point. Where you can access all the this reference data and, exit through these R package or Python package whole loop and then sort it up into this platform called evidence. Which I'll show next. But, yeah, we have a lot of data that's ingest so far. These are running on the triggers anytime there's new data set uploaded, then that triggers workflow and get the the latest reference set added. And so evidence, this is our our visual analytics. There are a lot of apps in here now. This is with our partner, David Vision. They're based in Austria. And what you're seeing here is a single cell app. Actually Rebecca also worked on. And it's there's a lot of apps in their pharma pipelines checking to see, like, what all the different clinical trials are. Yeah. There's a lot of We have a lot coming up still. There's always more due. But, yeah, some of them include this smart and a lot of paper version of our machine learning model registry that's available now for people to just access any their cutting base models that they're interested in that have been approved. By legal. And some work on saying about the regulated environmental theory. Always new pipelines coming in. With Lamin, we are trying to set up a just sort of get up with code, a code request for data. If you're working on a reference set that's someone else a reference set that other people are gonna use, then there's a way for the owner of that data set to approve any changes to the effort you're using. So, yeah, we have a learned. We've been doing this for about year and a half now. Adapting adopting system takes time for a community. You really need documentation tutorials, nothing replaced face to face, mentoring through a lot these things. And, really, the research and IT organizations for any of this, the work we have to work together. And our day products VCS leading assets, not something you just set up a time, walk away from, gotta keep improving on this. Hey. Go. So, yeah, a lot of people involved in all this. Team members are here today, so if you'd to talk to them. So, this. I just wanna end by saying we are hiring. So if there's anyone who's looking for a data engineering or pipeline engineer role, Look at the description here. It's not officially posted yet. So once it is, then, you know, you invite there, but please feel to reach out to me. Early in the morning. Thanks. Thanks very much. That's a fantastic talk. It's a very comprehensive ecosystem you have. It's really impressive. Any questions? I have some. Have some. You explored that Nextflow automatically registering that metadata in Laminate or is not made process. I agree that getting people to annotate that data is silliest, hardest part it's holding. So we haven't I mean, wouldn't say duck duck like within pipelines themselves, but with the linear tracking down inside the the language, it actually has more information than we're getting for in terms of, like, the modules that were being run. So it's it's a cool addition, think, to what we have. But we still need this standardized set of terms that that we have to enforce on our or otherwise, it's gonna end up, know, way way too history like them. Too much variety and too hard to actually, like, integrate different data sets crossing in. So we do really rely on that that upstairsing step of registering your your first. Then from there, then it can propagate into that starting gate products. Now one of the things we built into the lineage was the ability to ask kind of custom and it was a bit of a kind of We weren't quite sure how people would use it or what people would use it for. I had always always been it being tracking them through the interest. kind I just say that you are a cloud and medium. So I'm wondering that you run this everything on the cloud. How about the computational cost? For that? Great question. So, yes, all of all of these these capability platforms are running on AWS right now. And the cost is, like, you we were shipping from, like, on premises HPC computing to off prem cloud. You can you can save money by doing a cost saving plan upfront. And then the actual money paid because of that upfront savings plan is a lot less. But, yeah, I mean, we have monitor costs all the time. And I'm always looking to be wondering why people are using resources they really need and keep track of that. Because, yeah, possibly on something you One question for me. There's two Shared reference data Yeah. Actually, in our group, has just finished a new next level pipeline that we use. For making custom genomes. Or anytime there's a new reference to release, that workflow will get run. But we often have groups that, you know, have very special model systems for their for their mouse genomes. There's a gene added, or if it's oncology, it's a patient driving a graph. So getting of these these different flavors of reference data added can now all be tracked with this genome manager pipeline. And then those on so then it becomes a safe source of where those students are So, I mean, I I set up shared AWS I think it was also years ago, and Right. Last talk of the day before the panel Thank you. So first of all, thank you, Mr. Kara, for Austin thank you, I work in the Azure Storage Engineering Organization. I follow as a product but in general, whatever is an HPC workload is something that they tend to influence on Azure in terms of storage products. And in my experience, I also had the opportunity to interact with many of large scale customers running HPC workers on Azure, and some of this many of them are also working in the genomics domain incidentally using So the idea of this session is to share with you some perspective and high level learnings that has been the one that made our customer being successful in the cloud and also helping them optimizing both the performance and the cost at the same time. Just a disclaimer, if you're not the users of Azure specifically, this session of from my point of view, even if he's talking about specific Azure product in some parts, still contains underlying concept that you can apply Looking at the looking at the agenda, let's say, first maybe for many of the experts in this room. This is just a little bit of level set about what does it mean moving data terms of input data, output data, and then looking specifically on how this can be basically optimized. Another disclaimer I would like to say before we start is that the presentation will be talking a lot about the the details that are being seen in the next open source code, but many of these optimization and that orchestration are, in the case of Secret platform, carefully managed and simplified thanks to the work that the Secret platform does on of So you will see a lot of details to go into, understanding the the real all the specific of the data moment, and we would be referencing to the open source next row code. And say when we have a task processing next flow, at a very high level, you will have a script at the core of your your task that will need some input file to perform some operation on. And, of course, there will be output file that will be output of your final workflow or would be intermediate output consumed by your other process. Let's say that when you have a workflow that is composed of multiple process, you usually need to have an executor which can have different forms starting from logout to cloud executor that we'll need to execute in sequence and with the proper dependency, all your task, being sure that for each task, the script is able to consume inputs, and then is able to daily outputs to the subsequent stages or even as an output for the pipeline. Let's say that at a high level, if we want to visualize a pipeline in a sealer way, of course, we know that the this is a That this is a very simplified case of sequential processing. The main idea is that there will be a continuous consumption from a different task of input file and, finally, a production of output file in a working directory, which will be consumed by the subsequent task in a workflow. Let's say that you can see that there are also scratch directory options, and all of these sum ups in saying that for every specific executor inside the next row, there are basically what are defined five copy strategies. Five copy strategy that basically determine how the data are available to the task and how the data are eventually with the specific staging scenario as we will see in one of next example are staged for processing in the intermediate storage like to get the most performance, but, of course, with the staging and a stage of margin that you make to balance. MSAT is looking at the the copy There is basically the classical Azure Batch executor. Azure Batch is a managed service on Azure which is about work is straight. Using These executors are giving us specific strategy strategy relying heavily on staging. Again, we're talking about the open source case. Heavily not talking about what's secure platform that's very nice, for example, with secure fusion. And then we have all the a have usually managed with the Azure cycle, which is the service that that sitting on the side of a classical HPC scheduler is able to infuse a Azure which with the Kubernetes executor. But going into the details, why, let's say, we are looking for this? Because depending on the executor, you have different possibility on all can optimize the cost of the performance of your infrastructure on Azure. This is generally true also in other scenario. Usually, when we have a customer, this has been my personal experience, there is a strong shift in my set that needs to happen to find value to deliver value for the cloud, which is the underlying concept is that the we come with the mindset. I have this computational budget. How many petabyte or how many cores I'm able to purchase? When we are able to shift the mindset to say, what is the highest amount of science that I can produce from my computational budget, is finally what scientist wants in term of added value because nobody takes any value known in course or known in petabytes. The real value is basically what you can deliver. In terms of science. Next to that value is where, of course, we're doing the This Assessment and measure of the data so that you can address, have And at the same time, when you do data stage, that we're seeing in several of the scenarios, it's something that happens. Making some example. When we talk about When talk about It's a super Azure Blob storage centric. Azure Blob storage in the standard tier is the most efficient TCO storage that you can have, but need to be aware of an object interface. So you need basically to stage in data and lock in a mini processing and then to stage it out. And this can be used and is also the default for the working that allows you that can be used at least for input and output mounted directly to the copy modes. And that will, of course, depending on the that the usage of your or your pipeline may have different cost and performance you look at the other domain that we're talking about where we're going in direction of the simple five copy strategy, in this scenario, you have better not better, but you basically a different approach to the the storage in the pipeline because you can't rely on a shared working directory that is a process of system can While object storage is usually still used for input and output, but usually not used at least if we are still talking about the is not used as a working directory. This means that in this scenario, you are able to plug in inside the inside the storage option different solutions that are shared system that can go from super performance system like Azure Manager Laster to high performance NAS, like Azure Meta file, and even in plugging in Azure file mounted in NFS and so on. So all of this, as you can see, is a broad spectrum of different possibility that needs to be carefully tuned for the specific scenario. But other than these architectural examples, let's go effectively in what does mean optimizing in several of the scenario we have seen the cost and performance in a pipeline. In many cases, we're seeing customer that for the input file, they are staging in for each of the pipeline that run sample externally. From basically the cloud environment or Transfer account using solution like that are file based. For example, Azure in a way that is performance cost efficient is usually a good success. Another example is the use of staging. Staging in some cases is mandatory, like in the Azure batch case. In other cases, you can trade off depending on value pattern or the process. Or it's better to basically use a field system. And this goes into the Scripts and the tools and the application that probably need That Talk it a lot, and this For your troubleshooting or eventually for your scientific research, but at the same time, going to direction for long lasting data who use to use the All of these are things that we have seen and we need to optimize on every case and on every pattern that we've for our customers. And it is something that, of course, can reduce the cost an enormous amount. Going into specific, one thing that we have seen happening more and more is a storage parting that we define core core storage and storage quality. Some of the Azure's resolution has the possibility to be POSIX compliant system that have SST performance to be at the front end interacting with the compute, which are seamless integrated through the core storage layer, which is usually object storage. And when I say seamless means that you can see from a namespace and a data perspective a continuity between the two. Azure All of this is a great value because allows the users to optimize the layer above in terms of size and performance just to deliver what is needed in the working dataset while basically then tearing down storage whatever as a longer term retention. Another aspect that is very usually also pretty neglected but sometimes is very significant is how basically do we connect from a networking perspective resources. Neglected, Azure dot storage account have the possibility to go into the direction of being multiregion, meaning that you can access the storage account with different endpoints. Is okay, that gives you the possibility to avoid extra network cost that you'll have in that scenario. Similar cross region transfer, again, you need to understand if it's a one off, or it's better to create a copy in additional region to avoid the performance and cost reading from from another region. Other things that and the compute in a way that gives you the lowest lanes. All of this seems to be very complicated and overwhelming, but one of the things that work more with and, let's say, work very well with our customer is the monitoring that comes for free in many scenarios in the cloud that usually start to get on premise. So wrapping up the session, I would like to live with the some key concepts that are first of all, there is no solution that can be one size piece also. Depending on your workload, depending on the difference in your organization, depending you can have an architecture, but there could be deviation that you need to do for the specific cost and performance optimization of each of the packing you have. This goes reviewing that and you don't keep basically validating how your robot is changing and now needs to change the infrastructure, As a consequence to that, you will for sure go outside the boundary of the best cost and performance optimization. And last point, the groundwork is always the key. On prem, we're used to say we have one or two solution at most in many scenarios, In the cloud, you need to deal with the fact that to get the most intermost constant performance, need to a parting of solution that you need to plug in and manage in a way that you can get the most out of your budget in terms science you can produce. You so much for your time, and I'm really happy to answer if there there are any questions. Thank you very much. That's a really comprehensive talk. Thank Obviously, there's a huge amount of flexibility and ability to customize and different options that you've mentioned here. You said yourself, it could be underwhelming. On kind of theme of of this particular summit, there any way that AI can help people make these decisions? Absolutely. We one thing that we've seen in this amazing in my point of view is that the we have many MCQ servers. One of them is the Azure MCQ server that allows you to connect directly to the Azure control plane And for example, you can tell a specific job or a specific time frame in which you want to to ask him to analyze the resources and what he would suggest in front optimization. And we have seen very impressive results. Sometimes he's doing way better than me. The suggesting from we've Thanks very much. Thank you so much. Okay. With that, I'm going to welcome Evan on to stage. And for the final part today, we're have a really exciting panel. But Evan, think I'll let you the instructions if that's okay. Awesome. Yeah. Thanks. So first off Background microbiology in that country. Awesome. Hi. I'm Maria Enrique. I'm working at Ceres for almost five years. I've working on agent KAI in bioinformatics, but I also have a background in developing pipelines and running bioinformatics analysis. Hi, everyone. My name is Isha. I have been at the care for about three and a years. I lead our scientific engagement team, we're a team of working closely with pharma and biotech teams really across pipelines, infrastructure, and now newly AI, and making all these kind of three components work together. Awesome. Okay. Let me start off with the first question we have here, this is really around thinking about the day to day work that is going in terms of what, like, buying from physicians and computational biologists doing. How you can see their work change day to day? Well, when I think about when I started in BrainTraumatics in the late nineties, didn't have bioinformatics programs, and it was all kind of self taught. I started in And today, if I had these same tools, obviously, my trajectory would have been very different. Yes. It's amazing how quickly you can, you know, get almost, you know, production ready pipelines and capabilities. You know, it's a it's extremely exciting. Yeah. Awesome. Yes. And I feel like the role of the bioinformatation has changed a lot. So being like a hands on executor, now it's like mainly an orchestrator or an architect of solutions. But I still see that the role of information is really critical. Mainly, like, first, they need to find, like, a relevant problem and a relevant question that they want to answer, but then in also setting up the tools and setting up the environment with their domain expertise, seeing what would be necessary. And then also, it's really needed that they review the outputs. And with their domain knowledge, they can validate them. Yeah. Yeah. It's interesting. Think the role of biometition, scientist now, as we've seen from the talks, span such a wide surface area beyond just pipelines. It's the infrastructure, it's the tools themselves using containers and as that sort thing is growing, it's not a bit more difficult to wrangle all of that and tackle problems that come from each when you're just trying to get the results. I think AI, in a has made wrangling all those components a lot easier. So less time thinking about these problems more on orchestrating AI to solve them, then spend more time to get on the results to where the humans are really needed. I think if we think about those problems more broadly, some of them can be classified, like, quite, in fact, similar way to code. And, like, last four months have been huge advances in terms of coding. What are some of the tasks that are maybe not associated with coding tasks? What are things that scientists have to do data data not necessarily well adapted for the kind of code you know, code code analysis. Yeah. Well, let me know biology is really hard. Often not exactly deterministic. So I take any kinds of interpretation. Are still a little bit there's some improvement there. And obviously, you're currently with the tools. There's still issues around hallucinations. And so you want to trust the outputs, you need to verify. And how folks think about in terms of experiment tracking here because when you can run 100 pipelines in parallel, how are the folks also managing all of the huge increase in capabilities of 6Q? Yeah. That that's certainly a problem. Now, you know, now to execution and coding is cost of that is zero almost. And so think people do struggle with this, you know, this fire hose of data that we can generate today. If you have good practice, like, you certain things came from a software engineering background, reproducible research background, or adopt those practices, that helps. But, yeah, certainly, secure platform, good help with a lot that managing it. Yeah. That's why I think, like, it's really important that governance is at the beginning of the process and not at the end. Because if not, end up, like, running a lot of results. Validation takes more than running the results. And maybe, like, running those results would take less than validating all of the amounts of the results generated. So that is something we need to be careful. And yes. I think we feel a this week about almost the ability to verify. So if you really know an output wasn't filtered, this I a about verify. You really when Yes. I feel like it's difficult in biology to define correctness. Also, Like, it's not the same as in software development where you would run a code and get a result, you would have, like, test datasets and everything. So some problems in reality, it's more difficult to define what is correct. Like, for example, if you would add a model in the pipeline, then, you have already, the test, and you have something to verify against. For example, if you're trying to include, like, a new tool where you don't have anything to verify, well, that's a difficult part that's where humans are really relevant. I think that I really like that example because in software development, you write a tool, you write an application. It doesn't behave as expected or it crashes, you're able to go back and, I guess, now with agents and AI pinpoint exactly where that might be coming from. You also then include unit tests functional functions are working as expected. But I think the ground truth in science and biology is so much fuzzier than that. Taking a look at the results of an analysis or a workflow and trying to pinpoint why it's not looking the way it's expected to do is not as easy of a solution to tackle with with AI agents. If you can free up more time for, say, cognitive biology or scientists to focus on that part rather than fighting Docker container issues, that's that's kind of the value wanna unlock. And then what are you seeing in terms of the ability to to bring in tools and maybe take something which is, something you try at once, like an AI tool and sort bring it into that organization and sort the changes that are required there? But there's certainly lots of dabbling with what I'm seeing happening, which is great. Share IT kind of Yeah. Share IT now for sure. And and now I feel a lot of empathy towards IT departments because I think they're now facing this issue of how to govern and review, and and there's all kinds of issues around security. So, yeah, I think this you know, AI is just accelerating these things. I think there's still good techniques and practices that you can follow through. Establish even if didn't have AI. But it's, yeah, it's a really exciting time. Yeah. I mean, you mentioned Ad hoc analysis, and I think that is something where AI is bringing a lot is because people who weren't able to run analysis, like, for example, are now able to do it. In terms of bringing that into the company, that's kind of a separate question because it has to have governance. Has to meet with compliance I've seen mostly that sometimes it's more difficult, for example, with all of the bioscientists and the coscientists that are coming. Sometimes it's more difficult to use them to connect them to internal tools than maybe using a more general general system and that is already approved and it's already in the in their companies and connected to the internal tools. Yeah. Think it's it's nice to try out new tools in Sandbox in your your local environment. It's really sticky. And what we've seen with folks that we've been working with is when they start to then roll out some solutions across the board because you're now touching several other systems beyond just workloads. It might be LEM systems, ELNs, You're touching identity providers. You're touching audit logs. So and there's also teams behind those systems, humans behind the system. So trying to understand how to get those in place and then making sure things also so vulnerable. I I don't think I've seen a clear solution to you yet. Think also guardrails are really important and how he might look to approve or disapprove like this these like these decisions. It's really important, also having, like, sandbox environment as you mentioned. Yeah. And it's also really easy now to, you deploy really function functional tools or apps. And then you think about support and adding features and that's something I'm seeing a lot of for sure. I think that relates to that. Are you seeing that some tools becomes more valuable as it relates to new AI era where things are connected via APIs or are you seeing sort of we're focusing more value on that in terms of a decision buying decision being across legible. Well, certainly, skills and MCP servers and things like that are are really powerful. And then I guess going back to that previous question I thought, sure I think when you think about deployment in enterprise environments, know, with SCARA's building and other similar solutions, think that's where you wanna lean into. And so you're not recreating the same app in a couple different groups and you have a better governance layer and sort of support for organizations. Yes. I think actually applying for man six and more like what NFL has to skills for authentic AI systems. And, yes, I feel like it's a matter of more working into like, the skills and connecting it to the tools that are important to solve problems. The authentic system itself. Yeah. It sounds like AI agents will work best in structure systems where there's there's governance, auditing, and then but also access to tooling beyond that and standardization that we've been building on for years, again, unfortunately. Do you have any kind of maybe outside the box things that have really worked well or somebody you've applied AI to that maybe people not necessarily think in the room might be something helpful? Well, I know many people write like writing documentation, so but it's obviously been very good for that. And certainly, debugging, like, I've used, you SICK AI for debugging some pipelines in the cloud. That's been really, really powerful. Yeah. I've also seen a lot of value in maybe creating datasets or creating data data. There's a lot of value there because that is what is also what is also going to use the authentication system. So it's important to have a strong foundation in the datasets, tools. And I've seen DLLMs and the authentication system. Improved that a lot. Definitely documentation. No one wants to write documentation most of the time, but in a way that allows us to surface a lot of the tribal knowledge building in teams. Like, elements can very economical too. Capture a of that, then also be able search through it. It's almost like, like, knowledge knowledge is internally for experimental design, analysis and things like that. I know. It's okay. We've acknowledged based nature. Yeah. Yeah. It's actually on the same vein. We've been using obsidian almost as our kind of second brain. It's where we dump all our chicken scratch notes and how we're thinking about things. And because they're essentially just marked on, files behind the scenes, it's a great way to plug in LMs and do a search across our thoughts, across different categories of development. How we're talking to teams in a way that doesn't introduce Laura Campbell over because we've just been doing that already and keeps our workflows the same. And how was it, how was thinking about evaluation within organizations? Like, we've seen some the prices of some AI, as I told you, look at very, high there. Is it contractional to ROI? Prices Is contructional Yes. Right now, for sure, like I think OpenAI had an early advantage with ChatGPT, but, obviously, we're looking at cloud and using that. So I think we're still at a phase where people are it's very fluid. And tracking those costs are sometimes whether you're using Bedrock for doing some new work or cloud code or codecs, we're still trying to get wrap our head around that overall investment. Any kind of Any specific outside of models, a good question. I think certainly just in day to day activities, we also use Notion to a similar platform for just obviously, transcribing, you know, some things like that. That's been a huge help as well. But Yeah. Yeah. I feel like the tools and what they use various a lot across companies, there's not, like, a standard yet. What I think that they take into account whether or not to include that or not is mainly, like, first, the usefulness, Is that going to be useful for something, or what problem is going to solve them? The consistency. Like, making sure that when it runs the analysis or whatever it does, gets a consistent result. And you get, like, all of the provenance and all of the logs and reference packing results. And then, I guess, also, governance is important. So having guardrails in turning human in the loop, important decisions, having law ins. And also, I think that's something important in implementation of these tools. Is having a system where there's ownership. So, you know, who's, like, owning what. So there's someone maintaining the tools, someone maintaining the systems. So I feel like all of that is really important to get a ton of new tools. Yes. It's interesting you mentioned kind of ownership because there's, like, a hidden cost with using, like, AI agents code creation and generation is extremely cheap now. People almost start solving the same problems, but with similar solutions. And maybe we lack the same discipline that we have with handwritten code to have ownership, make sure you maintain it. So just kind of token consumption for solving the same problem over and over again. Definitely a conclusion. One way we're trying to address that is we have, like, this internal shared repo across the portfolio. So we're we're contributing skills in repo and other tools, making that aware. I think that's a little bit helpful. To address that issues. Yeah. Not sure much visibility to go into some of the larger customers, etcetera. But what I've seen is it's gonna vary vary in terms of adoption and and just the kind of tooling that's available to scientists in these large organizations. Is that something you've seen as well or, like, or any thoughts on where this is gonna go? For sure, there's, like, early adopters, like, people that are when ChatGPT got out, they're using it within days. And then there's probably a long tail of people skeptical Certainly more experienced software engineers in the early days were definitely skeptical of these tools. But it's it's clear it's gotten much, much better. Yeah. Think it's a typical thing. Any, like, adoption of new technology follows that, you know, that same patterns. Yes. I think that, something that, a problem with adoption, I guess. It's probably, first, that people need to be trained and need to be informed of what they do last. And what they can do with it. And then I think that there's a lot of trust issues as well. So I guess trust, for example, begins with knowing, like, what you don't know. And we've seen a lot with some AI systems that they maybe try to answer a question, where they don't have any data or any tools that would answer the question. So I think something that would help in adoption of these systems is also, like, work on that trust. So being able to be transparent, not only with what you to know, but also with you know. So being able to have, like, show, like, the versioning of the pipelines that you are running, show the parameters that you are running with the tools, show the different thinking methods, what steps they are running to solve the questions. Yeah. In a lot of teams that we work with, there is generally considered effort to start implementing AIs. Unclear exactly tactically in what specific area, but I think the general strategy has been start with a focused part of, their teams or workflows, build that trust, understand what is governable, is not, what is missing, justify that trust before going full steam ahead. It's generally a strategy we've seen so far. It's still varied. Yes. I think sometimes it also helps to maybe, like, start with just using, for example, AI to maybe assist and then, like, go upward. So maybe start with assisting you, then start, like, using it by maybe just accepting or rejecting the plan. And then, like, after you are sure of the guardrails and security and governance, then maybe lean towards a more autonomous system. And I also think it's great to have evangelists throughout your organization, kind of doing show and tell, and and that can really amplify AI capabilities. And we have internally a psychanalytical build of AI. I'm just really, know, highlighting things because there is so much, like, to learn to get inspiration from you know, I purchased it in Singapore. So okay. This team only making predictions this time next year around, aging designs. And it's very difficult. I feel like in science, where domain context is really important, it's not about, like, even if you're advances in the model, there are the context and having the expertise, the biological expertise is so important. That just the model doesn't solve the problems. It's not like getting software development. So I think that maybe in the future, there will be more systems that are capable maybe of doing results analysis or high process generation and high end systems that can really, like, co with the scientists. Yeah. Predictions are hard, but I think actually the the token usage cost will go up, think. Rob? Yeah. Don't know. That's right. It's just like some Well, you can install the investments in the raising around the models themselves and which models prefer for particular use case. Think at the end the it's be the system that will win over any particular model in its own. This system has access to a good base model with the reasoning and access to domain context context, and then harnesses on top as well as being able retrieve data and tools. This system has a whole wing over any individual model. This I Let's try this. So I think I wrap up here an amazing, you know, opportunity to bring folks in. If you've got people in your organization who wanna learn Nextflow, I'm really being able to bring them to things like this. Can get exposed both on the training side as well as the ecosystem. So we're trying to reconfigure that. It's regularly. We two days together. Looked a ton, got a huge amount of energy from the from the community, and, yeah, really gonna last a lot of ideas and a of to think about in the Hopefully, you receive the same. Absolutely necessary, harsh selfie, I guess I go along with all of these. This is a picture that we had. Necessary, And the the pixel art, we have a winner. Michael Soria. Hey. I I got I didn't pick the word. It's a secret. Can can we get an explanation of how? I mean, honestly, I grew up on Sierra, an infill company Okay. What's coming up? I've got some some things got There's gonna be a webinar which are useful for sharing your colleagues there really going deeper into secure coscientists, showing me some of capabilities there. We'll learn about that or share with colleagues. And then we also have a secure session, is a And finally, another question would be in here as well as, as ISNB and also So if you're if you're around for that, know if folks join there as well. So final thing I'll do is forget to say the date we have the next four summit, is, gonna be virtual for the summit as of it, in October. As well as we have the subterranean hackathon, which we meet in Boston it, which there. So that's from TS and from where is he going? Sorry. From You know, bring a lot of things together, lot of work that goes behind the scenes. Then, obviously, during this week, the folks on AV, folks in catering, etcetera. I really appreciate that. And final one from me, like, really thanks for all of who've been here It's a it really makes difference to have you here and see you in person. So with it, thank you all, and look forward to seeing you around. Reach out to us in the community as well. Always looking forward to, to catching up Thank you.