2026-04-30-nextflow-summit-day-1-am

Date: 2026-04-30 Source: nextflow-2026/transcripts/sessions/2026-04-30-nextflow-summit-day-1-am.md Word count: ~19,692

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: Vicksville Summit: AI in Science and Computational Workflows Date: April 30

Transcript:

Them: All right, Good morning, everyone. Good morning, and hello also to everybody joining us on the livestream. Welcome to Vicksville Summit. We got great program for you for the next day and a half. But first, before we in, I have a few opening notes. First of all, many of you were already with us for about two days of hackathon and training, you already know your way around. But to everybody else who's just joining us now, welcome, and most importantly, here's the WiFi password. Now I'll leave that for a minute. Have a slide for if you need them, they're just out of this room over there on the left when you come out. So, yes, we in in the unready events of a fire alarm, please exit the building in an orderly fashion. Out that way and cross the street over there and wait for everybody. We will have to count you, so please don't wander off. Unlike you, of course, we have to say. Yes. So we have have very interested in getting the questions. That's a big part of the point of a live event like this. So please don't be shy. Ask us your questions. We'll have runners with microphones in the room. If you're in livestream, you can use a Q and A function on And there's an online moderator who can help you with that if you're having trouble. Last for questions, please, you ask speakers. Anything else on the line? Now next, we are running an art contest. You can meet the Tamagotchi, which is very cutting edge technology. You You were born in the nineteen eighties. But, yes, you can you can win. All you have to do for that is go to some websites. Just click around in the black area. You'll see how it works. Once you're done in your art is beautiful and perfect, you'll see that there's a little question mark in the bottom right corner. You click on that, that'll export your art. If you're not on social media, good for you. Just come and talk to one of us Okay. All the talks are being recorded, they will be available online after the events on podcast or our live streaming platform as well as on YouTube, probably, I believe, next week because it does take a little bit of time to process everything. Your help is really important to making this event feasible for us. So without further ado, I have the immense pleasure of calling to this age our CEO and Co Founder, Adam Floydham, who's going to kick us off with some very exciting events. and it's sometimes I think the community is a really big part of what may get happen, and sort of all of you for the work that you've, you you're gonna do that. Let me start with, there was some some sad news overnight, with the parting of Winter, and a teenager, was personally very inspired by, what had Craig was, in some ways, a maker. It's also an inspiration to many folks. I grew up, you know, sort of told the idea of tracking around the world of boat. Spitting up the seawater, sequencing everything in it, and then having a huge to it. Really it was a driving force in the kind of new era of genomics that sort of took place off of this. I think it's within that spirit that we can we're kind of here today. Think about the kind of new era of science, is changing as it relates to AI. So for the next twenty five minutes, I wanna talk about the three kind of core things. I wanna tell you what the world looks like from from my perspective, what we said. Wanna share a little bit about what we've been building over the last year. And then I'll share view on where this is going and provide some, like, a new insights that that we had to raise this. We think that it's something special that's happy in science right now, and it's a little bit hard to fully absorb exactly what it is. Let me show you great So it's kinda hard overtake what's changing around us. Whether you think about the structure prediction, foundational models of biology, or think about how it science, some of autonomous labs that are coming out. Each one of those things could be like a conference by itself. But putting it together, really see that the scientific process is sometimes becoming regrettable. And we gave all these at once. This acceleration is linear. Really compounding. Process all of that data, and those feedback loops are really gonna start driving the next experiments. And the whole system can improve again and again, and that's really the iteration that's gonna be a key theme of my forte. Think what else it makes this moment feel a little bit different. It's not that we just spin up. Like, we're really compressing the entire loop. And just like in software, think a proxy for a major success in science is gonna become iterations. How fast can we learn? This is obviously happening across the industry. It's really, very fast, and investment is really flowing into this new era of life science. A way that we have to measure maybe twelve months ago or or twenty four months ago. Labs. Validation API from So let's get the world started a little bit at why this why this matters. So for the years, we had this promise of AI, applied science. And really, success in the clinic was somewhat limited. In the last year or so, we had our first AI powered drugs, so AI powered drugs, We applied to trials, And right now today, there are a 173 drugs which are ORI originated are So this is no longer a future tense conversation. This is this is happening now. Any examples about it just biopharma, of course, but this is happening in diagnostics, neurotechs, and theorebiology, public health, So everything is changing in some way. Well, kind of. All technological revolutions that have happened definitely allow us to imagine like, how we can do things differently. But maybe, like, one of the core incredible things about science is that fundamentals often remain the same. So, like, what will stay constant? What do we think constant in this through this massive acceleration? A lot of science depends on having a record of what was done. Who did it? Push data. More data means that we need little illustration. We need verifiable reproducibility. And if we connect together more systems, agents, we need accessible APIs which will define which software you will use. Liquidation efficiency, you'll be able It. But as we improve the efficiency, can do more analysis, more cycles, more iterations, and really the fundamentals that drive that become more important than less. Much more important. But even itself in compute This is, a framework to build on, that either one ecosystem can work with. With the wood, you have Linux, clouds, got communities, mobile And then you think here that the AI power science has mixed flow. You can see in the metrics. So mixed flow itself in terms of the execution plan is increased about 300% in the last twelve months. And in the last month alone, there's been a little further in Nextflow runs to launch. So this is a really accelerating trend that we're seeing across the community and ecosystem. It takes what is like the system for AI And I think this is true for individuals, for Python teams, organizations, So I'm gonna show you what I'll be building using that foundation, and I'm gonna free this through a couple of different The first of is is context. Then we go on to execution, and then see how this all comes together for this new system of AI power science. So I'm gonna start with context here. I showed a demo the other day about one of engineers at Sierra, and they've done is we like to strip away all of the all of the the agent, the harness, the memory of the LM. It should direct directly directly on the API again. And also it seems like without that scaffolding that exist, I think that all the moments that we've seen in last twelve months are often attributed to the which is true. Science is super dependent on that context. There's, you know, samples, your protocols, runs, parameters, etcetera. They all are really key. Think that that context can sometimes be the difference between novel finding and a file which is just, you forgotten somewhere. I think that whole chain really is designed to process. It's that link that takes inputs, it takes our our our analysis and turns them into results. That become verifiable. We did this a lot over the last year in really expanding the chain and and to keep it intact, to go from end to end, and be able to go from pinch through computational analysis. I think our partnership with Pinch is an example of this. We're able to connect the data that comes from the lab, and allow us to use that information that's in the lab for being able to apply to downstream analysis whether this is building moles or using any reporting. I'm super excited to announce this this new launch of the integration that we have today. Your analysis may live in various buckets across the tower. It's like two of the same experiment, but they've been fundamentally separated. May in cloud. So boot scientists can now launch pipelines, combined Sagira's can They can pipelines, pipeline results back to the samples and also the people who generate them. And this really breaks down all those silos. Okay. And when the pipeline has run, so execution's complete, maybe you finished this last night, Something you finished last year. This is this label context, is inexactly what happened. We think that every file in a way has its own story. Which analysis produced it? Which run? What are the inputs? We also use downstream of this. And it was last year on the stage that we announced is part of 25.4, but data linear is now part of the first class that is in within The idea behind this is that if we run every task, every output, it's assigned to this lineage ID. And that really provides us complete provenance after a provenance JSON file. But today, excited to announce that we've been surfacing all of that directly inside of secure. So it's visible from what runs or from the data explorer if the lineage ID is a leak. So that means you can follow the chain. You can either use API, click through the links, understand what's upstream and what's downstream of any given file. We think about verifications of this. Obviously, in areas like from QA teams, you might be able have a sense of actually evidence of the building around this. Obviously, in areas like you to And these advances like this highlight the value of having a system of science. Enables you to really build the analyses together and really build some of systems and processes that you have as part of your organization. Knowing what was produced, how it was produced, can only get there in one click. Was speaking to a a customer there last month and talking about some these new features coming out, and used case. There's normal folks who using Nextflow to be able to train in our models. I think we're gonna see some examples down today. And oftentimes, the fact that a process has happened and then, for example, there was a a bad QC result for one of the examples, or a patient who's removed consent. For the data that was processed here. Allows you to go directly through and understand which samples we've been taken through, which analysis they relate to. So you're then able to retrain the model, be able to remove that data and cut people complaint in there. Sometimes it's not it's just enough to identify file. May have lineage and data set up. You really wanna look into the results And this is the future when you're working the cloud. You may often have to download a six gigabyte file open in the subfaces ITV locally, configurable voice paths. So we've done in the last month or so at IGV and also to quickly enter browser. So this means you can view band trends, PCS, PCVs visualized in place your buckets. So it means no Really how you go from output to, like, visual QC in really matter of seconds. And when those visual checks are not enough, you know, sometimes we need get into data. Studio is an interactive workbench that we've released since Iqara last year, allows you to have custom built use custom containers. Tools, your editor, shell, etcetera. And today, we're super excited to announce So you can connect an SSH client, can use, for example, CoreCode SSH or Versus Code's remote SSH your your data stays in the cloud, but you have a local experience that I know we all want. So customer example of someone who's making most of this, is a a next generation person business company. They've been really working specifically on immune rated inflammatory diseases. Scientists across the organization, connect their resources, and dive into the results very quickly. Okay. So that's the I'll be Okay. All right. Perfect. Hi, everybody. To get a little bit of a demonstration, some of the features that Evan just mentioned, I'm gonna do a high level walkthrough of an RNA sequencing analysis starting from a very familiar place for a of scientists in the room watching, which is eventually electronic So here we can see that we have a table eventually that corresponds to, you know, some sample data and some of the data related to the past of these files. With benching AI in this scare MCP, we're able to actually prompt in benchling in a way that can trigger actions in Scare platform. So for example, here we can see that we're asking, to take this table in the LM and create a data set in secure platform. So we can see that this data set is created. It's making tool calls to secure API. And even more than that, we can go and actually pop it again, take that data set we just traded in secure platform and launch an RNA sequencing pipeline run using that data set as input. When we jump over to secure platform, we take a look at some of these datasets that have been created via the the eventually AI integration. For example, this particular dataset, was exactly like the one in in the Ventolin If we want to take a look at the runs that that particular dataset was used as input for, we can view the runs here. And we get some of these RNA sequencing runs. To show the data lineage feature, let's click into one of these runs, and we're taking to the task page. So here we can see all of the tasks. And let's say we want to get some more specific information about the alignment test. Here, for example, we could we could do the the start test for this particular sample. And not only do we get some some specific test level information, but we also get some linear data, as I've previously mentioned. So here we can see the inputs to this task. The outputs to this task, as well as the upstream and downstream tasks. That correspond to this. If we wanna look at particular BAM file, for example, we can click on the BAM file, and we're taking to Data Explorer. With the genome viewer, we can interactively observe some you know, and explore this particular BAM file. What's even more is this ability to have a full circle look at the data and the that it correspond to So here we can see that we're able to see the task that this was used in before, and the task that used it as input downstream. We can click, and we're taken right back to the task page. Let's say we want to not only, you one by one, but we wanna use an interactive downstream analysis, is very common. In studios, as I mentioned, we can create a studio session, such as this. Is built on a public Docker container. And we can see that it's running on an instance in AWS that we've already configured in our Workspace. As well as has some data mounted from an S3 bucket that we've connected to our Workspace. If we connect to the studio session, we'll be taking to a new tab. Where we have this shiny dashboard where we can view things such as volcano class related to the results of our pipeline. Even more with Studio SSH action, we can connect directly to that instance, that studio session in the terminal. So here, we copy that command. And we can open the terminal. We can see we are now connected to that particular studio session. We can view the data. That we mounted via S3, and we will see that the data for, let's say, RNA seq process now available for us to do. So as you can see, we kind of covered, like, at a very high level, end to end process of, you know, starting and mentioning all the way through the interactive downstream analysis. So thank you, everybody. Doing small use cases here, and it's really exciting to see how how folks take these in community Because you're running your environment with your own data, it really provides you, like, in ways that you can do things that maybe are easy Do things maybe are easy by yourself simply without any restrictions. But we wanna be able do that in a either a regular environment or a founder environment. We really can do that in a safe way. Okay. So we've talked a little bit up till provenance at the dates of year. And we'll talk about somehow the provenance of actions. Anyways, every result that has there is an author associated with it, This could be a person, could be a team, service account, be an agent. The needs are not the same. Organizations need to know who acted, what was changed, under what governance, and really it starts with identity. So if you want a regular environment, you may know the story. Your IT team uses Okta or they use Intra, but Someone has to replicate those permissions. There's hierarchies everywhere. Often people drift. Change, and words can get pretty painful. So today we're shipping 70 features to really fix this. First one is that IDB groups now map directly to Teams in Scrap. So someone can be able to a group in Okta, and they'll have access in SQA on the next login. If you move them, the same thing happens. So your IDP becomes a source of truth for them. Team's tools now use find their authorization. This allows you to have the exact same permissions and really find what those permission boundaries are. So now the RIC can be configured through the organizational settings. And really means that you have to can speak concerned about these credentials With this, it means that every event can be captured and have like before and after state. So you have the filter for users, time, organizations, etcetera, and we have export that. Obviously, highly important when you gotta think about keeping keeping records in a regulated environment. And so speaking of the environment, it's something that's something that a lot of our phone calls are supposed to be really waiting for. So today, announcing secure JXP packages. We're offering this as as two offerings. We have JXP documentation, is really a compliance foundation, including supporting, traceability, and verification documentation. As well as a more complete JSP validation package, which really adds the validation test evidence, sort of includes those test protocols, which we're all aligned to the GMPI framework. Means for customers that you can quickly get going with validated by informatics. Those related environments according to GMP, GCP, GMP, and we're all aligned with FDA and EMA guidelines. What we've seen is more more and more and more sort of analysis has flowed through on the critical part of not just early research and development, but flowing through into clinical and into commercial seats. Together, go from research, through the clinical, into manufacturing in some cases is a real is a real win. This is using Sydney Flow, same platform, just with the documentation and those controls that a lot of regulators expect. Also, the light Orchestration layer for manufacturing in clinical trials. And the process is really what production grade scientific computing looks like. You know, to start early on in research and then be able to bring those same through into a clinical setting. Okay. Let's wrap things up in terms of context. You can see how some context helps in terms of speed up the the process of science. Let's get up in just some of the speed upstairs that most new adopters are often and folks are familiar with about the execution, and typically execution for AI pilot science. This is one thing about agents agents that us do more. More pipelines, much faster, it's really harder, more prioritization, but has consequences. Has consequences all along the stack. We have pipelines themselves, the modules that are built from, the tools. We need to We need to rethink all of those layers about scheduling, the compute. Like, all those are gonna be able to run-in using agents effectively. We can rebuilding many of those layers as well be rebuilding many those modules and next door modules being launched liberally. So if you familiarity of core, modules allow you to compose a pipeline out of these individual building blocks. Today we're launching a new community module registry. So modules are now available for every NextLock pipeline, not just in F Core. If you could, you must go to all of the folks in your core. Installed. To this. And we can have this as a beginning, like, a a standpoint for all of the modules that can add into next door itself. So say you're able to be able to run or to install module inside of your your pipeline. But now the piece is very relevant for thinking about how agents can build pipelines. We wouldn't, like, run any module as a standalone tool. So we have one command. We've already containerized everything for that tool. We gotta say next row module run. We've A lot faster. This is looking at Sarek. You're in point 5,000,000 of you for the name of here we have to accelerate by applying GPUs. Same pipeline, same results. Actually, thanks to our partnership with Nvidia that is now available in both Army Seek. NVIDIA that Available We're super excited about. So preeminent They're starting by looking at processes in the brain, and their goal is to really apply this and to translate these into clinical outcomes. Thing that is So we now provide them with that pipeline infrastructure to process those data sets, give it a brief training. And they do the math. It's really great to see so many users who are using this flow for If you're going to use the GPU user scale though, it's always in your breakdown see what they what what they're doing. But today, I you GPU metrics as a first class ofism. It a So give a task that you a GPU, now records its type, the driver version, utilization, average, peak memory. And visible in your matrix graphs. So apply that per task across every workflow, and you can push your visibility into the GPU workloads that you're running. Okay. Just So it's a bit Everyone has a we have a moment, which is not using Internet. Six months ago, I opened my Slack. I saw a result from Phil was asking, well, what happens if we can combine Securus Agent with some very specific constraints and one demand? How and one demand, which is really make this pipeline faster? Well, that $10,000 in Anthropic API costs later. This is a result. Rewriting tools and pipelines from the ground up. Here, showing the QC sub workflow from RNC. Which is called Rust QC, able to do that in fifteen minutes. So it's really the same outputs, and it's a validated byte for byte. In most cases, this is it's gonna be a huge deal that we can apply to different pipelines as well. But in how many runs it per accounts to a year. It's a really meaningful number. This also opens up a a box of questions, though, and we published a manifesto join. This fits into the community process. I'd like you to read it. Phil is gonna be up here with Ron and talking about more of this in detail. I don't how do we apply to consideration for original authors, and one to maintenance of those projects over time. Do we authors Okay. Think about, like, cloud batch digitals, so Ing about ends of late twenty teens, you know, a real breakthrough. Allow us to get benefits of elastic compute. And if you look at Mixflow sort of uses of a computer, use who them. And if look at And many of those cloud schedulers, they designed to recopy the HPC model. So you define your queues. Group your jobs into the expected amount of resources that you that will use. It's basically a definition of what you're doing for compute. We have Console agents, and so ask, why is our schedule is gonna in the past in this in a static way? Decide to change this and to to build our own. So it takes to announce secure intelligent to a new generation of service which to this into to takes recover from spot preemptions, as well as self heal in situations like altimumiers. No new configuration here. We real talk about technology we've been building for for a couple of years. All experience we're seeing with hundreds of customers with have who have been using the best services and wants to optimize them. Three key things. First is around optimize spot. So you'll start with spot instances, recover from failures, and most importantly, look at the the price of the spot market and intelligently fall back. On demand is needed. Means that your Next up is is rightsizing. Here we wanna really predict what each task will use using its own information, so it's a mobile behind them. Instead of using default requests, their resources, we're able to right size the task request, that means that each task isn't over provisioning the resources that they requests. And finally, packing, is being a habitual problem with some of the best literatures is that you're up with these huge machines They've just got a small task on them that's run for a long time. Here's being able to really compact the pack the same types of tasks more densely on the fewer machines which seems to be quite efficient runs. Okay. I can suppose this is overall. Machines, Using spot infusion, we get down to four. And now with the judging compute, can hit $2 a sample. This is still your compute. This still your VPC. This is just not a schedule for that. Give you more durations, and we will drive more science. Intelligent people into preview today on iOS. Other clouds coming very soon. It really enables us to do more science modulations. As we improve our execution, the scientific loop can run faster. It's not just saving budget. Really buying more discovery. Okay. So we talked about context and talked about execution. I must stay back a little bit. A little away from the tools, the infrastructure. Think what we need now to do science in this this new scientific era. Kinda like work down problems that uses our constantly looking to solve. It's a huge variety. We often have them solve scientific problems, scientific reasoning. These are our questions or hypotheses. Interpretation of data. We are dealing with the physical issues. We wanna know the math, the amount of We're often dealing with algorithms, code, the tools, and the pipelines associated with those. Different breeds of scientists, We believe AI is really gonna change about this. So just as a line between developer, someone in products, a designer, they really have to get blurred. This is gonna happen with the same fundamentals. There's a 50 software. Gonna happen to computational science. The and the challenge here was to design a new protein binder for glycoprotein g. So NIP is a pathogen the to glycoprotein So maximum binding length that we could use the amino the similarity of those of new acids, there's constraints around that. I think the real problem was that we found another challenge quite late, like, a couple of weeks before it was due. And then the integration's slowing. So start to finish a few weekends with from from And then our 600 teams start finish, based on Insulinical results there. Sixteen days, not months. And and so the scoring, Floyd is gonna give out more information on exactly how we got this process Things examples like this are really exciting. Us, both what we see in next four community and Which I think what would happen if we could apply this if we could apply this to every problem? If we could give every scientist these tools, every biologist who's got a question or a hypothesis, many of them even don't have the time to learn next flow, but they have some intuition and some knowledge that's locked away. So it's really gonna be our guidance goal, and today, we're really pleased to ship a solution to them. So introduce things to characters scientists. And this is your context with that. Optimize execution as we plug it up between AI agent that understands the science. This This is gonna come Every step is your the learning process has has things evolved. And it's a key one is we're not meeting scientists where they are, where they actually work in their environment. Today, all comes together with coscientists. And we guys see it's working in three ways. So three main modes you might think of them. It does assist? Which is e scientist wants to launch a pipeline, explore the results, understand what came back, really about learning a new tool. There's building before the write, debug, optimize analysis, maybe work in CLI, IDE, They want the full space context. And finally, a one. Automation. Agents that can trigger pipelines to take failures against self fix and improve code twenty four seven the background. So turning to numbers here at if we're bench in statement conversion. Are These are many ways to just tell part of the story, though. It's hard to quantify some of the things like being able to work inside your environment. Being able to be secure in an environment with the governance, and also some things around efficiency. So considering cost, being Being able to look at, for example, the the number of tokens that you're using per tool or per eval in this case, we can really see, like, significant savings here because it's has the right context. Okay. These presentations are anything Okay. All right. So I'm excited to provide a first look here at ScareTire Sciences. This is something we've been working hard with the team on for the past couple months, and this is just a of view for bioprotections to create a self serve launchpad for scientists in a group. And so from this, can just have to go in and start to create projects for your scientists to run their workflows. And based on just tags in scale platform. And so you can see we're looking at the entire workspace here. But if we go into the project RNAC, for example, you can start to see that we we've called down the runs. We now selected for those. We have some datasets, and you can see all the reports of those. And so scientists can go in, and they can see the run details and pop active platform. Can relaunch a pipeline. They can look at reports that are available here and go download them and call the results themselves. We can also launch pipelines from this and configure these with selective CMS as well. So all the F4 options, can't expose those. Can just give them an input, for example. And then the other thing I love is you also go in and do explain your failure. And that will take off a a Scientist run, and it'll start debugging workload for them. Them. And so this is great first task for them to take because sometimes, know, it's not actually it's either error. It's not actually pipeline failure, it's not really something they need to ask you for. For, and they need be blocked by you waiting for for the results. And so I've got pre one here. And so you can just get sound to the end at the end of it, and comes back to what what the actual issue is. And then users can just ask relaunch the pipeline. And then that was relaunch for them. And so this is great. It kind enables you to be a little more hand off with your scientist. But all this as well is just kind of the actual automation of these things, I mean, faster and faster with the scientific loop. And so we've taken some feedback and on some new actions that ends as a preview of it. Coming up soon for us. And so this allows you to have different action events. Like a bucket event, for example. So one way you might use this instead of using a different automation tool, you might go in and do, we're not a sample sheet from a sequence or something, and then it might kick off pipeline. You know? Like, And then you might be able to in, go back up, and also have a pipeline status downstream from that. And from that, you might say that if this pipeline succeeds, you find a multiply pipeline that succeeds, then you might run right in RAC from that. And take off another pipeline. So you can start to chain your together, and this is all gonna be available to the API as well so your agents can set it up for you. And so from that as well, if you're like, man, you pick up your runs before you go there, then they they might, like, film, or then I be able to wake up to your results. From You'll be able to trigger an agent and say, take first looking results. And be able to kick that off as well. And so the last piece that we have also done is the actual Scare AI. Say, live, you're not aware. And so this is a two application, and this just allows you to actually talk to ScareAI. It also has headless notes so can start to automate that right in your scripts. And just inside the deterministic points of your automation workloading yeah, So that's it. Okay. Awesome. Okay. How we should be thinking about thanks So, like, say, top of the data, you are giving other agents as well in your stack the tools, the rules, the context for working there. Platform parameters like pipelines or datasets or studios that can be actioned on. And many of those rules those are kind of the built in aspects of the platform itself. So it provides submissions for your agents. Provides the lineage of exactly what they did, shows you the old model so you can go back and trace any piece. And I think that's kind really key as we think about bringing agents in to, particularly, labor or to more of the, restricted environments that we were working. The context is always always key as well. There's past runs, samples in your buckets. The results in reports, bringing all that together is really important. Is your data and your environment. It's model agnostic. So can bring your own key if you are have a particular preference for CLO or GMI, etcetera, you can bring your models there. Okay. Here, I'm Okay. So what's We've got faster execution. About the scientific reason we've But what's it for all this year? What can we achieve? I think you're not just bit of science, a whole new approach to science. We'll take a on something last couple months that combines all of what we've shown above. We think it's something we can really apply as a new way of doing science to to any analysis, and probably to many of the to pipelines and workflows that that you work on today. Maybe instead of running a once, we can run it a thousand times. Maybe we can run it across all the premise space. Maybe we can try alternative tools. Can rewrite those tools. And make that to the defaults now, they no longer need to be the method. We can use those defaults as a, like, a starting point. A starting point on analysis from which to explore. Exploration here is the key because the number of experiments that we have is infinite so many datasets, so many tools, so many parameters, so many ways to do processing of the data. And most of the time, we are just going with what works. Those defaults, those tools. Throughout all science, we kind of take human but prior as I face value. Not because they're optimal, because they're very expensive to explore all the space. And I think this is what allows us now using the fundamentals as well as the new tools that we've got before to kinda do something very new. So today, we're showing some of the good results of this new approach, and this is a closed loop optimization layer was built on top of Zigera. Think about the system we can combine some in our optimization, agent execution, scientific reasoning, with this algorithm that changes with rewrites. And with this approach, we can intelligently hill climb across the entire experiment space. The data, the processing, train models, using models, validation, the whole loop can be closed. Be run autonomously. For sure, ones are breakfast. Here, this is applied to a pretty much Here, is to the quantification. You can see that each dot there is really an inner loop. A tying run combined with the rationale. And you can see that at each point, we're able to get a record of what's changed. Understand, like, what the rationale was for changing particular premise in this case or particular tools. Also, another example here we can apply this here to model training. This is a prediction pipeline. Here, we're not changing this as just a parameter. We're able to change the model itself so we can change the model architecture. And then And then we'll apply that in training and see exactly where the results come. So a very good base so this we're really excited about where where this can go. Where results So through what we've walked through. We've gone through the context there, exploring execution, intelligent compute, coscientists, agents, etcetera. As as AI really accelerates, we wanna think exactly how we can apply this. I can't leave you this message of, like, use these tools, rediscover, communicate the science that you're doing, and really, we're actually asking you, like, what will you just come yourself? We're gonna have a panel on agent to biometrics. So we're looking forward to that as well, I will be chairing. So, okay, thank you very much for time. Really appreciate it, and I'll you on the screen soon. All right. Thank you, man. That was a lot of new stuff up That was really a lot of fascinating new content so much that it couldn't fit in forty five minutes. So I'm going to we're going to skip the q and a in If you were to talk Evan or any of people who mentioned, there'll be a coffee break She's going to share her experience with building a pipeline factory with AI. So without further ado, my friend, Laura is gonna be my mind. This is good. Thank you. This is the first time in Although I don't know if I should skip the talk now because I understand this is here. So should I give the talk or not? Okay. Describe the talk. Ya son de por This experience, I'm to call it, eight months ago in the intersection of decided that it was a very good place to go on how to build the platform that I needed. I'm going to start warming up with a few news that I shared. Really not the case, but I'm speaker, so I would share anyway. Jump into an nostalgic journey because I think that we should remember where we come from. To actually do the contrast. Things have never changed, and maybe they will remain like that forever. So you understand why I'm here and how I try to work in the new AI. Thing if you just want to play the game and you know everything I want to say, I use AI for many players' lights, and of course, game and, want to what I'm going to present. This is not about super robust pipelines that run everywhere. No. This is about fixing a problem in a specific environment fast enough for discovery to keep in faster. So no end of complying at all, but it works. Reproducible. Without data in our environment. So let's go with the news to share. This may be my last talk in the Postdoc Summit. Maybe for a while, maybe everyone who's like is After twelve years in Boston growing with all this community, really, I'm coming back to Barcelona. Share that if people are based on I would like to reconnect. And as well because I'm living, looking for somebody who continue this work in this lab. So if you really love to work in a hospital environment, will you impact patients directly, and you like the development, please come talk to me because we are okay for somebody. And move to nostalgic journey. So I have been in my pharmacy for almost twenty years now. That sounds crazy. So long enough that actually the complex between when I started to now is almost absurd at this point. So I would say in '28, I was fully into my informatics. Maybe for some of you that don't recognize this. But it was a time that you spent time knowing and looking for what workstation you should buy to keep in your desk next to you. And what to extend the workstation with memory RAM and all that. And in my monitor, my screen looks exactly like that. Where I had several terminals with the console, the code. Maybe you were using image and beans already. Remember still this plot from her. I got one, two, three that you have to be open. But now 2026, and it's really different. And this is all like three to five to seven Versus Code windows with code running behind the scenes. I jump minutes into each other to see what the agent is telling me how things are. I'm thinking about what agents do you what skills to use, what memory, how to prompting. Is XML promptings a real thing that works better than other prompting? And is really crazy. And I really cannot describe with words what it means to be this year in bioinformatics. I'm trying to transmit the feeling with two images. So this is the moon, right? Silver, beautiful moon. With the years, we have seen, like, better resolution of the moon. But then 2026 came, attempt is to went to the moon with the tribulation, played back. And we got this. Like, in the moon. He's like, what is this thing? And it's amazing. I haven't thought about that. And it's a transplantation. They are minerals. At the same time, it look like that really. This is like a very enhanced image. If you But nevertheless, it's something very new and fascinating. And I think that is the best metaphor I can find for these models where you really adding a new world but they make it look like brighter than it is. But it's pretty, I like the images. But all this change, they can never change. And this is why I'm here. So I joined this group. This is a group that data and infrastructure look like that. They stand CTR, like crazy, generating data location. All type of technology. Several postdoc instructors. And there was a time that they find where things were anymore. So I said, okay, I can do that. Can platform, so let's go there. And then, of course, it was much bigger than I thought. But I don't know if, you know, models at some point will do science by itself, but I know this will change ever. Maybe will promote human. From the beginning. Maybe they're really different. So what happened? That as well is the same. Is that it's expected that one person did every role. Like, a platform has a lot of different parts. Right? So you have sysadmin, developer, hats, engineer hats, analyst hats. And now it's even like what you call this mother can be your hands. So yes, the I can do it. And the particular problem I found is that did Lab really have 100 of samples, thousand of samples. They, like, make that up, mix up. And they don't do things, they came back and say, oh, we do this this tool because there's this hypothesis that could give us this target or this new science. Building new pipelines for these people to make discovery. Of course, we are now with this new technology. And I said, okay, let's try to accelerate that and remove that. And I submitted for this sent me the talk, and I said, well, I cannot come here and just just talk about my journey because I'd like to do a little bit more of a critical thinking here and do scientific experiment. So I really compare all the job I put it's a goal, by the way, the words things that were not available when I started to what I have now. I'm sure that yes. Everything because So I created this experiment that is, let's build a pipeline same prompt, in three different scenarios. And that was like in March with the new security client. So But It's not some tools or any of these things that is already solved. And it has documentation, but not super great. So I So I said, I want to try with Claude. Come on. I want to try with Grow once I get skills, and I'm going to put my agents on behind and see what's happening in three cases. Seven minute planning. I think that it helps. Four minutes behind, and I got into our whiskey in the loop. But I decide to stop because I use all my credit. But that's fine. What happened I What happened with Secret AI? Five minutes planning. Four minutes building, quite a lot lot less time. And, actually, it worked. I had the same But log around. I think 2018. It fixed everything. It goes running. And Google Cloud is running is just that with the configuration, it wouldn't work. Like, we need to configure the the input output for big files. And then I tried That is what I decided to call my infrastructure. If you get the reference, probably you are too. Young for that. But it's mainly needs a sentence saying, build this next flow. Pipeline. Planning because at the whole point, After fifty minutes, I saw this. Know how to do the things. I'll get back. I don't I know you're thinking right now, but I Making things up. I I didn't make anything up, but I went to the history of the chat. What happened is that I had this agent checking for Doctors available in my registry. And because I give the things to experiment. Docker was there. So the Docker agent just went up to the next two steps. And that was And that was what happened, just like context. So basically, is a script that has a template that works really well with our environment. Skills, Docker resolve skills to check for Docker, and four agents to make everything happen. It's all context. All my knowledge there for things that I know that would work in our case. So no magic, just information. Few technical tips, if you want to know. I discovered that the skills need to be put in the sub agent. If you don't want to the sub agent, just stop after the using a skill. And this is something, accommodation for Claude, but it took me a bit. It was like a It works. Would recommend it. More general takeaways is that really Cloud Secret AI streamlined at all the next role calling. Containers are Containers Rules are context rules. All that is what really matters to make your agents efficient. And in the case, autonomy versus efficiency was a goal. It was not about I want to do anything. It's like any department as fast as possible. So for other people, this may be a different case. But I got it to really build my plan in another, like, fifty to twenty minutes. Minutes. I'm get it running and give them a result in the next day. But it was a really a good thing to do. And now this timer comes down One minute. Okay. Have a minute. Want to I'm gonna let you go with one last thing since it's my last time something was done. For the people who maybe feel like I feel sometimes where work that you're doing sometimes, this one, maybe other, it just feels like maybe it's been impacting your career path, and what is your next. Move. And that triggers a feeling that is uncomfortable, right, because everybody wants to really move forward in their career. Wish I what I tell to myself. To make you feel that what you do is worth it. And it's that what you do matters because it matters to you. I really believe that this principle will make everybody to stay in the right path for everybody, a path that is unique to each of us. In great uniqueness, thank you for listening. Thank you so much, Lorena. That was very insightful. Because we are a little bit you did fantastic to be clear. I always invite you to talk to Lorena at the break. She's a wonderful person. So but we are now. Our next speaker, he's cofounder of He's prolific creator of Thank you very much for the introduction. I thank you all of you for joining both in person and online. It's a pleasure to be back on stage. It's been a few years since I've up here at the It's been a really fun experience just getting here this week and meeting so many funny faces, and just seeing what the excitement that everyone has for for building with these stores and seeing a This talk, like many by this, is kind of a of an unexpected side quest that took over my life for past few months. And trying to work out what we be doing with it, what the best use it is, and what new opportunities are there. And and this whole tour comes out just a little, like, silly idea, which which balloons complete out proportion. But they they tend into a really, really interesting film experiment. And and, fundamentally, I think the take home from me after all of this is is that software written by AI here. Right? We talked about last year or so about helping us to write code, helping us with analysis. We're trying to 36 meetings to be there And I think this is a really interesting time in the history for us right now in science, where we have this kind of transition period, which we're really right back in the middle of now. Where when we look back on this in the future, we'll see kind of all the the existing future. We'll existing new new pages are designed to be built today using AI. We have this kind of glutted existing pipeline existing tools where the huge opportunity now to use these new tools so you kind of upgrade the talk with things we have already. And kind of take that low hanging fruit and and material differences to analysis that we're running today. And that's exactly what rust you see as. So hopefully, some people have heard this already because we we can put this out on socials. A few to go and match with a couple of blog posts and I'll right to sponsor. And we'll start by taking you through how I came to this project and then I guess And what's still left to me for years is this little annoying It's not the alignment. Which is taking a long time. It's not the education not education towards analysis the screen part here is a QC sub webflow of ROCE going up in not really doing actual work. This is generous because there's some QC tools in here that we would love to run that just take too long. If we look at our CQC as a transport security number process, can just that task alone takes between, like, six to nine hours. And If look our So, you know, we'll be we're gonna So if some are the of They're And so, actually, I had rewrite tools without asking for it, which made my jaw job. And so actually, initially, she called it at all. Duke Duke Rust was her first name. My job. And so you know, generating metrics and plot sake. And by the end of the Friday so it's Friday lunchtime, get off. I think by about two in morning on the weekend Friday, I had a two page accountant, all of these are CQC lots. This one I effectively brought 16 different biphoratic commands into one, some tools across that on previous one because I need to name addressing the prerequisites. And this had a huge impact primarily on disk access. So now the RTQQT let's say, RTQT step up flow reads every single file about 18 times. Just for QC. There's two and a half terabytes of on, like, a band file that's at 10 or 11 gigabytes. Where's RescueSee to do it in under point one terabyte? And so that alone makes Rust QC vastly faster, like, hugely faster. But other observations are not being Rust itself is fast and other other observations we can make, but but there's the architectural efficiency of doing everything at one on CPU hours and as a result, like, energy usage and everything. And we just kind of agree that this is too good to be on the table. And so Rusty was born. Had to design a logo as a first party. And and we kind of started to really take So we're going into all. I'm gonna propose something pretty radical to this massive users that I'm gonna destroy half of the pipeline they're running in production on all my samples and replace it with this at all in Rust, and I don't really know And, like, is kind of opportunity is clearly huge, but, obviously, the risk here is pretty huge as well. How do mitigate that? How do we inspire trust? And I think myself, I'm happy to actually release this since I And, of course, the key to that is validation. Real scientists And the problem is that elements are built to cleaves like they are. Agents telling me that they found bugs in Upstream software. Like, I wonder if they drafted me an issue to submit a pre seed repo drafted pre How do we really know when we think about it or when when the model is wrong or when it's right? And this is what gave him this idea of kind of like rewrites. Right? Not building new tools per se, but aiming to emulate existing tools. Because if we do exactly rewrites, we have something very specific that we're aiming for. Something's very tangible, we can tell the model it has to do. Exactly the same as also it already produced by this other dog. It's been around ages, and it already has the trust that we want. Exactly the same as It's very prone to excessive extrapolation. It would go from, a 100 reads to 200 reads and be like, I've got a 10 x speed up And I assume that this would kinda scale to a 100,000,000 reads, which is just very wrong. So both of these cases are actually using large testator files frequently, was really important to ground the the process of the It was very clear in the long the industry. Mean, I said, like, said, like, of course, we tend to write And I'm like, nope. I didn't ask an eye. And it's like, so So even with some kind of contacting, like, such favorite, little phrase there, like, you know, this this work. What does get you out all the is thinking about having a deterministic verifiable aim. What I did is I used Nexo and and NFTs for this. Instead of saying, I set up a test line is using the pipeline for the upstream tools. And rewrite. And I said, these are these are the files. Are the Do what you write. Run it again and keep running until these NFS passes. And that's there's no way around that. Zero pass referral. So we're trying to snapshot testing of upstreams will help us get a real verifiable target agents. And we, as Kevin hinted about, very, good next row is the perfect harness the perfect framework and tooling for this kind of autonomous long range work with AI. So this went quite quickly, and we kinda realized that if I was able to do this so quickly, then pretty sure that everyone's doing this. And we can we'll start to see maybe people come back with some of projects pretty that everyone this. We can actually implementing it and a little bit about kind of acknowledging between tools and and how making sure that these are one off tools which are advantageous to community. And then we push this up, it was kind deliberately slightly productive because I felt like everyone was kind of in my little while thinking about this, I wanted to frequently, people questioning Quite with that of building up your own of the And you a new but start I mentioned other people doing some of the things that but I've had so many people reach out to me with some of the projects that they've either been working on or started as a result of of seeing seeing this kind of discussion. I've listed a few of them here. Rob Atreus, things his And it's also interesting with the rewrite on new architecture is able to solve and and what you've to do for years and not been up to. But I think this is a is we win that tool, and it just gets upgraded in place. Usually, because we're accelerating implementation we've but took a bait. And I've guided years of job experience, and managed to the agent I think there'll be a gold rush of people doing this, rewriting old tools in the pipelines, and it's gonna not last forever. But I think when you're gonna see also, we're gonna see this happen a lot. I am gonna go out to bit of a limb and say, think we're gonna see more drops of rust. It's got bit of a kind of fan base already. We'll finish on Helen Lee, one of the most famous by the mentions I've wrote a blog post on this topic just a few days ago. All So Thank you very much. Thank you very much. Think you've given all of us a lot to think about. We are going to break a coffee break. Please do talk to Phil at your leisure. We will be back with Palace Talk at 11:30. Please go ahead and have some caffeine, some oxygen, and come back in a little bit. Who is our cofounder, who is our chief architect, and the original brain before this What's new. Thank you, Jarvine. Good morning, everyone. And thank you for joining and holding the trust in Toki's technology. Today, I'm going to talk to what's new into NextPro 20 six-zero-four that we just released this week. And before we get what's new, big rounding what we have done so far and some key might still into the project, please. Computational framework to allow us to write pipeline a recklessable manner that we could deploy in different infrastructure. Maybe developing logging in there, running to HPC, then scaling the cloud, depending where, how we need to run this pipeline. And we could materialize this this idea that we could have a pipeline running different environment scale production rate on cloud like AWS and UAWS batch. This was an important point. Also bring more people, more user, and much of a beauty for the framework. That was in '19 with the evolution of the syntax at DSL two is super for models, we added a more scalable scripting language that allow us to write more complexable flow to support for models that's only the language more And also, course, zero four, and we continue to focus on these two main topics, the language and the engine or the language that allow us to run a scaled pipeline in different environment. Let's see. Let's Since this version now, Nextflow view the Syntax Parser version two, that essentially is a real implementation, the next flow language, in its own grammar and so in syntax parser. Are not depending anymore in the syntax. So this give us more better control than we can do in the language. We also have the ability to have some let's say a cleanup to remove some ambiguity in the language, made more current, still keeping a strong so far. That can be some little changes that may be adapted when moving to this next version. In case, if for any reason, your code cannot Or pipeline script, and we are much better ever reporting that had to you find out syntax error on some And then type system. Type system is a key component and question. We are pushing to have more strong static type system into next year that allow us to have more control the language of what we can do with the pipeline script and our code. One first part, we are introducing more support for types is the parameter of the pipeline script. So there's a new component, params blocking, allow what we can use to define a that are used by the pipeline, in which we can find for each of them the name and the file that is expected to file part, that's the train, William value, we can also define option up in default default default value for each of this parameter and the benefit is that we can validate when we run the code with the common line parameter that you can And, again, moving more data types and study top into the language, we are also having the support for static type to channel group for definition and even more the ability to to provide custom regular type into your Along with this, we also extended the support for we also extended for types that allow us to define a more structural manner each component and the type for each component. We have more control on the input in double section and then better validation around time and completion time. On the data that is receiving, the process and the implication of the process itself. Along with this, Table still exists. It is a classic way to define input in a true, but is extended to support the aesthetic type into the language. And then I'll jump a bit more models. As mentioned before, we're extending the support with models in next row, And there was important component, but there was not other not not sophisticated management for models in MaxRu. Just a separate way to run and incorporate existing script as modeling to next to a script. Is that now what you're doing to bring a real model to CD312 central service in which we have access to all models that came from the community and also have better control before entering this. Essentially, before what we were doing when using model in just including the model, referencing the model script in your computer. But this was a bit complicated because you need to maybe create a friction on meeting experience. We don't need to specify anymore the part of the model. We can just specify the name of the model. So when next is to be beyond the sea, try to look for for this model in the canonical part part, including the log of a system where the pipeline is, find it, just run. Otherwise, pull from the the registry and this is what I was mentioning before. under control and manage all model, the model life cycle, integrate into next row. We can search model, view the model middle data, run the model directly from next row. So there is a bunch of stuff that can be possible models now, and it's not that I can show you where in practice does the working system, this is the next law registry that was evolved from the previous plug in So now here we have ordered the plug ins as we had before. Now we have all models that we import from the network community automatically both so that are contributed by other contributors like this So join are already new new contribution to this model model registry. Here I can see I can navigate with the model. Can search them. I can find I see more details about the model, maybe in Zooming Elite. All the input and output are defined. Two that are used, and we keep track all the version that are important automatically from NFL collection. So I can even see the differences that are with the previous version. So we have much more controlled public model for our pipeline. And I think it's even more interesting how I can use these from the CommonLine. So now we have this new next level model command. It helps me to control all models, how can use model trade, store, round list, etcetera. One interesting one search allowed me to search for models in the registry You know, we are not using an example search, but a semantic search. So you can even say, I want something related to genomic assembly, and then find all matching models. Much on them, we show the first 20. Let's say I want to do this one. I can use model next to model view. And this show, essentially, same in data that will show you before in the UI, from the command line. And we have the definition, the tool that is used, the input structure. And what is nice is now that's me how I can run this model with the next row command line. I think it's very useful because, essentially, I can run any model that is in the network or the next row registry For example, Meta, I can just say example, and then fast view, testing. That is enough to run this model. So this makes it super, super simple prototyping how we're pipelining because I can use this system to find out the model that may need into my pipeline, that's with my real data or some sample data. Iterating. For the guy, made it much more simpler. Before, was then I was saying, oh, you're probably our container. Yeah. Here, I'm just using the container, I'm using the container, not building the container. I'm just using the wave containerization service that using the tool definition that there is in the model, building the container on the fly. Point that the Docker container are not the right abstraction for running a packaging software software. The right abstraction is the model because you give you all the data that is needed to run the TESO from the inputout the definition, the tools that are used, and also we have the tooling to shape and build on demand And what is cool is approach if you think you were saying, yeah, said, what scientist, using common line boots. Think it is one other big reason we wanted to bring the model command into next row because now essentially you can tell, you're a coding agent, I need these tools to write this pipeline, access all these metadata, understand how to use a composed data model to run, against your data. So make this this flow in Halo much simpler and so much more effective. And then what? Once again, I can even want to create my own models. We adjust a command and create a schedule for the model, and you need to specify your name, the name is composed by two For this example, name space and then just alone. And now now we are in this part, the model was created. Submit. Hello. I could inspect the main to see the basic implementation, but then I could iterate. I can change the input structure. I can modify what the script is doing. And then I can't gain. I can just use the command line to run it on the fly. When I that I register yourself in the next row registry, and then you can publish with the published command. Cannot next Here we go. Okay. Simple but effective. Just works. And that this is your model. This is accessible by the community and your pipeline suite. Okay. So I think we give you a good idea of how this model mechanism works. Let's jump back. Into the presentation of the second part second segment about the engine, how we run, how we deploy presentation the before that the use of batch allow Nextflow to scale the cloud was an addition to for the framework to scale, to ramp pipeline, all these services try to mimic the HPC model in which we were using pipeline. So this amplified the transition to the cloud. At the same time, we some kind of mismatch between two to these different work. What is the mismatch? The HPC model is essentially a And what on CPU architecture. And when you're running five running this environment, you're optimizing for a well defined data center, a well defined computer architecture. So it's simpler because it's much more predictable. How you are around the pipeliner. Moving to computation in the cloud is a completely different game because we need to we have the challenge to have much more dynamic environment in which all the computer resources are provisioned on demand. There's more network latency in all Think only about the adaptability of the different instance target there are in Amazon Cloud. Think only in the Amazon Global Cloud. Each of them with a different requirement for CPU, for memory, for disk, IO, So optimizing this pipeline at this scale is totally impossible from from a human point of view. This is why we think we believe that we need to re retain the computational layers for our data pipelines so that we can manage this now working for the determinants. the execution of this era intelligent computer layer. And the first thing that we do, we collect the task in we classify them depending on the compute profile that we can observe, also depending on what past execution history. So for each class, we can't create a profile being able to also consider the input that we observed are being going to be submitted during the pipeline execution predict the similar amount CPU and memory duration for each task that you're in the pipeline. Credit find the most efficient, the best machine time they can be used to enable in the market in that moment to run your task in your pipeline. Back in then dispatching Tuning and remember what you learned and continue to apply the next execution of your pipeline. The past rounds, essentially. Main capability that Arna mentioned so far. So one nice thing that using this system, we don't need to maintain infrastructure anymore. If you have used AWS batch, know, you have to create custom NER. Need to create compute So in this case, all cluster, our provision on demand while pipeline is running. We can't even speed. Multiple cluster for the same pipeline because each cluster is optimized for a segment of the We also use the system use of beam packing algorithm, try to ship multiple tasks in the same DM to reduce the unallocated or capacity of them shipped So for example, some things that integer compute is doing, they detect condition like other memory, re execute the task, example, using AWS batch. Are, let's say, best configuration probably subject to use the customer and then integer compute. And, instead, using Cajun Compute, we have two hundred and thirty CPU hours and nearly five hours. And both in compute time, execution time around 20%. The impact of the cost is even bigger. So the same pipeline always running AWS batch and better compute. AWS batch required $34 instead using Pager Compute less than 15. So we cut the cost more than half. And from where coming from? The setting. Well, actually, the setting coming from because we use less compute, but also we use more efficient compute. Because AWS batch, essentially, you are paying 0.1¢ per hour, CPU hour, And in Essentially finding the best available machine time for each task for your workflows. And okay, in conclusion, these technologies that we're going right now to also get our customer as a preview technology, and they are able for Amazon Cloud, and we are going to have this approach from cloud in the coming weeks and months. This preview program that we had here, just ask 20 of us about this technology. And, okay, in conclusion, I present too many times about how we are evolving the NextEra language and the NextEra modernization system with a new central All right. That was a really interesting presentation. A of new things. A lot of new things for us to document and get into the trains too for next level. So that's questions, do we have any questions for any questions in the room? We got time for one question. If if not in the room, I do have a question from people online, which I think will be, of interest to everyone. Is there any support for private module repositories? Would consult them. Support for Our next speaker this morning, who is Suriya Gose, who is a interdisciplinary scientist and bioinformatician at Imperial College London, and is going to talk about using Nexlo for more proactive vaccinology. Thank you. Sorry from a bit jet lagged from London. So I'm going to talk about how Greg Nextflow has been in creating some pipelines. To help us get very quick data that we need for proactive responsiveness. So I'm just gonna go through the difference between proactive and reactive technology. I'm gonna go a little bit into the biology, see what we're actually doing and what the pipeline we're looking at. And then we'll talk about the leak that's happened with all these lab processes. So during COVID, all of these processes were done purely in the lab. Computational modeling was happening, but it was NGS based, and that leak has occurred now. And because of that, because whenever you hear NGS now next closed becoming someone's social heroic competition, it's a go to tool. Used in our previous roles quite bit in our NAIC. So we always are heavily relied next loads. I'm very happy to continue on with that in this role. And I'll introduce a new pipeline that I've submitted to NFCall Group who has approved this pipeline. I'll talk bit about that in some future work. So the team behind the group at Imperial is led by Christopher Paul Kelly and Wendy Barclay. They've secured an MRC grant to basically be be a little bit quicker in case another pandemic happens. We're also really monitoring code variants with collaborators at the University of Cambridge, Francis Crick and Herbert, and some other teams. And this is a a very big group that's basically encompasses lab people and biopetitions. I'm the lead biopetition in my group. We've got clinical fellows like Ruth Ram and Katie is a lab assistant that's working heavily in the lab. And we have quite a big group. We have monthly meetings, we're basically everyone's tracking what's happening to COVID at moment and which variants are coming out or not. And the grant covers there's gonna be study that basically goes through which variants are gonna make vaccine for or not. So proactive versus reactive. Proactive is becoming a lot more in vogue at the moment. We've usually been reactive as you would have seen during COVID. Pandemic, a lot of us reactive, and it was pretty specific to regions and areas, and we became quite aware of how unprepared we were. Nevertheless, certain breakthroughs occurred at the mRNA vaccines, and within eleven months with the vaccine, whereas before years of clinical trials, And we basically don't repeat that. We've learned a lot from the last pandemic, and proactive vaccineology deals with what we've learned and basically mapping variants before they could and I'll talk about that next. So proactive technology is basically using animal models to create the mutations in vivo before they happen and testing them to see whether there's an antigenic fitness that occurs or not, and watching those particular variants closely in the world. So it's it's off road protection. We create vaccines protecting against multiple viruses and variants instead of one particular variant. Because that you know, viruses proliferate quite quickly as we're going with influenza. And the platform technologies are changing. So, yeah, there's a lot of preclinical work that's going on before pandemics in Oj. So previously, just a little bit about biology. Go too far into it. We've we've been doing these things called report essays in the lab. And what it basically involves is putting a reporter into the cell. If the cell gets infected, get this cell fluorescence, chromycithrase fluorescence. If it doesn't if the virus doesn't infect the cell, you don't see that. And basically, that's what happens. In the lab. And it gives you one curve. You can see on the far right side, it's one curve of one virus reaction. Say, okay. 50%, which is called the NICE 50. Sierra or whatnot, neutralize this virus at that level. And it's one virus that takes a lot. Time, each well measures a single pseudovirus construct. Sensor using like viruses to use a pseudovirus. And it depends on all sorts of different things, it takes quite a bit of time. You get a proper good signal and that's why it's been used for so long. What's happening now is being a paradigm shift, with Jesse Williams Lab in Utread Hutch, and he's thought about that bit of just that one curve, he thought, can we do many curves at once? So I'm gonna see many viruses react to same serum at the same time. That's where the NGS heart comes in. So basically, incubate full of Alcoated viruses within a within a serum. So you get many bivariants in one one well, you do hundreds of across the plate. You varco them with a particular varcoid variant. And then you put them a situation where they either compete or they express or do not express. So no more visitor arrays. For instance. And you basically count the barcodes at the end. You normalize, you get your IC50s. And those response curves from that. That's been the shift, and that's all based on NGS data now. And another thing they do is the mutation scanning, we'll be doing COVID. So you basically force limitations to occur. Before it does in a while and see how they react, whether they're very good at cell entry, practical variant, reciprocating. Or SCAPE. Those are the main three phenotypes we looked at. And then that's all sequenced and counted to see how many which really well. And then we can then develop vaccines. Based on the data. So because of this, it's multiplexed. So that comes in in multiplexed form. You get different sequence mutations. Can get different mapping tools and things like that. It's really good because you're gonna need small bounce of zero. Within a few months and recommend vaccines from there. Okay. So this is something that's preaching to acquire. We had to kinda justify why we're using certain tools. Last QC had explain why why it was useful and how accurate it is. So we're a very biology headed group. So I had to really come into why the rates are gonna be basically dependable. And trustworthy, as I showed them this. But we do have to also tell tell them the other side of the coin, which is a single to longer reading sometimes. And the the things are referred from counts. This this stuff happening in PCIs and library building. So it's not a 100%, but it's pretty close. So that has alleviated a lot of worries. Let's see. Sorry. We also use Nova six, so we're getting getting up to from 40 for it scores. Everything above our results above 40 is 13 excellent. As you all know in this room, doing these things one by one is a bit of pain, so I suggested to my team we use next flow, and it's made the whole process and make it reproducible. I had never heard of next flow serving. Trying to teach them. Well, it's fantastic in the five minutes world, and it's basically the go to thing for automation. Are other little pipelines out there, but they're not as stable or adaptable and dynamic as next flow. So with that, So, yeah, we we basically made the the argument that it's curated It's quality, standardized, reproducible, and it'll be available to the worldwide audience, and it'll be open source. And it'll be a collaborative effort with coders and by competitions worldwide. I went into the other positives of the pipeline. Could be based. So lot of biologists quite scared of Nextflow and other automation, and they don't know if they can run this on their PC. I explained that Docker allows you to get burden control on all these things. And the actual environment builds already for you and keep think it gets kept up to take. Can just pull it down. And basically run it. I'm trying to build it so they can just do a press play kind approach. We'll see how that goes. So this is the pipeline so far. We've called the product seek. And it's a it's a pretty standard pipeline as an F4 quote. Corgo's got fast Q stuffs like QC, fast QC, and all that. Occurs, and they're reporting that comes out of that. And I've and I've built some little modules in there, the barcode counter, which is like a full module And we've got the neutralization data. I've got the deep mutational scanning arm still being built. We've got third party tools that built for this kind of workflow called NuCove from the existing wind lab. He's got a DMS code as well that we can add into this pipeline and take out the maps of different variants through time. How antigenically distant they are. So it's been fantastic to have such a platform like NextWire to be able to use their party tools, build their own tools, put into this on a seamless automation pipeline. So this is some validation we're done with the pipeline. On the left, can see a a traditional port that you get from just pure lab work. And on the right now, NGSpace, getting closer. In a computational way, basically, sorry, to see the different code variants there on the right. It's almost close. We're still trying to get proper IC50s from it. We're getting there, this is purely from NGS based dot. Other kinds of output that we're expecting, other new plots like that, the antigenic maps that can come in three d. And a deep mutational scan data basically shows which binding mutations are relevant, so we would choose vaccine candidate. Based on this data, and we just needed a proper pipeline to produce these pretty quickly. This is our Jesse Bloom's data that we validated the And it'll be based the test data will based on his data that he's got already. And that's come through okay. See. So future work. So it's ready to process incoming house neutralization data sets. And it's also being validated on data sets, with flu and SARS go go to. And I just need to put it in a call for testing and publishing and basically get help from the guys there. And we're still developing and getting new metrics and new tools to be put into into the sorry, into the pipeline. So we haven't had any kind of hiccups at the moment, but we're open to your suggestions as well if if you see our tool or you build a tool that you wanna be you wanna showcase in that point, please speak to me. But I'll So, yeah, we've we've gone from a purely lab based response and proactive technology and reactive to an NGS based neutralization SAE format. It requires high resolution response out of needed. I thought it'd a fantastic way to use Nextflow in that kind of field just so we can scale it. If we get 100 if there's another pandemic, there's hundreds of samples coming in, it'll be fantastic. Just have that ready to go. And together, I think they'll make a really great response tool. But, yeah, thank you. Wonderful. Thank you. Believe me, early is not a problem. So thank you for that. Really interesting. A microbiologist, told myself that really speaks to me. I I was wondering, and this I was gonna say this is a this is exactly the kind of use case that I think I'm striking about something like NFC or something standardized. Are there any questions from the audience? We have a microphone. That you can speak into. Otherwise, I'd love to hear what what people are thinking. I'd love to hear what was your experience developing for NFCore. Had you developed anything in NFCore before? Or and and what was kind of the thought process of did you start developing it directly as an NFC or pipeline or did you start with regular mix flow and then convert that? I was I was used to the next slide pipeline because my previous role was STARC, and I realized it's MGS again. Gonna run fast two c even run trim below and do all q c, and I wanted to automate that. I immediately thought, I was added to that pipeline, again, I thought next load is gonna be much easier to attach onto the QC section modules. And I think I just built it from there. Think because you guys are so well known now in in five minutes world, it was a it was a very easy connect of automation equals Yeah. That makes sense. Thank you. One kind of follow-up. Because you had the secure platform up there. Is the plan to be able to make the pipeline available not just through the repository, but available through the platform so that people can use it in a kind of point in place kind of way. Yeah. I've just been learning about all that and some of other trainings as well. That would be fantastic if we could. Incorporate in any way. We can get community health as well. Wonderful. All right. Well, if there are no further questions, last call. Good ones, going twice. All right. And then thank you very much, Maria. That was great. Then I am going to call up our next speaker, Scott Norton is a senior bioinformatics scientist at Merck, and is going to talk to us about MG3 and doing more microbial genetics with Thank you. Right. Thank you, everyone. So, yeah, I'm Scott. Gonna be talking about entry tree. I want to give a little bit the biological background of the specific project problem that MGTree is attempting to solve. I'll introduce the pipeline and, show off how it works. No light on those. Sorry. So on a relatively small biomthematics team embedded in medical research IT. John, Lee, I are based in Cambridge, just across the river. Sam is based in West Point, Pennsylvania. Elaine is a rotation specialist based in Rawway. And at the time we were working on this project, we had a summer intern, Alfredo, who contributed a lot of color for the early cooling. So for this project, we were initially approached by a team in the vaccine development group They were working on a vaccine candidate for norovirus. If you ever had Norovirus, I'm so sorry. There is currently no vaccine on the There are a few candidates in other groups in various stages of clinical trials. But the best method for prevention is always washing your hands with soap. Regardless, there's, at the time we did this slide, we're 900 reporting address in August 2025. Probably closer to a thousand. Haven't been past that. One of the challenges with designing a vaccine for inverse is that it's highly variable. Virus in particular has about a 100 different genotypes, and this phylogenetic tree clusters them into eight distinct geno groups. Of which one of nurses in gene groups one, two, four are known infect humans. So this team had a candidate targeting some of these unit types working back here within genetic group g two. And what needed from us was a way to test a sample for what genotypes of neuritis present in that sample to make sure that the vaccine is actually suppressing injection of the target, genotypes. Organization came towards a very similar problem this time for HPV. Now we do have an HPV vaccine in Gardasil nine, which provides protection against nine different variants of HPV. But there's over 200 different variants of this virus. We're working to expand the coverage What did I just press? Okay. So we're working to expand the coverage of the of guards of the extinct differentiation variants. Again, we need to be able to monitor which variants are actually present in trial samples and in environmental samples. Now both of these problems roll down to the same question of, you what is the composition, for example? And this is, by and large, a metagenomics question. Now traditionally, metagenomics is you take a sample from the environment, you want to know what species are present in this sample. Usually, bacterial, you could find it feral as well. And in popular 12 for doing that is Kraken, Kraken two, Kraken unique, represented here by this champion. The way that Kraken works is is an exact tamer matching. It'll chop up your NGS sequence reads into short, KMer's second effects links. And back them against a polygenic tree of your plate of interest. And classify the composition of your sample based on that mapping. So when you throw tracking true at this viral gene type meeting, it will provide an input an output rather, In this case, it says that this example is g one seven, which is finely dandy except we have to know that this sample is a pure isolative genotype g one one nerve And this raises the major issues with the parametric approach is that it kind of breaks down what we have a lot of genetic similarity into your reference dataset. And we can throw at this problem as much as this one. As much as we want, but it is sort of a task. So what do we need from a viral genotype? Pipeline? Well, we needed to be correct. We needed to be able to handle multiple references per genotype, particularly with the neurovirus, the sequence database has multiple different isolates that were all assigned to the same genotype. Any one of those could be the interlinked infection, for example. We need to be able to handle that at not only the genotype level but also the isolate level. It needs to be able to handle different regions for genotyping as the VP1 protein in enterovirus. Or a antigen on HV. Should be able to assign the genotype at the correct tacitonoc level, whether it's gonna type of gene group. And it should be accurate, automated, and high throughput. So now I'll go to introducing our solution and your new champion. MGTree. So MGTree is an excellent pipeline. Of course, workflow starts from SQ. Does some initial QCs in the engine merging adaptive mode. And then it performs o type two mapping against a multi reference FAFSA. So the specified will contain reference differences for all of the genotypes that you think might be present in your sample. So both type two in the state is running the dash page flag. Which tells it to output all possible alignments instead of just a exact number of best alignments. And we chase that with a custom script to parse them. Which will perform a dynamic filtering to select only the best alignments from the band file. There's all alignments with an alignment score matching the number one best alignment. And there is a configurable wiggle room. We typically run this with a maximum difference of one. And then we go into the core of m p three. This does is it takes the inline reads, and maps it onto a solution that's encoded in new format. That's the NGTtree can process. And then all that goes into a customized multi q c report. This is just ASMRD of research usage. Most of these scripts are very low resource usage. That's largely due to the size of the task that we see for this kind of pipeline. Particularly with more of us are dealing with stool samples where most of the genetic material is not neurovirus. So just to show off the multi q c record, we, of course, have the standard FQC to assess sample quality. As well as a custom graph Well, we have the custom view of what the genotype predictions are, each color band represents one gene type that was detected in the sample input. Alright. So time for battle. So how does MGTree compare against practicing two? How does perform? On benchmark data and on real data? So going back to that, John, our first data set that tracking two predicted wrong. Again, we have a known sample of g... Of neuroverse g one dot one. And NT3 calculated that the sample was g one dot one. Hooray. Scaling up to multiple samples m p three, gets the correct genotype Most of the Gravity two hardly gets it correct at the genotype level. It does get most of the samples directly at the group level. But it also makes a lot of mistakes that m v tree just does not make. Additionally, Gradient two has much lower rate of classifying reads based on the NodeRx type of origin. Relative to MG trees. So that's all those white bars on the Gradient two side. Again, showing these samples in the MLDTC. When you open it up in the browser, get all the typical interactive features, like masking over the different bars of the agent type detections. In this case, this is an HPV gene type assay with known infections. And in all cases, m p three is able to correctly classify the samples 100% of the reads. Are classified in all of these cases. And, again, it's more often than correcting you. There was this one sample on the fourth one on this table. Where we had a known cone section, two different stereotypes of HBV. Present in the sample. And these were all validated using traditional desktop methods like QPCR and sync sync sync. And in this case, we didn't initially detect the coinfection using PCR and Zander. This was picked up by m g three, and we have good reason to believe that this call is correct. Of course, tracking two also call the code section. Not the primary section. So yeah, I am a loaded fast. Not a bad thing. So a few conclusions. Strategies I'm employed by tracking two are two and four cases where your references are highly diverse. And fall apart when the references are very closely related such as with myelogenic typing. MGTree overcomes those notations by taking advantage of the full context of the reasons of breaking it up into short gamers. And gives a much more robust overview of your sample composition. So with that, I'd like to again thank the team. Let's extend thanks to the vaccine development group who raised this problem in the first place. And a huge congratulations to the Victor. While oh, so sorry. Are there any questions in the room? Oh, yes. Phil over there. We can get him a mic. The hope Thanks for the talks. Great great application and great results. Yeah. So, you asked for the reasoning short read sequencing, whether we can get, better resolution from long read sequencing. The short answer is yes and are using short way sequencing. These viruses tend to have relatively small genomes. So a 150, 200 base pair paired in to read usually gives us enough to accurately type the virus in. So I don't see more than a marginal improvement in accuracy from switching to long reads. All right. Very good. Question in the back there. Okay. Very good talk. Thanks. We're doing something similar. So I'll come and find you if that's okay. Can we can we can we can talk about it. I was just wondering, in terms of the phylogenetic tree, is how diverse can that be in terms of the size of the tree? It very specific to the target? We're doing bacterial work, it's a bit different. Yeah. So you're aware of how diverse the genetic tree can be. We don't put any caveats on that aspect of it. It'd be as diverse were closely related as the nature suggests. In the case of Norviz, lot of the genotypes are genetically similar to each other. And if you're using muscle three multiple signals alignment, it's able to pick up those small differences and place them on the tree in such a way that entry tree is able to then directly classify. Okay. Cool. Can I just also ask? How does it handle the absence of strain? So if it's not in the file generic history. So how well does it handle the absence of a strain? That is a good question. So for the most part, if you don't have a string in the database, it's probably not going to map your reads. That string may map them to some other friend that shares genetic sequence, or it could fail to map read at all, and that could go into the unclassified bucket. Great. Thank you. All right. Any more questions for Scott? Oh, yeah. Over there in the back. I think following up some moisture into what Tom was just asking about how would it handle the combination? Would it be able to kind of treat different segments of the genome, distinctly enterprises, for example, or flu have more combination? Think some enterovirus as well. How would it handle recombination? That is outside of the scope of this pipeline. So it makes a assumption that you are tree is a representation of what the virus is doing. Thank you. Alright. Any more questions for Scott? No. If not, thank you very much, Scott. Yep. Before I go, got to foresight to put a QR code, but the GitHub link for this pipeline should be in the agenda. All right. Perfect. Thank you. All right, then our next speaker is Lara Yanov, who's co director of the biological data science score at the University of Alabama, Birmingham. And she's going to tell us about NX spatial, which is a reproducible framework for standardize the preprocessing clustering of xenium data. All right, everybody. Thanks for having me here today. I'm excited to be here. Is something I'm I'm going to go through introduction where I'm going to discuss two of components of this talk today. I'll then give you all the review of the actual control. Some sample outputs, and then end it with two directions. Alright. So for a simple moment, stage of just coming for those of you who have not heard about this yet. Alright. So first foremost, that has for instance, Or for instance, if I am working with cancer, this can be a microenvironment or tumor microenvironment. With that being said, there are a lot of different methods that are available for stage transcognomics. But they generally fall under two general branches. You can either have sequenced days approaches such as the vision or the but they under And for today's talk enough spatial supports The next item that I'd like to discuss sends a bit of a different AUAB, and this is not unique to core facilities. It's also true for instance, companies have to provide support to buy informatics services at scale. But as you like to know, there are a of logistical means that come with this Right? We need to go through project intake and scope. We need to think about communication, stakeholder management, data management. Obviously, we're execution, robust visualization as well as reporting and analysis and And we're bringing in a new service, especially one that's subject to I doing this, couple of years ago with the Xenium, we tried to really kinda highlight what were the process or the that we could automate versus those that really have to be within more the expert driven domain. For those that we could automate, what would be standardized to make them broadly applicable to multiple projects? And then we still wanted to provide interim or intermediate results So with that in mind, the items I was actually have out in mind is what we're internally calling on a tier one analysis, understand that there's going to be more analysis after this. And highlight specific components now, but then in the next few slides, going highlight specific once somebody brings in data, the very first step that we have is this one called annotations. And has the name suggests it essentially allows you to bring in any metadata that is maybe important for your analysis. Maybe you try of the section that you don't want to rethink in your analysis because that can actually affect the customer results. This serves that dual purpose. Beyond that, we have quality control at very different steps of the workflow. So we'll SQC at the raw data. After we filter the data, after we normalize the data, also after we cluster the data as well. I'll show some examples of this Once an And once all that is done, results are compiled. They go into RDS files, and we do a lot more visualizations after that so they can interpret those those results robustly. And we end this with a report. So here's some examples of the quality control steps. For those of you who are familiar with single cell analysis, just be aware that all that is already part of workflow. But what I want to highlight today was some of the customized steps that we put into this workflow and some of the Zune specific metadata as well. So most of things figures to in the middle and then to to your left as well are they are figures that are highlighting essentially the area. So this is something that we'd like to look carefully because if you're analyzing data of a similar cohort, you normally would expect the distribution of cells to be similar. If you find deviations away from that, that can be indicative of a quality concern. And what we've done as well there is also compiled the information so that multiple samples can be visualized together for the sake of simplistic here and this workflow of removing cells is up to you as a user. Next up to your right, have that stack bar plot. And what we're doing here is very straight to classify the cell shapes. So we do a simple cell shape classification. We classify them by circular in on data in polygonal. Obviously, cell shapes are far more complex than that, but the idea here is also to leverage towards The idea behind that is to leverage market expression as a way to, tell us a little bit about the segmentation quality. Segmentation is quite challenging. And you do want to ensure that it's at least sufficient for your downstream analysis. And the assumption is that if you were to plot these markers, in a in your data, markers that should not be coexpressed between two different cell types is that you should find very few cell that actually coexpress them. Ideally, none, but that is a little bit difficult at this time. And while show here is actually an example of this, you can see the majority of the data follows either the y axis and the x axis because these plus should not be quite express. But for those purple cells there, those are cells that have co expression and then can be of the problem. Now this is a small percentage in particular they said I think on 1.49%, but I think the goal is to be aware of what's happening with your deal. Next up, we have normalization. But I think I'm not gonna go through a of normalization in detail. All they've been And for those of you who are not familiar, air normalization essentially scales, the gene expression by the median cell area. And we go ahead and we do As I said earlier, though, you do have a choice of any one or the other or both. Once So we are not just looking at principal components in resolution only. We also need to look at spatial way or spatial So for so for for do the Even though you can go ahead and run quite a lot, I would advise don't go ahead and do 200. We've done that, and it's a lot of data that you get back. My recommendation here is to start, you know, with a sensible guess, review the results, identify which parameters you need to find to, and then can also take a advantage of next full resume capacity. Simply change the configuration file, and if you if you resume the pipeline, the pipeline will force only perform those additional parameters that you're now including. But review process in this particular step is critical because all the downstream analysis that you do on that tier two, tier three are dependent on the identification And then finally, of course, we have the report, really compiling all the information in a way that is easy to is critical. We made a report that compiles everything that I point those from the raw metrics. So we have things such as the numbers And we came to the cluster results because we expected the videos to run a set of perimeter sweeps, we'd actually compile these into videos that are part of split cost plus, which is the Humamps slip by sample with your sections detail as well as integrated Humamps and also the violin plots or the top plots, which is what is showing the market expression so that you can make decisions about self typing. So that's our review of workflow. There's still rule ahead. I told you to touch on this a little bit in detail by first talking about what is out of scope. One point of component here is to understand that the input data should be post segmentation already. If you're using an out of the shell solution, Solution, my recommendation is to use a 10 x multimodal capacity because relying on that is saying alone is certainly going to be problematic for segmentation purposes. With that being said, you may still need to resignment. And if you do need to resignment, I just wanna give a shout out to the NFCore space to workflow because that's exactly what that workflow is addressing. But what is in scope for us as you think about data growth? These data sides are growing very rapidly. This workflow was built in our environment because in ideally, this majority of the users are gonna be our users, we want to make sure that the were something that was immediate with. Duplicate effort. I'm curious to see if the cosigners AI will help us here so I can maybe record on that in the future. But importantly as well, beyond making this by conversion, I also want to include cellular transfer. This is just to allow you to transfer biases in reference data sets. That's why I kinda stay away from this at this time because I'd like to give you those to be aware of these biases. If we can think of methods that enable curation of the references first, that would be ideal. And then there's cosmics. My team matches a lot of experience in cosmics. And I would say that even though these technologies may be be similar at high level, this particular technology has some specific quality control steps that need to be taken, and I'd to share that experience in the future. In a next flow workflow as well, whether it's part of this or another one, is something that we can sign in the future. So with that And of components. Did a lot of work initially in understanding the technology, understand the limitations, and also developing the scripts that is not part of this particular workflow. Also, want to acknowledge my colleague, doctor has been championing a lot of these technology bringing speech expertise and math is to UD. Jeremy Day has been working with her as well to have So we're going to do it. Do I have any questions from the room? Yes. While you're thinking about it, I was interested another question to when I asked Rhea earlier was, did you come to Nextflow to implement these kinds of solutions? Was it an obvious thing? There other things that you considered? And how's been your experience of NextLow and then, of course, specifically? That's a good question. So I've been working with workflow managers for a long time before Nextflow. I used to develop a lot of things make workflows, but once an effort community really started to grow, it became an obvious choice for us that it made sense to instead use Nextflow so that we could depart this great community and really, again, minimize the location of fabric. It's nothing worse than, like, you know, just seeing this effort be dedicated. So for us, that was primarily the main driving factor there. And, of course, for this particular workflow, I would say that, yeah, that was natural because of our experience we already had with testing on a seek. And we hope that this can also be part of of the other ongoing efforts. Actually just recently reached out to Florian from and, of course, to see if this something I would like That's amazing. You so much. Any questions? I have question over there, etcetera. Or the, for the labeling process, I guess, like, we guys do it now, I guess is, like, is AI part of the future because I haven't seen any people say that labeling is the favorite part of their job. So, yeah, the Yes. Absolutely. So the cell label transfer, a lot of that is already leveraging some. Of the underlying engineering algorithms. What I have noticed in the past though is that the problem is not so much on the transfer itself. It's with the biases that present in that reference data set. So if you have a reference that may have, for instance, a But they may not actually be true. They may actually be done cells. So somebody is just somewhere of cleaning up that data in advance, also being aware of any of, like, any biases. So does your reference have these e samples? Are those even applicable here? And also being aware that our reference data set may not be sufficient to identify all your cell types because we have hypothesis specific clusters too. So, yes, I think they actually be part of that, I hope to be able to leverage what we kinda do more manually these days to validate with no markers and be able to also clean that up. So possibly. Aspect. And then we also have the fact that R can hold all that memory in all that data in memory, which can also be very resource hungry. So the motivations there is to make it more efficient and allow us to essentially make it more efficient towards that future in So it's not so much the pipeline as it's within pipeline picture. Exactly. Yeah. Some of the tools, I do think there could be some tools that I'm it's pipeline would translated Python and that might be where we mainly try to do a level rewriting as we have indicated as well. Or maybe fill in this rewriting and rest for you That would be great. I would take it. Please. Bueno. Or Thank you. Practical question, how long does it take to run a typical data set that you may have? The second is, how many samples do you think that you can run before it gets into maybe too big for the report? Yeah. That's a really good question. So that depends on the panel size. So when we first start the topic, we're working with a lot like, the four three hundred to 500 panel sizes. We ran the data with dozens of samples, that can typically be done in, like, a day. But something when you have five ks, that is a different piece. That is not the same one time we ran this workflow with I would say, close to 250 per The resource heavy component was fine is a spatial cluster analysis. Actually, for the most part, everything else is not time. Yeah. Yeah. From my experience with lazyum data, find that biologists often come to me with regional interests, predefined that they are interested in making those types of comparisons across. Does the pipeline currently have functionality to use data from the biologists with help other computational team? Is that part of annotation step that you mentioned earlier? You make those higher level Yeah. I guess I'll break your question to ensure that I understand. So yes, you can include I guess, on the confirmation for you as part of the workflow, and that's gonna be part of the However, if you do add that in there, we'll still do culturing. It's not the culturing is not leveraging any many annotations, if you will. It's just part of the workflow. We'll culturing then you can probably minimize the amount of spatial costing you need to do. Right? Because at that point, already have the information that you need. Finish the workflow with minimal runtime there, and move forward with that. So that's exactly what's going. Yeah. All right. Well, thank you very much, Laura. That was fantastic. Like to ask for one more round of applause for all of our speakers from this morning. And now we're going to break for lunch. Lunch will be outside there. You've already seen the table?