Listen on
Overview
Spatial biology and new molecular assays generate data at an unprecedented scale—terabytes per run, hundreds of thousands of cells, and billions of transcripts. This explosion of complex, high-throughput data creates immediate bottlenecks in R&D, delaying experimental cycles and demanding the re-architecting of computational biology pipelines. For data leaders and executives in biotech, this means current infrastructure and analytical approaches are breaking, costing time and money, and hindering the pace of discovery and therapeutic development.
Host Ross Katz speaks with Kenny Workman, co-founder and CTO of LatchBio, who offers a unique, battle-tested perspective on this challenge. After years of direct experience building data platforms for biotech R&D, LatchBio identified a critical market shift: the most urgent data infrastructure problems now lie with the “solution providers” – the companies designing and distributing the next generation of high-throughput assays. The conversation explores LatchBio’s diagnostic journey, from initial observations of computational waste to building a platform that powers these assay providers, ultimately accelerating scientific insight for their end-users.
Kenny details the LatchBio platform’s core components—workflows, distributed data storage, contextual metadata, dynamic compute, and high-performance visualization layers—designed to handle escalating data volumes. He also discusses how these solution providers use LatchBio to white-label data analysis and visualization, driving faster kit adoption and clearer ROI for their customers. The episode further delves into the surprising application of large language models for standardizing fragmented public molecular data, and the crucial need for faster technology adoption to keep biotech competitive globally.
Key Takeaways
Direct sales of data infrastructure to biotech R&D shops face unique challenges.
Due to market volatility, internal R&D structures, and a culture of vertical integration, biotechs often default to building expensive internal tools rather than adopting vendor solutions. This economically irrational behavior burns millions and delays progress, making it difficult for external platform providers to find consistent usage and demonstrate ROI.
New assay solution providers are the true drivers of advanced data infrastructure needs.
Companies developing next-generation molecular measurement technologies—like spatial genetics—generate data at a scale that necessitates specialized, high-performance platforms. By serving these providers, LatchBio enables them to white-label reliable data analysis for their customers, directly impacting kit repurchase cycles and scientific understanding. This shifts the focus from individual biotech R&D shops to the source of the genuinely big data.
Spatial biology data volumes will break current computational systems without purpose-built architecture.
Modern spatial assays generate terabytes of data with hundreds of thousands of cells and millions of transcripts per run. Rendering these complex, multi-layered datasets in a browser demands specialized file formats (like PMTiles), in-memory SQL engines (DuckDB), and low-level system engineering. The industry’s rapid data growth (2-3x every four years) will magnify these issues, demanding continuous re-engineering to prevent experimental delays.
Large language models effectively clean, structure, and standardize fragmented public molecular data.
While often overhyped for predictive biology, LLMs excel at transforming unstructured scientific papers, study metadata, and raw files into consistent, structured objects. This significantly reduces human labor for data curation and harmonization across massive public datasets like GEO, accelerating the accessibility and utility of historical research for new discovery.
Related: CorrDyn helps biotech and life sciences companies build reliable data engineering solutions and ensure data reliability for complex assay data. We also provide strategic guidance on AI strategy and offer specific insights on gaining data value in biotech manufacturing.
Full Transcript
Jason: Welcome to Data in Biotech, a podcast from CorrDyn where we explore how companies leverage data to drive innovation in life sciences. Every two weeks, we sit down with an expert from the world of biotechnology to understand how they’re using data science to solve technical challenges, streamline operations, and further innovation in their business. This week, we’re joined by Kenny Workman, co-founder and CTO of Latch Bio, to get his take on how the company is transforming data infrastructure for biotech solution providers. During the discussion, Ross and Kenny unpack why spatial biology and new assays demand scalable, high-performance data infrastructure, the five core components of the Latch platform and how they power a scientific insight, and the essential role of data in modern biotech innovation. Kenny also addresses the lag in technology adoption within the biotech industry and the importance of evolving approaches to data-driven research. Here we go.
Ross Katz: Kenny Workman, welcome to the Data in Biotech podcast.
Kenny Workman: Thanks so much for having me. Appreciate it.
Ross Katz: Awesome. Well, just to kick us off, would you mind giving us an introduction to your background and what brought you here today?
Kenny Workman: Yeah, totally. Have always had a fascination with living things for probably as long as I can remember. Greenhouse full of carnivorous plants, collections of frogs and newts, that kind of stuff early on in middle school. I actually studied EECS and BioE at Berkeley. Coursework was mostly math heavy with a focus on the underpinnings of machine learning systems, so statistics, linear algebra, convex optimization, this stuff. Thought I wanted to head in that direction, but was also working in a wet lab pretty much since I was 15. Started at local community colleges, transitioned to different groups in and around Berkeley, basic research in a variety of domains. Lots of painful hands-on experience with rudimentary laboratory work, PCR, cell culture, Golden Gate. Around sophomore year, I transitioned from basic research into computer stuff. Got an internship at a cell therapy company working on both early machine learning and a lot of cloud infrastructure. Worked there while I was studying and during the summer. The following summer, I worked at another SynBio company called Asimov building computer-aided design systems for cell line engineering. Around that time was when we dropped out and started Latch.
Ross Katz: Yeah, awesome. So that’s a big decision. You decide… so, sorry, how many years ago was this when you dropped out and started Latch?
Kenny Workman: Yeah, it’s a good question. I was 21. This is 2022. At the very beginning of 2022. Or sorry, sorry, the very beginning of 2021. Timeline wrong.
Ross Katz: Yes. So a few years ago. So, what was the insight that led you and your co-founders to start Latch?
Kenny Workman: We didn’t really have a deep and prescriptive insight into how biotech should be handling data differently, at least at first. Our early observations really were those of repetition and waste. My co-founders Kyle, Alfredo, and I, we were all friends at Berkeley, several years prior to actually founding Latch, and we spent many late nights hanging out, drinking, talking about all kinds of things. Coursework, science and tech news, company ideas. Basically I would riff on the cool things I was doing with computers at biotech and Alfredo at the time was working at Google and Facebook, and just talking about the gap in computational tools. I was 21 at the time, they were a year older, and we decided to send it and fill it. I also will note I think many of these founding stories are over-glamorized in American pop culture and suspect somewhat retrofitted to establish continuity with a unique insight that carried forward to current success. The only reason I bring that up, it’s not to be a contrarian, it’s to highlight what Latch is not yet. We didn’t have a specific insight.
Ross Katz: Yeah. Well, what did you do? Obviously you’ve developed this platform over the last few years, so can you just take us on a tour of what that journey was like, back to the beginning when you were just trying to think about the intersection of computation and biology and try to figure out a place where you and your friends from Berkeley could make an impact.
Kenny Workman: We had conviction in something more like a theme. That biotech, primarily through data generation methods and a host of improved tooling, was going from more of a science to an engineering discipline. Both to manage the volume of information, but also because new methods of discovery and engineering would emerge with scale of data, and a lot of the industry would reorganize around computers. We all understood computers really well. Most of us studied together at EECS. We met in the machine learning club and developed an independent interest in low-level systems, file systems, operating systems, languages. We didn’t really know exactly how to approach the beast that was the biotech industry. We just understood that this theme was playing out, we could solve technical problems at its intersection. We approached every aspect of the company as an engineering problem. Particularly when it comes to product development, what insight did we have? We ran lots of experiments, collected lots of data, and redirected resources based on data. That can segue into how we started with conducting hundreds of user interviews, etc. But that is how I would describe Latch and why I push back against we didn’t have original good idea that is still a good idea.
Ross Katz: Yeah. Well, it sounds like you had sort of a scientific approach to identifying the company that Latch was going to become. Tell us about those user interviews. Who did you decide to interview with and what were the sorts of questions that you were exploring with them to try to figure out the direction you all wanted to head?
Kenny Workman: Yeah, looking back, it was really anyone who was doing biology or programming at a biotech that wanted to talk to us. There was a crazy period where we were just spamming every possible channel on LinkedIn and email, got banned from multiple… shout out to Bluebird Bio, apologies for absolutely destroying your email servers. I have specific stories like this. Asking scientists, hey, will you meet? We had a limited amount of experience. I worked at a few companies, but we were just trying to talk, gather information. What stood out… very not focused, very not smart, but ingesting data. What stood out was heterogeneity. What I mean by that is there were almost no patterns in the types of data generated, in the way that different experiments are strung together or combined in the tools or infrastructure waiting downstream, and the number of types of teams waiting to interpret it to close the loop. Nothing about this problem showed any sort of consistency from biotech to biotech at first. There’s definitely common threads running through the industry that I can see now, that just came four years after reading conversations, building, and 21-year-old me and co-founders in early engineering team were trying to connect dots prematurely. It was challenging.
Ross Katz: No, that makes a lot of sense. Can you give us an insight in what were some of the early dots that you connected that sort of started the platform that you started to build back then?
Kenny Workman: Yeah, I can explain what we built out and how it encapsulates the quote what I’ll call a skeleton of a generalizable life cycle of data in biology. We definitely started with the obvious and most computationally taxing pieces because it’s most aligned with the thesis of data generation causing issues. And then we count lots of biotechs who needed a lot more and then we would either reject them or they would reject us and we would just add a note to reach back out when we expanded. But in a nutshell, there’s five major technical components to the platform. We started with workflows, of course. It’s the most obvious, makes the most sense. It’s well-defined batch processing of a large molecular data. We then added a place to store data and this became a distributed file system that we wrote that just ingests blobs of molecular data and then mounts it to the rest of the platform, initially just workflows, but everything else. What became important next was, hey, you need contextual metadata. Information about the cell line you’re using or the temperature in the lab that you need to associate in a structured way with files. We built a registry, which is a tabular database type scientists can interact with, but then can be pulled into Python. That theme of different levels of control, I’ll touch on later. We then needed computers. Bioinformaticians need to write whatever code they want. So we built a computing system called Pods. And then over the last year, we’ve developed the creme de la creme, which is Plots. It’s not just plots. It’s actually the most ambitious thing we built. It’s a combination of scientific notebooks and dashboarding, with a custom reactive framework and Python-generated widgets that are backed by large computers. It’s kind of like a GraphPad for really, really large molecular data sets that can synthesize everything else on the platform. Very long-winded. You’ll see this as a theme when I answer your questions, I try to distill the complexity of what we do into something that is simple, but I feel like in that process it doesn’t do what we do justice. I am often caught between a rock and a hard place in trying to explain this stuff, I think.
Ross Katz: Yeah, no, that makes sense. And you should feel free to go into whatever degree of detail you want to go into. For our audience, I think we can handle the detail. So, you built this biological data platform with workflows, data storage, contextual metadata, the ability to spin up and spin down compute for things that you wanted to run, and then now a data visualization layer that’s really high performance for the large complex data sets that are being introduced. What is Latch working on now? And how does that relate to the customer discovery process that you’ve undergone throughout this whole process?
Kenny Workman: It’s interesting. In a nutshell, what we were working on over the past I would say four years is a infrastructure that stores, processes, and analyzes experiments. Ingesting streams of data coming from different machines or experiments, processing them into a state that’s human-readable, and exposing them, often layering or synthesizing them to different members of a biotech organization so they can draw scientific conclusions and make progress. I’m sure many of your listeners can relate, but rarely in early work did we encounter issues consistent with the original thesis that the scale of data was delaying experimental cycles or straight up stopping them because existing software systems were unable to handle load. In reality, most of the problems in analysis with the largest return on the pace of science come from basic team and project organization. Modern biotechs are I like to describe them as human engineering projects that need to coordinate very diverse teams around scientific consensus, but everyone has some different training, speaks a different language, and key to analysis has various levels of computational fluency. The example I tend to give is grouping lots of things that shouldn’t be grouped, but for the sake of this example it will make sense. Immunologists, molecular biologists, chemists, they all need graphical components and click controls. Computational biologists don’t really do well in AWS, they need to have a sandbox where they can write basic Python and R code and use their unique mix of hey, basic programming ability, but heavy statistics and domain knowledge. Bioinformaticians and software engineers need to basically run whatever computer they want with their favorite tools, otherwise they’ll freak out, they’ll lose control. In a single leg of a cycle where you’re developing a therapeutic, you have to bring all these people together. If you’re building a platform where they’re communicating, you need to build product for all of them. That is initially what we were working on and what problem we were trying to solve, and we had a modicum of success there. We’ve actually shifted away from this in the past year, and I think the story there is pretty interesting. We can get into that.
Ross Katz: Yeah, no, please. I’d love to hear what drove that shift and what the shift is towards.
Kenny Workman: I think this is a pretty interesting and very much earned secret that only came from navigating the edges of the industry and throwing things against the wall for years. But biotech companies actually make poor customers. There’s a few things here, and some of them are perhaps somewhat controversial. Let’s start with the least controversial. Experimental structure between companies look very different. Already established this. What this means for a cloud infrastructure company is finding companies that have unique challenges at scale of data is hard because of the heterogeneity. We just a consistent poor customers and any attempt to group these by therapeutic focus, disease, etc. in practice fell up short. Ecosystem’s volatile. I think biotech’s very unique in that R&D is very overrepresented in startups. For a data analysis products, all the interesting and useful problems are in R&D. All you’re working with startups, they fail a lot. Or they get merged, but then the merging company has no interest in working with you. It just makes them challenging in that regard. Here’s the most controversial bit. Compared to most industries, the dollar amount raised on average is exceptionally high and the feedback cycles are quite long. Market forces are pretty weak, and this leads to what I’ve observed as very economically irrational behavior. I have seen teams burn millions building internal tools instead of running proper vendor evaluations because they could. And I saw many of these companies, the same companies, shut doors years later where it’s entirely unclear to me… I’m not saying I would actually never suggest that the reason was because of this decision. I would say that this decision is a symptom of a broader culture of vertical integration without need. But no one knows when you see big NewCo going out of business, that’s some that I’ve observed and that waste never made its way into the public dialogue. A huge part of this is you have to understand startup physics, innate truths about your market when trying to build a company, and there’s certain things you can and cannot control. We could control technology. We can’t control the market education about an entire group of companies and what they will and will not buy. For all those reasons, we moved away from biotechs. If you think about it, the really gnarly problems come downstream of the new class of molecular measurement technologies that generate actually big data, terabyte per run data. You could think single cell and spatial epigenetic, proteomic assay. This is where the need for data infrastructure was high and actually urgent. Because the biotech customers that did adopt us, I will say we did have early success with biotechs. There are some biotechs that are absolutely on the forefront of the adoption of new technologies and have this need, but repeatably finding them and ensuring that we would have consistent usage… their compute usage was spiky and moved up and down with an actual cadence of their experiments. It was hard to find more of them, for the reasons I said earlier. We just did what made sense to us. If you think back to the thesis of the company, we found the companies that were generating the data. We went right to the source. These are the solution providers building the new assays themselves, and packaging them in kits and services that distribute them to the industry.
Ross Katz: Yeah, that makes a lot of sense. Can you give us some examples of what those new companies are and what are some of the assays that they’re bringing to market and the scale of data that’s being created?
Kenny Workman: We work with quite a few and they all do really interesting things. I would say the true data-intensive scale comes with spa… mostly with spatial. In that camp, we have spatial epigenetic, AtlasXomics is a phenomenal flagship example of this. They’ve built this technology spun out of Rong Fan’s lab at Yale called DBiT-seq. It’s called deterministic barcoding in tissue. You basically take a section of tissue, stick it on a purpose-built microfluidic chip, and then flow molecular barcodes over overlapping X and Y coordinates such that you know their code ahead of time and what channel they went in and then you can resolve them computationally post facto. Because these are just generic library adapters, the barcodes end up being generic library adapters for NGS, you can extend this to any NGS-derived assay including CUT&Tag and ATAC-seq, etc. that as listeners may know are epigenetic measurements that just happen to use some hack with NGS to get the readout. Curio Bioscience does true single-cell spatial by somehow you guys should read the paper, managing to get their barcodes inside the nucleus of a cell, providing that compartmentalization. We work with a company called Broken String that doesn’t do spatial and is actually not an example of scale of data, but they provide kits to let people evaluate the off-target or just editing statistics of genetic editing proteins for all the host of companies that are building things there. Focusing on the solution provider bucket, the reason they make such great customers is they concretely understand the ROI of Latch. It drives down the kit repurchasing cycles by helping scientists understand data and scientific questions faster so they can rationalize why they’re purchasing it.
Ross Katz: So if I understand correctly, let me just try to connect the dots here to make sure I’m understanding. You’ve got these companies that have got new primarily spatial but just high data volume assays that they’re bringing to market that require high computational load and new computational approaches combined together to get the insights that customers of these solution providers need in order to get the value from the assay or the kits that are being sold. And Latch is bridging the gap between the hardware providers that are providing the kits and the assay and the scientists or whoever the CROs or whoever the customers are of the assay providers that are then trying to get value from the data that’s coming out of there. Am I thinking about that right or where would you update that understanding?
Kenny Workman: Yeah, I’d say that’s exactly correct. It’s B2B2B. And that sounds kind of fake, and at the outset if you asked us, hey, is that what you’re going to do? I would have said that makes no sense. That’s stupid. But just because of the amount of complexity involved there, that just turns out that’s where the highest density of problems related to data generation exist, are the people distributing the things that generate the data. And the add-in advantage of, A, they have these economic models that package this idea of kit repurchasing is also beneficial, but also they don’t want to build two products at once. I think they understand that viscerally. For a variety of reasons, that was a bit scattering. They just make a really interesting segment to work with, and are closest to this original thesis we had as a team.
Ross Katz: Yeah, that makes a lot of sense. And these are very different assays solving very different biological problems. How does Latch work with one of these new solution providers, assay providers to bridge that gap, to stand up the workflow that the end user needs?
Kenny Workman: It’s a good question. The advantage here, and this is why I think the history of the idea of Latch is unique and helpful, we originally built this for biotechs. Our intention was a full replacement for AWS or GCP or etc. for a biotech, everything you’d need from that complete data life cycle to all the relevant controls to modify things, having good dev tools that work with the various levels of computational fluency for the people working on them. Those were all things that already existed prior to us making this pivot. The answer to that question is we set up the domain practitioners within the solution provider to have what they need to encode their domain knowledge into the tools and establish their own feedback loops in a variety of very specific ways. Everything from uploading a workflow, which is a pretty solved problem. We can get into that later. To having pre-configured environments for customers to write code for pre-configured dashboards, which is very much not a solved problem, especially as the object the count objects get really large. Actually building intuitive browser-based applications that are performant is a pretty challenging systems problem in 2024, go figure. But the point being is for each of these components, the solution provider controls it and has some amount of internal engineering that’s dedicated to Latch. They treat Latch as their data infrastructure.
Ross Katz: So if I understand that correctly, they have developers on their end who are creating templates within Latch that their end users can then build on top of to get the insights that they need from the assays that are going in, and then the entire Latch ecosystem comes pre-packaged with those templates for the end user when they receive the assay or the machine, the workflow, the kits that are generating the data as well. Am I thinking about that right?
Kenny Workman: Yeah, that’s exactly right. This actually caused us, as you can imagine, to really, really focus in on white-labeling the platform, and making the entire platform feel like not Latch and exactly what the solution provider wanted. We did this for a few reasons. The first one was pretty obvious, but each solution provider is a domain expert in their relevant branch of technology, they have deep expertise, extensive relationships, trust that we will never have. We want them to leverage their own distribution and be in full control over their brand. And then the other thing is you have to make these folks happy from a productivity standpoint. They need full control to make the changes that they want because that just matches their pace of development. White-labeling in that sense is kind of psychological, but they should feel like it’s their platform so they have the license to make those changes.
Ross Katz: Yeah, that makes a lot of sense. Did you consider trying to get the Latch brand out in front of end users and get into internal biotech customers that way, or had you already come to the conclusion that internal biotech R&D shops were not where you were going to want to focus your intention anyway, and so white-labeling and making your preferred customer happy was just the best way to go?
Kenny Workman: I think it was pretty much the latter from the start. It’s pretty clear if you dig in that Latch is powering things for the power users that are reading the documentation or getting in the details. I think that’s the correct division of responsibility because it lets the solution provider defer a lot of cognitive load and wasteful activity on debugging things or redirecting requests to our team. For folks that really dig in, that split is clear. But it’s an interesting question.
Ross Katz: Yeah, that makes a lot of sense. When solution providers develop these new assays, and they get to the point of thinking about, how are we going to enable our end users to get value from the data that we’re generating, and they’re looking at, what are the we can build it, we can buy it, what are the options out there to buy it? What do you see as being the aspects of the Latch platform that are driving the most value for your customers today? Or just making them most excited to use Latch as their white-labeled data platform of choice?
Kenny Workman: I really think we’re first in class in this category, which is a data infrastructure specifically for folks servicing their own array of customers. The alternatives really are to build in-house, and there’s nothing else I would say. When we drill into why Latch, the layers of stuff over the past year we built on top of… this was already an infrastructure that a team of 100 scientists could use to analyze data, but now we’ve layered in admin controls, kit packaging mechanics, billing management. The lift of doing that and starting from scratch or even extending your current system, especially in the current climate where it’s really focused folks to sharpen their pencils and be precise with resource allocation, I think it’s been somewhat clear. It’s us versus internal teams, and the lift for internal teams is just getting greater as our team gets better at our job.
Ross Katz: Yeah, that makes a lot of sense. I saw that you just released some updates to your visualization layer. You have an H5 viewer and the ability to explore these highly complex high data volume spatial assays in the browser has increased. Can you just introduce us to that feature and walk us through what are some of the parts of that problem that are hard and where you think what Latch has done is most valuable?
Kenny Workman: I think for a long time our team stayed away from digging into the details of assays and especially spatial because it just requires writing a lot of custom code that’s specific to the problem. But actually as we started watching things unfold a bit more, we realized that a lot of the structure in spatial was here to stay, and that as we focused… in terms of file formats, operations you want to do on the data, and the differentiation would be from just slightly nuanced differences between solution providers, a difference between their technologies that we could pull out and let them extend. But it was very much worth it for us to start building closer to the details of the assay. Spatial biology is I think even from domain practitioners mis-underrepresented in its scale. We’ve been working with a lot of CosMx data, and I think the example we wrote about a few weeks ago that was shocking to me when I discovered it is that a single run in 2025 can be close to a terabyte in total file data. And the number and types of layers are incredibly fragmented on top of that scale. If I remember correctly, it’s 450,000 cells. For context, a 10x Chromium run produces 10,000 cells. Something million transcripts, multiple layers of high-content diff images that each measure a different fluorescent emission wavelength that represents antibodies bound to different proteins. You’re layering transcripts, you’re layering images with phenotypic information, proteins… wait, did I already say that? You’re layering cells. Those three things on top of each other. And then the browser has to somehow, based on the geometric location or the area in the tissue that the biologist is navigating, fetch all three of those things from where they lie in a file somewhere and then performantly send them back and layer them on top of each other, downsample them, and render them. And the interesting thing here is the laws driving how these assays work just point to further gains in throughput, molecular throughput, and costs driving down something like 2 or 3X every four years if you look into it. These problems are just going to magnify. All this to say is that is a great case study of the type of systems engineering challenges that are uniquely here at with some of the solution providers. They really have issues with computers.
Ross Katz: Yeah, no, that makes a lot of sense. And it strikes me as you’re talking that these… when you talk about geospatial computing and you think about a tool like ArcGIS, this has been over time solved in the geospatial world, but our Earth remains relatively fixed in terms of the nature of the shapes that you’re rendering and the nature of the data that you’re layering on top of it. But when all of these new assays are coming out, and the tissue samples that you’re running are different each time, and the nature of the data that you’re layering on top is different each time, and the volume is just increasing and increasing, it’s just a different type of problem. Did you take any inspiration from the geospatial approaches or how do you characterize the differences between what you’re doing and the other computational methods that are out there?
Kenny Workman: Yeah, by geospatial approaches, do you mean the tools that emerged around processing satellite images? Yeah, exactly. Mostly. Exactly. What’s super interesting is a lot of early computational, classical computational results, either in algorithms or distributed systems, were motivated by molecular biology. And I promise I’ll return to the point about Latch. A great example I’d give is dynamic programming with the advent of early NGS, the Human Genome Project in 2003 with shotgun sequencing. It’s a classic algorithm taught in any undergraduate course is completely motivated by these problems. And then a lot of early distributed computing problems, MapReduce, fanning out data over many computers and bringing them back together, motivated by spinning up sequencing search servers. This dropped off around this time and then the new problems that started pushing the frontier of scientific computing became what you referred to as geospatial, what I understand is satellite imaging and particle collider, essentially particle collider outputs. I don’t want to get into this too much, but a lot of the file formats we use that we repurposed actually for the machine learning revolution that store large matrices of numbers and then the ecosystem for doing algorithmic operations on those matrices of numbers came originally from Bio then transitioned to what you said and are transitioning back. The point is that yes, we use an archival image layered format called PMTiles, it’s just a specific example I’m sure these things are fluid and will change, but specifically built to render large maps performantly, and then we looked at how it worked and then repurposed it where necessary, hooking it into Plotly, which scientists love as an example. And then on the other components of this specific system we built use DuckDB, which is an in-memory SQL engine, lots of low-level programming ideas and how to move data around efficiently on a server. But yeah, these are some of the ideas that went into making this browser and I’m sure will change but have roots in that field that you mentioned, which is cool.
Ross Katz: Yeah, no, that makes a lot of sense and the fact that you’re using an in-memory analytical database like DuckDB, back in the ball of all of this just highlights the way that you, by focusing on the compute layer, you can leverage the best tools that are available today and stay up and the nature of the problems that you’re solving forces you to stay up on whatever the new paradigms are in data or computing that can help you to provide the value to the companies that are producing all of this data. Thinking through the problems that you’re able to solve for the solution providers, where do you see the next frontier of challenges that are going to come up for these NGS and high-throughput assays, spatial assays, and the end users of those assays over the next call it, I don’t know, three to five years?
Kenny Workman: Yeah, the three to five year timeline, I think it’s pretty simple and doesn’t take an incredible amount of specific foresight. Take those numbers I just gave you for CosMx, they’re pretty similar across most of the competitive spatial vendors. And now just double those numbers and then half the cost. These are going to break almost every constituent of the compute… in some cases break and some cases just make actual concrete delays in experimental cycles from computation, and will just demand a team that has pretty deep knowledge of how to build these things to re-engineer them for the assays. That is the prediction. But I don’t think it’s particularly far-it just makes sense.
Ross Katz: Yeah. You mentioned that in some cases it’s going to cause things to break, in other cases it’s going to cause delays in experimental throughput, and that there’s going to need to be a re-architecting of the way that computational biology is done in order to make space for all of the data that’s going to be coming through. Does the nature of R&D or the way that biotechs do business change as result of that or what are some of the second-order impacts that you would identify from that?
Kenny Workman: That’s an interesting question. I would bucket that stuff as second-order impact. Single streams of experimental data becoming hard to manage won’t really have the structure of the impact won’t be as you described. But what will happen is in pretty much every field of science, at some certain threshold of data generation, the entire approach to reasoning about the system or manipulating it just changed. It won’t be because CosMx or Curio just starts generating five terabyte readouts, but because molecular molecular measurements are so cheap and abundant that you no longer have to do the weird hacky stuff we’ve been doing in biology and you can just understand how systems work in a data-driven way instead of using rational design and small-scale experiments that are really perturbing small sets of molecules at a time, etc. One thing I will say tangential to this point is that currently we don’t see this. We are… I’m in San Francisco, where are you located?
Ross Katz: I’m in Cleveland, Ohio.
Kenny Workman: Okay. I’m currently in San Francisco. I spend a lot of time in Boston. I would say that today, and it’s happening less as the market has been as it is, but tech-adjacent venture capitalists have long been bullish on data generation. I think we talked about this a bit prior, but what we see in practice is that while this thesis is directionally correct, in the same thesis that Latch is pursuing, the actual deployment the scale and the fragmentation of the machines generating all this data is not quite there. As it really ramps up, some of those things I was alluding to will begin to happen like data-driven, model-driven discovery and perturbation of systems. That’s something that all kinds of folks are working on.
Ross Katz: Yeah. Well, what do you think is the bottleneck that’s keeping this… I don’t know, maybe it’s just my own newsfeed, but I see lots of news about how AI is transforming biology and about how the data that gets generated is going to be able to create these foundation models that enable us to solve problems that were never solvable before. And it feels like the solution providers that you work for and the types of computational methods that you’re applying fit very neatly in that. I’m having trouble squaring that with what you just said, which is that VCs are very much bullish on data generation in biology, but you’re not seeing the on-the-ground impact or adoption accelerating like one might expect. So, first of all, am I understanding that correctly? And if I am, why do you think that contrast exists?
Kenny Workman: Yeah, and to be super clear, I’m just suggesting there’s a lag. The direction is totally correct, there’s just a bit of a lag. And when it starts happening, it’ll really ramp. I think people forget when there is something that blocks a process, whether it be in the real world or digital, it blocks a process and the impact of that… here I’m specifically referring to computational bottlenecks due to the volume of data… will be felt. The reason you probably don’t see that, and this is another slightly controversial opinion from again the vantage point that we have is pretty unique because I’ve seen internals of hundreds of companies over four years, there’s a huge gap between what is marketed and what is actually going on within biotechs, especially with respect to machine learning based platforms. The reality is that most of these companies are not working nearly at the edge of developing models, and that is because they’re not nearly working at the edge of generating data. It’s second-order problem. That is because they are still leaning on old-guard rational techniques when rubber meets the road, because that is what has the highest probability of success to fill a clinical pipeline and return dollars to investment, which makes sense. It’s just the pitch that they gave was not that in most cases. A lot of these things fell back to the conservative approach when funding dried up and it was clear that you couldn’t keep telling the story of a platform without data. And a lot of the platforms didn’t work. Oh, these are very broad-stroke observation. At the same time, there are companies doing incredible work here. The virtual cell point, a virtual cell is basically a language model that ingests lots of well-structured molecular data to aggregate the statistics across different tissues or different, whatever the focus is, tissues, diseases, it’s an aggregation of the statistical state of a molecular system so that you could predict a perturbation in the presence of a drug, for example. There’s some incredible companies in terms of go tackle this problem in a very prescient way. Most of the way you tackle this problem is really innovate on the data generation side, build platforms that generate high-quality data. So hopefully that, again, a little bit scattered, but hopefully that answers your question.
Ross Katz: I think it directionally answers the question. And it also strikes me that the capital intensiveness and the investments that you need to make in the data generation infrastructure causes you to want to get the most you can from the historical capital expenditure you’ve made in the data generation infrastructure, and so there’s always going to be a lag between the creation of these assays and the adoption of these assays just because of the scale of the expense that needs to be made. To your point, not just in the infrastructure itself to generate the data, but in the computational pipelines and tools that will need to be there to ingest the data.
Kenny Workman: Yeah, to that specific point, I would say it’s a little more deep than a lag. These companies never had the culture or the correct infrastructure in place to legitimately be a data-driven biotech. There are companies that are emerging that are data-driven biotechs and have this culture, and I think it will begin to show in very strong top-line clinical data from programs that are the result of these efforts. I don’t know, we’ll see it play out. A great example of a company who’s actually reinvented themselves here is Genentech’s, the gRED. What Aviv Regev is doing there with respect to the scale of generation, the use of the data generated to drive unbiased discovery, and the integration of those wet lab systems with researchers and computationalists is a great example of probably the model that will drive us forward into an era of biology that is too hard to reason about in context of a human mind. That should be a good example for folks to look at. Very inspiring stuff.
Ross Katz: Interesting. And all of this makes me want to get into some of the things that you say in your manifesto. So one of the things… you seem to be calling for the continuous mining of biological data and the ability to better teach computers about biology. What do you think that means over the long term?
Kenny Workman: Yeah, it actually doesn’t mean that much. I will say we were young, we’re engineers, and we wrote this from more of a place of just inspiration and hand-wavy romantic poetry than any concrete idea that people should take seriously. I think at the time we were all systems nerds. We’re all reading and studying the internals of databases, programming languages, these things. And we saw a future where if these themes continue to play out, you’d have to rewrite all these components for biology to get more performance, indexing in a database is quite easy. At the storage layer, the layout of sequencing data and the way it’s compressed on disk should be purpose-built for biological applications. Those are the kind of things that we were thinking about at the time. In practice, it was just not necessary. I think the one place where this idea does extend to very concrete efforts now is on the reuse, repurposing of public molecular data. We’ve all established that data generation is a bottleneck in biotech, above anything you could do with computers because computers are just making what comes from the data interpretable. But that doesn’t discount the fact that there’s a lot of unused experimental information scattered across NCBI servers and GEO that really cover a broad space of diseases, tissues, unique patient identities that are just not being used because the amount of human labor that goes into structuring them, and making them consistent is quite large. Very, very recently, there’s an entire cohort of solution providers their whole job is to organize, or clean, curate, organize, and then sell data to biotechs. We’ve started working with a few of those folks. The older ones are Rancho and BioTuring, but there’s a new crop, Meromix and Phythio. We started building tools for them. And very excitingly… a tangent, but I think the advent of language models in technology broadly, but especially in biotech has been overhyped and a lot of the promises of their concrete capabilities overstretched. But specifically in the problem where you are ingesting papers, study metadata, and scattered author raw files and trying to make them into a consistently structured object with a set of rigorous tests, we have found incredible success using latest ideas in language model engineering that is the subject of a full hour of conversation that we’re basically equipping these data providers with to label data more efficiently, put it up on the platform, and distribute back to biotechs. And in some cases the data from original labs, the revenues from proceeds of those sales were trying to redirect back to the labs for basic research. Just a cool branch and probably one of the threads from the original manifesto that we’re continuing to act on.
Ross Katz: Yeah, that’s an amazing use case for large language models. It’s funny for as many times as I hear people talk about the off-the-beaten-path use cases of large language models, just the core use of text cleaning, standardization, converting unstructured data into structured data that can be translated into outputs that are standardized across different places is just a huge value. And that’s a really interesting use case.
Kenny Workman: Yeah, and you have to think about why it works. For molecular prediction tasks that need to build some model and learn statistics of physics and molecular interactions, an incredible amount of both architectural changes but also domain knowledge need to go into building new models. Today if you’re talking about reusing in general reasoning models who have been trained on text, the exact substrate for them to perform are those of text manipulation and human reasoning. And there are very few concrete examples of those that actually move the needle in experimental campaigns. But the one I described is potentially one.
Ross Katz: Yeah, that’s awesome. Well, as we head toward the end, I just want to ask you a few final questions. Is there a misconception that people have about biotech and the way that data is used in biotech that frustrates you, or ways that you think people should think differently?
Kenny Workman: I think a lot of my frustration comes from the pace of adoption of new technology in the industry. That really explains most thing. That explains why, as we talked about, data isn’t moving as fast at first glance as you would expect. That explains why we did not have success with a lot of biotechs. You can have other reasons for explaining that. I legitimately think that’s the explanation. You look at a lot of current problems with American biotech, cannibalization of clinical market share clinical assets by China, the amount of these companies trading low cash etc. I am here for this industry. I’ve been here in many ways since I was 15, I’ll be here the rest of my life. We need to get our act together in terms of adopting new things and experimenting and trying. Relegating to doing what we know because it makes sense will not allow us to compete in a fast-moving world of innovation, especially when the results have not been outsized. Not to get into another tangent, but most of the drugs we’ve been developing just continue to balloon in cost with very marginal impact in their given disease area on survival etc. because the actual mechanisms that we’re tackling in terms of disease are just exceeding the complexity and the reasoning abilities of a human mind. We need to adopt new technology to make progress in the engineering of these systems. That just extends beyond software. But probably not the direction you wanted me to go with that question. But…
Ross Katz: No, that’s absolutely a fantastic direction to go with that question. So, thank you. Is there anything in particular that you’re really excited about, either for Latch or for the biotech ecosystem at large over the next three to five years?
Kenny Workman: Yeah, man, it’s really these tools. You really study any industry and whenever the constituent tooling gets better, things start to explode. In computing, the analogy for me is always the resources in a personal computer and what that did for developers. There’s an entire class of problems that I as an engineer didn’t even have to learn about because I have 32 gigabytes of RAM on my Mac. The struct packing and bit twiddling and getting things to work on a 64 kilobyte machine, etc. It just let you focus on new interesting problems as resources became abundant. And I think you’ll just find many of our experimental approaches, many of our reasoning about how to design an exper- a campaign of approach will become restructured as data generation and the second-order effects of that… because everything in biology is a derivation of the core readouts, we think of molecular interventions or molecular manipulations, they’re always a way of perturbing the original NGS readout and that will probably continue to be true with new technologies. But the way of thinking about all these things will just drastically change as data becomes cheap and abundant.
Ross Katz: Yeah, that makes a lot of sense. And for anyone out there who is founding a company either a biotech company or founding a company to serve the biotech industry, is there any advice that you would offer to people having gone through the journey that you’ve gone through?
Kenny Workman: Nah. Everything is fake. No one knows what they’re doing. Especially a 21-year-old kid who has yet to IPO his company. I think just take everything with an enormous grain of salt and develop an internal reasoning mechanism where you can evaluate things for yourself and that’s…
Ross Katz: I love that.
Kenny Workman: Including everything I said here, right.
Ross Katz: So for people who are interested in learning more about Latch Bio, where should they go?
Kenny Workman: We have a great website, latch.bio. Pretty active on socials. We try to write about what we’re working on in as much detail as possible. And especially if you’re a new solution provider and you are getting off the ground building basically the hub and spokes of the industry, the things that will let all these scientists measure their systems in greater throughput, reach out to me directly at [email protected] or go to our website and put in a request to talk to our team. We’re mostly engineers, we want to talk to you.
Ross Katz: Yeah, and for anyone interested in learning more about the high-throughput assays that Kenny’s been talking about on this call, I can’t recommend highly enough their Substack that provides a survey of these assays and the way that they work. It’s a fantastic resource for anyone who’s out there.
Kenny Workman: Thank you, sir. And shout out to the whole team for that, particularly to Hannah, amazing product manager who’s really been writing a lot of these pieces as of late. That’s a lot of her work.
Ross Katz: For sure. Thank you, Kenny, for joining us today. Really appreciate it and look forward to connecting down the line.
Kenny Workman: Sounds great. Thank you so much for having me.
Jason: And that’s it for this episode of Data in Biotech. If you enjoyed the episode, please subscribe, rate, or leave a review in your podcast platform of choice. See you next time.
