Building a SaaS Company in Biotech with Abhishek Jha

Welcome to the Axial Podcast. Axial is an early-stage investment firm based in San Francisco. We partner with great founders and inventors investing in early-stage life science companies, often when they are no more than an idea. Axial is fanatical about helping the rare inventor who is compelled to build their own enduring business. We are recording, Abhishek. Thanks for doing this last minute. So I'm really excited to talk about your story, how you built Elucidata into one of these leading data companies in biotech. And then also, thanks for the tip on Up. I'm going to look into that after this conversation. Just for context, Abhishek runs Elucidata, co-founder, CEO, and Elucidata is actually used by one of my portfolio companies, Serenova Bio. And really awesome product in terms of harmonizing datasets in bio and making them useful for machine learning. So great to have you here. I'd love to maybe a brief introduction and we can go from there.

Yeah, no, Josh, thanks for having me. I keep— you keep showing up in my inbox and I'm part of a Slack community, so finally see your face and put a face to the name. And then, you know, I appreciate the opportunity that you extended to us. A little bit about, you know, my own story. So I was trained in academia, you know, the usual grad school, postdoc. I was in at UChicago and then later at MIT to do my postdoc. It was in computational biology. And for the longest time, like, you know, I aspired to be a professor in academia. And, you know, for a variety of reasons, at some point it did not make sense to pursue that. Luckily for me, I learned about a small company in Cambridge, very close to where I was at MIT, you know, that was coming up. And one of my friends who was in the lab that I was at postdoc, he was— his PI at Harvard Medical School, he was starting, or he had started a company already, and they were looking for someone with my background. And, you know, that worked out really nicely for me. That company was Argos Pharmaceuticals. And as I look back, right, I think that was a huge inflection point for me to switch from academia to industry and still, like, you know, do quite cutting-edge science, publish, and at the same time, most importantly, like, you know, do something which has a very, you know, intimate and profound impact for patients, right?. And I learned all of that at Argus. So in many ways, I feel that Argus is where I started the second phase of my career. I was there for 5.5 years, 6 years. It was incredible. Great team, great science, great mission. And one of the things that I tried to pick up and learn from my CEO at Argus was to remember that it's all about patients, you know, at the end of the day. And all the slides that you make and all the code that you write or all the funds that you raise, right, at the end of the day, it's important to remember that you're serving a patient and the family, right? And he, David Shenkin, who's now, you know, I think he's moved on to other things, Google Ventures, he would take a lot of pain and effort to remind us of that, right? So there were some physical reminders. For example, when you entered Agyas, there was a— in the foyer, in the lobby, there were pictures of, you know, friends and family members of colleagues who were lost to cancer. It's a very visceral reminder of why do you show up on Mondays, right? Why do you work late nights? And I remember in one of the company retreats, he had invited a patient. She was a peak— she had an inborn error of metabolism and She was fine, you know, she was teaching, and but her lifestyle was heavily compromised. And she was very brave. She shared a very intimate story with, you know, the entire company. Everyone was like, you know, teary-eyed. It was quite a remarkable story. And it's funny because, you know, scientists are trained to be cynical, right? But this was a very emotional— like, for an hour everyone was quiet and listening to her. And one of the things that she said at the end, which has stayed with me, is like, hey, for someone like me, like, you know, she, all she has is hope that some scientist somewhere is working on that rare disease or, you know, that hard-to-cure disease. And that's the hope that she's hanging on, right? And if, you know, she requested us, right, if you can do one more thing for one more hour, right, do an experiment, do an analysis, one more meeting, right, do that because that hour means a lot to me and my family. And, and that was quite a, you know, uh, remarkable call, right, to me and my peers, right. And I, I fully understand and appreciate that as a data scientist myself, which is what I was doing at Argus. I would suboptimally use many hours in my week, right, like doing some things which were just not, you know, the best, you know, usage of my hour. And in many ways, I think that was the inception of what I wanted to do next. I started thinking about it, and, you know, a couple of years later, I started Elucita in 2015. You know, I was trained in computational biology. My training was in statistical physics. But I had been working on biological problems of different, you know, nature in my PhD, protein folding, and in my postdoc it was computational biology. But I had to learn, you know, a lot of biochemistry when I started Argus, so I learned it on the job. Cool. And biology is very, like, you know, broad, right? There are very specific things, and you just have to be ready to embrace grunt work, right? So I remember When I got my job offer from Argus, there was a few weeks of break because they had to, you know, take some time to process my visa and whatnot. Yeah, and I remember those days because I just memorized all the biochemical biochemistry of metabolism, cell metabolism, because, you know, Argus was a metabolism company. And, um, at that time, I don't remember it anymore as well, I knew all the enzymes and their isoforms and, you know, what reaction started glycolysis and how many ATPs were consumed at these stages. It was just, you know, you had to just dig deep and be ready to embrace the details, you know, which are quite important, right? For one, they give you the language to talk to your colleagues, right, to understand. And two, like, it also educates your work a lot, right? So yeah, I learned a lot of that on the job, and, you know, I was very glad to have an environment where that was encouraged.

Cool, that's awesome. And then, and then like for Agios, you know, Agios is like one of the most successful biotech companies in the last decade. I think there's like 3 drug approvals or something.

Yes, yes.

So, so like, how did you— do you feel you got lucky to just find Agios, or did you actually like run a process as you made that transition from academia to industry? Like, did you kind of like, like, did you like compare companies and pick Agios, or Do you just feel that, oh, we just kind of came to your network and came kind of easy, it came natural to join that company?

Yeah, that's a good, great question. So, um, you know, I— it was very opportunistic. So I think, you know, I was lucky to know the right people at the right time. Um, Jason, uh, he was my friend, um, he's my friend, um, he was postdoc and his PI was starting this. So that's how I learned about it. I did not run quite a process, but there were a couple of options that I was evaluating quite seriously. One of them was with D. Shaw and one was with Argus, both equally, you know, exciting in different ways. For one, I had to move to New York. One, I had to stay back, you know, in Boston. D. Shaw, that job profile was very similar to what I had done in my PhD, and that was, you know, not quite thrilling. Argus was less familiar to me in terms of the science that they were doing, and I thought that would be a good place to, you know, spend the next few years. And in hindsight, you know, I feel, uh, you know, life does not give you a chance to do control experiments, but, um, I am glad I made that choice. So it was, you know, fairly opportunistic and lucky.

That makes sense. Yeah, I think if you joined Esha, you probably got— would have got paid a lot more money, but, uh, you probably would never have left.

Yeah, yeah, exactly.

So I don't think Elucidata would have got started.

If you— maybe not. Yeah.

So, okay, so you started Lucidita in 2015. Yes. Um, you know, what do you— can you just explain what you do first?

Uh, yeah, at Lucidita?

Yeah.

Uh, well, personally, I, uh, don't do, don't do much at this point, but, you know, um, what the company does is, uh, greatly influenced by my own, uh, experience at Argus, which was not unique by any means, you know, it was quite universal. And I have learned this over and over again through our customers, through my peers, where, you know, most of the time I would spend was just, you know, cleaning the data, right, putting it in the right format. There's a lot of scrubbing, you know, showing off your Excel skills, or writing that, you know, hacky script in R or Python to play with data frames. Excel pivoting. There's a lot of that, right, that went on. And just to give you a flavor as to how these things started, right, at Arges, often my peers, my boss, or some colleague would ask, hey, can you look up at all those experiments that we did in last 4 months where, you know, this gene was up or that metabolite was down? In plain English, that's a very straightforward ask. And, you know, to be fair, it's a very reasonable ask, right? Like, you should be able to do that because you did those experiments, you have all the information. But in practice, that was insanely hard.

Yeah.

And, uh, one was just getting access to data. Well, let's say you have access to that, then, you know, they're all in different formats, put them together, and finally, you know, you derive some insight out of it, right? And insight was the currency that, you know, everyone understood, everyone, you know, was excited about that, that That's where my glory was, right? But what was under-recognized and underappreciated is all the time and effort that went to get the data ready for that insight analysis, right, generation. And to be fair, even I did not appreciate that it took so much work, right? Because I would not even talk about it. I would not even acknowledge it. I would not even know that I could get help for clean data, right? And this is, you know, in my Argus days, but increasingly we saw that over and over again in different companies, in different stages, in academia, like, you know, labs at Yale, or in small biotech companies like, you know, Argus, in large pharma companies like, you know, Pfizer, Genentech, Novartis, you name it, where the proverbial, like, you know, the urban legend of data scientists, right, where 80% of your time is just cleaning the data and 20% is analysis. And that has shaped what we want to do at Lucidata. We want to flip that ratio where you spend 20% of your time getting access to clean data and 80% of your time on analysis and insight generation, which is where the value is, right? So we see ourselves as enablers for all the scientists, like, you know, I was at Argus and others who are again trying to, you know, bring drugs and therapies and diagnostic tools for patients. Cool.

I mean, that's like, honestly, I, I, I used to, used to do a lot of bioinformatics when I was a professional scientist. And, uh, yeah, just like you spend a lot of time on spreadsheets or different genomic files or, um, yeah, using— what was that? What was the programming?

Awk?

Awk? Yeah, awk. Yeah. Different, like you're just like probing datasets and trying to align them. It's a, yeah, it's a pain in the neck. And so, um, maybe it's backing all up a little bit.

Josh, I don't know if you, uh, have seen this, uh, sorry to interrupt you. There's a, there's a meme that's quite popular and I like this a lot. Um, in the first panel of the meme, um, is, you know, someone at the front of the room is asking Hey, who wants clean data? And you know, everyone in the room puts up their hand. And the next panel is, who wants to clean data? And you know, there's radio silence, right? No one wants to do that. That's what we saw over and over again. In my own experience, I would suck it up and do it for like 5 datasets, you know, a week, and I would not sign up for doing it for 50 datasets, right, for historical datasets, right, or new datasets that were coming out in public domain. I would not raise my hand for that, right? And I believe that at Illicita we learned that's a huge opportunity and that's where we are focused and committed to, and we have made some decent progress to, you know, address that problem.

Cool. So it's 2014-15-ish, and then you're solving these problems at Adgio's. I think Adgio's had already gone public. Yes. So that was nice. Congrats. And so What was the kind of the trigger or the spark for you to say, hey, let me go start this company? Was it like finding a co-founder? Was it customer pull? What was the kind of the spark to say, let me leave Agios, pretty risky, and just start a startup and like, you know, go all in?

Yeah, I mean, again, I'll be very honest, I think a lot of that was naivety on my part. Right. Argus was a very good gig for me, like, you know, they compensate you well, great colleagues, great science. As you said, you know, it was one of the most successful companies in the last decade. Being part of that from early days was, you know, huge. And I could have, you know, stayed there, I feel. But I also realized that I was at a different place personally, right? Um, the company was at a different place, and, um, I, I did not do any kind of, you know, market research or such, but I felt for sure that, you know, the things that I wanted to do was essentially make tools, right, for drug discovery, not necessarily make the drugs. The— myself would like be in that, that organization, right? So the agenda that I had to pursue, I think Argus would have been a great beneficiary of that, but not the best place to build those things in terms of recruiting that talent pool or, you know, having that mandate, right? It does not quite align. There was an intersection, but not quite aligned. There are a few other things in my personal life too, right? Like my wife, she was insanely supportive. She still is. And, you know, I cannot thank her enough.

That's awesome.

I think if our conversation gets to that point, right, I'll, you it's a huge ask from your family to support you in this. And she was very supportive. She had a stable gig at MIT, so that was helpful. We did not have a mortgage. We did not have any kids then. Aajius had gone public, so that was good for my public health. So I did not do any kind of Excel analysis, like, should I do this or not? I just jumped in it. And I had my co-founder who I could rely heavily on. I had, you know, deep trust in him, and I thought like, hey, let's, you know, uh, come into this and explore. And, and that's how it started.

That's awesome. Okay, so then let's, uh, I think one thing too is like, you know, you built one of the few pure-play software companies in biotech that has had success. And it's super duper hard. So maybe we can talk a lot about the challenges and opportunities to build better software in biotech. I think we talk a lot of early stage companies when they get started. How did you find your first customers? Did you just hit up agios or did you focus on a particular category initially? How did you convince them to pay you money? Was it initially services, or did you kind of get them to agree to like subscriptions or something?

Yeah, no, great question. And, uh, let me first acknowledge that, you know, it is hard, period, right? Uh, building a software company in biotech, building a software company, or building any company is hard, right? This is harder. Um, and then, you know, uh, when we started, um I think there are two options ahead of us. One option was to raise some capital and build a product. We were very intentional about being a product company on day one, that I think both me and my co-founder were very intentional about. The other option was to just get into market and have our customers fund the product development. And capital aside, right, I think We really wanted to make something that the customers actually wanted and were willing to pay for it. And none of these options are easy, or, you know, we just decided to go with the second one. And again, Agios was, you know, a great supporter where we had a customer, Agios in this case, pretty much on day one, right? And for first 3 years, we bootstrapped the company. Doing services which in hindsight, you know, where, you know, variety of services, building dashboards, building pipelines, building, you know, doing this and that, just to pay the bills. And but again, we were very intentional from day one to build a product. And we learned a lot from the market, we learned how to, you know, identify a need that was quite critical where, you know, our customers would pay us, hopefully substantial amounts of money to be worthwhile, right? Like we should in return create enough value for them to demand that kind of contract size, right? And I think that educated us quite a lot. And as we got closer to launching the product, then we started knocking on the doors of what I call friends, families, and fools. To give us some capital, right, to accelerate the growth. And that was when I think, you know, our ambitions as a company, as a group changed quite dramatically where we did a seed round led by Hyperplane. They're a Boston-based fund, early-stage company, and they have been spectacular partners for us. And then that round closed in 2018. Summer. So the first phase of the company was just like going to market, learning from the customers, getting paid, bootstrap, very lean, you know, and then build the product.

Cool. I think so. I think that's kind of— it seems to be all the case studies of the successful SaaS companies in life sciences. You know, it took a few years for them to people productize services and find product-market fit. Because why do you think— maybe one quick question is like, why do you think it takes a little bit longer to find product-market fit in biotech? Is it just the number of customers are smaller? Is it like some sort of like maybe skepticism on biotech If you go back in time, do you think Elucidate could have just launched the product and been successful, or did you need that space in the beginning to find product-market fit?

Right, again, that's a very good question, very worthy of like an hour-long conversation just on that. And I'm certainly biased and limited by my own experience. I think there's definitely a component of the market, right, where Multiple things, like I would say, you know, there's a very classic conflict of build versus buy, and you don't want to be in a situation where you're competing with your customers, right? You have to define what is complementary and where you can help them, and that's, you know, definitely part of the dynamics. And I have said this often, and I certainly believe it very deeply, that all of us are builders and buyers. And my own company at Elucidator, right, we have to continuously make a decision as to what we will buy and what we will build, right? We cannot possibly sign up for building something like Jira, a ticketing issue, right? Like, we just buy it. And, you know, all of us and our customers also have to answer that question for themselves, and that, that's definitely a part of the dynamics. The other part is also, you know, I think clarity or lack of education. For example, if I was to just go back to my Argus days. As a data scientist, I did not even know that I could get help, right, for data cleaning, something that Illustrator does now, right? And I believe there are a lot of data scientists like I was at Argus who are still like, you know, doing what I was doing at Argus, you know, manually in a non-scalable way, but, you know, which works quite well for them with their artisanal, you know, awk scripts or Perl hacks. And are, you know, pivoting commands and whatnot, melting the table. It's just that they don't even know that, hey, you know, this could be automated, or, you know, you can hire a vendor like Illustrator— there are those too— to take it off their plate so that they can focus more on, you know, high-value problems, right? So, uh, this lack of education, there's, you know, this dynamics of build versus buy. This is, you know, market size not number of companies that you can go to is not huge. But I think that's true for other sectors too, like financial, you know, industry, right? With around 5,000 banks, right? Few of them are big, huge, most of them are small. And a lot of very successful fintech companies have been built to serve that few thousand companies, right? I think biotech or life sciences at large will have that too. We are in early days of that. We feel that we are very fortunate with the timing. And you're seeing a lot of activity, right, from VCs, from young companies that are coming up like Latch, where you mentioned, or CySpot, many of them, right. We certainly see ourselves as not— we're still learning and we're still getting started, but there's definitely that and there's a a question of when will a first unicorn or $100 million revenue company, SaaS company, will come out in life sciences. And Benchling seems to be on that track. I don't know if they have already crossed it, but they have been very successful. They're a SaaS company. But I remember when I was in Cambridge, there was a group of folks from MIT and Broad There was a, you know, they would meet up every few weeks, Hammer and Nail, that was the name of the group. They would just talk about problems. And I remember John Kilty at Third Rock, he had extended that invitation to me. And one of the themes that would come up all the time is when will be there a billion-dollar software company in life sciences. If you just look at, you know, names, right, there are only a handful of them. Schrödinger is one. Benchling is another, if you, you know, from whatever I can gather from publicly available reports, you know, their market cap is greater than billion. So, you know, there has to be first one and we have a couple, right? This will change more in days to come. If you look at, you know, AWS adoption, right, that all has happened in last, you know, 10 years or less than a decade, right, in life sciences. Now it's a default. When I was at Argyus, you know, we had to go a lot of— there was a lot of back and forth in just you know, getting an AWS account, right? So things are changing and changing in the right direction. We see that, you know, data scientists are hired in the first 5, you know, team members in a small biotech company. That is not true even just a few years back, right? That roles like Chief Data Officer and Chief, you know, so things are changing, but, you know, at the same time, biotech companies are not like Netflix or tech companies. We believe, and that's one of our key thesis, that biotech companies of the future will look more like tech companies, and, and that's what we are here to enable too.

And that's pretty insightful in terms of like comparing biotech SaaS to fintech progression, um, in terms of the market structure is pretty similar in terms of like Number of companies are lower, and it's in the big ones, the small ones, and very few in between. Um, and so I think that build versus buy mixture shifting, where pharma companies, you know, we're willing to buy more tools, um, as long as they're better than what they can build, is important. Maybe back to Elucidata. So maybe we talk about the first customer, Agio, raised some money, and then you're You're in the thicket of building a SaaS company in biotech. So you're in this, you, you live this. And so could you talk about like what was the first product? Because the product you have now is pretty expansive. RNA-seq, proteomics, metabolomics, like you can ingest a lot of different datasets, harmonize them, and then enable a company. Yep. What was that first product? Was it focused on like, I'm assuming RNA-seq. But maybe it was, maybe what was the first product focused on?

So that has again been a journey and evolution of sorts for us. So our first focus was not even on the product, just on the technology, right? And the technology that we were, we are working on, and I think we have a very differentiated technology here, is has two pillars. Pillar number one is what I would call data engineering, right? Where we deal with data in, you know, let's say a dozen formats. And our goal is to bring it into one consistent tabular schema every single time, okay? And that's pillar number 1. Pillar number 2 is, which is very heavy on NLP, and we take the metadata, and metadata of 3 things, like the dataset and the rows and the columns, so features and samples. And we harmonize the metadata of all of them, right? When we have done both of those two things, we call the dataset to be ML ready, right? And, you know, just to give you a little more color, right, you would have diseases where, you know, scientists would name hepatocellular carcinoma, or next day the same scientists would call it HCC, and two days later it would be liver cancer, right? Now, as a human expert, you know that these things, these things mean the same, but the software does not, right? So certainly if you want to do some, you know, AI initiative or ML modeling, you know, you're compromised. And who is going to clean this up, right? So this is the tech that we are trying to build, and this is our pillar number 2, where we would link testicular carcinoma, HCC, and liver cancer to a particular ontology. And then, you know, when you find it, you will see all those 3 hits. Now this is true for your dataset-level metadata. It's true for your sample-level metadata. It's also true for features, like for example, you use gene symbol or Entrez ID or spot ID, and you know, again, you use it different days, different IDs, and you're compromised. So we are trying to harmonize that piece. And the first act was to go to data that is in public domain and just test our technology, right? And we did that with around 90 terabytes of biomedical data from 30 different sources. And that included, you know, NGS data, that included mass spec-based proteomics metabolomics data, that included non-omics data like flow cytometry and such, some clinical notes too. But when it came to product, I think one of the challenges was not to do a whole lot, right? So we have narrowed down the scope of the product quite massively. And again, for a stage, you know, a young company like Illiciter, the The challenge is to do less, and often the temptation— and personally, you know, for me being a scientist, right, trained as a scientist, I always go back to, you know, signing up for this and that, right? That's not quite helpful, and that's, you know, a lesson for me to, you know, remain focused, keep my company focused. So the product question that you had, you know, we are focusing just on two data types, uh, RNA-seq bulk and single cell, and we are taking data from public domain. And making it, you know, findable and usable. Right. That's the— anything else, you know, we enable that by providing our services. And that has, you know, been very helpful for us to get, you know, quite an exciting traction in the industry. So 4 of the top 10 pharma companies are, you know, working with us, and a couple of dozen, you know, smaller biotech companies are leveraging our product and services to accelerate the discovery programs.

Cool. Hey, Serenova, definitely Leon loves you guys. So, you know, Elucidata really sped up a lot of Serenova's, uh, you know, discovery work. So I appreciate that.

I think if only you can ask him to write a testimonial for us, because, you know, we always run into this legal contracting where our customers, our champions, are very keen to give us, you know, uh, but you know, we don't—

okay, I'll talk to Leon this Friday. I talked to him every Friday, so I'm gonna bring it up with him and say hey to Aaron and the whole team on Friday. But, uh, cool. Okay, um, uh, one thing about Lucidata, the comp I make to your company is like Fivetran. Um, yeah. Familiar with Fivetran?

I'm quite familiar with Fivetran. I mean, well, as much as their website would tell me, you know.

You know, actually, the CEO of Fivetran, he's a PhD neuroscientist, uh, just, uh, just on a side note. But, uh, he, you know, for context, Fivetran is an enterprise software company, more for more general, focused on integrating datasets from like CRMs and a bunch of other sources and harmonizing them. Yes. And it's actually pretty hard. I think that company was founded like in 2012.

It did happen. Yeah.

And then it took them like 5 years to get successful. Yeah. And only in the last 2 years have they become like really successful.

Yeah.

So it was a long, long journey. And I, I just remember like they have these really cool stories of like how difficult it was to like make sure to— they, they delivered data from CRM that was exactly the same. Yeah, for Lucid data, do you have any like funny stories like that when you were trying to be— when you were structuring all that public data? Was there like little issues that would like— that were a pain in the neck for you?

Oh man, tons of story. It's funny that you bring up Five Ten because I, I go on their website almost every week, uh, just to check them out. And you know, uh, and it's funny because, you know, I've shied away from using that we are Fivetran of biomedical data because most of my audience does not, do not know about Fivetran. But we, as much as we can gather, you know, we love what they're doing and we see ourselves like, you know, playing a very similar role for our customers. And in some ways, you know, if you think about it, it's the challenges are more nuanced and more specific to the vertical that we are in. Where, you know, okay, if you're an enterprise, you have a CRM and more or less all your customer data is in that CRM, right? Salesforce or Pipedrive or HubSpot, whatever that may be. And guess what? Those CRMs have some programmatic access, right? But if you go back to around a dozen resources that are in public domain that our customers use, Geo or TCGA or something released by Broad. All of them have, if you're lucky, some of them will have, you know, programmatic access. The interface of that API is very different, can change anytime, right? So in some sense, ingesting the data at that scale is massively harder, right? We had to, like, you know, first I'll give you some numbers. We have around 40,000 datasets from GEO, RNA-seq datasets that are harmonized, cleaned up on Poly. And there are around 20-25 datasets that are being added on GEO every single day. And we have a real-time, you know, update where we pull the data. And now these datasets are processed differently, and we had to standardize that. Now processing you know, 20 RNA-seq datasets daily, which, you know, roughly has around, you know, let's say 400 samples, is a non-trivial task. And if you were to do that at scale that we have done, you know, it's highly non-trivial where we had to build a computational infrastructure for our own sake, process all of them through one single pipeline. For single cell, it's even harder. So we have published some of the blogs, we'll publish more, but the point I'm trying to make is that, you know, the heterogeneity and the domain expertise that you need for single-cell bulk RNA-seq is very nuanced, and that makes it challenging, right, to build something like this. But in an abstract sense, we are trying to do, you know, at least we think we are trying to do something very similar to Fivetran. And of course, these comparisons are always tricky because there are always nuances that, you know, Fivetran is dealing with which we are not, and we are dealing with and they are not. But, um, yeah, that's a great parallel.

Yeah, I think, yeah, Fivetran has to deal with like so many different like inputs. Yeah, it seems like Lucidata, not only that, but you also have to deal with like so much variability in like data quality and like I'm assuming I just know for a fact, like an RNA-seq package from one institution is definitely not the same as RNA-seq or a dataset from another institution. It's a lot of, especially on the academic side, like much variability in terms of quality annotation. So you have to kind of, I'm assuming that's especially, it took me a few years probably even to like, you know, make a dent there and do that for 90 terabytes. That's insane. So maybe we can touch upon the other pillar of Elucidata, which is the NLP side, the ML side. I have a friend, Jacob Oppenheim. We actually did a podcast with him last year, and he had this really cool framework of for biology, you need to train your models to do unphysical things. Yep. Which means like you need to be able to give them datasets that are helpful for them to know what to do, what not to do. And sometimes biologists only do the experiments that are positive.

Yeah.

For Lucidata and maybe in general, like why is data structuring so hard for like application, diagnostics, and other things.

Yeah, no, so the two threads here that I hear, one is just the retraining the, you know, large language models or NLP models. And, you know, this is a very timely conversation. There's so many LLMs that are coming out, large language models, and our thesis is that Well, one, it's hard to make an LLM. It's very expensive. It requires a lot of expertise. It requires a lot of time. And I think what we have seen, even if you put all those aside, right, let's say we had the time and the money to do that. I think what we are seeing is that there's a lot of value in retraining that, and that gives you bigger bang for buck, you know, and that's how we have approached this problem. So More specifically, we have taken BERT, which was published by Google a few years back, and we have retrained that on experimental designs for RNA-seq, right, where we extracted all that information from PubMed, and we retrained BERT, and we call that PolyBERT. And that's where we have, you know, created value, right? So now PolyBERT outperforms BERT and other models when it comes to very specific tasks like recognizing entities like cell type, cell line, disease, regression of disease or survival. These are very specific fields, entities, that we do much better, almost as well as human curators themselves. We definitely outperform a more horizontal, flat tool like BERT. And the magic was in retraining, right? Continuously retraining. Now, the second piece is we also don't see human curators as being replaced. We see them as being enabled by Polybird. So we have, you know, around 40 curators who constantly use our curation platform to curate these datasets from public domain or, you know, from your proprietary sources at the 3 levels that I mentioned, right, the dataset, sample, and feature. So these 2 have been like lessons for us where we retrain the model, not build a new one, and and keep humans in the loop, right? Which we call, and others have called this too, right, active learning. And as human curators, you know, curate data, the models get better and smarter. With these two things, like, I think we have been able to hit this kind of scale of, you know, pushing 90 terabytes over the last, you know, couple of years, which would not have been possible otherwise.

Interesting. And so, like, On the— hmm. So what I'm hearing is that there's an interplay between training models and then— maybe the better question is this. How do you think data structuring keeps up with model progression? Because it seems like models are always updated every year. And then the underlying datasets, it's at least from my perspective, it seems like the models are progressing much faster than our ability to structure the data, especially in biotech. How do you think— and I think Elucid is playing a role here— how do you make sure that it has parity to making it easy to structure data to make better models? Because it seems like knowing the AI field just a little, there's always new models coming out. I know LLMs are really the hot trend now. You know, a long time ago it was like, you know, TensorFlow and other types of neural networks. So how do you think data structuring keeps up with all this progress on the AI side?

Yeah, so one thing that we have, you know, seen this over and over again, and others have talked about this too, right? Where, you know, we have to almost get on the same page where having access to clean and linked data is, you know, central to predictive ML, right? And there's a famous quote from Vas Narasimhan, who's the CEO of Novartis, that we use a lot in our slides where he says that, you know, often organizations underestimate what it takes to clean and link data, right? And they start an AI initiative and 6 months later, 9 months later, they realize that, oh, some very highly trained ML modelers, ML scientists are, you know, cleaning data. If you take, you know, more successful and public story of AlphaFold, right, I think the kind of success that they have had would not have been possible if RCSB did not have, you know, PDB files from all the labs over many years as structured as it is, right. It's not perfect, but it's pretty good. And a lot of, you know, stakeholders like those labs and NIH came together to make that resource available to the scientific community at large, right? So the value of having structured data is essential, and we have to get on the same page, right? And we also have to acknowledge that we don't often understand and appreciate the kind of expertise and time it takes to do that, right? It's very— again, if I go back to my own experience, Rajesh, right? Very savvy scientists very, you know, deep domain experts, but they almost, you know, trivially ask for like, hey, give me this 40 datasets, right? And that's not an easy task. You're not quite set up to do that, right? If you had the right tools and infrastructure, yes, then it becomes, you know, almost like a snap of finger and you have that. But that's not where we live. That's not what the reality is. And, um, uh, you know, even if you look at the you know, all the excitement that ChatGPT has generated, right? I believe there's a lot of work that must have happened to make those prompts, right? Clean them up or put them in a particular form and such as to educate that model, right? So as an industry, like, and as a community at large, we have to acknowledge and resource and, you know, invest in making sure that the infrastructure is there where, you know, your AI initiatives will succeed. Otherwise, you know, we'll be running a big risk, right?

Cool. One quick question from your purview. What's like, what are some of the fastest growing datasets in bio? Is it proteomics? What are you kind of seeing from your customer base? Like, is it, are like people producing more and more protein sequencing data, or is it just single-cell sequencing? Just like, you know, keep on, is it like, you know, is that growth just still torrid?

Yeah, I think the usual culprits, right, genomic data is huge, but often they're not quite available. You know, they're often behind some kind of regulatory— for good reasons. Transcriptomic is another one, bulk RNA-seq, that's growing quite rapidly. Single cell has generated a lot of interest, it's quite, you know, very eagerly sought after. Then you have, you know, a bunch of mass spec-based omics like proteomics, metabolomics, lipidomics. But there are some inherent challenges with mass spec, where, you know, it's not as quote-unquote well interpreted by a broader community, like, you know, let's say transcriptomic data is. So there's, you know, some challenges there. But these are the data types that we hear a lot. And there's, you know, tons of data with not as much, you know, size, But a lot of velocity, right, that comes outside of this space, like flow cytometry dataset is being used in discovery space, in manufacturing, and a lot of, you know, really powerful questions can be asked, but it's compromised because the data is not, you know, quite clean and linked. And again, we see us playing a role in non-omics data as well as we move forward.

Oh yes, this would be the 5th trend for bio. Yeah, yeah. 5th trend took, you know, to be honest, 5th trend took like— what, 7 years to get really, really successful. So yeah, 2 years from now, I think Lucid Data is going to be a billion-dollar SaaS company, but it's on the 5-transaction trajectory. But to kind of wrap up the conversation, any final lessons, or at least at this point in time, any lessons you've kind of— I think throughout the conversation, we've kind of alluded to a few, but any kind of lessons for maybe aspiring entrepreneurs or people who are trying to make that jump to industry?

Yeah, so, you know, I, I, uh, I do not even realize, but it's been 7 years since I've been doing this. And as I talk to, uh, new CEOs who are thinking about starting a company or have started, I suddenly feel like, you know, I'm on the other side where I have things to share with them, which is funny to me because I, I believe we have just started and we are still learning. To me, um, you know, It's the best self-improvement project that I could have signed up for. You cannot hide, you have to face conversations and challenges and issues. It's a constant learning and it's a huge privilege. I'm very lucky to have a support system of my peers, my family, my team, my investors, and my customers to let me do what I'm doing. It's a huge privilege, right? But there's also, you know, a huge cost to it. It's, you know, it can be quite overwhelming. Your family, you know, signs up for this ride. They don't get to experience the highs that you do, but they do get to experience the lows that you do. Because, you know, you need someone to go and talk to, and those things take it all, right? So anyone who is thinking of doing this, I would suggest Do it. This would be the best job that you would, you know, do in your life. But also make sure that you have a support system, right, from your family, from your team, customers, investors, because every bit helps, right? And you need a lot of support. It's, it's a tough gig, but again, it's a huge privilege to do what I am doing. I'm very thankful.

That's awesome. Well, thanks for doing this. I think it's gonna be really useful. I know it's gonna be really useful., and we'll do this like 2 years from now. But just, I really appreciate, uh, kind of taking the time to talk to us. And, uh, yeah, we'll talk soon.

Yeah, Josh, thank you again. Uh, this has, uh, been a wonderful conversation, and, um, uh, I hope this can help someone. But I would love to connect with you again in 2 years and see where we are.

It would be a great— gonna be super successful. Honestly, Fivetran took 7 years to get, yeah, moderately successful. And then the flywheel just became huge around like 2020. Yeah, yeah. So like, I think Fivetran is gonna be worth like tens of billions. So yes, I think Elucidata is in good company there.

Yes, indeed. Yeah, I'd love to, uh, learn more from them, compare notes.

Um, next time you ask, let me know. I have a personal connection to them. And maybe dinner or something.

Yeah, I'll take you up on that.

Cool, man. Hit me up and we'll take a walk and get some dinner. But I appreciate it.

Awesome. Thanks, Josh.