DNA Computing, Search, and Data Storage with David Turek, CTO of CATALOG

(Intro Narrator at 00:00:04) Hello, my friends. Today, Joel is talking to David, the CTO of Catalog, and they discuss David's journey from running the supercomputing division of IBM to CTO at a startup working on DNA computers, use cases for DNA data storage, search, and computing, and insight into what it looks like in practice to store and manipulate data using chemistry. All of this right here, right now, on the Modern CTO Podcast.

(Joel Beasley at 00:00:41) Here we go. This is the Modern CTO Podcast. So I want to give, you know, all of what we do. We hang out. We talk.

(Joel Beasley at 00:00:58) You know all of that. So we had a great first episode where we talked a lot about DNA storage. But real quick, just to set the tone, can you give a little bit of your background and how you got involved with Catalog and what it is?

(David at 00:01:13) Sure. From a background perspective, most of my career is spent working with IBM. I was managing the supercomputer business for IBM for the last 25 years or so. So very much high-end scientific computing, leading edge, cutting edge kind of stuff. A lot of work with IBM Research to take research ideas and turn them into product.

(David at 00:01:37) About a year and a half ago, I joined Catalog because I saw a new opportunity to really help transform the computing industry by taking this so-called blank sheet of paper that occupies what goes on in chemical computing, and through Catalog, really create a different model for how people can approach problems associated with data, search, compute, and so on.

(Joel Beasley at 00:02:02) And then how did you meet the people at Catalog?

(David at 00:02:05) Well, actually, they were looking for a CTO, and so a recruiter reached out to me and we started having conversations, and then the rest is history.

(Joel Beasley at 00:02:15) Nice. So you have a pretty good risk profile, right? You saw that this was an opportunity, and you saw where the market was going, and you said, "I want to be a part of this."

(David at 00:02:24) When I think about it from that perspective, I think of it as the adrenaline rush from working in an integrative domain and doing something that has a material impact on the future of computing. So from that perspective, it was a pretty easy call on my part to engage on that.

(Joel Beasley at 00:02:45) Nice. Well, man, I'm excited. I firmly believe this is the future. I'm pumped. I've been talking with everybody about, I think you said it was called "write once, read never," one of the use cases for DNA storage.

(Joel Beasley at 00:02:59) And then my producer told me something. You guys were working on DNA computing. And at first my mind was blown, but then later I was out on a walk and I thought, you know what? We talked about being able to search DNA, to be able to do all of these different things with DNA. And so computing would be a natural progression to move up from storage.

(Joel Beasley at 00:03:20) I don't understand it, though. How do you differentiate computing and storage with DNA?

(David at 00:03:27) Okay. So we'll define a couple of terms here with respect to information first of all. So imagine, if you will, the encoding of a whole bunch of data that you have available to you in DNA, and it sits, in our case, in a pool or somewhere. So it's in liquid, just molecules floating around representing all the data that you've encoded. If you want to search on data, all you're doing is retrieving what you already have.

(David at 00:03:55) If you want to compute on data, you're applying some sort of transformational idea to the data you have to create new data. So simple example: You encode the numbers 1, 2, 3, and 4. You can search and you can pull out the numbers 1, 2, 3, and 4. But now I apply a transformation to this data called addition. So it's 1 plus 2 plus 3 plus 4.

(David at 00:04:20) And now if I'm doing my addition correctly, I can now withdraw the number 10 or create the number 10, which didn't previously exist in my data store, and I can now add it to my pool. I've created a new piece of information. That information is labeled 10. So that's a trivial example of what we mean by compute, separate from search. So what we're doing is both. We have the means chemically to search for data out of encoded DNA. But we also have—we're beginning to develop chemistry and, crudely put, chemical instructions, if you will, that can operate on data and produce some result as well.

(David at 00:04:59) So a trivial example would be: I've encoded a whole bunch of numerical data into DNA. Question: What's the biggest number I have encoded? Answer: We've developed chemistry that can go and search through all that and identify what the biggest number is. So that's a trivial example, but it's an example of the direction that we're taking. Now, "instruction" is a bit of a stretch here.

(David at 00:05:25) When we work in the chemistry world, we use language that's evocative of what we're all familiar with in the electronic world, except the words have slightly different meaning. So we don't have compilers. We don't have languages. File systems don't exactly look the same. You have DNA just floating around in a liquid.

(David at 00:05:46) They're not tied to a physical structure like you have with tape or hard disk or even flash memory, things like that. But we use that language because it's the easiest way for people to get a grasp on what we're doing. So in essence, yes, we're demonstrating the ability to operate on data. We're doing it entirely with chemistry. Our media for facilitating this is DNA, and we're developing automated tools to do this at scale.

(David at 00:06:17) So this is beyond benchtop kind of experimentation, but to do it at scale on problems of conventional size. And what I mean by that is not a toy problem. Toy problems have been referenced in the literature. You'll find people do it in universities and so on. Now we want to solve a problem using the amount of data that you would come in contact with in the real world and to solve it using DNA instructions, if you will, in a timeframe that resonates with the needs in the real world.

(David at 00:06:47) And that's what we're doing.

(Joel Beasley at 00:06:50) I'm a super visual person, and so there's a huge knowledge gap. I'm not an expert in this topic by any means. What is the—I guess the best question I could come up with is, what's the interface? So back to your example, 1 plus 2 plus 3 plus 4. What does it look like to input it, and what does it look like when you get the result back out?

(David at 00:07:08) Okay. So from a chemistry perspective, we would create a chemical process, if you will. It might be multiple steps in nature, but the chemical process would be applied against this body of DNA that encompasses the data that we've encoded. So think of it this way. Think of one beaker consisting of all this data embodied in DNA, and another much, much smaller beaker that also has fluid in it and some so-called DNA instructions in it.

(David at 00:07:39) We would take that second beaker, dump it into the first one, and then we would invoke a chemical process that would cause the removal of the right kinds of results from that body of information to produce the answer. Now in practice, that's the way you would do it on the bench with just chemists. In practice, in our case, we would use software, conventional software, to orchestrate the operation of our data encoding machine called Shannon to produce a particular set of DNA molecules that would then find its way into a pooler that exists at the tail end of the writing process that would cause this chemical reaction to take place. Then we would reduce the volume of liquid in the pooler. We would extract the DNA molecules in a very conventional way that would represent the answer, and we would then convert that answer, if we wanted to, in electronic form so that you get a readout on a piece of paper, on a disk drive, on a tape, or something like that.

(David at 00:08:47) But all the computational activity, all the manipulation, would take place in a world of chemistry, not in the world of electronics.

(Joel Beasley at 00:08:54) Okay. So I'm going to tell you one of my perspectives of it, and you can tell me how I'm wrong or right or if I'm on the right path.

(David at 00:09:02) Sure.

(Joel Beasley at 00:09:03) I've been noticing emergence of very specific, sort of like GPUs or specific chips and processing. For example, there's this company Groq and they make this very specific—and so you're seeing these application-specific, you know, the quantum computing type devices. And it's almost as if the process is creating these little application-specific organic things, but it's just in a much faster cycle. You're not actually manufacturing this whole chip and doing production, sending it out. It's like you're making these little physical things that can then process the data, like filter through the data?

(David at 00:09:47) Right. So DNA is physical, right? Just as, you know, 60, 70 years ago, when you looked at computer storage, you have little pieces of magnetic material, and you flip a unit from one side to another magnetically to represent an action or a piece of data or something like that. All we're doing is manipulating DNA in the same kind of way.

(David at 00:10:11) And this is taking place in a physical sense. It's just that the physical media are these molecules of DNA. Now, with respect to your comment about the specialization, that's actually an important observation you've made, because over the last 10 years there has been, in the computer world, the observation that you look at workflows and workflows are capable of being decomposed into different pieces. So for example, there might be a phase in the workflow that's data manipulation, then there's the application of an algorithm, then there's an application of more data manipulation, the invocation of another algorithm, visualization, result, et cetera. So the compartmentalization of these activities has given rise to an interest in the deployment of specialized kinds of devices, like GPUs as you pointed out.

(David at 00:11:04) And now people are doing acceleration by chaining together Arm chips and this and that. But the ideas are fundamentally the same regardless of the technology. It's the decomposition of the workflow that's critical. Another way to say it would be the following: You're not going to use a quantum computer to do payroll application, because the nature of computation and so on that goes on in payroll is perfectly fine for a classic Von Neumann kind of architecture that is embodied in your laptop PC or what have you.

(David at 00:11:35) There's no advantage that will accrue to you by virtue of the invocation of a quantum capability. On the other hand, there are other applications—encryption, decryption, things like that—where Von Neumann architectures are constrained in terms of their utility, and then you invoke a quantum computer. It's the same with GPUs. GPUs are an amalgamation of vector processing units. So if you can find a domain in your workflow that invokes that, and certainly visualization was that for a long time, but now people are using GPUs for very heavily quantitative algorithms, the opportunity to invoke that gives you a real advantage.

(David at 00:12:18) It speeds things up quite dramatically. But it speeds it up quite dramatically in a context of a workflow, and, you know, we've talked about Amdahl's law for many, many years in the world of scientific computing. You can take an application and you can render parts of it parallel and invoke a supercomputer, but the overall time spent in the application is still going to be capped by the parts that you can't render in parallel. So if it's some serial I/O process, for example, well, that's it. I mean, you're never going to get faster than the fastest you can make that particular step run.

(David at 00:12:58) Now, this has been going on for 10 years, more or less. And what it's done for us in the DNA world is it's taken away the intrinsic resistance on the part of clients and customers to experiment with new and emerging technology. Ten years ago, you'd have not seen anybody say, "Well, what I want to do today is I want to implement GPUs, and I want to implement quantum, and I want to do all these other things." No. Back then, 10 years ago, people would say, "Moore's law governs all. I'm just going to use commodity processors, and that's it." Well, what's happened? Well, Moore's law has been shown to be dramatically constrained, and there's this gradual revelation in the minds of a lot of people that if you focus on workflows and understand how you can apply the best tool or technique to different elements in the workflow, in aggregate you'll get a much better result. So the removal of that reluctance to embrace new technologies pioneered by the GPU and quantum guys bodes well for the chemistry guys, right?

(David at 00:14:03) So now people would say, "Well, why not look at DNA computing? Oh, it's exotic." No, no. Exotic went out the window when you guys started experimenting with GPUs and quantum computers and so on. You've already embraced exotic, right? So this is just another form. So why not look at this? So that's benefited the DNA community quite dramatically.

(Joel Beasley at 00:14:26) Do you have any good examples of how people could use this or specific industries?

(David at 00:14:31) Oh, sure. And Catalog in particular is engaged with a number of commercial enterprises to both showcase and demonstrate the technology, but also to collaborate with them on particular applications. So it falls into three categories. It's DNA as storage per se. So there are some companies—think of companies that are engaged in long-term archival kinds of needs, maybe for historical purposes or so on.

(David at 00:15:01) And they want to keep things around for 100 years or 200 years or whatever the case might be. And maybe it's art, maybe it's film, maybe it's music, doesn't matter what the case is. But people have a need to just preserve things effectively forever. And the longevity of DNA supports that. The density of storage encoding makes it reasonable and cheap.

(David at 00:15:23) You don't have to build a warehouse. You know, a single room might handle it. Low energy profile might support what your need is, et cetera, et cetera. And if you need to keep something for a year, two years, 100 years, 1,000 years, there's also no obsolescence. So, you know, it's different than what we have today where you go from LTO-6 to LTO-9 and tape, and suddenly you can't read a cartridge with a new technology, and you've got to do all sorts of updates on what you have installed.

(David at 00:15:55) DNA molecules are DNA molecules. As long as we preserve the means by which you decode it, which we can do also in DNA, you're fine. You know, 1,000 years from now you'll say, "Oh, it's a DNA molecule. Here are the decoding instructions. Let me decode it. I see what we have here. Thank you very much," right? No problem with that. So that's category A.

(David at 00:16:16) Category B is search. It says I've encoded data. Now I need to find it. Well, imagine, if you will, a world where all data for archival is on tape, and you're looking to find one little piece of data out of all your archival tape. The way tape works is you process everything serially until you find what you're looking for.

(David at 00:16:38) One of the advantages of DNA storage is serial processing goes out the window. I just create a probe of what I want, I throw it into my massive data, and I pull out exactly what I'm looking for. I don't have to examine every piece of data. I can simply extract the one piece of data that I'm looking for. So it's non-serial in nature.

(David at 00:16:57) Random access is, you know, the standard that we have in the DNA world. So from a search perspective, it holds a promise of searching data in whatever form that takes, by the way, you know, compound searches, whatever, being able to do that very directly, very efficiently. And then the third category is transformation, where you're actually manipulating data quantitatively. And here what we're looking at are applications like digital signal processing. There's some evidence that we can make some progress in training models and machine learning.

(David at 00:17:32) There are examples that we're working on where it's sort of a combination of search and compute and what we call an identification problem. Is the person that just walked into the airport a terrorist or not a terrorist? There's information I have to search. There's information I have to process. Can I do that as well?

(David at 00:17:53) And DNA is particularly good at looking at things that don't require extraordinary precision. So, for example, a photo. Do I need to be absolutely precise with every pixel that represents your image? No. I can allow some things to fall out.

(David at 00:18:07) Not a big deal. I can extrapolate the differences, and I have a clear identification of who you are. But digital signal processing, for example, is a precursor to how you would do transformational mathematics. It's the kind of thing that you would see, for example, in seismic processing. So now I have a seismic database file, and I operate on it using seismic processing techniques, which would incorporate these kinds of approaches.

(David at 00:18:35) And it's a different way to do analysis on that data, contrary to what you would do in a conventional computing sense. Why would that be advantageous in seismic processing? Well, now maybe I can put a substantial amount of computing onboard a ship. I don't have to wait for that ship to dock, transport data to a data center, and so on. I can do it essentially in near real time, right, at low power.

(David at 00:19:02) Because in a DNA world, I might be able to amass not hundreds or thousands of computing units, but maybe trillions of computing units where every unique DNA molecule that I create actually has some computational capability or in combination has tremendous amounts of computational capability. So scale, parallelism, low energy, low footprint—all those things conspire to help me deploy computing into the field, if you will, as opposed to having to go back to a brick and mortar classic data center. And it may not have to do complete processing. Maybe it just does some filtering on the data, capturing the field before I have to send the residue back to the data center. But anything that can speed up the total time spent in the workflow is going to be advantageous.

(Joel Beasley at 00:19:53) Can the models learn in the sense that, you know, you produce these chemicals or the structure to maybe filter the specific data or have this specific algorithm? Is it fixed in the sense that there's some algorithms or processing people will do, and then they'll put it on a local device, like in IoT, so that it's really close to where the data is actually happening, but they have to just keep updating that one? It doesn't really change or store or do anything. It's just this specific algorithm. Can they change after you make them?

(Joel Beasley at 00:20:25) Can you write them to change after you make them and learn?

(David at 00:20:29) We think so in the same way that a classic deployment of the machine learning algorithm will learn depending on the data that it's exposed to in the wild. So what do you do in classic machine learning and deep learning? You train a model with historical data, and then you deploy a model into the wild. But that doesn't mean the only data that model ever sees is the same as what it was trained on initially. So it has to evolve over the course of time as it begins to observe new phenomenon that was outside the realm of what it saw initially.

(David at 00:21:03) We think that can be echoed in some sense with DNA as well. We're hopeful that the low amount of energy that we require and the density of footprint will eventually have some advantage in IoT kinds of devices. Our approach to DNA computing is not going to have utility in every context that one could imagine as well. Right? And so we're in the process as we explore these initial forays into compute.

(David at 00:21:31) We're looking at very specific domains of application to make sure that we don't inadvertently create some hype that never gets realized. It's just like the quantum stuff. Right? The quantum stuff was motivated by one application, and now people are spending time looking at other kinds of applications. But they know a priori—again, as I said before, you're not going to deploy quantum to displace all Von Neumann kinds of architectures. You know, it's not going to provide you a material advantage to try to do payroll with quantum. So we have to find the boundaries and the limits here to make sure the promise of DNA computing is not oversold. Now from my perspective, the way you do that is you look at language and make a determination of how informative language is to the plans and strategies you create. So I'll give you an example.

(David at 00:22:26) People say storage market. That phrase has zero value to me the way people use it. Storage market is, you know, pick a number, $50 billion a year or something like that. Our interest is how you look at behaviors within that gross market that provides insight to us that are discernible, observable to us, where we can actually provide some real value. So for example, we don't want to do hot storage, the stuff that you need access to instantaneously.

(David at 00:23:00) Because in the world of chemistry, you have a latency problem that you can't readily overcome. Similarly, when you look at archival, you don't look at it and say, "Oh, we can solve all problems in archival." That's probably also not true. So you have to burrow down under these gross appellations and find those domains of applicability where you can demonstrate real concrete value. And that's why we put a focus on trying to flesh out these value parameters that govern DNA, and it has to do with energy, density, this notion of preservability, longevity, the lack of obsolescence.

(David at 00:23:42) All those kinds of factors and how they come to play begin to filter the marketplace in the general use to give you those categories where there's some real utility in the short and intermediate term. I will give you a concrete example. In Europe right now, there's an effort underway, corollary to what was done with the seed repository in the Arctic Circle to preserve seeds in case catastrophe occurs. You know, where would you get the seeds to restart agriculture and plant life and things like that? There's a similar kind of activity going on right now where they're encoding music into tape to put into this repository north of the Arctic Circle for the preservation of music should some catastrophe occur.

(David at 00:24:32) That could also be done on DNA. Right? And maybe it could be done in a way that's cheaper, more effective, more preservable over thousands of years or tens of thousands of years or whatever the time dimension is. So people are beginning to entertain how they can begin to take technology and carve out domains of applicability like this music example today, which are kind of the way we think about things as well in CATALOG. We stay away from the grandiose pronouncements of markets and so on.

(David at 00:25:07) And with our collaborators in industry and commerce, we try to auger into the nature of the way they're running their business today, how they expect to run it in the future, and see what is it about DNA that could be helpful to their future strategies.

(Joel Beasley at 00:25:25) I've got so many questions. I fully get the hype thing, and I'm acting like a small child trying to find the edges of where this is going to be useful so I can better understand it. It seems to me—well, I think everyone would agree that there's just more and more data all the time. And as you get more and more data, like, let's say we were monitoring the entire seismic activity of the whole Earth, and it was generating, I don't know, exabytes a second of real time seismic data. And then you wanted to somehow filter that.

(Joel Beasley at 00:25:58) Right? It sounds like this DNA computing would be a good way to do that. My question, one question I have—well, first of all, I'll ask that. Is that a good example? You have a tremendous amount of data that you need to filter, and this would be a good application for DNA computing?

(David at 00:26:15) The answer is yes and no. If you could affordably encode all that data into DNA, then filtering it with DNA capabilities would absolutely be a brilliant way to do it because the computational behavior of DNA is such that it's unaffected by the size of data. So in other words, the amount of time it takes you to run an application, a DNA application on a megabyte of data is not going to be discernibly different than the amount of time it takes you to run on a gigabyte or an exabyte of data. The part that you have to worry about is, what does it cost you to encode all this data into DNA? How much chemistry is involved, et cetera, et cetera, and how long would this take?

(David at 00:26:58) And I think that that's a long term ambition to the DNA storage community. But the fact of the matter is most computing applications today do not worry about operating on the totality of the world's data. So it's very intriguing to say, "Well, you know, the world has yottabytes of data or zettabytes of data, and so this will be a panacea to that." Most work is done on megabytes of data in the world. You know, an MP3 file for music is about three megabytes.

(David at 00:27:27) Three megabytes. Not three exabytes, not three yottabytes or anything like that. And the question is, can you run applications against three megabytes of data with DNA that would be useful? And the answer is yes, you can. And there's a lot more application activity going on on data in that relative domain of quantity than there is people operating on exabytes or even petabytes of data today. So there's sort of this segmentation in the marketplace that people need to come to grips with in terms of the predicate of data volume as being the motivator for why you do things with DNA. And sort of think about that in terms of time horizon of when DNA will be able to operate on megabytes and gigabytes and terabytes and petabytes and exabytes, et cetera, before everybody gets wrapped around the axle saying, "It's the panacea to the world of yottabytes." No, no.

(David at 00:28:25) No. Don't worry about that. That'll take care of itself down the road. What we need to worry about in the short term is how do we manage megabytes, gigabytes, terabytes, and so on. And so that's the way we think about this and how we focus on it.

(Joel Beasley at 00:28:41) Now when you gave the example of maybe, you know, some historical storage, or maybe if I'm a company and I have you come in and instead of using all of these tape backups, you come in and you write a bunch of data to me or data for me. How do you actually physically give that data to me? Is it—I'm assuming it's not like in a bag that says, like, "data." What is it in?

(David at 00:29:06) It's in a bag with liquid and a goldfish, and it's yours. Well, it could be that way, but we'd probably give you a tube, or we would give you a tube encapsulated in a physical device for preservation. There are companies, by the way, that are beginning to emerge in the DNA industry that are focused on exactly the containers that you would put DNA in for transport, preservation and so on. Not our business. It's the business of other companies.

(David at 00:29:37) So that's one way to do it. That's if we were preserving it in its sort of liquid state ready for manipulation in some way. We could also dry it out, desiccate it, and we could actually deposit it on a piece of paper, if you will, something like that. And you could reanimate it, if you will, with the right kinds of chemistry and so on. Or we could do some post processing for you, and we could actually render all that stuff into some digital form if you wanted to preserve it in some digital form.

(David at 00:30:09) All those options are available to you. But most people, I think, will first—for the sake of density—because, by the way, once I transform it back to digital form, you're back into, "Let me build a warehouse to hold all my data." So I think most people will have it contained in some sort of liquid, semi-liquid, or desiccated form, and they'll manage their volumetric amount of data in that fashion. So, you know, I can take all the data in the world. I can put it in the bread box, if you will.

(David at 00:30:39) And so that's pretty cool. And, you know, you can keep your bread box in your kitchen, and that'll be all the data in the world. Be careful what you do with that bread box.

(Joel Beasley at 00:30:50) I know. When, like—are there any pieces of DNA, sorry, in museums or like a tube of it? Like, can—how can somebody go see it?

(David at 00:31:07) Well, DNA is not visible to the naked eye. So what you would see is you would see a tube of clear liquid, and somebody would say there's DNA inside. That's why our goldfish is named DNA. So when we put that in your bag, you can say DNA's inside and actually be truthful. So you really can't see it that way.

(David at 00:31:25) And we have, by the way, for other purposes, we actually have desiccated DNA and put it on a piece of paper. But then all you have is a piece of paper. So the molecules are too small to be observable to the naked eye. And that's the intrinsic advantage of DNA in terms of really dealing with the volumetric problem of how you store data. So you have to sort of trust us in a certain sense that, yeah, there's DNA in this test tube.

(Joel Beasley at 00:31:53) It's a good business idea. I'm going to start selling DNA on my website. Just—I love the fish thing. Yeah. I love the fish thing because—yeah. Alright. Last time we talked, I may be wrong, but I think you had said something along the lines of, like, when you're processing this data or you're looking for a specific result, you can change chemistry things like maybe a color, like have it come out blue or red. Was that something true? Okay.

(Joel Beasley at 00:32:21) So I would say that is like a very basic visualization of data that you wouldn't actually need any sort of computers to see. Right? You would just program it to be that way. You would mix it, and you would see the result red or blue in the vial. Am I on the right track?

(David at 00:32:37) Not quite. Okay. The idea of fluorescence or colors is still taking place at the molecular level, so you still wouldn't see it with the naked eye. You'd have to use scientific instruments that could detect it. And the instrument itself, by virtue of detecting it, could draw a conclusion about what it's seeing.

(David at 00:32:57) And that conclusion might be about the content of the data that's being stored, or it might be something with respect to the computation that's taking place that suddenly a color has changed. And so, therefore, you know, it's gone from positive to negative or whatever the case might be. We're doing a lot of experiments with fluorescence and colors and so on to find ways to more efficiently decode DNA molecules.

(Joel Beasley at 00:33:25) So there's nothing that I could—you know, I have you have the data over here, like the source data, and then you have the tube with the algorithm or whatnot in it. The—what do you call them? What do you call them when you put it in there? Like a program? I don't know what you would call it.

(David at 00:33:41) We call it a DNA program. We actually haven't agreed on language yet.

(Joel Beasley at 00:33:46) Oh, wow. So you have a DNA program, a DNA model. Right? There you go. And then you pour the data into the DNA model and you can't physically see a color change if you had programmed one. You would need some sort of instrument. Correct? Is that what you're saying?

(David at 00:33:59) Yeah.

(Joel Beasley at 00:33:59) Even fluorescence, like under fluor—you just—the genetic structure for the blue color is inside there? I don't get it.

(David at 00:34:08) So think of it this way. DNA is composed of four letters: A, G, C, and T. Right? And the order of how these letters appear in the double helix have an impact on the nature of what that molecule is all about. The current devices that read DNA read these letters, A, G, C, and T. And they do it with great precision because that's the way you're able to do genomic analysis and so on. In our particular case, we don't need to go down to that level of specificity in part because of our encoding scheme using a predefined set of DNA snippets, oligonucleotides, which we call components, that we tie together to build a bigger molecule that carries the nature of what the data is in the bigger molecule. So we know what we're building these DNA molecules with. It's all synthetic.

(David at 00:35:08) It's not from a living organism. It's not biologically active or anything like that. And all we're doing is we're experimenting with colors and fluorescence so that we don't have to go down to the decoding of a molecule to discern what the sequence of AGCTs are. We can simply say, you know, every 30 base pairs, if you will, if I see the color red, it means I've used this piece of DNA. And if I see the color yellow, figuratively speaking, of course, I've used this other thing out of our DNA alphabet that we use.

(David at 00:35:48) So the exploration of color, fluorescence, and so on, for us, is very much a function of the scheme that we've used to encode data. This is not for everybody. Right? It's our—it's mapped to our specific encoding scheme. If somebody wanted to encode data a different way, they might still experiment with colors, fluorescence, and so on, but it might be for an entirely different purpose.

(David at 00:36:13) Net net to what you're saying, however, is this: the quantities we're dealing with and the variations we're dealing with are so great in number that having colors available to you isn't going to offer you much because you're not going to be able to discern trillion different molecules with a trillion different colors. Your eyes can't tell the difference between a trillion different colors. You have to use scientific instruments that can make those small discernments of difference to help you understand what you're really seeing. So for us, it's a means to an end.

(David at 00:36:46) It's not an end in itself. It's not something that's going to be visible to you. And even if it were visible to you, you wouldn't be able to have the ability to discern things at that level of detail that would help you at all.

(Joel Beasley at 00:37:00) Okay. So let me help. I'm going to try to figure out the stack. Here we go. The base layer of the raw is the actual four letters, right? The DNA sequencing letters.

(Joel Beasley at 00:37:12) Then you have some sort of abstraction, which I think you called the snippets of how you sort of encode specifically with them. And then on top of that, you have some sort of domain-specific layer with colors that help you understand those. Or is that wrong?

(David at 00:37:29) Let me explain it this way. First of all, let's agree that none of us can visibly observe a DNA molecule without the aid of a scientific instrument of some sort. If we can't observe a DNA molecule, we can't really observe the A, G, Cs, and Ts in any effective way either. They're smaller than the DNA molecule by implication. Okay.

(David at 00:37:54) Now in our case, we operate at a level of agglomeration of about 30 base pairs at a time. Each of these is a small piece of a DNA molecule that we refer to—our nomenclature is a component. Think of it as an alphabet. All right? So, you know, the English language has an alphabet from A to Z, and I always get the count confused between number of letters and number of teeth, so I don't know whether it's 24 or 26. But I have to count them. But it's some number of letters. And by ordering the way the letters appear, you create words, and that's information. What we're doing is we have an alphabet, 144 in number of these little components of DNA.

(David at 00:38:44) Each differs from the other just like A differs from B. And by virtue of the order that we link them together, we create a word, if you will, that depicts some information. So if it's component one coupled with component three coupled with component four, that's different than one coupled with two coupled with 12. Right? It's no different than if you said ABC is different than AZAAML.

(David at 00:39:14) Right? So it's the same length, but because the components in both of those cases are different, they connote something different. Okay. So far so good. So our ability to operate is at these collections of DNA snippets.

(David at 00:39:35) We call these identifiers. Now we know all the letters that are invoked because we supplied the letters. These are not just random pieces of DNA. And for us to say, "What word am I representing here? What piece of information am I representing here?" I don't need to go down to the level of the A, G, Cs, and Ts. I just need to know whether I use component one, three, and four or component one, seventeen, and 144. And if I can put an identifier on each of these 144 letters, right, I can stay away from having to go down to the level of all these A, G, Cs, and Ts. So it's an aggregation of representation that I'm searching for here. And if I can put a color on it that says, "Well, the letter A in my DNA alphabet, the component one, is going to be labeled with the color pink," figuratively speaking.

(David at 00:40:38) There's a little more to it than just conventional colors. And if component two is purple and component three is yellow and on and on it goes, then if I can look at a molecule, instead of going to each of these A, G, Cs, and Ts—meaning I would have to go through 500 of these—now instead of going through 500, I can go to maybe 16 different colors. And I can say, "Oh, that color scheme of 16 says that I'm looking at this piece of information." If I use a conventional genomic sequencer, I have to look at all the A, G, Cs, and Ts in detail, meaning roughly 500, to identify what information I'm looking at. So it's a matter of automation and scaling that is causing us to look at, "Can we identify snippets of DNA without having to go below the surface and look at all the A, G, Cs, and Ts in what order they're in?"

(David at 00:41:39) And one way to do that is to append something to each of these snippets—a little bit of fluorescence, a color, what have you—that says, "Ah, now I can look at something at a much more aggregate level and discern what it is."

(Joel Beasley at 00:41:52) Oh, I got it. Yeah. I'm good now. Yeah. So now I've got a clarifying question.

(David at 00:41:58) Sure.

(Joel Beasley at 00:41:59) Okay. So we said before, you have the raw data, you have your snippets, and then you have these—these colors are a way to identify a collection of snippets.

(David at 00:42:10) Mm-hmm.

(Joel Beasley at 00:42:11) Right? Yes. Okay. And so not the specific data within a snippet. It's used to identify a collection of snippets for color. Okay. Great. Now, in the—I think you said your alphabet was 144 characters, or the snippets were 144 characters. So you would associate a different color or fluorescent or something with each one of the 144, and you'd step back and look at it and have an idea of where you're at with your data.

(David at 00:42:42) Yeah. Another way to think of it is think of a Lego stack. Okay? So imagine, if you will, a collection of Legos in a dark room. You can't tell their colors, but they're all the same size. And you can stack them to a reasonable height, let's say 16 high. If you turn the light on, assuming that you bought conventional Legos, there would be different colors in the different stacks you composed. And you can discern one stack from another by the sequence of colors in the stack. You start at the top.

(David at 00:43:13) You would say, "Oh, this stack of 16 Legos is red, yellow, green, blue, et cetera. Next one is white, purple, red, yellow." And by virtue of that, you could tell one stack from the other. As soon as you can find a scheme to identify one stack from another, you can effectively have a scheme to encode data into these stacks of Legos. Right?

(David at 00:43:37) And you would decode it by virtue of having a key somewhere that says, "These colors correspond to the following pieces of Legos, and that dictates what I'm looking at here," and that's it. Right? So it's nothing different than looking at Lego stacks, except here instead of Legos, each individual Lego is a piece of DNA, and each piece of DNA has a color mapped to it. And by virtue of that, I can then look at a stack of DNA, or a length of longer DNA composed of these smaller pieces. And that sequence of colors will tell me exactly what I'm looking at.

(Joel Beasley at 00:44:17) That is a great example because there's Legos all over my house. Yeah.

(David at 00:44:23) These are encoding data, and you don't even know that.

(Joel Beasley at 00:44:26) I know. Last night, they had made some camel. They made a camel, and my daughter brought it up to me and I said, "How did you learn how to do this?" And she goes, "Instructions." It was in—and I go, "In—" but she said it wrong. I go, "What do you mean instructions?" And she brought over the side of the bag and pointed at it. Right. I said, "Oh, you just looked at that and built it?" There we go.

(Joel Beasley at 00:44:43) That makes me happy.

(David at 00:44:45) That's right. So it's a recipe, and we have a recipe in our world, which is algorithmic software and everything else that instructs our writing device, which we call Shannon, to build these longer length DNA molecules. So it's the same thing. You know? We have instructions. They're encoded in software. We hit go. It orchestrates the chemical activity inside the machine, and we produce out of that this collection of composed molecules, each of which contains a unique piece of data.

(Joel Beasley at 00:45:22) I just thought about—I was thinking about security just now as you were talking about this because you could have keys that are terabyte-size keys.

(David at 00:45:32) Mm-hmm.

(Joel Beasley at 00:45:32) Right? And the security on this, and especially, so you got that aspect. Then you have the concept that, you know, this could—each company or each customer have its own unique colored alphabet, like the 144 characters be in different colors. Or is it the language at Catalog?

(David at 00:45:50) There's a lot of ways to manipulate that for sure. And it would be a matter of cost, scale, and effort. But, yeah, absolutely. We could change a lot of schemes that are specific to an individual. But also, we're also creating an environment which is secure in a different kind of sense.

(David at 00:46:09) Right? So I'm doing computing data in a chemical construct, in a vat somewhere. How do you hack that? Right? You don't know the encoding scheme. You don't have physical access to the chemistry. You don't have any idea of what the stochastic elements are that are going on in this soup of chemistry going on. It'd be a pretty interesting kind of conundrum a hacker would face to figure out what's going on. Now eventually, you have to produce results, and you could say, "Well, maybe I can hack the step that involves producing the results because I want to see a report that I want to read or something like that." And you can invoke conventional encryption for that kind of step.

(David at 00:46:52) But the compute process per se, the encoding process per se—that's pretty hacker resistant.

(Joel Beasley at 00:47:00) The work that you're doing, do you think it's going, or do you think it has the potential to help us understand our own DNA as humans better?

(David at 00:47:11) So we operate at a scale quite a bit removed from human DNA in the following sense. The DNA molecules we synthetically build—again, synthetic means they're not biological in origin, nor are they capable of being rendered biologically active because we put stops in the molecules and so on that would make it impossible effectively to turn into a biologically active device. But for thermodynamic and other reasons, the length of the DNA molecules we construct are about 500 base pairs in length. Okay? Human is 3.5 billion.

(David at 00:47:48) So we can't use these techniques to build human DNA, for example. On the other hand, the pioneering work that we're doing in terms of synthesizing DNA molecules and so on might be very good for people doing classic biological research in terms of providing faster ways to create probes at scale and this and that. So it might be an adjunct capability that one could use to study human DNA, but we operate at dramatically different scales.

(Joel Beasley at 00:48:21) Yeah. And I didn't want to confuse the two, but to me, it just seems, you know, I've been watching Catalog for at least four years, very excited about the progress that they have made as far as the speed of timing and encoding and decoding DNA.

(David at 00:48:36) Mm-hmm.

(Joel Beasley at 00:48:36) But, you know, I would just—it would be—I would find it hard to believe that all of these advancements that you guys are making in the world of DNA isn't any use at all to us as humans understanding our own DNA. There seems like there has to be some, at least inspiration if no less, to help people who are working on human DNA to see how you guys are manipulating DNA or understanding it.

(David at 00:49:02) Yeah. I would make the following nebulous comment. Nebulous for what I hope will become apparent reasons. There are research-based rules of thumb that govern our understanding of how DNA behaves, how it can be manipulated, and so on. I would submit that it is likely that an operation like ours, which is operating at scale, it goes quite far beyond what people do in conventional DNA research, might display phenomenon or make visible phenomenon that would put some of that rule of thumb theory in question.

(David at 00:49:47) Right? Now why am I saying this? We do 500,000 DNA chemical steps per second. Okay? Now you could argue we're actually doing 8 million a second because when I say what I'm talking about here is the ligation or the connecting together of these pieces of the DNA alphabet we have.

(David at 00:50:13) And we do 16 of these at a time, 500,000 times per second. Okay? Now that's at a scale that goes beyond what anybody else is doing with respect to the manipulation of DNA, regardless of how complex a DNA molecule is and so on. I think that the prudent thing to observe would be the following, an idea stolen from the world of computing. Because certainly, this was true during my career at IBM as we were building supercomputers. And that was that every design we did was perfect on paper, and then you built it, and then you found out there are all sorts of phenomenon that manifested themselves that you never anticipated because things happen at scale that you don't see in the absence of scale.

(David at 00:51:02) Right? It's no different than, in my personal case. So my resting heart rate is 50 beats a minute. Okay? There are electrical anomalies in my heart that are irrelevant to my health that if my resting heartbeat was 70 beats a minute, no one would ever see.

(David at 00:51:23) Right? And believe me, the anomalies are absolutely irrelevant. This is not—I'm not sick or anything like that. But as you move between these different sort of domains of measurement, you see things that you might not see otherwise. Think of the tides in the ocean.

(David at 00:51:42) You know, the tide goes out, and suddenly you see structures on the floor of the ocean that you never knew were there, that never bothered you before. But now that you see them, you pay attention to them because, "Oh, well, I can't, you know, sail my boat at this hour of the day because of the tides." But for most of the day, you're sailing on the surface of the water totally oblivious to what's below you because it doesn't matter. Right? And all I'm saying here is that in the world of DNA, DNA research, human DNA, health applications, and so on, it would be prudent to expect that as we start pushing the boundaries of DNA chemistry at levels of scale that no one ever has experienced before, you are bound to observe something anomalous to what the theory has told you to believe.

(Joel Beasley at 00:52:34) I got it. I'm good. Change of question here.

(David at 00:52:39) Yeah.

(Joel Beasley at 00:52:40) You know, as we're sitting here talking about DNA, quantum computing, traditional computing. When I—I have less than 15 hours of education in quantum computing, so I'm very naive. I don't know a lot at all about quantum computing. But when I got to talk to Robert Sutor about one of his books he was writing on quantum computing, he had said that one of the things that could be potentially useful—because I find our conversation with Robert very similar to you. Everyone seems to try to pull you to be something that you're not and hype it up. But so to bring it back, Robert was saying that for things like maybe modeling molecules, right, it might be exceptionally fast there because from what I took, the current way the scientific community will do it when drug testing is they use a traditional computer and they try to make this virtualization of molecules and how they interact.

(Joel Beasley at 00:53:38) It's very, very basic and it contains very little of what the actual molecules are and how they interact. But in quantum computing, you'd be able to do that in a much richer environment. Right? So taking that for what it is, and right now, people are comfortable with this concept of using traditional computers to interact with quantum computers because of things like, you know, I think Microsoft and a couple other companies, Honeywell, they did quantum APIs. We can actually go run things on quantum computers in the cloud type deal.

(Joel Beasley at 00:54:09) Okay. So with those things in mind, do you ever think that maybe the company will end up as a quantum software play? Because when quantum technology advances far enough to be able to richly create and interact with these representations of these molecules that they could then take all of the Catalog DNA knowledge and put that on top of it, and now they can do that in some quantum space? Or is that just insanity?

David at 00:54:41

I wouldn't go too far down the path of the merging of what we're doing with quantum at this point. But I would answer you in the following way: there are a lot of lessons being pioneered in quantum that weigh on us in terms of how we think about the development of technology and DNA for reasons I alluded to at the beginning with respect to what's presumptively an exotic technology at a certain point in time, but holds promise to do something quite dramatically important somewhat down the road. So we're thinking about software a lot in terms of, for example, what the software paradigm would be to invoke these chemical instructions that we're creating. How would you do that?

David at 00:55:30

What is the nature of the API that would allow you to orchestrate that kind of activity? And so in calendar year 2022, this year, our hope is that we'll create, much like the quantum guys did at IBM, an emulator for how to do computing with DNA in the context of the CATALOG encoding scheme. We don't know a priori whether it's generalizable to an arbitrary set of DNA encoding schemes, but we need to do it at least once to stimulate the community to begin thinking about how one should do this, how we should go about it, where should standards be, where should they be avoided, and so on. And I think IBM and Bob Sutor, an ex-colleague of mine, did a lot of right things here with respect to stimulating the community to really help, in aggregate, explore a lot of things with respect not only to the nature of what SDKs should look like and APIs should look like, but also the application domains that this technology could be applied to. We hope to follow in those footsteps to a certain degree and leverage a community to help us and the community at large get a better understanding of how you do things with DNA as well.

Joel Beasley at 00:56:51

This is such an exciting time to see all of these new technologies progressing at such a rapid pace. I'm super happy to be alive right now.

David at 00:57:00

Better than the alternative.

Joel Beasley at 00:57:02

Right. Oh, man. This is good. So I want to make sure that we get everything covered that you also wanted to get covered here on the show. I had a bunch of questions. I was like a kid in the candy shop. Thank you so much for indulging me. What sort of message do you want to get out to the world?

David at 00:57:22

I think from the perspective of CATALOG, the message we want to get out to the world is that you will see material progress on computing with DNA in 2022. And by material progress, I'll define that as working on a real world problem at scale. So different than a toy problem. There have been academic papers on toy problems and so on using bench chemistry. We're looking to do something at scale so that one can look at it and evaluate the progress on its merits and not have to get a pencil and paper and say, "Well, if I extrapolated this to the real world, it would look like this."

David at 00:58:00

Now we want to take that away. We want to solve one real world problem at scale in '22. And I think that's an important goal that we've set for ourselves. We have a long ways to go. We would expect to begin to publish results prior to the end of the year in terms of the path we're taking and where we're headed. But that's '22 for CATALOG. We want to show people what the possibilities are with compute.

Joel Beasley at 00:58:29

That's exciting. And then you get to come on next year and tell us about it.

David at 00:58:33

Oh, next year we should tell you about a lot of different applications in the DNA compute space. And hopefully, in '23, we're on the verge of getting to a commercial offering of some sort.

Joel Beasley at 00:58:45

So just so I understand—for this year, and we can edit whatever we want to edit, but just for me being a nerd—did you, like, this year, by the end of the year, you're going to release what the problem will be that you're going to solve the following year, or you're going to actually have solved a problem at scale in the year?

David at 00:59:02

We're going to solve it this year.

Joel Beasley at 00:59:04

Oh, nice. So you'll be able to, like, see it operate, right?

David at 00:59:07

Yeah. Yeah. We'll publish the results and show you everything.

Joel Beasley at 00:59:12

Just so you know, as we talk about all of this, the hardest thing for me, and I'm sure you ran up against it with just talking to the general public, is the visualization because we're inundated. I've got screens all around me. We're so used to "here's a screen, I click a button, I do something." And so I'm constantly trying to, like, you know, use the best reference I know, the one I'm most familiar with, and tie it back to how will DNA computing be like that. And so that's just one of the things I'm sharing because I'm pretty good at articulating the problems I'm having wrapping my mind around an idea.

Joel Beasley at 00:59:49

And that's something I struggle with is the visual representation of how this looks. But if you say, you know, potentially, you know, I put instructions through a traditional computer, it uses Shannon to do its movements around, it uses another thing to pick up on the colors or whatever little information you've encoded to understand the results, and then it kicks that back around to me, I'm completely comfortable with that. Is that the right workflow?

David at 01:00:14

Yeah. We have software pipelines that govern the conventional manipulation of data coupled with the orchestration of the activity in Shannon. So for example, you know, our Lego model—we have a lot of inkjet printheads in Shannon that are depositing these little droplets that incorporate these snippets of DNA, my alphabet, if you will. We have software that governs the sequence with which all that happens. And so there has to be this sort of oversight mechanism that goes on.

David at 01:00:48

It's all electronic in nature. But, you know, at a certain point, it turns into chemistry and the chemistry produces the stuff in a pool at the end of the machine. So we haven't, to my satisfaction, yet come up with a scheme to represent the way we're doing compute with DNA. The Lego model is fine for helping people understand how you combinatorially apply different pieces of DNA to create a depiction of information. I think that's reasonably straightforward.

David at 01:01:19

But compute is different because, well, there are an infinite number of ways you can compute something, and it's not simply a one-to-one mapping of classic compute instructions to chemistry. It's different, right? And we have to come up with a different paradigm to depict that in fact. Now the paradigm that we use internally is by looking at trees.

David at 01:01:43

And trees are just sort of this cascading set of nodes and, you know, lines connecting the nodes that you could say represent logic paths through a sea of information in a certain sense. And that's effective for what we're doing at the nerd level, but I don't think that's going to be palatable to a layperson to understand exactly what's happening. It's kind of like, you know, sometimes you see these carnival machines where, or toys where, you put in a ball or something at the top, and it goes down through this maze of things. And it bounces off this and bounces off that, and it ends up at one particular location at the bottom. And maybe you can bet on this or something.

David at 01:02:27

Maybe you have 20 or 30 different locations where it can end up. And by virtue of what it hits along the way, it gets routed one to another. That's kind of what we're doing here. But I think that's relatively obscure in terms of translating that into a compute instruction. But that is actually a compute instruction of a certain kind.

Joel Beasley at 01:02:48

Have you—I mean, you're going to have something commercial, you said, by 2025. I think I read in the press release, there hopes to be commercial things by 2025.

David at 01:02:56

Yeah. We'd like to do something experimentally in the marketplace in '23, but, yeah, we would expect to be pretty commercial by '25.

Joel Beasley at 01:03:04

Given the fact that I'm an impatient person, I always think farther ahead. So I'm thinking about scaling and talent. You just brought that up, right, because of the way the education system is set up. And I was just curious—when I was searching to try to prepare for the interview, and I was trying to search for, you know, CATALOG DNA computing versus DNA storage to try to understand these differences because everywhere in the press release, it just said CATALOG DNA and compute or CATALOG storage and computing. And I was like, "Well, where's the and computing?

Joel Beasley at 01:03:36

This is just a press release about them raising money." I didn't pick up on it. And then I started trying to figure out what it was in general. And then pages of results, and then finally, I found some YouTube of some person trying to explain DNA computing, and they just explained what storage was. And I was like, "Okay. There is a huge opportunity for education." I was curious. Have you guys talked internally about, you know, how you're going to handle the education gap to gain talent at a rapid pace?

David at 01:04:11

Yes. So in many, many different ways. You'll see, for example, active participation by CATALOG this year in a variety of venues, both domestically and internationally, where we begin to unveil some of the nature of the things that we're doing. You'll see us publishing more things, at least in lay press, if you will. And the reason I say that is not because we haven't been thoughtful to, you know, refereed status kinds of publications, but we're moving at a much different pace than what university researchers are and so on.

David at 01:04:50

And so we're not that interested in refereed journals, etcetera. This is all pragmatic. So there's going to be a lot of outreach in that fashion. We have some activities queued up beginning pretty soon with universities reaching out broader into the university communities to let them know what we're doing to both intrigue them as potential collaborators, but also to motivate students to begin to get thoughtful about how they orchestrate their own educational programs to be able to do things like this. So, yeah, it's going to be a long slog.

David at 01:05:26

But at the end of the day, nothing breeds interest more than success. And so the extent to which we can produce results and demonstrate value and so on will go a long way towards motivating interest for people to move in this direction. You know, the quantum field, listen—IBM's been working on quantum for five decades, right? But it's really only been in the last ten years or so where, you know, you've got quantum devices and you got SDKs and you've got emulators and all these other things, and you've got notable applications that can be pursued with quantum that has suddenly sort of sparked interest beyond where things were in, you know, 2001 or 1991 or 1981, right? Because now people can taste it. They can touch it. They can feel it.

David at 01:06:18

They can begin to play with it and so on. And I think that as we get to that level where we move away from the abstractions that have been so prevalent about DNA computing and storage today, that's it. Now with respect to DNA computing per se, I think the seminal paper on this goes back maybe to 1996, traveling salesman problem, okay? And there have been some other academic things since.

David at 01:06:45

But one of the reasons why we went out of our way to mention computing in our announcements last fall was to signal to the world that we're here, and computing is coming, and people should start thinking about this. We think DNA storage as an endeavor per se is nothing more than a stepping stone to create the notion of truly active devices that link storage and compute in the same environment operating continuously at low power all the time. That's where we're headed, right? The idea of having passive storage, okay, fine. The idea of having passive storage in support of compute, okay, fine. But these classic paradigms, there's no reason they should exist that way. Why not have everything in one environment doing everything at the same time?

David at 01:07:40

And that's us.

Joel Beasley at 01:07:42

Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague that you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email: [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.