Ideas: The journey to DNA data storage

This, along with the fact that DNA is itself a  relatively rugged molecule—it lives in our body;   it lives outside our body for thousands  and thousands of years if we, you know,   leave it alone to do its thing—makes  it a very attractive media.

we've contributed to steady and promising progress  in the area. We've helped push the boundaries of   how much DNA writer can simultaneously store,  shown that full automation is possible,   and helped create an ecosystem for the commercial  success of DNA data storage. And just this week,   we've made one of our most advanced tools  for encoding and decoding data in DNA open  

source. Joining me today to discuss the  state of DNA data storage and some of our   contributions are several members of the DNA  Data Storage Project at Microsoft Research:   Principal Researcher Bichlien Nguyen,  Senior Researcher Jake Smith, and Partner   Research Manager Sergey Yekhanin. Bichlien,  Jake, and Sergey, welcome to the podcast.

and I heard things like, “Wow, why are you  thinking about that? It's so far off.” So, first,   please share a bit of your research backgrounds  and then how you came to work on this project.   Where did you first encounter this idea, what do  you remember about your initial impressions—or the   impressions of others—and what made you want  to get involved? Sergey, why don’t you start.

it just so happened that at that moment, I  was thinking about what to do next. Like,   in California, I was mostly working on coding  for distributed storage, and when I joined here,   that effort kept going. But I had some free  cycles, and that was the moment when Karin   came just to my office and told me about the  project. So, indeed, initially, it did feel a  

lot like science fiction. Because, I mean, we  are used to coding for digital storage media,   like for magnetic storage media, and here, like,  this is biology, and, like, why exactly these   kind of molecules? There are so many different  molecules. Like, why that? But honestly, like,   I didn't try to pretend to be a biologist and make  conclusions about whether this is the right medium  

or the wrong medium. So I tried to look into these  kinds of questions from a technical standpoint,   and there was a lot of, kind of, deep, interesting  coding questions, and that was the main attraction   for me. At the same time, I wasn’t convinced  that we will get as far as we actually got,   and I wasn't immediately convinced about the  future of the field, but, kind of, just the depth   and the richness of the, what I’ll call, technical  problems, that's what made it appealing for me,  

and I, kind of, enthusiastically joined. And  also, I guess, the culture of the team. So, like,   it did feel like a startup. Like, we all work  for Microsoft; we’re all Microsoft researchers.   Microsoft isn’t a startup. But that team, the  team that drove the DNA Data Storage Project,   it did feel like a startup, and it was  something unusual and exciting for me.

this is a crazy, crazy idea. I definitely was  not sold on it, but I was like, well, look,   I get to meet and work with so many interesting  people from different backgrounds that, one,   even if it doesn't work out, I’m  going to learn something, and, two,   I think it could work, like it could work. And so  I think that's really what motivated me to join.

you know. You look at it every now and then  to see what information it can tell you about,   you know, what maybe your drug might be hitting  on the target side, and it's, you know, that   connection—that the DNA contains the information  in the living systems, the DNA contains the   information in our assays, and why could the DNA  not contain the information that we, you know,   think more about every day, that information that  lives in our computers—as an extremely cool idea.

rise of large language models and these other big  generative models. The data that we do produce,   our video has gone from, you know, substantially  small, down at 480 resolution, all the way up to   things at 8K resolution that now take orders of  magnitude more storage. And so we really need a   denser medium on the other side to contain that.  DNA is extremely dense. It holds far, far more   information per unit volume, per unit mass than  any storage media that we have available today.  

This, along with the fact that DNA is itself a  relatively rugged molecule—it lives in our body;   it lives outside our body for thousands  and thousands of years if we, you know,   leave it alone to do its thing—makes  it a very attractive media,   particularly compared to the traditional  magnetic media, which has lower density   and a much shorter lifetime on the,  you know, scale of decades at most.

So how does DNA data storage actually work?  Well, at a very high level, we start out in the   digital domain, where we have our information  represented as ones and zeros, and we need to   convert that into a series of A's, C's, T's,  and G's that we could then actually produce,   and this is really the domain of Sergey. He'll  tell us much more about how this works later on.   For now, let's just assume we've done this. And  now our information, you know, lives in the DNA  

base domain. It's still in the digital world. It's  just represented as A’s, C’s, T’s, and G’s, and   we now need to make this physical so that we can  store it. This is accomplished through large-scale   DNA synthesis. Once the DNA has been synthesized  with the sequences that we specified, we need to   store it. There's a lot of ways we can think about  storing it. Bichlien’s done great work looking at  

DNA encapsulation, as well as, you know, other  more raw just DNA-on-glass-type techniques. And   we've done some work looking at the susceptibility  of DNA stored in this unencapsulated form to   things like atmospheric humidity, to temperature  changes and, most excitingly, to things like   neutron radiation. So we've stored our data  in this physical form, we've archived it, and   coming back to it, likely many years in the future  because the properties of DNA match up very well  

with archival storage, we need to convert it back  into the digital domain. And this is done through   a technique called DNA sequencing. What this does  is it puts the molecules through some sort of   machine, and on the other side of the machine, we  get out, you know, a noisy representation of what   the actual sequence of bases in the molecules  were. We have one final step. We need to take  

this series of noisy sequences and convert it back  into ones and zeros. Once we do this, we return   to our original data and we've completed,  let's call it, one DNA data storage cycle.

forms of media, the errors are bit flips. Like,  you store a 0; you get a 1. Or you store a 1; you   get a 0. So these are called substitution errors.  The field of error-correcting codes, it started,   like, in the 1950s, so, like, it’s 70 years old  at least. So we, kind of, we understand how to   deal with this kind of error reasonably well, so  with substitution errors. In DNA data storage,   the way you store your data is that given,  like, some large amount of digital data,  

you have the freedom of choosing which short  DNA molecules to generate. So in a DNA molecule,   it’s a sequence of the bases A, G, C, and  T, and you have the freedom to decide,   like, which of the short molecules you need to  generate, and then those molecules get stored,   and then during the storage, some of them  are lost; some of them can be damaged. There   can be insertions and deletions of bases on every  molecule. Like, we call them strands. So you need  

redundancy, and there are two forms of redundancy.  There's redundancy that goes across strands,   and there is redundancy on the strand. And so,  yeah, so, kind of, from the error-correcting   side of things, like, we get to decide what kind  of redundancy we want to introduce—across strands,   on the strand—and then, like, we want to  make sure that our encoding and decoding   algorithms are efficient. So that's  the coding theory angle on the field.

that enabled us to get to the densities that  we needed. And then on the reading side, we   used different sequencing technologies. And it was  great to see that we could actually just, kind of,   pull sequencing technologies off the shelf because  people are so interested in reading biological   DNA. So we explored the Illumina technologies and  also Oxford Nanopore, which is a new technology  

coming in the horizon. And then preservation, too,  because we have to make sure that the data that’s   stored in the DNA doesn't get damaged and that we  can recover it using the error-correcting codes.

storage itself. And that's a very different beast  than, you know, the traditional storage mediums.   And so we worked with teams who literally create,  you know, these little tiny micro- or nanocapsules   in glass and being able to store that there. It's  really interesting, this multidisciplinarity,   because we're, in a way, bridging software  with wetware with hardware. And so you,   kind of, need all the different disciplines  to actually get you to where you need to go.

the Molecular Information Systems Lab there,  to build. And really, at that point, you know,   there had been work suggesting that DNA data  storage was viable, but nobody had really shown   an end-to-end system, from beginning to end, and  in fact, my manager at the time, Doug Carmean,   used to call it the “bubble gum and shoestring”  system. But it was a crucial first step because  

it shows it was possible to really fully  automate the process. And there have been   several interesting challenges there in the  system, but we noticed that one particularly   challenging one was synthesis. That first system  that we built was capable of storing the word   “hello,” and that was all we could store. So  it wasn't a very high-capacity system. But in   order to be able to store a lot more volumes of  data instead of a simple word, we really needed  

much more advanced synthesis systems. And this is  what both Bichlien and Jake ended up working on,   so do you want to talk a little bit about that  and the importance of that particular work?

to meet the demand. And so Bichlien started out  looking at a thing called a microelectrode array,   where you have this big checkerboard of small  individual reaction sites, and in each reaction   site, we used electrochemistry in order to  control base by base—A, C, T, or G by A, C,   T, or G—the sequence that was growing at that  particular reaction site. We got this down to   the nanoscale. And so what this means practically  is that on one of these chips, we could synthesize  

at any given time on the order of hundreds of  millions of individual strands. So once we had the   synthesis working with the traditional chemistry  where you're doing chemical synthesis—each base   is added in using a mixture of chemicals that are  added to the individual spots—they're activated.   But each coupling happens due to some energy you  prestored in the synthesis of your reagents. And  

this makes the synthesis of those reagents costly  and themselves a bottleneck. And so taking, you   know, a look forward at what else was happening  in the synthetic biology world, the, you know,   next big word in DNA synthesis was and still is  enzymatic synthesis, where rather than having to,   you know, spend a lot of energy to chemically  pre-activate reagents that will go in to make   your actual DNA strands, we capitalize on  nature's synthetic robots—enzymes—to start  

with less-activated, less-expensive-to-get-to,  cheaply-produced-through-natural-processes   substrates, and we use the enzymes themselves,  toggling their activity over each of the   individual chips, or each of the individual  spots on our checkerboard, to construct DNA   strands. And so we got a little bit into this  project. You know, we successfully showed that  

we could put down selectively one base at a  given time. We hope that others will, kind of,   take up the work that we've put out there, you  know, particularly our wonderful collaborators   at Ansa who helped us design the enzymatic  system. And one day we will see, you know,   a truly parallelized, in this fashion, enzymatic  DNA system that can achieve the scales necessary.

encounter in DNA data storage. Like, in DNA  molecules—just the sequence of these bases,   A, G, C, T, in maybe a length of,  like, 200 or so and you store a very,   very large number of them—the errors that you  see is that some of these strands, kind of,   will disappear. Some of these strings can be  torn apart like, let’s say, in two pieces,   maybe even more. And then on every strand, you  also encounter these errors—insertions, deletions,  

substitutions—with different rates. Like, the  likelihood of all kinds of these errors may differ   very significantly across different technologies  that you use on the biology side. And also there   can be error bursts somehow. Maybe you can get  an insertion of, I don’t know, 10 A’s, like, in a  

row, or you can lose, like, you know, 10 bases in  a row. So if you don't, kind of, quantify, like,   what are the likelihoods of all these bad events  happening, then I think this still, kind of,   fits at least the majority of approaches to DNA  data storage, maybe not exactly all of them,   but it fits the majority. So when we design  coding schemes, we are trying also, kind of,   to look ahead in the sense that, like,  we don't know, like, in five years, like,  

how will these error profiles, how will it look  like. So the technologies that we develop on the   error-correction side, we try to keep them very  flexible, so whether it's enzymatic synthesis,   whether it's Nanopore technology, whether it’s  Illumina technology that is being used, the   error-correction algorithms would be able to adapt  and would still be useful. But, I mean, this makes   also coding aspect harder because, [LAUGHTER] kind  of, you want to keep all this flexibility in mind.

coding theory, in our algorithms, we use ideas  from very different branches. I mean, there are   some core ideas from, like, core algorithm space,  and I won’t go into these, but let me just focus,  

kind of, on two aspects. So when we just faced  this problem of coding for DNA data storage and we   were thinking about, OK, so how to exactly design  the coding scheme and what are the algorithms   that we’ll be using for error correction, so,  I mean, we’re always studying the literature,   and we came up on this problem called trace  reconstruction that was pretty popular—I mean,   somewhat popular, I would say—in computer science  and in statistics. It didn’t have much motivation,  

but very strong mathematicians had been looking  at it. And the problem is as follows. So, like,   there is a long binary string picked at random,  and then it’s transmitted over a deletion channel,   so some bits—some zeros and some ones—at certain  coordinates get deleted and you get to see, kind  

of, the shortened version of the string. But you  get to see it multiple times. And the question is,   like, how many times do you need to see it so that  you can get a reasonably accurate estimate of the  

original string that was transmitted? So that was  called trace reconstruction, and we took a lot of   motivation—we took a lot of inspiration—from the  problem, I would say, because really, in DNA data   storage, if we think about a single strand, like,  a single strand which is being stored, after we   read it, we usually get multiple reads of this  string. And, well, the errors there are not just  

deletions. There are insertions, substitutions,  and, like, inversive errors, but still we could   rely on this literature in computer science that  already had some ideas. So there was an algorithm   called BMA, Bitwise Majority Alignment. We  extended it—we adopted it, kind of, for the needs   of DNA data storage—and it became, kind of, one  of the tools in our toolbox for error correction.

So we also started to use ideas from  literature on electrical engineering,   what are called convolutional error-correcting  codes and a certain, kind of, class of algorithms   for decoding errors in these convolutional  error-correcting codes called, like, I mean,   Trellis is the main data structure, like,  Trellis-based algorithms for decoding   convolutional codes, like, Viterbi algorithm or  BCJR algorithm. Convolutional codes allow you to  

introduce redundancy on the string. So, like, with  algorithms kind of similar to BMA, like, they were   good for doing error correction when there was no  redundancy on the strand itself. Like, when there  

is redundancy on the strand, kind of, we could do  some things, but really it was very limited. With   Trellis-based approaches, like, again inspired  by the literature in electrical engineering,   we had an approach to introduce redundancy on the  strand, so that allowed us to have more powerful   error-correction algorithms. And then in the end,  we have this algorithm, which we call Trellis BMA,  

which, kind of, combines ideas from both  fields. So it's based on Trellis, but it's   also more efficient than standard Trellis-based  algorithms because it uses ideas from BMA from   computer science literature. So this is, kind of,  this is a mix of these two approaches. And, yeah,   that’s the paper that we wrote about three years  ago. And now we're open sourcing it. So it is the  

most powerful algorithm for DNA error correction  that we developed in the group. We’re really happy   that now we are making it publicly available  so that anybody can experiment with the source   code. Because, again, the field has expanded a  lot, and now there are multiple groups around   the globe that work just specifically on error  correction apart from all other aspects, so, yeah,   so we are really happy that it’s become publicly  available to hopefully further advance the field.

it was really a comparison, right, between the  different storage medias that exist today. And   one of the issues that have come up through the  years of development of those technologies was,   you know, hard errors and soft errors that were  induced by radiation. So we wanted to know,   does that maybe happen in DNA? We know that DNA,  in humans at least, is affected by radiation from  

cosmic rays. And so that was really the motivation  for this type of experiment. So what we did was   we essentially took our DNA files and dried  them and threw them in a neutron accelerator,   which was fantastic. It was so exciting. That's,  kind of, the merge of, you know, sci fi with sci   fi at the same time. [LAUGHS] It was fantastic.  And we irradiated for over 80 million years— The equivalent of … NGUYEN: The equivalent of 80 million years.

Yes, because it's a lot of  radiation all at the same time, …

which based on the error-correcting codes,  we can recover from. So essentially,   there's not much … one, there's not much  interaction between neutrons and DNA,   and second, we have error-correcting  codes that would prevent any data loss.

technology to thrive commercially. We did bring  together multiple universities and companies that   were interested in the technology. And one thing  that we've seen with storage technologies that's   been pretty important is standardization and  making sure that the technology’s interoperable.  

And, you know, we've seen stalemate situations  like Blu-ray and high-definition DVD, where, you   know, really we couldn't decide on a standard, and  the technology, it took a while for the technology   to be picked up, and the intent of the DNA Data  Storage [Alliance] is to provide an ecosystem   of companies, universities, groups interested in  making sure that this time, it's an interoperable  

technology from the get-go, and that increases  the chances of commercial adoption. As a group,   we often talk about how amazing it is to work  for a company that empowers us to do this kind of   research. And for me, one of Microsoft Research’s  unique strengths, particularly in this project,   is the opportunity to work with such a  diverse set of collaborators on such a  

multidisciplinary project like we have. How  do you all think where you've done this work   has impacted how you've gone about it and  the contributions you’ve been able to make? NGUYEN: I'm going to start with if we look  around this table and we see who's sitting at it,   which is two chemists, a computer architect, and  a coding theorist, and we come together and we're  

like, what can we make that would be super, super  impactful? I think that's the answer right there,   is that being at Microsoft and being in  a culture that really fosters this type   of interdisciplinary collaboration is the key  to getting a project like this off the ground. SMITH: Yeah, absolutely. And we should  acknowledge the gigantic contributions   made by our collaborators at the University of  Washington. Many of them would fall in not any  

of these three categories. They’re electrical  engineers, they're mechanical engineers,   they're pure biologists that we worked with.  And each of them brought their own perspective,   and particularly when you talk about  going to a true end-to-end system,   those perspectives were invaluable as we were  trying to fit all the puzzle pieces together. Yeah, absolutely. We've had great  collaborations over time—University of Washington,  

ETH Zürich, Los Alamos National Lab, ChipIr,  Twist Bioscience, Ansa Biotechnologies. Yeah,   it’s been really great and a great set of  different disciplines, all the way from coding   theorists to the molecular biology and chemistry,  electrical and mechanical engineering. One of the   great things about research is there's never  a shortage of interesting questions to pursue,   and for us, this particular work has opened the  door to research in adjacent domains, including  

sustainability fields. DNA data storage requires  small amounts of materials to accommodate the   large amounts of data, and early on, we wanted to  understand if DNA data storage was, as it seemed,   a more sustainable way to store information.  And we learned a lot. Bichlien and Jake,   you had experience in green chemistry when you  came to Microsoft. What new findings did we make,  

and what sustainability benefits do  we get with DNA data storage? And,   finally, what new sustainability  work has the project led to?

digital information in DNA, that we would have  benefits associated with carbon emissions. We   would be able to reduce that because we don't need  as much infrastructure compared to the traditional   storage methods. And there would be an energy  reduction, as well, because this is a passive way  

of archival data storage. So that was, you know,  the main takeaways that we had. But that also,   kind of, led us to think about other  technologies that would be beneficial   beyond data storage and how we could use the  same kind of life cycle thinking towards that.

Storage Project, you know, is something that can  be much bigger than any single material. And where   we think there's a, you know, chance for folks  like ourselves at Microsoft Research to make a   real impact on this sustainability-focused design  is through the application of machine learning,   artificial intelligence—the new tools that will  allow us to look at much bigger design spaces   than we could previously to evaluate  sustainability metrics that were not  

possible when everything was done manually and  to ultimately, you know, at the end of the day,   take a sustainability-first look at what a  material should be composed of. And so we've   tried to prototype this with a few projects.  We had another wonderful collaboration with   the University of Washington where we looked at  recyclable circuit boards and a novel material  

called a vitrimer that it could possibly be made  out of. We've had another great collaboration with   the University of Michigan, where we've looked at  the design of charge-carrying molecules in these   things called flow batteries that have good  potential for energy smoothing in, you know,  

renewables production, trying to get us out  of that day-night, boom-bust cycle. And we   had one more project, you know, this time with  collaborators at the University of Berkeley,   where we looked at, you know, design of a class  of materials called a metal organic framework,   which have great promise in low-energy-cost  gas separation, such as pulling CO2 out of the,   you know, plume of a smokestack or, you  know, ideally out of the air itself.

really try, to do the most important work I  can. And this is what attracted me to these   other areas of environmental sustainability  that Bichlien and Jake covered, where there's   absolutely no lack of problems. Like them, I'm  super interested in using AI to solve many of   them. So how do each of you think working on  the DNA Data Storage Project has influenced   your research approach more generally and how you  think about research questions to pursue next?

end-to-end prototype and all the requirements  that you need to get there. And then also,   really the importance of having, you know,  a background in computer science and really   being able to understand the lingo that is used  in multidisciplinary projects because you might   say something and someone else interprets it very  differently, and it's because you're not speaking  

the same language. And so that understanding  that you have to really be … you have to learn   a little bit of vocabulary from each person  and understand how they contribute and then   how your ideas can contribute to their ideas  has been really impactful in my career here.

that goal? And this is something that we've  taken with us into the sustainability-focused   research and, you know, something that I think  will affect all the research I do going forward.

how long does he keep the videos. And he told  you that he keeps it for about a couple of days   because it's too expensive. But otherwise, like,  if it weren't that expensive, he would keep it   for much, much longer because, like, he wants to  have these recordings if later somebody is upset   about the ride and, I don’t know, he is getting  sued or something. So this is, like, this is one   small narrow application area where DNA data  storage would clearly, kind of, if it happens,  

then it will solve it. Because then, kind of, this  long-term archival storage will become very cheap,   available to everybody; it would become a  commodity basically. There are many things   that will be enabled, like this helping the Uber  drivers, for instance. But also one has to think   of, of course, like, about, kind of, the broader  implications so that we don't get into something   negative because again this power of recording  everything and storing everything, it can also  

lead to some use cases that might be, kind of,  morally wrong. So, again, hopefully by the time   that we get to, like, really wide deployments  of this technology, the regulation will also be   catching up and the, like, we will have great use  cases and we won’t have bad ones. I mean, that's   how I think of it. But definitely there are lots  of, kind of, great scenarios that this can enable.

we can do in terms of gene synthesis, in terms of,  like, fundamental biotech research that will lead   to that next set of drugs and, you know, give us  medications or treatments that we could not have   thought possible if we were not able to synthesize  DNA and related molecules at that scale.