a16z Podcast
Original podcast
Where to listenCopy share link
3,547 words · 1 speaker
So, my name's DJ, and my company is Emerald Therapeutics, and today I'd like to tell you about our chief offering, which is the Emerald Cloud Laboratory. So the Emerald Cloud Laboratory or the ECL is a fully virtualized life sciences laboratory. So we have an application, so scientists anywhere in the world can design an experiment as if they are standing in front of the instrument themselves. That experiment then runs at our central facility in South San Francisco. You're all welcome to come visit. And then once it's completed, we send them back the data. But it's not just the data that we send back for that particular experiment. It's all of the other meta and ancillary data around that. So things like for the instrument they ran it on, when was the last time we ran controls on that instrument to verify that it was operating nominally?
When's the last time that instrument was maintenance? What are all of the, what was the humidity in the room when the experiment was run? What were the three experiments that were run on those samples before the one that they just ran? So all of those variables are captured and put into a database that the scientists who are using the system have access to. So this is something that we've built over a number of years, and there's an arbitrary number of ways that you could do this, and many of them are awful. And so the way that we've decided to design this system is was very organic. So just by way of introduction, the company is called Emerald Therapeutics because five years ago we were founded by myself and my co-founder as an antiviral therapeutics play, specifically ways of curing persistent viral infections.
And when we started the company, we were just two freshly minted PhDs, which got like resounding crickets from most traditional life sciences investors and most investors in general. So we were, we were still able to raise some money. The fact that we were able to raise any is a minor miracle. I'm still convinced five years later. But the thing that we realized was that we were going to have to take a radically different approach because for the type of platform we wanted to develop, nowadays a biotech company might raise 20, 30, 40 million dollars for a first round of financing. And we raised a tiny fraction of that, which is okay. I mean, the bar is still at the same height, right? Like you can't develop a worse drug just because you raise less money.
But you can do uh the same amount of work if you just make your scientists that many times more efficient than a company than those who work at a big pharma company or at one of these other biotech firms. So the way we decided to do that was by utilizing laboratory automation and computer science. My co-founder and I were both trained in both uh the life sciences and computer science. We went to undergrad together at Carnegie Mellon at a time when CMU was one of only two schools in the country that had an undergraduate program in computational biology. This was long enough ago that people were still saying things like, why would you study those two things? Like they don't go together. So suffice to say it was a little while ago, and thankfully times have changed a little bit, but the idea was that we were going to utilize uh the learnings that we had from that education, plus the fact that as graduate students, we both sort of became autodidacts when it came to lab automation.
So the idea was that every experiment at Emerald was gonna be automated such that it was push a button, walk away. You would never have to have scientists standing in front of an instrument babysitting it, which is generally how things work at companies and uh in a lot of academic labs. Just as like a side note, one of the first instruments we bought was from a it was an instrument for synthesizing short pieces of DNA, and the company that sold it to us was called bioautomation. Uh again, the company's name was bioautomation. And when they installed it, the technician who was installing it asked who the employee was who is going to be running that instrument, and we said, well, no, we're just gonna set it up and let it run.
He was like, oh, well, like the waste bottle here where all the like excess reagents get flushed needs to be emptied every two hours, but like a standard run might take eight or ten hours. So you need to have someone standing in front of this instrument while it's running. Again, the name of this company was bioautomation. Uh so anyway, the so we decided to automate every experiment, so it was push a button, walk away, and we decided to teach every scientist we hired. So organic chemists, molecular biologists, geneticists, you named it, how to write code. And we didn't want them to have to worry about things like memory management or caching or anything like that. So we taught everyone how to write code in what's now called the Wolfram language, but used to be called Mathematica.
It's a very high-level language, um, sort of similar to MATLAB, except that Mathematica is sort of the Cadillac of data processing languages. So it has all these fantastic tools already built in for visualization and plotting, things like that. And so it was a great platform to start our scientists off on. And then we told them like build build tools to solve your own problems. And so the first thing we built was just a data handling platform so that you have all the this great diversity of different instruments in the lab, and usually the data that is generated by them just sort of lives on their hard drive, or maybe you have like a file server that you're just like dumping Excel plots on or something like that. Instead, we built a database where you could push or pull data from any of these instruments onto this database and then access them from you know any connected computer.
So that was nice, and people liked that for about a week, and then they were like, well, once we have this data, then we need to analyze it. So in this relational database, then we built objects for analysis that were then linked to the data objects. And over the course of about three years, this just grew and grew as we needed more functionality to make our research move faster and more efficiently. So we added things like sample tracking, instrumentation diagnostics, uh, inventory management. All of those things were built into the system until the last thing that sort of closed the loop was protocols. We started viewing the protocols that were being run in the lab as just another piece of data in this vast data network.
And once we did that, the lab just emptied out. Now people were only coming into the lab to set up the instruments with the plasticware and samples and buffers that they needed, and they were just running everything from their computer at their desk. So at that point, we decided, okay, there's probably more value in this system than we can capture with just this one internal tool. So we should make this available to all scientists. So we spent the last year and a quarter, year and a half turning this internal tool into an external tool that's accessible all over the world for people to run experiments. So that's the high-level look at what we're doing. And there are a couple important like distinctions to make. So it's been mentioned a little bit earlier, but there are these contract research organizations, which are labs that have scientists that will do contract work.
Uh, it's utilized a lot by pharma companies, and they are it's like a very big business. But it's important to understand that what I'm talking about is not being, it's not outsourcing, it's virtualization. So we are much more similar to Taiwan Semiconductor or Amazon Web Services than we are to any of these CROs. Because with a CRO, you're getting experimentation and consulting sort of bundled together because you're just sort of describing in English to another scientist what you want done. And they're making all these sort of bottom-level decisions about what parameters they're going to select on a given instrument, what's the setting on this photo multiplier tube going to be, things like that. Whereas what we're doing is just turning over all controls of the every sort of scientifically relevant control in the lab to you and letting you design the experiment.
So our goal is we will faithfully execute the experiment that you design, and that all our instruments will be operating nominally when we do so. So that's it's virtualization is this very distinct bucket from outsourcing. The other important thing that is not necessarily obvious is this helps solve the reproducibility issue that we have in science. Because you read these studies like the one the mouse study that Joe cited, or a similar one from Bayer, where they were specifically looking at uh like uh pre-MAUS, so the cellular and molecular biology studies that Joe referred to, and it was similarly like difficult to reproduce. And the issue there is not malfeasance, like maybe a tiny fraction of that might be malfeasance, but most scientists are trying really hard to like convey the results they have gotten and how they got them.
The problem is that when you're doing that, when you're doing everything by hand and you're just writing down the parameters that you think are important to the experiment, you're bound to miss something. Or to just not take into account like what the humidity in the room is or what temperature it is when this reaction is running. There are many different things that people could mean by room temperature. So our goal, because everything in our system happens robotically, every scientifically relevant variable is captured in lab automation and robotics, we have the fully detailed look at what your exper what the world looked like when you ran that experiment. So therefore you can go back and get push button reproducibility because you can just replicate all of those results.
So even things like where a plate is on a liquid handler or how long it is between different aspirate and dispense steps. Maybe none of that matters, but unless you're going to do the grinding of investigating the hundreds of variables that go into every experiment, the important thing is to just record them, and then you know exactly what you need to do to get back to that result. So enough like talking, I'll just sort of show you what this looks like when you're going to design an experiment on the ECL. So there are four relatively obvious steps. So the design step where you are sort of expressing your intention of what experiment you want to run, then we run the experiment, we package it up in this very uh in this linked data network and hand you back access to that so you can not only see the results of your experiment but all the ancillary information that I referred to earlier, and then finally we have a whole suite of analysis tools, because if you're running an order of magnitude or more experiments than you normally would be, just plotting everything in Excel doesn't really do it for you anymore.
So if you're going to design an experiment on the ECL, we have this interface where there's sort of two key parts to it. The first is this graphical user interface here, where you can just design an experiment by point and clicking. Everything reduces to a single line command, which you can also type in over here. And the system is built to educate people as they use it. So as I'm going through and clicking in the graphical interface, you'll see the command being generated here. It's not unlike any other application like Photoshop, where you're going to use the menu system to start, but any power user is just going to learn the hotkeys. So we think of this sort of command line here as the hotkeys. And the idea is that this all needs to be human readable so that it's obvious to you, even if you have no experience with code, uh, what is going on when you're issuing these commands.
So for this example, I'm going to compile a Western blot experiment. That's an experiment for analyzing the protein composition of a cell. And so the system, so you can see over here, after I selected Western blot, it has typed experiment western blot. And now it's going to ask me for the samples and antibodies that I need for that experiment. These come either from samples that you've sent us. There's a whole shipping and receiving logistics aspect to our business, or things that came out of a previous experiment. And now here is the key part, and in many ways the secret sauce of what we've done. Once you've sort of selected the basic things, it presents to you all the possible variables that you could control. And there are lots of them.
This is actually a rather cut down list. For any experiment, there may be 80 or 100 different possible variables that you could change. It's very important that you be able to change all of them, but we didn't want to have to play 20 questions every time you wanted to run an experiment when you had to fill in each field. Instead, the system's relatively intelligent. It's looking at, for example, the antibody you've selected and saying, like, okay, I know that these are pretty good defaults for the wash volumes and wash times based on the antibody that you selected. So it's giving you like intelligent defaults, and we have an entire team called scientific development who are in charge of bringing these experiments online and designing those designing this, the experiments and designing the defaults that are used.
But you do always have the ability to reach in and change things. So let's say you want to run a little less sample for a little longer and with a slightly higher voltage. And then all we have to record over here is the deltas. And this is the entire command that you have to give this system to run that experiment. And it's worth pointing out that this is modeled after scientific communication. If you, if I were to tell someone, like if we worked in the same lab and I were to say, hey, like run a western blot for me, I wouldn't like explain to them how a western blot worked, I would explain the things that need to be changed from our traditional way of running a western blot to go forward. And so that's what we've done here.
We've just specified the diff. So once you've done this, a database succession number gets created, which is the number for this protocol, it gets sent to our queue, and then within 48 hours, it will be running on one of our instruments. And that cycle time is really important. This needs to be like the Amazon prime of life sciences research, because if there's too long between when you submit your experiment and when we do it, it's not going to be useful to you because we want this to be a research and development tool for scientists to use for exploring for exploring possibilities. And if you have to wait a month to get your results back, it's not going to be useful. So I'll so once the experiment is run, the next step is for us to hand you back the data, which is actually the significantly more difficult portion of this exercise, at least from our end.
So here's an example of a data object. This happens to be flow cytometry data. And there are a couple key things to notice about this. So everything is a key value pair, and we have a strict ontology so that regardless of what instrument it comes out of, and scientific instrumentation is sort of notoriously like everyone thinks that they're inventing the standard, and so you end up with 10 different instruments that do the same thing that call everything something a little different. So we use a strict ontology so that we don't have that problem. If it, if we call it side scatter, it's called side scatter, regardless of what instrument it comes off of. And this is all indexed, so it's searchable. So you could say you could do the type of searches on this network that you could never do on the web.
You could say, show me every flow cytometry experiment that was run at a temperature greater than 37 degrees centigrade, right? Like you can't do that kind of query when you're just doing text-based searching. So that's part of the value of the system, but really the important thing here is that everything is connected, because a piece of data in a vacuum is essentially meaningless. And so this piece of data is connected, for example, to the experiment that generated it, and you can get to that just by clicking links. So you should think of this sort of like Wikipedia. You can explore the network of your results just by sort of clicking through links and following lines of inquiry. So here's the experiment, this is the protocol that was run, which includes all the things that we would have selected in the design experiment step, plus the machine readable file that was generated to tell the instrument how to do this experiment, and things like environment that you didn't specify, things like environmental data, what was the temperature and humidity in the room when you ran this experiment.
So all of this is captured here, and then from there, there are two general lines of inquiry. One is about validation, making sure that the instrument is like everything was good with that instrument, and the other is about understanding where your sample has been. So for the former, you can follow this back, and here's the object for the actual instrument that this experiment ran on. And this not only shows you like all the variables you could have set and what the maximum and minimums are, which are important for informing the design, but also all the controls and maintenance experiments. So when's the last time the lines were flushed? When's the last time the detector was changed? And all of this is important because if we're asking people to run experiments with us, it's not just that we have to give them the same experience as they would have using their own lab, but a richer experience.
And so knowing like when the last time the instrument was maintenance, how many that the instrument was operating nominally when you started, those are things that we're able to offer because we can run a research and development environment as if it was a production facility because we're doing so many runs at once. If you were just had one flow cytometry run to do, you might not run all these controls because there might be five or six controls for that instrument, and you just have one sample. So it just does you should do that, but it just isn't practical in, I don't know, any like graduate student or postdoc that did research like that. But because we're running an RD facility like a production environment, it gives us that ability.
And then the last thing I'll show in the remaining few seconds I have is just that this is actually a very dense scale-free network, and we've totally cut it down to make it to make it something that you can like visualize here. But here's the sample that went into that experiment. And here's the experiment where we stain the cells before we could run this flow cytometry experiment. And then here's what the cells looked like before we stained them, and you can just follow this back. So you get this very extremely exquisite detail over where your sample has been and also where it's going. So you'll be able to see the sample as it progresses through your pipeline, and also any analysis that's done on it again is linked into this network.
So we're just at the early Stages here. We have just started letting the first scientists online basically at the very end of Q1 and are sort of slowly ramping up to get capacity so that we can allow more and more scientists online. So that will be happening sort of in the coming months. Thanks.
Page updated August 25, 2026 · transcript and summary are machine-generated