Summary
Notes
Instructions for HW3 are here:
https://docs.google.com/document/d/1ek2LEZNJg_uxMsX2MXm4123-0NTT-65CAwqqr0-YiXA/edit?usp=sharing
Transcript
and set the stage a little bit. Our goal for the next hour plus, we are gonna go and actually go check out Nebula, which is the world's largest automation system and is the future. It's going to blow your minds.
Again, it's a really epic space. So, and shout outs in advance to Sydney, Michelle and others that are helping. We're going to have for folks online, there's a whole, uh, video experience that we're going to bring you on as we go check out this tour together. So that's great.
And then we also are going to spend some time talking specifically about the homework. So I mentioned this yesterday in class, we're going to be on a multi-week journey between now and the week of October 12th. We do not have class the week of October 12th. So we're basically going to be structuring the homework in the lab this next few weeks around this CloudLap homework. So again, as Ronan said, Rachel shared yesterday, but we're going to kind of say this over and over again.
We're doing a lot of this self-referencing synthesis design. And because of our role, I think, in competition as this kind of education group, I think all of us, and my hope too, is that we all can do a good job together, both here at MIT Harvard, but then also globally in your nodes as it relates to these CIPS designs. So anyway, so we'll spend some time on that. Ron, do you want to do that now or do you want to do the tour first and then do that at the end? I think the tour is kind of fun.
Let's do the tour. Let's just go get going with the tour. Do we feel like we can get an audience that's comfortable? Yeah. You need a few minutes to get this set up or are you ready? We're ready to go. We can kind of start heading over there. We have to get geared up anyway, but on last notes, we can start.
All right, so let's turn on the camera. Let's spotlight whatever camera, if it's Sydney, is the camera. Are we recording this? Are we actually recording this? We are recording. Excellent. Okay, great.
And then. Yeah.
Wait, I'm just saying. Yeah.
Up to you.
Whatever you want. We can pin two cameras.
Two perspectives is fun. Then we'll have both then. I think that'll work really well. Perfect. All right. Should we start moving over to the... Let's do it. All right, everybody. We're going to start heading into the lab.
Okay, so I'm going to remove this spotlight everyone and we're going to switch over to phone spotlight this right in.
Yeah, so we're going to put on lab coats and we're also going to put on some PPE like gloves and goggles and things like that.
We lost contact now.
Okay, whoever is. Okay. Can you guys hear me?
Yeah, we can see it now.
Are we recording this camera? Are you sure we are recording this camera?
Can we spotlight? You can just spotlight that. Yeah, the best way to do it is leave audio. But you need to leave your audio because we're going. Well, it's just got to spread.
Yeah, so I think everyone can hear me right now. I think so. Does someone mind turning off the audio in that room? Hey, Alex. Can you turn off the audio in that room? Thanks, y'all, for going on this journey with us. Give it a shot. Okay, we're going this way.
Going this way? Yeah. This is a funny site.
-We're not going in here.
We're not going in there. No, no, no, no, we gotta gown up. We have to gown up over here.
Yeah.
Come on in. So grab a lab coat over here.
Or you can see the live stream. Go ahead to the rack on to the right. Yeah, those are the slide cards.
Yes, everything is, yeah, please video as much as you can. So behind here, this is called Nebula. So this is Ginkgo's autonomous lab. So it's trying to automate wet lab work in general so you can actually turn scientific experiments into code. So this system has like over a hundred racks now. So this is called a rack. Each of these like individual units, they're all connected together with this MagnaMotion track.
We'll walk around and you can see it. But anyway, each of these racks contains at least one piece of equipment, like lab equipment. In this case, it's like an X-peel, so it's like peeling microplates. And it's also got this robotic, like... six axis arm that can actually pick up from the track and put it onto the piece of labware. And then what's cool is that every piece of labware is connected through a centralized system that's all programmable through one interface.
So you don't have to know every vendor software to program this entire system. It's actually really fun, right? You can vibe code an experiment in a weekend and have it running the next day. Anyway, above this is like a HEPA filter. It's blowing air down to keep your samples clean. Yeah, lab coats. Let's start taking like a loop around because I think you'll start to get a feel for what it looks like.
That's Piyon, he's an automation engineer here. He's very cool. You can ask him questions if you want. OK, so you can kind of see down this track, right? This is the MagnaMotion track. There are little pucks that travel around. And that's what actually move your samples from each piece of equipment. So, and then each setup has little like buffer nests. So if you have like intermediate steps in a protocol, you can send your plate to like little locations and get easily accessible to the lab equipment.
Also, we can like talk, you can ask questions. It doesn't have to be me monologuing the whole time. Um, Yeah, so there's like 105 racks on this system right now. I think some of them have like multiple pieces of equipment. So it's around like 157 or something unique instruments. Yeah.
That's a great, yeah, that's a good point. Yeah, so you can, all of the like experiments that you submit to the scheduler, get interleaved together. So you can like lots of scientists, like on any given day, we have 50 to 100 scientists, right? Submitting protocols, all of those experiments are happening simultaneously. And like all, you know, handled by the scheduler system, which is like a non-trivial, right? Like computational problem.
And when you think about, you know, like pathfinding algorithms and also like what happens if something goes wrong, right? Can you have AI handle that error or does a human have to be involved in that loop? And then how do you schedule things around an error? So anyway, there's lots of considerations. We'll talk about a bunch more during the semester. But like the whole idea is that you can have lots and lots of scientists collaborating, running experiments at the same time on the same equipment rather than having labs where just a single like, you know, expensive piece of hardware just sitting around not being used half the time.
And so that's like one of the major benefits is this runs 24-7. So like every night we're launching experiments that are running into the morning. On the weekends, we're getting data back all the time and analyzing it. which is the loop that you all will be part of, right? You'll be launching cell-free experiments, getting the fluorescent data back And then you'll analyze it. We'll plan the next round together.
So what we'll talk about is like some of these liquid handlers, for example, like right here, this is a multi-flow. So this is for like bulk dispensing liquid. So this is set up every day with fresh like ethanol and PBS and water. And so then at any given moment, you are free to send your plate programmatically via the API over here to have it like dispense liquid into. rather than having to like go load that yourself or like prep this equipment.
So can you guys hear okay back there? We're good. Is everybody hearing okay? So this, what we're building out at MIT, there's a hundred of these robots here. Like there's a lot of robots here. At MIT, we're going to have six robots. Okay. Which it's not a hundred, but If you think about a general kind of biology experiment, Every experiment has a certain number of instruments that you're going to take a biological sample and put it into, right?
Here's a Ferris star, here you go. Perfect. So it's not that exciting to look at, it's just a box. The magic of what is inside this box is, in fact, a... a system where there is a Different frequencies of light that are being exposed to the sample. And fluorescence basically involves what's called an excitation frequency, where there's a certain frequency that you're exposing the sample to, and then an excitation frequency.
Like the sample then will, and in this case it's usually proteins, will generate a frequency that comes out that you can read. In our case, and in the example of what we're doing with the Ginkgo homework, one of the first proteins that we're most interested in is called SGFP, which is, that's his first... Super folder, yeah. Super folder, yeah. So green fluorescent protein, again, is maybe the most important fluorescent protein in the history of the life sciences because of how widely it is used for just about everything.
Disulfide bonds, yeah, that's right. What I'm getting at is it's not only just that you translated it, it's actually folded into the appropriate three-dimensional geometry. Okay. And also what's called the chromophore, which is the part of the protein where the excitation and expression parts of the process are working. So if you get a green signal and you're trying to make GFP, that's what that tells you.
And we'll use the same technology, right? For retiplase, it's like a clinical therapeutic enzyme using a different excitation and emission wavelength, but it's still fluorescence. Yeah.
So sometimes you'll wanna like start a culture of E. coli or something, right, in a plate. So you could send your deep wall plate, say, here, and then you could get a breathable seal. So that gives you that aeration that you might need while it's incubating, right? So that's what like this is just for sealing plates with a certain type of seal. This is another plate reader. These are thermal cyclers.
So if you wanna run like PCR reactions, right? Or anything that's like needs a thermal cycler, or maybe you just wanna like do a heat shock. Maybe this is your mode of doing a heat shock. You could set these in advance, have your plate arrive, go into the cold or heat, go between them. It's pretty open-ended, which is the fun part. In 10 minutes, you can help us take a picture.
This is a plate shaker incubator. Thank you so much. No plates in it right now. But anyway, you can have this set to like 30 C and it just shakes them together. This is another liquid dispenser for ball-free agents. You can kind of see on the track right here, these are like the pucks and they're each carrying a certain micro plate. So like these could all be for different experiments that are running simultaneously.
This is a stereo store. So this is keeping all the plates at four degrees. So you can kind of see it like spinning actually. So, I don't know. We'll see if there's a good play in there. But all of the plates will be barcoded on the system. So at any given time, like in this stacker, it gets scanned when it enters. And so you always know where your labware is at any given moment. It's actually kind of fun because when we run a protocol, we'll be able to watch the camera footage back and see the entire like...
This is Tundra store, so this is like20 it was like cold storage and So in terms of the protocols we'll be launching, we'll load our lysate plate and our substrate plate in at this temperature. So we can keep it like Semi-frozen This is another plate reader. This is like a plate elevator. They were using this during like COVID testing days. They could send samples across the other room. Not to this but there's like a like that other high side of the hallway. Yeah.
These are a little bit harder to see, but these are ECHOs. So these are acoustic liquid handlers. So this is what we'll actually use to go from like a reagent plate with like 50 plus reagents at like 50 microliter volumes. And we'll like use the acoustic dispensing capabilities to do 25 nanoliter droplets into the destination plate, which will be your custom designs. So these are really fun actually.
Probably my favorite instrument on the system.
This is a centrifuge. There's a lot of these high Gs on the system. So what's also nice is they come with like balance plates next to it. So if you send a plate over, it'll just choose an appropriate balance for your plate and spin it down. More X-feels, so if you're unsealing plates, that's here. Lots of Bravo's on the system. So this is another Bravo for liquid handling. There's 96 well variants and there's 384 well variants depending on the workflow you're doing.
There's also a variant that we're using right now called like a 96ST head. It's really cool because it lets us send like a 384 tip box, but then we can take quadrants out of that tip box. So like for our experiments, we'll actually set up the reactions in a 384 well plate, and then we'll do like stamps into a deep well plate of 96 well deep well plates. And then all of those will go separately to get incubated. There's another multi-plow.
One of the important like EHS aspects of this is that you just-EHS is environmental health and safety. Thank you, yeah. So like for the safety aspects, we need to dispose of the plates and the liquid and the media in appropriate locations. So we actually have like different trash racks on the system. It's like this one I think has like solvent waste in it. That's like going in here. Um, Yeah, it's nice because you can just say like, you know, go to this trash and it'll just figure out the right trash bin to put it in for you.
You all are going to be the first people to use them. Yeah, for sure.
My hope is before the end of October to run an initial experiment with the class in the MIT lab.
Let's do it. We got 10 minutes. We'll get this in. by sharing the actual homework assignment linked on So let's start. Okay, so this is where you are actively designing a brand new custom reaction composition for cell 3. So you guys are going to set the stage for a gold standard of production of these enzymes. Here's the rough timeline that we're planning on. So you had Rachel and give her lecture. This is October 1st. We've done the field trip now. We're gonna talk through the assignment.
And really quick, Kate Onamala, she's a professor from the University of Minnesota, expert and legend in self-referencing synthesis. And it's also leading one of the labs that is part of Teen Humans in the co-op edition. So she's actually one of our, it all just worked out timing wise. We scheduled Kate in advance anyways, but it turns out she's one of the labs. So, well, this will be like lecture slash initial group meeting amongst our team.
Yeah, this is going to be so much fun. Anyway, okay. So then Thursday next week, we'll have a dedicated recitation time where I'll be on a call. I'll talk about the ideas you all are having. And then following that on the 20th, we'll have all the designs due.
And also as stated here, the week of the 12th, there's no class, but we're going to schedule a bunch of opportunities for everybody to be able to do some virtual office hours, share designs with each other, et cetera. We may just take the existing lecture time and also even the recitation time on Thursday just to be optional time for folks to gather and discuss. But we'll find that out a little bit next week.
OK, so thank you Michelle for writing this nice introduction to the experiments we'll be planning. This is the high-level overview of what's going on. So we're going to do around 1,280 reactions in pre-84 well plates. We will take the best performers, the medium performers, whatever we decide, and go into 96 well plates. So it will still be 20 microliter reactions, but they'll be in deep well plates, so they'll get more oxygenation.
And then the final test, like after all these iterations, we'll do like two mil tubes, which is kind of the like conventional industry standard. This is your overview. I think all of you are familiar with the central dogma now going from DNA to mRNA to protein. So this is what's happening in your cell-free reactions. And yeah, so Here is some, you know, Discussion on each of the proteins that we'll be talking about, degrading plastic, red atlase, like therapeutic or.
I think the key nuance to know is that there's a different type of bond that we're trying to have fold correctly in Reddit plays. which makes it like conventionally a very challenging protein to express in a cell-free environment. So we actually need to be in oxidizing condition, which is different from like the typical, when you break open a cell, the cytoplasm is in like a reducing state. So that's why even the lysate itself is prepared differently.
You're not familiar with reduction oxidation chemistry. Hopefully there's a link to something at the bottom of this, but that would be good for you to look at.
But don't worry, it gets more exciting because you will also have a chance to come up with your own ideas for what we should add to these reactions.
You can also suggest your own new Lego blocks to add in there. Here's the actual assignment. So this is where you're gonna start. You'll start by just looking over some literature, comparing some of the different types of common reactions, right? There are some that are more expensive that are like there's trade-offs, right? You can make a lot of protein up front, but then the reaction kind of collapses or you can build systems that are more sustainable, but maybe don't contribute the most production at the start. So it'll kind of think through some of those trade-offs here.
I made a dedicated web interface for us to all collaborate on ideas. This is nice because you can just log in with your Key Cloak account and can all share ideas.
The reason it has to be behind an off login system is because we don't want OpenAI seeing and scraping all our good ideas. It's not competition.
but we'll need some sort of voting system. You can get creative too. So that's the last part of this assignment is I would say the most difficult part. The challenge, okay, start by like making the plasmids that we'll be expressing in Asmall Kernel, since you all have access to this really cool platform. But the nuance and most interesting aspect here is what if you could design DNA that encodes for a different protein or enzyme chaperone to facilitate the folding of reddit bytes?
That I think is the most interesting aspect here is if you could design a really interesting, let's say a mutated version of a disulfide bond isomerase to spike into this reaction. Now, it's trade-off because some of the cellular machinery is going to start producing that instead of your protein. But if that's helping you produce more actively folded protein, it could win out in the end.
proteins that are not the protein that you want to bold that help the protein that you want to fold. So they're like bites. They're like in there helping out. So what I'm saying is you can also, as a part of this plasmid, express some chaperones that are going to be helper proteins to actually help
So what I'm saying is you can also, as a part of this plasmid, express some chaperones that are going to be helper proteins to actually help fold the thing that you want to fold. So you don't just have to add salt or energy supplements, right? You can also add DNA. Mm-hmm. Please use AI, documented use of AI. That's like encouraged in this assignment, right? We want to show how humans are collaborating with AI systems. So yeah, come up with cool ideas.
Yeah, so this assignment is gonna set you up just to get used to the website, the interface, how do I submit designs. Submit your first pass at maybe a variation on a known reaction formulation.
Probably multiple. Because, yeah, so the way we have this right now is we're going to do 320 unique designs for each protein. So we'll have somewhere in the range of-960 total reactions split across the class. Okay. And then, and then is there like, as the students generate their designs, is there a way that we are then evaluating those reactions? Like we as either instructors or like, you know, Well, can you say a little bit more about that, what that might look like? Yeah, I think so. This is part of the process that we have to define, right? How does a large community of scientists decide on what are good experiments to run?
Typically, what we've done with the opening eye models is an adversarial agent upgrades the ideas, and then it will down select two of the ones that wants to run in reality. In this case, I think we also want to like maximize human contributions. So we will also lean towards, you know, the most amount of scientists running experiments as possible. But we will need to define this and likely I think what the scenario ends up being is that the TA team on like perhaps a call with everyone goes through and identifies which compositions will be the ones that we move forward with. Based on what?
Probably based on the breadth of search here. We don't want to run reactions that are almost identical. If you have the same concept, and it's like 25 nanometers in one agent, probably not worth testing both of those. Especially we'll want to lean towards testing new ideas. Reactions that use new reagents or approaches will be favored.
right exactly what i'm saying about the selection of the actual reactions that go in the final robots okay and um for sure i think the instructor team we will all be chatting about this but if any of you as students have ideas on rubrics metrics ways to evaluate the collection of student designs. So this is a meta-science problem. If you're interested in that, let us know, because we will definitely be doing some meta-science, certainly at the TA level.
But if you're interested as a student in helping with that, think about that, then awesome.
Have these three proteins become your best friend. You learn about them. know what they like, know what they don't like. They're your new buds. And really, because each of them, they have a personality. They genuinely have things that are like, oh, too much salt is not good for me. I like having these chaperone best friends around. They go to the party together. It's like they all have their own personalities.
I could see that being a really interesting extension to this project, but in this case, we are keeping the YSAE consistent. Okay.
So this I would describe as more quote unquote domesticated, right? Cultivarium, which is going to be our last formal... Henry Lee from Cultivarium is going to be our last speaker, but Cultivarium is actually going to come to class in like two weeks, I think, or three weeks. They are a company that is working specifically on what are called non-model organisms. Because again, if you look at the whole spectrum of the living world, I mean, there's...
But it's not like there aren't other things that could be as good or even way better. So, Caldivarium, they're dedicated to trying to find these non-model organisms. And so, actually, they're going to come in a few weeks, and they're going to give to you all, again, this is like very, you know, kind of valuable, interesting industry things that you're all going to get free access to.
So, If you feel inspired to write about this in the homework, don't be limited by the questions that are in section three. These are the ones that we want you to definitely think about. But if you do want to think about maybe different types of life saves, definitely feel free to write about that and do some thinking. And it might even be useful for your final project there. Well, and even just to say too, it's like, let's say, Caleb, to your question about mammalian, you might go down a rabbit hole with mammalian in one of these proteins and then realize, hey, actually a mammalian system might express better than a bacterial system. And then you might be like, well, what features of the mammalian system are actually making it better?
10 minutes past the hour. Any other closing questions, comments, thoughts? This was actually related to the lecture last Tuesday. So Speaking of the fact that you're in the BioWorks, you have a collaboration about an AI, right? And most recently, the news was the idea of a hugging face incident that these agents ran free. And I guess the question is, what would be... I mean, at some point, what happens if they run free and try to synthesize some sort of thing that you can kill a lot of people? I don't know, a very effective COVID strain that could be 10 times worse or 100 times worse. Yes.
Is that like something that, maybe not for ginkgo biobots, but something that as we use AI more and more, something that would be sort of important? Yeah.
Because there's a lot of data out there for like even strains of things that could be very dangerous, taking advantage of a biology immune system. Yeah, no, I think it's a very pertinent and timely question. In this case, I'm not worried about it because we have constrained the experimental space to not be dangerous. We know that we're only modulating these reagents which are safe at hand, though. We know that we're only loading this amount of E. coli on the system.
We're not allowing for like different recombinant proteins to be made here. Of course, you all will be able to suggest DNA, right? But that's all being screened at the end of the day by a human. So there's always a human in the loop right now. But I think your point is very valid and important going forward when we think about fully autonomous AI-driven labs, right?
And then that's like the start of the movie. But in a very real way, though, like there are a couple of things that I think are important. Number one, as you all are. like learning and will continue to learn biology is hard biology is like really really hard like it's like getting things to work in life and just to do stuff is really really hard and like like this whole like working with domesticated organisms to do stuff and like trying to do stuff that is like not with this very very small set of like finite things that we work with in biology it's just all really hard like darpa used to have this program called bricks which is like biological robustness and complex settings i was in the program specifically because you can have you can throw millions and millions of dollars and like on planet Earth to engineer some strain, a bacteria, do some really cool thing.
And the thing, as soon as you put it in any other environment other than the highly specific media, growth conditions, you have set up there, it dies instantly because it cannot survive relative to anything else. So for now at least, we still exist in an era where a lot of biology is really, really hard. So that's just one point. Second thing is, to Rona's point about what's actually run in these labs, it's not like the labs can automatically, I mean, for now, you can imagine a future where they could synthesize their own DNA. That actually is obviously crazy.
If you had a lab that actually had a storage thing with a huge number of different types of cells and a built-in DNA synthesizer built into the automation, that's That would be problematic. We do not have that now. We're still, I think, a long ways from that future. For now, it's still like you still have to put stuff in the system and humans are putting stuff in the system.
And so for me, you know, it's like a lot of the risks that people are talking about, certainly related to autonomous labs. or at least cloud labs. Cloud labs, in my view, are so much safer because they are safeguarded by people. It's not like you've got molecules all around the state of Massachusetts or all over the world. They're in one room. They're over there. Or in our case, like in Texas, they're going to be in one physical location.
So the molecules can have a tremendous amount of IT security, biosafety, biosecurity built around them so that both from the design perspective of getting anything into that lab, and then also if anything starts to get produced, that you can shut the thing down, know what's happening, et cetera. I think far scarier are just people in labs all around the world right now without any robot that just have access to all kinds of crazy foundation models and whatever that could just start doing things.
So I think the autonomous lab thing, it's like, yeah, it's a concern, but relative to other actual security threats, I think it's just not... For me personally, it's like on the radar like yet now if you do your point start building in DNA synthesizers into the thing and like have super shoddy network security so that like anybody on the internet could pack the thing or whatever. But then even there it's like containment, how to get it out. You still need like some like human co-conspirator like you're saying that yesterday.
So Ginkgo's position right now is data in, data out. So if you suggest something, the experiment gets run, you are not receiving biological material at the end of the day. Exactly. It's all saved here. It's actually kind of contained.
Instructions for HW3 are here:
https://docs.google.com/document/d/1ek2LEZNJg_uxMsX2MXm4123-0NTT-65CAwqqr0-YiXA/edit?usp=sharing
Pertemuan 7 — Study-Ready: Nebula & Desain Reaksi Cell-Free