#67 - How AI Helps Make mRNA Vaccines Thermostable

WEBVTT
00:00:00.000 --> 00:00:09.119
The next breakthrough in mRNA vaccines might not be what they protect you against, but how well they withstand room temperature.
00:00:09.359 --> 00:00:11.359
Can AI help us get there?
00:00:20.000 --> 00:00:24.960
Hello and welcome back to Coding Cure, where we discuss decoding health in the age of AI.
00:00:25.280 --> 00:00:26.879
My name is Vasant Sarati.
00:00:27.120 --> 00:00:31.679
I'm an AI researcher and cognitive scientist, and I'm here as always with Laura Hagopian.
00:00:31.839 --> 00:00:35.119
I'm an emergency medicine physician and I work in digital health.
00:00:35.600 --> 00:00:38.159
So I'm very, very excited about this paper.
00:00:38.240 --> 00:00:40.799
I know I say that a lot, but I'm You're excited about every paper.
00:00:40.880 --> 00:00:41.920
You're like, there's a paper?
00:00:42.079 --> 00:00:42.719
I'm excited.
00:00:43.039 --> 00:00:51.759
Well, the yeah, this one I'm excited because it's sort of uh a little like a fresh break from a lot of the discussions we've had recently about LLMs and safety and security.
00:00:52.079 --> 00:00:52.799
Doom and gloom.
00:00:53.039 --> 00:00:54.799
A lot of doom and gloom discussions and such.
00:00:54.880 --> 00:01:02.079
And this is just a genuinely clean, interesting paper with an interesting approach and providing something that's super useful.
00:01:02.399 --> 00:01:12.799
It actually feels like a breakthrough and it feels like um it has the potential to impact so many people at the population level in a really positive way from like the public health standpoint.
00:01:13.120 --> 00:01:13.760
Yes, exactly.
00:01:13.920 --> 00:01:19.359
And I just want to alert the re-uh the listeners that there is no LLM in this approach.
00:01:19.519 --> 00:01:21.760
Uh it is called AI, but there is no LLM.
00:01:21.840 --> 00:01:24.959
There's no, you know, worry about chatbots or anything else.
00:01:25.359 --> 00:01:29.359
I mean, sometimes like the best ideas are some of the older ones.
00:01:29.519 --> 00:01:44.239
Like I think people use artificial intelligence and large language models interchangeably, but a lot of the the older stuff that's out there, I know we're gonna be talking about Bayesian algorithms today, is like, hey, it's really cool, it's really good.
00:01:44.400 --> 00:01:46.879
It fits this use case in particular very well.
00:01:47.120 --> 00:01:53.280
And um, and I I feel like we're just like skirting around what happened in this paper, but yeah, let's get right to it.
00:01:53.519 --> 00:01:54.400
But let's get to it.
00:01:54.560 --> 00:01:54.799
Yeah.
00:01:55.040 --> 00:02:03.040
What they what they did in this paper was they figured out a way to make mRNA vaccines thermostable.
00:02:03.280 --> 00:02:04.400
What does thermostable mean?
00:02:04.879 --> 00:02:05.920
Stable at room temperature.
00:02:06.239 --> 00:02:07.200
Stable at room temperature.
00:02:07.280 --> 00:02:07.439
Okay.
00:02:07.680 --> 00:02:13.039
So the whole so so basically, like if you have, I don't know if you remember back to the days of COVID.
00:02:13.680 --> 00:02:36.719
Uh I know COVID's still around, but when the vaccines were first being developed with mRNA, one of the biggest problems with distributing the vaccines was these, were these cold chain requirements, where the vaccine had to be stored at like very cold temperatures, which means that like it has to be transported at those very cold temperatures.
00:02:36.960 --> 00:02:43.120
When it gets to the spot it's going to, whether that's like a clinic or a pharmacy, it has to be stored at those temperatures.
00:02:43.360 --> 00:02:46.879
You don't want to have it stored at a higher temperature, or it could go bad.
00:02:47.120 --> 00:02:50.319
And so it needs like all of this special packaging.
00:02:50.560 --> 00:02:54.400
It needs like transportation conditions that are amenable to that.
00:02:54.560 --> 00:02:59.439
And you have to monitor it the whole time to make sure it has stayed at those temperatures.
00:02:59.759 --> 00:03:04.240
And so that's like a big deal because not everywhere has that, right?
00:03:04.400 --> 00:03:08.800
Like there are some countries where like you're like, well, I don't even think we could do this.
00:03:09.039 --> 00:03:13.840
And places where they do have that, like we do have this in the United States, it's expensive.
00:03:14.000 --> 00:03:15.280
It's like a big deal.
00:03:15.520 --> 00:03:28.479
And so the COVID vaccine is a great example of a of an mRNA vaccine where if you store it at like a little bit above freezing, it can last for two months.
00:03:29.039 --> 00:03:35.680
And if you store it at a very cold temperature, like negative 50 to negative 15 degrees Celsius, it can last for 10 months.
00:03:36.080 --> 00:03:39.039
But that's the storage temperatures that you're looking at.
00:03:39.360 --> 00:03:51.759
And what happens is when there is this ultra cold temperature requirement, it's really hard to get the vaccine into limited resource settings.
00:03:52.159 --> 00:03:58.159
And there's a lot of waste that happens because things happen along the way where the storage may not be perfect.
00:03:58.400 --> 00:03:59.280
And then you lose the vaccine.
00:03:59.520 --> 00:04:01.439
And now you lose or the vaccine expires.
00:04:01.599 --> 00:04:02.000
Exactly.
00:04:02.240 --> 00:04:02.639
Right.
00:04:02.879 --> 00:04:12.719
And so the thought has always been hey, can we make an mRNA vaccine that is stable at room temperature?
00:04:13.039 --> 00:04:17.120
Because this would make it not only safe, but like scalable.
00:04:17.360 --> 00:04:19.600
Now we can easily transport it.
00:04:19.759 --> 00:04:29.839
It doesn't, it's not so expensive to do so, but we can get it to places that don't have these ultra-cold storage, you know, special coolers or whatever it is, right?
00:04:30.000 --> 00:04:30.480
Right.
00:04:31.120 --> 00:04:37.199
And we want it to be stable at that room temperature for, you know, a period of time.
00:04:37.360 --> 00:04:39.519
So we can transport it and get it there.
00:04:39.920 --> 00:04:47.040
And that way we can have a longer shelf life and it would it would be better, right?
00:04:47.279 --> 00:04:48.319
Right, right, right, right, right.
00:04:48.399 --> 00:04:48.959
No, exactly.
00:04:49.040 --> 00:04:51.519
And so that's the space in which they're working.
00:04:51.759 --> 00:05:01.040
And, you know, I think that one of the challenges is that you have to do a lot of experimentation to figure out how to make it last in room temperatures for a long time.
00:05:01.360 --> 00:05:01.519
Right.
00:05:01.600 --> 00:05:03.839
Like people haven't really figured this out yet.
00:05:03.920 --> 00:05:09.519
There have been ideas, they've tried different things, um, but they don't have an answer.
00:05:09.759 --> 00:05:16.399
And so so the main reason to do this is that you're gonna make it easier to store, you're gonna make it easier to transport.
00:05:16.560 --> 00:05:30.720
But there's also like a secondary reason that goes along with this, which is that um based on the way that they are doing it in this paper, you could actually deliver the vaccine without a needle.
00:05:31.040 --> 00:05:32.079
Oh, interesting.
00:05:32.240 --> 00:05:32.480
Yeah.
00:05:32.639 --> 00:05:33.120
Oh, yes.
00:05:33.199 --> 00:05:36.000
That's that's the part about it being solid state, right?
00:05:36.160 --> 00:05:39.439
So we're versus being a liquid form or whatever.
00:05:39.759 --> 00:05:40.079
Yeah.
00:05:40.240 --> 00:05:50.560
There the solid state version of this can be integrated into these little patches that you can just like put on someone, which is cool on many levels.
00:05:50.800 --> 00:05:55.120
One is like some people do not like needles, so they don't want to get a vaccine with a needle, right?
00:05:55.600 --> 00:06:01.839
But two is when you give a vaccine with a needle, you need someone who knows how to give a vaccine with a needle.
00:06:02.079 --> 00:06:14.240
If you do it with patches, well, now you've expanded vaccine access because you don't need to have a sharp waste container and you don't need to have people who are trained in how to give a vaccine with a needle.
00:06:14.480 --> 00:06:14.800
Got it.
00:06:15.040 --> 00:06:39.600
And so this is really cool from multiple angles because we could, in theory, get the vaccines to more people more easily because they're stable at room temperature and easier to transport, and potentially deliver them in an easier fashion that doesn't require trained personnel or sharp swast and get them to people who are scared to get vaccines because they don't like needles.
00:06:39.839 --> 00:06:40.639
Right, right.
00:06:41.199 --> 00:06:47.920
So I think from like a public health perspective, globally, this is like a really cool concept.
00:06:48.079 --> 00:06:49.519
This is what we want.
00:06:50.000 --> 00:06:55.759
And I'm interested to hear from you, Vasanth, like, how do they use AI to do this?
00:06:55.920 --> 00:07:03.519
Because one of the things that we talk a lot about with AI is like, oh, you need to give it lots of data and it's gonna figure out an answer.
00:07:03.600 --> 00:07:08.079
But in this case, there's actually like not a big, big data set to feed it.
00:07:08.160 --> 00:07:09.920
Like we haven't really done this before.
00:07:10.079 --> 00:07:11.439
Yeah, it's not been successful.
00:07:11.600 --> 00:07:14.560
They've tried some things that didn't work very well.
00:07:14.720 --> 00:07:19.279
So it's like, well, how do you go about thinking using AI to solve a problem like this?
00:07:19.600 --> 00:07:19.839
Right.
00:07:19.920 --> 00:07:21.759
And so I think that's a that's that's it.
00:07:21.839 --> 00:07:23.279
I think that's at the core of the question.
00:07:23.439 --> 00:07:43.600
And I think before we dig into the AI, there's one little tiny piece of biology I think that we have to we have to hit on, which is that the notion of uh excipients, excuse me, uh, that are uh chemicals that are added um or additives that are added to the formulation of the vaccine that isn't part of the drug.
00:07:43.920 --> 00:07:53.680
And um, you know, I I think the example would be things like sugars and salts and polymers and other things um that basically allow you to protect the active ingredient that's in there.
00:07:53.920 --> 00:08:08.079
Um so that's a very cool, important point because the uh paper itself talks about uh four or five different key of key uh varieties of these things, including things like glucose and sucrose and some other things here.
00:08:08.240 --> 00:08:15.839
Um but why I'm talking about these uh incipients um is that that's where the design is.
00:08:16.160 --> 00:08:28.079
People have to figure out that the point of the uh getting the uh vaccine to be the right consistency and the and handle be thermostable and all that stuff depends on the right formulation of these other chemicals.
00:08:28.319 --> 00:08:36.080
And how much you put off those other um, you know, uh compounds is what drives the entire design of this.
00:08:36.480 --> 00:08:45.600
So then in theory, like what we're not changing like the mRNA, you know, the thing that is actually you're vaccinating with, you're changing like what's around it.
00:08:45.840 --> 00:08:46.080
Yes.
00:08:46.240 --> 00:08:52.480
You know, I I I in my head, I don't know if this is the right analogy, but in my head, I think of this as being like when you pickle something, right?
00:08:52.559 --> 00:09:00.240
You're trying to keep it for longer, it's at the very core of it, it's a piece of cucumber that you put in, but then you add all these other things to make it last longer.
00:09:00.720 --> 00:09:05.120
Okay, but like in theory, then also you can pickle like celery and carrots and whatever too.
00:09:05.279 --> 00:09:11.519
So in theory, if like we get a good pickling solution, then it should work in other settings too.
00:09:11.600 --> 00:09:14.240
We should be able to pickle other mRNA vaccines.
00:09:14.480 --> 00:09:15.200
No, no, absolutely.
00:09:15.279 --> 00:09:16.720
That's I mean, that's part of the point, right?
00:09:16.879 --> 00:09:22.720
So you have these, I mean, outside of this, I don't even know if this pickle analogy actually works, but right now it does.
00:09:22.799 --> 00:09:24.000
So let's just go with it.
00:09:24.240 --> 00:09:27.919
But uh, but the core idea is you have all these other things, these uh ingredients.
00:09:28.159 --> 00:09:29.519
It doesn't, you know why it doesn't quite work?
00:09:29.600 --> 00:09:32.399
Because we're talking about getting things into a solid state formulation.
00:09:32.639 --> 00:09:33.919
Oh, I suppose it doesn't really work.
00:09:34.000 --> 00:09:34.080
Yeah.
00:09:34.320 --> 00:09:36.399
And like I'm just thinking about the pick pickle juice.
00:09:36.720 --> 00:09:37.200
Oh boy.
00:09:37.519 --> 00:09:37.759
Yeah.
00:09:38.080 --> 00:09:40.559
Okay, let's let's forget the pickle analogy for a little bit.
00:09:40.720 --> 00:09:42.320
It's I think it's done its job.
00:09:42.480 --> 00:09:53.039
Um, so we have all these chemicals that we want to add, and and I call them chemicals, but it's just these additives that you want to add to the mRNA um core that makes the formulation have certain properties.
00:09:53.279 --> 00:09:55.759
And but you don't know how much to add to these things.
00:09:55.919 --> 00:09:57.360
And that's the design space.
00:09:57.440 --> 00:10:06.320
And what I mean by design space is that you want to decide how much, what you know, like levels you want to add each of these, and different levels perform differently for different purposes.
00:10:06.559 --> 00:10:14.720
And so it's almost like you have all of these knobs that you have to turn to kind of let in the right amount of these different additives uh to make the right recipe.
00:10:15.039 --> 00:10:21.120
So you're like trying to figure out you're like, oh, I have these five additives, like what's the correct ratio to get this in the solid state?
00:10:21.440 --> 00:10:21.600
Yeah.
00:10:21.759 --> 00:10:37.840
So basically what you do is you you you you you you have this um machine, let's say you have this pretend pretend machine where you throw in all the ingredients, if you dry it, bake it, and um you bake it at 37 degrees Celsius for weeks, and then you measure how the drug still works or not, right?
00:10:38.000 --> 00:10:41.759
Because that's that's typically the You're just like going on with the food analogies, yeah.
00:10:42.320 --> 00:10:43.840
Maybe I'm hungry, I don't know.
00:10:44.159 --> 00:11:00.559
Um But but the problem is if if say you came up with some set of ratios of these additives and you put it into this box and you made the uh perfect mixture, you think it's right, it's going to take you weeks before you know that it even worked.
00:11:00.799 --> 00:11:10.399
And then after 37 degrees for several weeks, you realize you get the thing back and it's you have to physically run the thing, and then you figure out that maybe it doesn't work, and then you have to change it again.
00:11:10.720 --> 00:11:14.159
So what you're telling me is it takes you like two months to make cupcakes.
00:11:15.840 --> 00:11:16.159
Yes.
00:11:16.399 --> 00:11:17.679
Which which is about the above.
00:11:17.919 --> 00:11:21.519
I think you need to go back to the pickle analogy because that actually takes a long time.
00:11:22.240 --> 00:11:22.639
It does.
00:11:22.879 --> 00:11:33.120
And so the point is that these evaluations of your design choices are expensive uh because uh each evaluation costs time and money.
00:11:33.279 --> 00:11:35.120
I mean, sorry, uh yeah, time and money.
00:11:35.360 --> 00:11:38.799
And so you have you have to wait a while for it.
00:11:38.960 --> 00:11:50.559
Um and it's not like your traditional machine learning problem where you get a result and you can sort of feed it back through the neural network and update its uh uh values and then do it again really fast.
00:11:50.720 --> 00:11:53.039
You can't do that, you can't do multiple iterations so much.
00:11:53.120 --> 00:11:55.919
You have so the pro so then you want to get it right from the beginning, right?
00:11:56.159 --> 00:12:06.639
You want to either get it right from the beginning or do experimentation fast in a smart way that gets you to the answer as fast as you can possibly get to the answer, right?
00:12:06.879 --> 00:12:10.720
Uh and so you you can't just like blindly search all the possibilities.
00:12:11.039 --> 00:12:13.440
Okay, so how do they sprinkle magic powder on this?
00:12:13.679 --> 00:12:13.919
Right.
00:12:14.000 --> 00:12:15.679
And I mean, that is the question, right?
00:12:15.919 --> 00:12:23.039
Um, you you also don't necessarily, I mean, although you want the best recipe, you don't know what that is upfront.
00:12:23.360 --> 00:12:34.879
This is where I'm this is where the question mark is for me, because a lot of times we talk about inputting tons of data, and here there's like you don't have tons of data to input, and what they've done so far hasn't worked.
00:12:34.960 --> 00:12:38.799
So how is it like magically like figuring out what work what does work?
00:12:39.039 --> 00:12:52.240
Yeah, so their core move is to build this sort of probabilistic model, right, off this otherwise really expensive operation of testing and uh and then do all the searching inside of this model.
00:12:52.799 --> 00:12:57.200
So the probabilistic model is a piece of software, it's also called a Gaussian process.
00:12:57.440 --> 00:13:05.679
And what it is is that at every um measured point, so every time you measure a point, you can add it to that, uh to that model.
00:13:06.159 --> 00:13:14.559
But for every unmeasured point, which you don't have an answer to, whether this when I say measured point, I mean a particular recipe, right, of these additives.
00:13:14.799 --> 00:13:23.679
Uh for every unmeasured point, it gives you a prediction of whether or not it's gonna work and a degree of uncertainty of how much it knows.
00:13:23.759 --> 00:13:31.039
And you can think about it this way the closer you are to the measured points, the more, the lesser the uncertainty and the easier the prediction it is to make, right?
00:13:31.200 --> 00:13:36.720
So you can be a little bit more confident about the points that are near the ones that you have measurements for.
00:13:36.879 --> 00:13:38.720
But at the beginning, you have no measurements.
00:13:38.879 --> 00:13:45.759
So all the points are equally available, which means every single design choice is possible, and everything is equally uncertain.
00:13:45.840 --> 00:13:48.159
So this is a big blank space you start with.
00:13:48.320 --> 00:13:59.840
And there are metrics, there are measures to get things going, but oftentimes there's expert knowledge, there's some um uh data that they may already have to kind of populate that initial set of dots, so to speak, in that space.
00:14:00.080 --> 00:14:04.639
So then once you have that, then the question is how do you get to the next step, right?
00:14:04.799 --> 00:14:06.159
How do you get to the next point?
00:14:06.320 --> 00:14:12.480
Because every time your algorithm proposes a new recipe, you have to go about testing it.
00:14:12.559 --> 00:14:14.799
So it better be proposing really good recipes to be tested.
00:14:15.120 --> 00:14:15.279
Right?
00:14:15.360 --> 00:14:16.960
Like you don't want it to be just a guess.
00:14:17.120 --> 00:14:18.879
You want it to be like an educated guess.
00:14:19.120 --> 00:14:19.360
Yes.
00:14:19.519 --> 00:14:20.960
So now let me ask you this question.
00:14:21.120 --> 00:14:22.799
Have you played the game Guess Who?
00:14:23.120 --> 00:14:24.159
Yeah, obviously.
00:14:24.399 --> 00:14:24.639
Yeah.
00:14:24.799 --> 00:14:34.960
And for our listeners who don't know that game, it's a it's a kid game where you have a bunch of pictures uh on one side where you you have this board where you can sort of pop up different characters.
00:14:35.120 --> 00:14:43.519
There's a grid of characters, and the characters are male, female, some have beards, some of them don't, some of them have blonde hair, brunette hair, some of them have glasses, some of them don't.
00:14:43.679 --> 00:14:50.879
And the idea is the other person asks questions, kind of like 20 questions, asks questions um to eliminate possibilities.
00:14:51.039 --> 00:14:55.039
And the smart questions are the ones that eliminate the most number of cases.
00:14:55.200 --> 00:14:56.559
And the, you know, you know what I mean?
00:14:56.639 --> 00:15:01.279
Like so a first question might be in that game might be, is your person a boy or a girl?
00:15:01.519 --> 00:15:01.679
Right?
00:15:01.919 --> 00:15:02.879
Actually, it's a yes or no.
00:15:02.960 --> 00:15:04.000
It's a yes or no answer.
00:15:04.080 --> 00:15:05.519
So you have to ask, is your person a boy?
00:15:05.759 --> 00:15:06.159
Oh, right, right.
00:15:06.240 --> 00:15:07.200
It's a yes or no game.
00:15:07.279 --> 00:15:07.519
That's right.
00:15:07.600 --> 00:15:08.720
So is your person a boy?
00:15:08.799 --> 00:15:15.120
And the answer to that is informative because uh potentially half the people on the grid are boys and half are girls, so you just right away eliminate half.
00:15:15.279 --> 00:15:21.120
But then it gets increasingly tricky because then you have to ask questions, you know, and and you can't ask very specific questions.
00:15:21.200 --> 00:15:23.840
You can't be like, Does your person have blue eyes?
00:15:24.080 --> 00:15:31.519
Well, the problem with that is if there's a lot of people left on the on the uh to be guessed, then that might just eliminate one person.
00:15:31.840 --> 00:15:37.120
Because the non-blue eyes includes green, brown, gray, black, all the other colors, right?
00:15:37.360 --> 00:15:41.600
So you have to be it's like it's smart about which questions you're asking.
00:15:41.840 --> 00:15:42.080
Yes.
00:15:42.559 --> 00:15:56.159
And the same kind of goes here when they're trying to figure out okay, what's the ratio of glucose, sucrose, L arginine, HCL, et cetera, that goes into the product that we're trying to create that will stabilize the mRNA vaccine.
00:15:56.399 --> 00:15:56.639
Yes.
00:15:56.720 --> 00:16:05.679
So the model needs to be asked the question not where is the best recipe or the best point in that space, but where is the most informative point to measure next?
00:16:05.919 --> 00:16:06.240
Right.
00:16:06.399 --> 00:16:10.240
And there's a there's a something called the acquisition function that kind of captures that aspect.
00:16:10.480 --> 00:16:21.679
So you want to get its predicted value, but you also want to get predicted uncertainty and you want to maximize uh getting the best, most informative answers back when you ask your questions.
00:16:21.919 --> 00:16:30.480
So I think that's kind of the key part here, um, is being able to kind of explore the space in a way that gets at the most informative answers.
00:16:30.879 --> 00:16:45.200
So even though you don't have a ton of data by using this Bayesian probabilistic whatever, you're able to ask the most appropriate questions to then feed the right things into the pipeline to test.
00:16:45.519 --> 00:16:45.759
Yes.
00:16:45.919 --> 00:16:47.120
And and they did exactly.
00:16:47.279 --> 00:16:54.559
And they did uh and so this is there's a mathematical formalism around Bayesian optimization that captures what I just described, kind of sort of in in general terms, right?
00:16:54.639 --> 00:16:58.399
But there's a formal mathematical process to go about getting at this.
00:16:58.720 --> 00:17:08.720
Um, and it's it's so it's under the broad subfield of multi-objective optimization, um, where you're trying to optimize many different parameters at the same time.
00:17:08.880 --> 00:17:13.200
Uh, and in this case, you're trying to optimize for the most information that comes back from the model.
00:17:13.359 --> 00:17:22.000
Um, but there's another problem, which is that let's say the model proposes one point, says, after all of this, I think this is the next experiment you guys should be doing.
00:17:22.160 --> 00:17:24.240
This is the recipe that you guys should be trying.
00:17:24.559 --> 00:17:26.960
Well, each experiment takes weeks again.
00:17:27.039 --> 00:17:27.200
Right.
00:17:27.440 --> 00:17:29.759
And a lab doesn't need to run them one at a time.
00:17:29.839 --> 00:17:31.759
They can run 50 plates at a time, right?
00:17:31.839 --> 00:17:32.319
In parallel.
00:17:32.640 --> 00:17:32.720
Right.
00:17:32.960 --> 00:17:33.920
So give me the top 50.
00:17:34.319 --> 00:17:37.119
So you want to propose the you want to propose 50 at once.
00:17:37.279 --> 00:17:39.599
Now you could take the top 50, right?
00:17:40.000 --> 00:17:45.119
But here's the problem with just taking the top 50 is that it'll always cluster around one promising idea.
00:17:45.440 --> 00:17:46.000
Oh, yeah.
00:17:46.079 --> 00:17:46.319
Yep.
00:17:46.640 --> 00:17:47.920
You want to make sure they're different in that.
00:17:48.160 --> 00:17:50.000
You want to make sure they're diverse enough, right?
00:17:50.079 --> 00:17:58.079
And so there's another technique called diversity guided selection, which they kind of carve out the space uh and deliberately spread the batch a little bit.
00:17:58.240 --> 00:18:05.440
Uh, and so you might not get the ideal, most optimal solutions, but you kind of cover the entire space a little bit better.
00:18:05.680 --> 00:18:07.279
Um, so that's another piece there.
00:18:07.359 --> 00:18:13.200
So then you get back 50 samples that are going to give you information, whether they fail or succeed, right?
00:18:13.359 --> 00:18:15.680
And that's more important than anything else.
00:18:15.839 --> 00:18:26.880
Um, and to find some of these candidates, they can also run some of these things on the model itself thousands of times so that they can get simulated environments, simulated evaluations to come at these uh final candidates.
00:18:27.039 --> 00:18:32.319
Uh, but only those final 50 or whatever gets to actually be turned into a real experiment.
00:18:33.279 --> 00:18:38.160
Um, in general, this falls into the family of AI algorithms called active learning.
00:18:38.400 --> 00:18:47.759
Um, this is the kind of type of things where the AI model, um, and I'm using AI again very broadly, it's not an LLM, it's a machine learning, statistical learning model.
00:18:47.839 --> 00:18:54.400
Um, but then it gets to choose its own training data in a sense, because it's deciding what to do what what next experiment to perform, right?
00:18:54.559 --> 00:18:54.640
Right.
00:18:55.039 --> 00:18:56.880
And that's going to come back as data for itself.
00:18:56.960 --> 00:19:06.240
So it's kind of select, it's kind of choosing the create the the best training data that's going to give it the most information to get to the best model eventually.
00:19:06.480 --> 00:19:13.039
Um, and and so I think that that's kind of the key here in this whole system, uh, the the whole approach here.
00:19:13.200 --> 00:19:14.720
And I think that it's a powerful approach.
00:19:14.799 --> 00:19:17.039
It's been used in other other settings too.
00:19:17.119 --> 00:19:30.799
Um, but like you mentioned at the beginning, it's got that generalizability aspect to it because really the recipe of four or five things could be anything and the active ingredient could be anything, and it's about, you know, it's about manufacturing and doing the right experiments.
00:19:31.119 --> 00:19:31.759
Right, right.
00:19:32.000 --> 00:19:36.480
And what's really cool is okay, they took this from theory into practice, right?
00:19:36.720 --> 00:19:41.839
What they did was they ended up ultimately stabilizing mRNA vaccines.
00:19:42.400 --> 00:19:46.720
That's incredible from this research in ways that they hadn't been able to before.
00:19:46.880 --> 00:19:47.200
Yeah.
00:19:47.440 --> 00:19:54.640
And they were able to prove that they could store it not at two degrees Celsius, not at negative 50 degrees Celsius.
00:19:54.799 --> 00:20:01.440
They were able to store it at 37 degrees Celsius, your body's temperature, for more than two months, and it still worked.
00:20:01.599 --> 00:20:01.759
Yeah.
00:20:02.000 --> 00:20:04.720
And you know how they told they could tell that it still worked?
00:20:04.960 --> 00:20:13.039
Is they gave it to rodents, they gave it to primates, and they said, Hey, are we gonna get an immune response here?
00:20:13.200 --> 00:20:22.640
And they were able to get an antigen-specific immune response, not inferior to what you would get with like a normal vaccine delivery.
00:20:22.799 --> 00:20:31.599
And this included, you know, the regular type of vaccine, but also like we talked about at the beginning, the like special patch, the one without the needles.
00:20:32.000 --> 00:20:42.799
And so I see this as like a really cool study because it has so many practical applications for vaccine distribution throughout the world.
00:20:43.039 --> 00:20:43.359
Yeah.
00:20:43.519 --> 00:20:44.799
Yeah, yeah, exactly.
00:20:44.880 --> 00:20:49.039
And this kind of squarely falls into the world of automated experimentation, right?
00:20:49.119 --> 00:20:53.440
So it's across science, I think, in general, this kind of approach would be powerful.
00:20:53.680 --> 00:20:54.000
Yeah.
00:20:54.160 --> 00:20:58.960
And it's like such a smart use case for AI because it's like, hey, there's a very clear problem.
00:20:59.119 --> 00:21:04.240
We've chosen this specific tool to help solve the problem in this context.
00:21:04.319 --> 00:21:05.759
And it's not an LLM, right?
00:21:05.839 --> 00:21:07.839
We chose a problem that was specific for this.
00:21:08.319 --> 00:21:13.759
We designed our experiment around it, and then we tested if it worked as intended.
00:21:14.000 --> 00:21:15.200
And the answer was yes.
00:21:15.359 --> 00:21:23.680
And now I'm curious, I'm excited to see what the next steps will be in terms of hey, can we ship vaccines to more locations?
00:21:23.839 --> 00:21:28.960
How much does this save us on transportation and refrigeration and ultra cooling costs?
00:21:29.119 --> 00:21:35.440
And also, how many more people can get this vaccine now that we have a needle less option potentially?
00:21:35.599 --> 00:21:39.920
And also, we may not need, you know, trained personnel to give the vaccines.
00:21:40.160 --> 00:21:41.359
Yes, exactly.
00:21:41.599 --> 00:21:42.160
Awesome.
00:21:42.240 --> 00:21:44.720
Well, thank you for joining us today on Code and Cure.
00:21:45.200 --> 00:21:46.079
Thank you.