Can AI Beat Cancer? Special Interview with MGH Researcher Peter Mikhael

Added: 2026-05-08

[00:00:02][music] [music] Okay, good evening everybody. I'm back behind the desk. It's been a while for variety of reasons as you guys in the audience in the regulars understand, but it feels good to be back behind the main desk. And I have a very special guest tonight with me and his name is Peter Michael and he is he is a researcher uh Harvard I'm not sure if it's Harvard actually we'll talk to him in a second but he the area that he researches is something that is very very interesting to me and has been for a long time because as you guys know I'm interested in all things AI and then I have personal reasons for being interested in things that are AI related to cancer answer. So putting those two together is something that's fascinated to me and especially with recent developments in the last few years with the large language models like chat GPT and stuff like that. So let me bring in Peter and introduce him. And I apologize for anybody who I apparently slipped up and put the wrong date. So we may not have it it'll probably take a little while for the live audience to get caught up to us here. But let me let me bring in Peter. Hey, Peter. Good to see you.

[00:01:30]>> Hello. I I can't actually hear you for some second for some reason.

[00:01:34]>> Uh, you can't hear me now. Can you hear me now?

[00:01:41]>> Can you hear me?

[00:01:43]>> I say for a second.

[00:01:53]Peter, can you can you hear anything at all?

[00:01:57]Uh, I'm going to switch.

[00:02:00]>> Should I go out and come back in?

[00:02:02]>> No, stay right there. I'm going to switch to a different computer. Um, >> hello.

[00:02:18]>> One second.

[00:02:20]Yellow cottage tails, can you can you hear me or is it my am I the problem?

[00:03:36]Hey, Peter. Can you hear me?

[00:03:40]>> Can you Can you hear anything, Peter?

[00:03:54]this. Hold on a second.

[00:04:06]You can't hear anything.

[00:04:17]Can you hear anything at all?

[00:04:22]I don't >> Are there people here or is it um let me go to email?

[00:04:38]Yeah, I can't hear Kevin, but you but you guys are able to.

[00:04:45]Maybe I leave and come back.

[00:06:01]Check. Um, Peter, I know you can't hear me. I think the pro. Do you hear this now?

[00:06:14]Okay. Can you hear me?

[00:06:17]Uh, let me see. Oh, can you hear me?

[00:06:19]>> Yeah.

[00:06:20]>> Let's see if we can hear you.

[00:06:21]>> I can hear you. Can you hear me? I >> don't seem to hear you, but I'm not sure if you're saying anything.

[00:06:26]>> I can hear I can hear you.

[00:06:27]>> I don't hear you.

[00:06:29]>> Oh my god.

[00:06:29]>> You got to make sure you hook You got to go into settings and make sure your mic you hooked up to the right mic and the right speakers.

[00:06:36]>> Yeah. Um, >> so settings is the down at the bottom there's like a gear wheel and if you click that then you open up audio.

[00:06:53]>> Audio I can I'm It says good.

[00:06:56]>> See I'm not hearing Are you guys hearing Peter?

[00:07:00]>> So Peter can hear me. So we've made some progress but I can't hear him anymore.

[00:07:04]And I did hear him before. Yeah.

[00:07:08]Hello. Can others hear me?

[00:07:11]>> And I've emailed him my phone number a few times, but he's not calling me, so I can't communicate with him.

[00:07:18]>> Maybe I'll just uh >> um if you check your email, Peter.

[00:07:24]>> Yeah.

[00:07:33]>> You here, Peter?

[00:07:35]Yeah, I can I can hear you guys.

[00:07:39]>> Hey, Peter.

[00:07:41]>> What What was the problem with you just not hooked up or >> went to the >> Let me try. Uh, >> let's see. It says I'm hooked up. I don't know why.

[00:07:55]>> So, they can't hear us, but you can't hear me.

[00:07:59]>> I can't hear you. Yeah.

[00:08:02]Um, >> go back out.

[00:08:09]>> Let me try take Let me try going Let me try one more thing and I'll try to go with the computer phone speakers.

[00:08:29]Okay, try that. Can you Hello.

[00:08:33]>> I can hear you. Okay.

[00:08:34]>> Oh, fantastic.

[00:08:38]>> All right. I can hear you over the uh >> Thanks, Peter. Thanks for taking over the technical difficulties.

[00:08:47]>> Yes. Well, I mean, I think it it has to happen when you're at a tech school, you almost always have to fail in technology.

[00:08:55]>> I know. I know. When I used to go on Court TV, they had technical issues all the time, too. Sometimes we were trying to work them out right up until 30 seconds before I would get on the air.

[00:09:04]It was uh nerve-wracking. Um, okay. So, everybody, this is Peter Michael and he is a researcher. Why don't you tell us just a little bit about yourself, Peter, and then I'll get to my first questions.

[00:09:18]>> Sure. Um, yes. Well, uh, nice to to see to see you and whoever is listening. Um, uh, yeah. So, my name is Peter Michael.

[00:09:27]I'm a sort of a grad student at MIT. So I'm a PhD student there. Um and um you know we I'm in a research group that works in a lot of AI applications to healthcare and drug discovery um essentially trying to bring that technology to um improve whether it's diagnostics or treatments or what have you in that space. And so part of the projects that we've been incredibly passionate about over the many past years has been sort of you know saying well how far in advance can you detect cancer um and so that's where some of our work uh that does that for mamography for breast cancer and CT for lung cancer where we developed a bunch of um you know computer vision models that look at images and try to forecast essentially the risk um for for and like you know I've been always interested in the medical field um and you know didn't necessarily know that I'd be doing a like working in it from the perspective of technology but here we are today um and it's very exciting.

[00:10:34]>> Yeah I see that there were two basic general types of areas where AI is helping you do the research. one you just mentioned with the uh early screening and so you could go into how AI helps you with that and then we'll go into the area that really fascinates me which is how AI is speeding up research into drug development treatments and things like that. So how is AI being used now generally but more specifically with the tools that you're using or developing to help with early screening?

[00:11:06]>> Yeah. And um so like the screening is sort of like a I think a problem we we all are generally aware of. The idea is that like if you can for many cancers specifically if you can detect them early on your chances of surviving them is is pretty [clears throat] high right and so u I mean at the same time we have developments and on on the treatment side for different stages but it it remains to be at least today the fact that like if you can uh detect as early as possible then just survival rates are really really high because of the interventions we can do today. And so the question then is how how can you detect things early on and for a a few types of cancers we have these im you imaging that is done for for women you have mamograms that you're supposed to do every year in the United States um for lung which is not as sort of prominently screened but if someone is a smoker that you're supposed to go get a chest CT um if you're a heavy smoker >> what age um Peter >> um also For mammo, it's usually, you know, 30 to 40 depending on your, you know, history up until you're in the 70s, 80. And then for [clears throat] for lung cancer is sort of similar sort of 30s um up until you're like in the mid70s usually. And but that there is also the criteria because the way that this has been studied in the past where we focused um our efforts on smokers because they are at an elevated risk. Um and but you you can imagine that this would be true for many people, right?

[00:12:45]Some people get cancer without ever having smoked.

[00:12:48]>> I was never a smoker myself.

[00:12:49]>> Yeah, >> exactly. Um so clearly we we have a lot of sort of room to grow in terms of how we screen, who we screen, how often. Um and so that's sort of where the technology comes into play. Uh and and if you think about risk sort of which is the flip side of this is like you screen people who should be gaining the most from the of the fact of your screening right because we screen a lot of people many of them will never benefit from this they just go and get an exam but will never get any disease and sometimes we are not catching the right ones so some people just fall through that screening program and so the question is how can you identify who really needs it and how often they should be getting Um and and this isn't necessarily a new idea. So like risk as a general concept has always existed, right? We say people have a family history are at risk and maybe we should be looking at them more often. Um people who have particular sort of exposures are at risk, right? Um you know, if you're someone who works within the um like a fire department, you are exposed also to more carbon generally. So you don't have to be a smoker. Um asbestous, right? So all sorts of things that we sort of consider as risk but those have been generally not very predictive you know their performance so like how good are these um heristics is relatively weak and the advent of sort of a lot of the AI technology that has been developing over several past years has allowed us to say well what if we take these images that people already do so you go to your doctor and you get screened um and then can On top of just saying whether you're healthy today or not, sort of assess to what extent your chances of getting cancer over the next few years um is is it elevated relative to the population or or you know average or low. Um, and what it it turns out that like these these sort of models when you train them on the right data and large data and um with more modern techniques can really detect a lot of small stuff uh that that you know indicate that someone is at a high risk versus not. And and when you compute, you know, how how well how well does that do? Uh it tends to do very well uh in the sense like it can catch cancers that maybe sometimes um you know the eye doesn't um it can determine that some people are low risk and uh they are generally healthy and so and so you can start to think about we can deploy this and use this to get the right people through the screening program. Um follow them more often when we need to. Um there's also talk of you know in the breast cancer world there's all sorts of interventions like chemoprevention um as well. So that's sort of where we're at.

[00:15:37]Um and you know we can go through how you develop such a model with the performance and how you test it and all these things but the technology has allowed us to go from sort of heruristics where you know you maybe use all you know is your family history and you say well I know this my cousin got cancer and so I'm going to just take care of myself versus looking specifically at your you know medical data and being able to say well you are at a higher risk and we're you know we have good performance on that. Um, [snorts] >> right. That's like we've seen some women that are told they're extremely high risk for breast cancer just as a preventive measure go ahead and have their breasts removed.

[00:16:17]>> Yes.

[00:16:18]>> So you're that would be something similar to I I had never heard of preventive chemo, but that sounds like what you're talking about there.

[00:16:25]>> Yeah. I mean, so that's like quite extreme. Uh, chemo prevention does exist as well. So you can give there are some drugs that are given um because someone is at a you know high enough of a risk um that they will take some drugs to presumably lower their chances of getting cancer.

[00:16:42]>> Yeah. Again it's it's very specific to certain diseases but but that's what we're talking about is like allocating those resources and figuring out who really needs them is the question that we're trying to go after. Um and and it a lot of this ends up become being uh a question of just how good can you be right so right are you 100% accurate or are you 60% accurate and at what level um do you start changing your guideline >> uh to reflect the sort of improvement that we have gained. So in the past, you know, if we look at 10 years ago, 15 years ago, the accuracy of the models that we had was around maybe in the 60s, maybe, you know, high 60s, low7s at best. But now we're starting to put like, you know, 70s, maybe in the 80s as well as how accurate these models are.

[00:17:30]And so that's like a that's a step function, right? Going from an accuracy that's like decent to solid or strong.

[00:17:38]Mhm. Now before I get into drug development, I wanted to talk to you about your you have a paper that you co-authored on the amino acid called methionine >> methionine. Yes. Yes. Long >> Yeah. Please talk about that a little bit if you could.

[00:17:54]>> Sure. Um so I mean this is like stuff that I did in my undergraduate uh and this is a bit um I mean it's a very interesting field you know it's and it touches on many things touches on cancer touches on diet touches on all sorts of um topics but the core idea there was and we're talk we're in the field of metabolism so there's no AI at this point this is like a 2018 19 project >> [snorts] >> um where we weren't working on But the core idea was you know as as as far as as you know how medicine does it right now is we think of these drug like changing uh your health by giving you sort of you know drugs that are very different from what you're used to taking right like you don't and your natural food are not taking a um a heavy drug um as part of it >> and so some of the questions have been well we know that in in cancer specifically how cancer uh sort of utilizes its nutrients is different from how a healthy cell uses its nutrients.

[00:19:03]So literally the metabolism the what a cancer eats and how it uses what it eats is different because it's growing so fast it require it has many different needs than a healthy cell. And so once you start teasing into well what are those kinds of needs right it needs maybe more of these particular nutrients than others it's you know it's making um it's dividing very quickly so it needs to build up itself very quickly um and so it starts hijacking the machinery that you would usually have for a healthy cell in order to make DNA very quickly to divide make the cell sort of components to make many more of it and so one of the things that gets affected by this are, you know, these molecules, the nutrients, and one of them is methionine. Um, and so in this in this case, we're we're trying to figure out, well, if you remove methane from a diet, um, what what effect does that have on the cancer versus healthy sort of state?

[00:20:04]And and you can you can wonder why you would pick that. You would pick that because you see some differences between the two sort of kinds of cells cancer versus disease and normal. But you also actually there you know papers in the past where people used to do mice studies and they would do methane restriction. So you would put them on a diet where they don't get this amino acid. It's one small molecule and you'd say and you'd see some longevity kind of uh uh you know effects where the some of the mice would live a bit longer and a bit healthier. And so the question is what happens when you do this um >> with humans >> in humans and in cancer. And so that was an interesting study because it was studied on cell on like cells in a petri dish as you do and maybe biologists do.

[00:20:49]It was also studied in mice. And then we all it also was studied to some extent in humans where you know we had human subjects who would just eat particular foods that had lower methane versus you know the normal amount and you can try to like see does this effect sort of um you know translate or is there any consistency >> um and usually >> when I looked it up it looks like it it looks like it's pretty difficult to cut that out of your diet.

[00:21:15]>> Well yes it is >> in a lot of things. Yeah, I mean so that's what that's why you you go from cells to humans and the you know the how consistent cells are vary but humans are different and how much you can you know really change your behavior is is is a problem but you can imagine if you find sort of things like that where changes in diet can make you it wouldn't obviously it wouldn't be necessarily the cure but maybe it could help the cure maybe it can be sort of make someone predisposed to responding well to a treatment or not. That's sort of the core idea there because maybe you're starving the cancer itself and now in addition to starving the cancer now you hit it with some strong drug and so the chances of succeeding is higher and that was sort of the idea there. Yeah.

[00:22:00]>> And and has since you wrote that paper has there been any new progression on that any new advancements on that?

[00:22:06]>> Um not at the I mean this was like written by uh you know grad students and it's a big big team at that time. Um, and not that I know of too much. Um, there were some follow-up studies that try to essentially target similar parts of metabolism, uh, but not necessarily through dietary restriction.

[00:22:26]>> Yeah.

[00:22:27]>> All right. I want to get into the part that interests me the most. And this is something that you would be able to talk about for hours and hours. So, I'm going to make it's going to be a challenge for you to kind of pick and choose what you want to talk about here. But before we get directly into how the your AI the tools you're working with and developing will help you with this, we have to give the audience a little bit of background on I need you to take us into the cell and what is going on there with cell signaling and biomolecular condensates, protein sequencing especially um you know uh protein folding, protein manufacturing and all that stuff because this is a lot of the you know where the AI where the where the the drug devel velopment takes place, right? So, can you give us a little bit of a taste for what goes on inside the cell and how, you know, and then we'll get into how some of these >> AI programs would help develop drugs that could, you know, treat specific cancers.

[00:23:25]>> Yeah. I mean, it's it's a fascinating world and it's really sort of gets complicated, you know, the more you get into it. Um, but if we take sort of a high level view to begin with, right? I mean we if we're talking about drug discovery then obviously we're talking about diseases right and so usually we see diseases on a very high level so like we see someone uh maybe their their weight goes down or they have some something that appears that wasn't there right a rash so there something that you observe but the whole point of drug discovery in sort of the modern era has been trying to sort of identify what is happening inside the cell that's causing this thing to to sort of emerge on the organism level.

[00:24:09]And and so part of this has been understanding, you know, what makes up essentially a cell and how it works. And um you know, if just to give an overview, what we have is we have cells and there you have this DNA which is like the people usually describe as this blueprint, right? So it has all the information that you need um to for this cell to survive, to keep going. And and what does that really mean? It means that you have this long sequence or a long sort of string of of uh of letters um your genome and from those strings you can sort of define the the proteins that exist in the cell and the proteins are really the machinery. So that's like if you imagine you know these um fibers that hold cells keep their shape and they're holding them. Uh that's like a lot of those are proteins. Or if you think about what what is actually taking in sugar and decomposing it into smaller like you know smaller molecules and giving being allowing to convert it from the thing we eat to the energy that we can expend all those those things are are all proteins. Um so you know proteins are sort of these like essentially small machines that exist in the cell.

[00:25:23]And now how is this related to disease?

[00:25:25]Well, a lot of the time when when you when someone is um or their cell rather is like has some mutation that happens or maybe they're born with it, >> then these some of these machinery is is not working correctly. So the you know the thing that maybe helps you eat or decompose um or metabolize nutrients is not correctly doing it. So you know uh if if someone doesn't have the right protein for uh metabolizing you know the milk sugars then they're lactose intolerant right now it's thankfully it's not like a a sort of death situation for them we can deal with it but that's just one example and we have hundreds and hundreds of these examples >> and and sometimes these again are things that people are born with. So they their genome just has that mutation that makes that protein not work this the way it's supposed to or um it happens through a like you during someone's lifetime. So they didn't have it to begin with but they acquired it over time and and a lot of this so these are like sort of genetic kind of things but also you can you can think of things that you get infected with. So you get a bacteria that you're infected with um or viruses and again uh they're also made up of very similar kind of proteins and molecules and the whole point of drug discovery has been developing sort of you know the biomolelecules so small molecules or even bigger ones so different scales to them um to sort of target that that mutation or that the thing that broke. So if you have a protein that's causing you to um to divide very quickly, causing the cell to divide very quickly, maybe you want to target it so you stop its action so it's not active anymore.

[00:27:14]>> Um and so that's where the so understanding this whole process sort of is important for you to be able to design something against it, right? You need to know what your what people say is like your target, >> right? and and and then the complexity is is added because this is not just like there's one thing I need to go after and that if I hit it and I take it out then I'm good because all these proteins and all these molecules are talking to each other. So when you you know mention cell signaling and and condensate so >> you know maybe you get um some some hormone is going through your body right because of when you wake up or when you're hungry or whatever that goes and sort of binds to some other protein that protein then goes and binds to another one. So you have this cascade the signaling link from one to the other um all the way to the end and this could be a very long chain. this chain could sort of branch and take on many different routes through the cell. And so, you know, when you hit one, your effect sort of can propagate in ways you don't know and you don't sort of expect.

[00:28:17]>> And those those um those signals are made of proteins themselves, right?

[00:28:21]>> Ex. Exactly. Right. So a protein will bind to another protein and will cause it to change in some way, change its shape in some way and that has an effect to another protein and and another one another one until uh whatever the sort of end state is. Um and and again and this is we're talking about proteins are also sort of the small molecules sort of more like the foods that you eat as well that are also part of this, right? Um so you know when you eat sugar that causes a rise in insulin, right? So how how do you go from eating sugar to that? Well, there's again similarly that has to bind to something which which at the end of the day causes your cells to produce more insulin and through some long chain and and if you don't right if you don't produce insulin or your insulin is malformed then you have diabetes right um so the all these sort of diseases you think of them in like this is the mechanism that's sort of happening in the cell and the body and and the whole point of Drug discovery is to one say can we understand this mechanism uh as it as it exists in a normal state and then when we see it being different in a disease state not normal state uh can we sort of as a result you know make some just modification. Can we we either usually make a drug that goes and binds to the non um normal protein. So essentially inactivate it.

[00:29:51]>> Um or sometimes we can do the opposite where we can activate something more or give you supplements. So because you don't have the thing that you're supposed to have. Um so we're trying to sort of replace a function that you've lost.

[00:30:04]Um but yeah so understanding this entire process is sort of a big part of the the research into drug discovery.

[00:30:13]>> So where does artificial intelligence fit into this? What tools in particular are you guys using? And then just also one other throw on the end of that question. What are the things that what's what are the exciting things that are breaking through here either that interest you or that you might be working on?

[00:30:29]>> Yeah. I mean so then like is you know we've all we've talked about has been biology in here and so like one might question you know where where where does technology come into place? Where is AI in this story? Um and AI sort of comes into into the story in many many ways.

[00:30:45]Um on the very basic sort of level when we talk about proteins and this is what we've been sort of discussing this entire time. We say these are the small machinery. Well knowing what this machinery looks like has been very very difficult. Um, traditionally people have to go take this out of a cell, sort of purify it and literally like have to throw rays at it in order to get a a 3D image of what does this object look like, right? I mean, it's very hard to do. It's very labor intensive.

[00:31:17]And so a a breakthrough sort of was in you know 22 would people would say uh when this alpha fold model this protein structure prediction model came out that could tell you what a protein looks like um very accurately just from your genome right we said your genome sort of is the blueprint so it tells you what what the this protein is made up of but it doesn't tell you what it looks like so essentially alphaf can read the genome or essentially like what's downstream of the genome the amino acid sequence and and tell you what would be the 3D structure of this object >> which is very important to understanding what it could do because once you see it you can then much more easily sort of say okay well it's likely to bind to something or maybe will have this function and also once you see it you can say well I can maybe if I want to target this if I want to inactivate it I can just hit it here um and make something sort of bind to it here and just make it completely inactive. But all this stuff requires you to have a 3D uh visualization of of these proteins.

[00:32:25]And so the number one thing was being able to predict the 3D structure of proteins. Yeah.

[00:32:31]>> So that was like a major breakthrough is this is a long-standing problem, you know, like has been for decades and decades where people have theorized whether it's possible to begin with. A lot of people have worked on it with you know some amount of success but been very difficult that has been was considered unsolved and today you know a few years later um we more or less call it a solved problem to to some extent you know with some caveats but like a huge amount of progress in a few number few years.

[00:33:02]>> Uh [snorts] so that was like number one right and so what can you do with this information you say what technology do you then develop? Yeah.

[00:33:08]>> Well, then you can develop the kinds of technology that says um can you design proteins? So, can you design proteins that you know do the function you want but do it better? Can you design proteins that do functions that don't exist in nature, right? Um so just so imagine this is a big not disease related but a big dream is to say well you have all this pollution you have all these plastics in the ocean can you degrade them by just creating these proteins that will go and degrade all this plastic and they're very you know bio safe so like the because they're natural they're part of you know proteins come out of nature they they are we can if we put them out there it'll be fine but we can clean up a lot of stuff maybe you can make proteins that also we talked about insulin, we talked about you know other kinds of diseases that can go in you give it to someone and they sort of replace a function that you've lost um and and sort of help you sort of recover and and so like once you know what something looks like you can then start asking can you design something that looks like this right um and so the the ability to design came out of the ability to predict what it what the structure would be um and then also knowing what the structure would be can tell you well what can I the kinds of molecules I can create as that would be sort of drugs um >> right >> in the past you wouldn't have the structure um or you well yeah if you didn't have the structure you would essentially have to create a lot of experiments to just test hundreds of thousands of molecules you just have to you know random randomly test a lot of stuff see what sticks and the ones that have like some activity uh whether you know it kills was a cell or not, you would keep them and you just iterate on that process.

[00:34:55]>> Right.

[00:34:56]>> But this is >> trial by error, just something that takes >> trial and error. Yes.

[00:35:00]>> Yeah.

[00:35:00]>> And then you would maybe gain some intuition and you'd be able to say well actually if I make these changes it'll be a better drug. But that takes a lot of time >> versus saying I here's a computer. I can say this is the protein. This is what it looks like. We know exactly what it looks like and we just want to create something that makes it inactive. And from the data that you have from the methods that we you can build today that is actually a problem that you you know is doable.

[00:35:25]>> Um there's still a lot of sort of you know challenges here. So there is a challenge of people say well you know is the drug is it is a problem with drug discovery sort of the drugs not being good enough and that's usually not the case. It's like drugs can be very potent. The problem is they can be too potent. So, you know, this is where the all the side effects that come along where people like are rattling them off a long list, right, in in an advertisement.

[00:35:52]>> And so, you have all these side effects, these toxicity sort of effects that happen because you give someone a drug and it does what it's supposed to do, but unfortunately also does a lot of stuff it's not supposed to do. So it hits a lot of different targets and and the question then becomes well can you make drugs that are much more specific in their uh in what they're targeting and they're not just going around you know hitting all sorts of things. You want to hit the thing that's that's malformed that's sort of disregulated and nothing else right. Um, so that's like a big question being able to say, well, right now if you take a pill, it will likely go many places in your body, right? But maybe only a specific part is the where the disease is right now.

[00:36:37]>> And so can you make drugs that specifically go to different parts of the system? I think that's sort of a problem I worked on, which is localization. So can you make it specifically go to a part of the human body and then maybe in a cell, can it go to a specific part of the cell itself?

[00:36:52]Because uh you know the target exists in one location but not the others and >> essentially >> if you can do that then you can give someone much lower doses right.

[00:37:01]>> Yeah. How do you go about doing that?

[00:37:04]Well, you have to so many different ways through you know again uh this as you get deeper it becomes more complicated and you can say well maybe I can target a specific part of the body because I know those cells like so like I say you know the liver has particular receptors um so particular sort of tags that I can uh target and make the molecule that I like [clears throat] I give you sort of go and stick there and other parts of the body don't have those tags. And then similarly as for inside the cell once it makes it inside there are different compartments and to [clears throat] get into different compartments you need different tags. And so can I sort of design the right tag so I can make force that molecule to go to the place I want it to deliver it to so that it has the best effect and is most specific. So like narrow in scope in what it's sort of targeting and doesn't doesn't just distribute all across um you know a human body. And so those are the kinds of then questions you have to design for. And >> um they're still open questions by the way.

[00:38:11]>> And I'm thinking maybe so like what if you pre-treat it like let's say you were trying to do that localization with cancer. Could you pre-treat the cancer with something else and then when you send through the targeted drug it falls the power it only goes where that >> that pre-treated cells were. I mean there are there are many many different strategies and this really depends on you know the kind of molecule you're designing. So we hear about small molecules which are like essentially like you know the things that you can make into pills and people just take them. Um there are also bigger molecules. So things that you inject with a needle for instance >> um or or antibodies that you can you can give to someone. Um and they're also cell therapy. So you can give someone you know a bag of cells essentially you know trans and transfuse it. uh people use like these CARTT cells that people use for some cancers. Um and so all these things operate at different levels. So some of them can easily make it into the cell. Some of them uh have to go through the bloodstream and have to like be given at maybe injection site and they cannot get into cells or also can only operate at the outside. And so all these things sort of matter when you're thinking of what delivery system can you make. So, can you make it sort of go and stick to the outer part of the cell and then the cell takes it in?

[00:39:34]>> Is it going through the bloodstream and it just going to hit everything? So, then you need to make it specific through some other mechanism. Do you put it in like a essentially like a ball like this like ball of fat more or less that you pass through and then that has so the drug itself isn't specific but the packaging that you put it in is specific. So all these things sort of then depend on what kind of drug you want to give. Um and but and for each one of those is can be something that an AI sort of model can tackle to some to an extent. So it can tell you what kind of designs, what kind of delivery system to use um and how the mechanism of it will be uh where will localize what will it be not target. So this off offtarget effects be all these things are like each one of you know we are essentially going through maybe a dozen problems and each one of them there's a role for for like AI to to play um in drugs better yeah >> what specific AI programs are you using or are you guys using on your team >> I mean you develop or are they just you know >> Yeah. Yeah. I mean it's well first of all it's a very fast-paced sort of time uh to be in this uh field right so AI is moving very quickly and there's a there's a good amount of interest um for people to use it for the betterment of human health which hopefully is one of the good sort of products of this um and the research the the group that you know I've been part of um has has developed more on the design sort of thing so um understanding How do molecules come together? You know, if you have two of them or maybe more, three or four, how do they come together?

[00:41:23]>> And these can be two proteins, but it could be protein and DNA or protein and molecules. So, it could be many different kinds of things that exist in the body. And understanding how when they do come together, how do how does that happen? What does it look like? And again, understanding this allows you to then design against it. And then the other sort of flip side is well if I have a protein and I want to make something stick to it what how can I design so give me a method that can sort of design that um that partner for that protein. Um and similarly when we're talking about RNA or small molecule or essentially any kind of human u you know biomolecule uh so those are the the kinds of technologies that the group has sort of mostly focused on in this realm uh of discovery.

[00:42:11]>> Are you building these things from scratch or are you using existing tools as a starting place like Alphafold or maybe some other open- source software?

[00:42:20]>> Yeah. Um well as as a matter of principle everything that the you know the research group here does has been open source. Um and and then but in in AI we sort of try to learn from each other as much as possible. Um >> and so you know over the years when when you see a particular kind of model do very well you tend to you know adopt it or adopt the things that make it very strong and then you use that as a as your base sort of layer and you build on top of it. So pretty much you you know you would never today want to build a model from scratch. Um you would [clears throat] take sort of maybe the kinds of architecture that AlphaFold has built on top of it. Um or the sort of tricks that they use you know that that make the model good in terms of its performance and you use that as your base um and use add to what whatever your functionality is because sometimes it's it just doesn't fit your particular problem. So you need to modify it >> um and you need to like develop maybe new techniques to train and all these kinds of things. So there is a lot of sharing of many of the same architectures across the field. Um and then you you know you try to make it as good as possible for your problem >> and there's obviously a lot of similar work going on with various types of AI at MIT. So you're at a good place to be where you can share ideas with somebody across the hall probably.

[00:43:48]>> Yeah. I mean and and not just that I mean you the the there is a lot of transferability between problems right so we've we have been discussing biology and drug discovery and you know healthcare but then there are also people who a lot of the the the model development that was successful happen in in language right so these models that can understand language and and generate sort of new text very well and many of those sort of methods also do very well when you apply them to biology problems for instance.

[00:44:22]>> Um so there's a lot of transferability between kinds of problems because like at the end of the day we were modeling data that it it you know it means something different but a lot of it looks the same from a machine perspective.

[00:44:37]>> Have you ever heard just by chance of the of Dr. David Sinclair >> and he's doing re research with reversing cell aging? um a little bit. I don't know too much about the research to like have a strong opinion.

[00:44:53]>> Yeah, I was just interested because it's I've been reading more of his stuff and watching some videos with him lately and it's fascinating, you know, and how the how the effect of what they're developing could be used on cancer and other diseases, you know, to >> to try to rejuvenate the cells. and they're using AI to try to speed up um what they would like just as you said what the adverse effects might be of a treatment like that or like the ones you guys work on on on the cells which is obviously a huge thing to get through that because you're trying we're trying to get these drugs especially like if you're someone well pretty much someone in my shoes but in very many people have someone in their family that may have some kind of a disease like this and you know it takes so long to get these drugs just to get to the clinical trial stage and then it takes many years from that.

[00:45:40]>> Yeah.

[00:45:41]>> So yeah, that's always was interested me for um his work.

[00:45:45]>> I mean there's all sorts of things there. So I'll point you to like for instance a similar kind of topic which is not regeneration but if you think of regeneration >> so right if you someone I mean the >> in the case of a you know other not humans so like in lizard you cut off the tail tail goes back >> right why can't you do that for for a human being and so the question is what kind of regeneration can you do so you take cells from from a a human so like literally skin cells and then can you make those skin cells you from those grow you a heart right or or at least heart tissue right so you can you can scale the problem um but that question >> Dr. Michael right at MIT >> no he's at TUS Dr. Mike 11.

[00:46:32]>> Okay. I'm not I'm not particularly sure, but we had sort of labmates that worked on this problem and because you say again, how is this an AI problem? Well, it's a reprogramming, right? So, you take cells and they have been already been programmed to be skin cells, let's say.

[00:46:45]>> And the question is, can you modify them in the right way so that they're no longer skin cells, but they're something else. Yeah. And >> and what is how do you pro reprogramming that is a can be a computational problem. So what is a combination of of factors that you essentially give these cells? I mean we talked about proteins.

[00:47:04]You literally give them a combination of proteins a cocktail um to make them revert back to you know bas being not skin cells but something earlier in the development and and then you push them towards something a different organ. So you again the a cardiac tissue and then maybe that could be so you see these cells that were skin and then in a petri dish now have like this small beat electric effect and they're sort of beating like a like a human heart. Um so that's like a very similar kind of idea right um >> yeah yeah >> and it is again a computational problem to your point about you at the end of the day you you do have to do all these tests and studies and they do take a lot of time and labor and part of it part of AI is to discover new things but part of it is also a you know can you accelerate the discovery process so we talked about the fact that you had to in the past and even today screen a lot of data. So you have to go through do hundreds of thousands of screens so that you can say well this drug is the best one and the question of of like screening with AI is to say well can you do that very quickly so can you say you don't have to try the one a million molecules I can run it through on a computer and we'll select the top whatever number that is much more manageable number >> and I can and people have done this for let's say antibiotics right so you can go through hundreds of millions of of molecules if you really needed to and select top, you know, 100 thousand of really good drugs that you think are good antibiotics and those a thousand is very manageable and then you sort of test it and and you and people have and our group has has been one of those um to discover new antibiotics um in that way where you don't have to do that screening and you can most of it go through what would be essentially not possible humanly um in the lab, but obviously we can on a computer.

[00:49:06]>> How how far along are coming things like computer simulations, for instance? I don't even think computers will ever be able to simulate a cell because there's probably billions of proteins in a cell, but maybe they could simulate certain parts of a cell, right? Or functions within a cell. Are you seeing a lot of progress in that area?

[00:49:24]>> Um, I mean, it's so there is there are different fronts of this. So I think if you know if people go and and like look this up um you'll find uh you know a lot of buzzwords around virtual cells. So you know making making these like cells that you can essentially simulate and there are many ways to go about this problem. So far we've people have been going at this from the high level genome sort of uh scale. So saying can we say how much essentially of of proteins does any cell have? So the the levels um uh that a cell sort of uh exists in and then trying to predict how would those levels change when you give a drug or make some mutation or something like that and um that is a very important problem but that still is missing some information. So it doesn't tell you how those proteins interact with each other.

[00:50:23]What is the system look like? It just tells you you have you know um a lot of protein a not too much of protein D and so on and so forth. Uh there is another sort of world and I've been really interested in this for a long time which is sort of simulating uh the actual sort of you know network in some sense. So how do these molecules come together?

[00:50:46]What is what is the effect of uh if you take this protein out, how would how does the metabolism let's say change?

[00:50:54]>> Yeah.

[00:50:54]>> And in a more dynamic way and and that actually has a rich history of people doing simulations from like mathematical simulations. So there's no AI and and but it's you know typically more basic um and it's it predictive power is there's a ceiling to it. So it only can take you so far, >> right? Um and I and I think the coming years are should be where a lot of that now transforms into an AI kind of problem that can predict what that system that is very dynamic changing all the time looks like um to simulate quotequote simulate the cell.

[00:51:34]>> I think [clears throat] some cells are easier than others. You know we can bacteria is much easier thing to simulate than a human cell for instance.

[00:51:42]Um and you know so bridging that gap will also be part of the progress.

[00:51:47]>> It would take a lot of power just to simulate one second of the life in cell right. So >> yeah I mean again like this is these are very you know interesting because they get complicated in different ways. So you can think of the life of the cell right that's one particular time scale.

[00:52:05]Um you can also think of the proteins themselves right. So uh when when proteins are interacting with each other and changing confirmations that's a totally different time scale.

[00:52:17]>> Yeah.

[00:52:18]>> Right. So you're trying to actually a simulation is simulating things happening on micros secondsonds things happening on milliseconds and things happening on days >> right >> all at at the same time. So yes, it's it you know I think it'll be difficult um but but so you try to take each problem and try to simulate it um somewhat independently and then hopefully at some point you can pull this together.

[00:52:44]So I guess we can finish with one last question then. Um Peter, what do you see big things that we can look forward to happening say in the next five years?

[00:52:56]>> Um I think I mean I don't want to you know forecast too much because that's uh five years ago I would have been totally um surprised. Yeah. Yeah.

[00:53:06]>> That thing Yeah. happened today. Um but I think you know there's there healthcare is very interesting because um there's a lot of good to be done in that space uh with technology that has been being developed some partly by us but many many groups around the world um you know I think there has been more accepting towards um and like sort of an appreciation that this technology needs to be deployed in some in some way into the hospital systems and healthcare systems. So if we take the models that we developed these breast cancer screening let's say or lung screening it's been a few years right it this was this didn't happen last year this happened 2019 2020 21 so it's been a few years since they've been developed they they haven't changed they've been sitting there this entire time um but there there is a lot of you know there's a lot of change that needs to happen on the health care system side of things that accepting models, testing them, validating them, the regulation of you know uh of of any new technology >> that that is there's a delay in when the technology is developed to when it actually can impact people and I think we we should be entering a time where more of these things make it into the actual healthare system where you know hopefully the next time someone gets an exam it's not it's you know there's a human reading it but there's also an AI sort of reading it as Well, and and hopefully that means that we we catch, you know, more cancers and and miss fewer people, >> right?

[00:54:47]>> Um and then on the drug discovery side of things is I think it's sort of a lot of big pharma um already sort of utilize many of these technologies. Uh they haven't made it all the way through to clinical trials because that's really the bottleneck is once you have a drug running it through, you know, doing all the human studies that you need to is still very painful. Um and that's where mo a lot of drugs will fail. Um but we're at least are able to nominate you know more drugs newer ones maybe safer drugs um uh more eff like effective uh drugs that um [clears throat] just h have the therapeutic effect that they should have and not anything else and that process is uh I think is getting sort of be being escal like sort of accelerated >> and we're able to design new things that we weren't able to before so things that were hard to design in the past hopefully is like now actually possible.

[00:55:44]So there's a lot of companies that are trying to do these antibbody therapies because there's been success where AI can design new antibodies relatively well. Um [clears throat] so hopefully that also is will be making its way more and so the the prediction is you know we'll we'll hear more about drugs that have been designed by you know a computer obviously there's a a lot of human >> a lot of humans in there. Yeah. and and the clinical trials too the people that worked as the guine you know the patients that served as the guinea pigs for it. So >> of course yes but more more and more of that um and and again I think understanding this space is also to understand all the limitations uh because that's really where you can be able to say okay this is this is where its use case is this is where it benefits us but we need to regulate it in this way because it can fail >> um in in many ways and so we don't want to we don't want to like introduce a technology that causes more harm than than Good.

[00:56:44]>> Okay, Peter, first, thank you very much for your time and being with us here today. And second, thank you very much for your work in this field. You know, this is something that's going to benefit a lot of people and you're just getting started in your career, but you know, I you're going to do a lot of big things and that are going to help a lot of people. So, thank you for all that work.

[00:57:03]>> Of course, and and as always, I should say, you know, this is a takes a village. It's a lot of people working on this. Um, >> and I think it's also in an era where it's important to say everyone that I know is working on this is doing it for from a place of good and trying to like sort of, you know, >> contribute something positive umh to the world, >> right? Okay. Well, you've been a great guest. Thank you very much, Peter. Sorry about the technical difficulties at the beginning.

[00:57:31]>> Yes, of course. Thank you, guys.

[00:57:35]>> All right, guys. So, we'll be back tomorrow night at 700 p.m. to go over I have two guests on the two writers for the book, excuse me, called Killing the Lieutenant.

[00:57:47]And it's about Lieutenant Ral Diaz, who kind of was the original Miami Vice character, very similar to that. I'm sure there's been many many people that who were the would have made good models for that, but he is someone who is through the the heat and the heart of the drug wars in Miami, the cocaine wars during the 80s and in the 90s. He was right in there. So, they will be on tomorrow night at 7 o'clock. So, I'll be back again right away. All right.

[00:58:16]Thanks, guys. And we'll talk we'll talk to you then.