Aired:
October 24, 2024
Category:
Podcast

Using AI and Synthetic Biology to Go Where Antibody Therapies Can’t

In This Episode

This episode of the Life Sciences DNA Podcast takes listeners on a journey through the future of biologics—where AI and synthetic biology join forces to surpass the limitations of traditional antibody therapies. It's a powerful look at how programmable biology is reshaping drug development.

Episode highlights
  • Breaks down where current antibody drugs hit their ceiling—like reaching targets inside cells or within complex biological system.
  • Showcases how scientists are now building custom biological tools to interact with diseases in precise, programmable ways.
  • Explains how AI is used to design, test, and refine new biologics in silico—dramatically shortening discovery timelines.
  • Introduces exciting modalities like cell-free systems and synthetic peptides that operate where antibodies cannot.
  • Looks ahead to how these innovations will fuel new classes of personalized, high-impact therapeutics.

Transcript

Daniel Levine (00:00)

The Life Sciences DNA podcast is sponsoredby Agilisium Labs, a collaborative space where Agilisium works with its clientsto co-develop and incubate POCs, products, and solutions. To learn howAgilisium Labs can use the power of its generative AI for life sciencesanalytics, visit them at labs.agilisium.com.

Amar, we've got Venkatesh Mysore on theshow today. Who is Venkatesh? Dr. Venkatesh Mysore is the co-founder and CEO ofAikium. It's a biotechnology company focused on developing therapeutics throughprotein engineering, synthetic biology, and deep learning. He founded Aikium inOctober 2022. He spent his career at the intersection of AI and drugdevelopment. He has held positions at NVIDIA

working on accelerating drug discoveryusing machine learning, at the Shaw and at the AI driven biotech Atomwise. Heholds a bachelor's degree in computer science from the Indian Institute ofTechnology, a master's degree in computer science from the University of WisconsinMadison, and a PhD in computer science from NYU. what is Aikium? What is itseeking to do? Aikium is a startup that is using its proprietary AI platform

to develop a new class of drugs to targetGPCR proteins. These are targets of great interest to drug developers. And whatare you hoping to hear from Venkatesh today? I would like to understand how thecompany's platform technology is able to tackle the specific challenges to finddrugs against GPCRs. I would also like to understand the novel proteins thecompany is pursuing as therapeutic candidates. And I would like to talk

about Aikium's business model as well.Well, before we begin, I want to remind our audience that if they would like tokeep up on the latest episodes of the Life Sciences DNA podcast, they shouldhit the subscribe button. If you're enjoying the content, please be sure to hitthe like button and share your thoughts with us in the comments section. Withthat, let's welcome Vankatesh to the show.

Venkatesh, thanks for joining us today.We're going to talk about protein engineering, AI, and the potential to applythis technology to a target class of proteins called GPCRs. But before we getinto that, I wanted to just talk about the announcement today about the NobelPrize in AI. Can you talk a bit more about that? Absolutely, Amar. It's apleasure to be on your show. Yes, today is a big day for deep learning inprotein engineering

because the pioneers of the field have beenrecognized with Nobel Prize, specifically Professor David Baker. And then fromDeepMind, we have Demis Hassabis and my one-time colleague, John Jumper, who weoverlapped briefly at Shaw Research. So it's a revolution in that AI forbiology has often been overstated but under-delivered. But this is an examplewhere

the impact has been so much that withinfive years of developing the technology, essentially every researcher in theworld in the space of proteins is using it on a daily basis. So it's a trulywell-deserved Nobel Prize. Absolutely. That's just really fantastic news. Solet's get to the protein engineering that you're doing and the proteins, right?So let's start with some basics here. So what is really meant by a target?

And can you please explain what's a targetin the language of drug developers? Sure. So if you think about what causesdisease, so many different things can be wrong in our body. But ultimately, ifyou pinpoint it down to a cell, within a cell, you often have certain proteinson the cell. And very often in a disease, those tiny proteins sitting insidecells could be the drivers of the disease

or they could be why you have the symptomsor why the disease progresses. So, a target is a protein that the scientificcommunity studying the disease pinpoints and says, hey, if you could onlymanipulate or modify or modulate this protein, then we could potentially comeup with a treatment or a cure or a way of managing the symptoms. So, a targetin this context usually refers to a single protein that has been implicated inthe disease.

And we are most of the time trying toinhibit the target protein from doing its job. But in some occasions, we aretrying to enable it to do its job better. Yeah, so we're basically trying tomodify the function of the protein, right? Either if we think that if there's toomuch protein and overdoing that, then we're trying to inhibit that. If there isnot much protein that we're trying to beef up if it's function, right? We'reusing the drugs that we get. And so there are some, like,

these families of proteins, So there's proteinlike families of proteins are the proteins similar to that. So there are someof these families of proteins that are more popular. Can you tell us a bit moreabout like what are some of these most popular families, protein families? Soif you think about a cell, the way I try to see where proteins are is whetherthey are outside of a cell, whether they're on the cell surface or whether theyare inside the cell. And when you go inside the cell, you can ask, is it in thecytoplasm

or is it inside the nucleus? So the targetclasses that are of biggest interest to Aikium are the ones that are hiding inthe membrane. So these are protein targets where they are not accessible. Theyare buried, embedded in the membrane by design. So they are receptors, signalsthat respond to signals in the extracellular environment and pass on thatsignal to the intracellular environment. So we have

protein families called ion channels,protein families called GPCRs, and those are the ones, broadly speaking,multipass membrane proteins that build the next big frontier to be attacked ortargeted with protein biologics. OK. As a multipass, you mean like they just goin and out, in and out like that, right? So within the membrane. Just being -crossing the membrane once,

they go in and out of the membrane. So theyexpose very tiny regions outside the cell and now we are trying to bind tothose tiny pieces of the target protein that are accessible inside the cell.And so these are signal passers, right? So they basically get some signals fromoutside the cell and then they pass the signal inside the cell based on what'shappening outside - what the cell should do, right? That's correct. They arekey in cellular signal transduction.

Okay, okay. So tell us about these GPCRsbecause you have a lot of focus on those. How many proteins are in this GPCRfamily? How many already have drugs against them? Are there some likewell-known drugs against GPCRs that are out in the market? Absolutely. So theGPCR class, it stands for G-protein coupled receptor. It's one of those wondersof evolution, let's say. It has created this family of machines that sit on themembrane

and they are attached to what are calledG-proteins on the intracellular side and in response to various signals, theyundergo slight change in their shape, also called as conformation and that issensed by the G-protein that is attached on the intracellular side and thenthose G-proteins go as second messengers and go and deliver the signal. So inthis category of targets, in this category of proteins, there are more than 600GPCRs in the human proteome

and about 300 or 350 of them are consideredvalidated drug targets. And within that, what you would not believe or what isnot common knowledge is that we have approved drugs for about 150 of those. SoGPCRs in a 1 in 3 or 1 in 4 drugs are actually targeting some GPCR or theother. The most popular one in 2023, 2024 is the GLPs

that everybody is talking about in thecontext of obesity and weight loss. So there, it is a small drug that goes andbinds to a GPCR to achieve its function. That's fantastic. So it's really a lotof revolutionary drugs coming out of the GPCRs. I mean that's just a fantasticstatistic. So many of the drugs targeting GPCRs.

So tell us, with so many drugs againstGPCRs, what is Aikium trying to do by finding more drugs against the GPCRs?Great question. So to understand that, we first need to ask, what are thedifferent kinds of drugs, and when do we need which class of drugs, and what isthe underserved or unmet need? So broadly speaking, you want to, let's say,let's think about blocking a GPCR for now. Let's say you're

targeting a GPCR with the goal of blockingit, you can either use a small molecule. And when people say a small molecule,it's much smaller than an actual protein. And that's your average pill that youpop in. A second category of drugs are called protein biologics. And really thefield is dominated by one particular protein biologic family called antibodies.So antibodies are a naturally occurring protein serving a certain function

in our immune system. Whereas about three,four decades ago, humanity figured out how to repurpose antibodies into drugs,specifically make the antibodies so that it binds to the target that we aretrying to block. And that becomes the basis of a therapeutic application. Butnow if you ask, when do we use protein biologics and when do we use smallmolecules? For many indications, most of us would much rather pop a pill

than take an injection or an IV. So proteindrugs in general require to be administered intravenously or through the skin.And so the most common drug in this case is actually insulin. Insulin is anexample of a protein, but it's not going after GPCR. But the advantage withantibody-based or, broadly speaking, protein-based drugs is that they stay inthe body for much longer, which means you need

far less frequency of dosage andimportantly there are certain disease classes such as cancer where you needmuch more powerful drugs and usually protein biologics have a higher safetyprofile, have a higher chance of getting cleared in the clinical trial processand ultimately they can enable certain mechanisms such as ADCC which smallmolecules cannot. So ultimately protein biologics are preferred for certain chronicconditions.

But yet given their overall advantages,there are only three FDA approved antibody based drugs for the entire family ofGPCRs and zero FDA approved antibody based drugs for ion channels. That'sanother 600, 700 targets. So this is because raising antibodies to thesetargets has been hard. So that is the unmet need. And so what Aikium has doneis rather than trying to rectify

the limitations of antibodies or makingsmall technical improvements, we have come up with a completely new protein family,also inspired by a protein family in nature. And we are predicting that as thenext big industry, meaning it's another vehicle that can become a universalbinder and go after these multipass membrane proteins using a different.

So these are not antibodies, these aredifferent types of proteins that will bind to the GPCRs and ion channels.That's correct. Okay, great, great. And you're calling them SeqR proteins, isthat right? That's correct. So we are calling them S-E-Q-R, SeqRs. Okay. theshort mnemonic is they are sequence-specific binders, or you can think of themas seeking a certain sequence and then binding to it

and then enabling the modulation of thetarget. Okay. So can you tell us a bit more about that? I haven't heard muchabout the SeqR protein. So if you can familiarize us with them. So Seeker isfirst of all based on a naturally occurring protein family, but it hasn't beenemployed for therapeutic purposes the way we have considered. So we have takenan inspiration from nature and re-engineered it. And this re-engineered proteinfamily is what we are calling SeqRs.

So these SeqRs have a fundamentallydifferent approach to binding to a target. So if you think about a protein, infact in light of the Nobel Prize, the Nobel Prize was about being able topredict the three-dimensional structure of a protein given the sequence of theprotein. The sequence is the chemical composition. But it turns out only about50 % of the protein surface has a rigid three-dimensional shape.

The other half on average is a highlyflexible region by design and they are called intrinsically disordered regionsor IDRs for short. These intrinsically disordered regions do not exist in asingle three-dimensional shape. Instead, they are meant to sample severaldifferent shapes and they are highly flexible or floppy. The problem is all thedrug discovery and drug development machinery

both on the small molecule front and on theantibody front has been heavily focused on these structured regions and theintrinsically disordered regions which represent about 50 % of the real estateon protein surfaces have largely been underserved. Our scaffold, the SeqR,actually targets these intrinsically disordered regions. So we are unlockingabout 50 % of the opportunity which was previously, which previously didn'thave

any reliable robust technology forattacking. Okay. And so these SeqR proteins that you're talking about, theylike, do they, like, can you specify the sequence that they should bind to orhow does that work? That's exactly right. So if it's a general, if it's auniversal binder in the sense it binds to all disordered regions, such aprotein would not be useful as a drug because it would

hit not just the target that you careabout, but everything else as well. So instead in the world of drugdevelopment, selectivity, being able to bind to just the target that you careabout is very, important. And I just told you that these disordered regionsdon't have a shape. So what is it that the SeqR protein can recognize and bindto? It can actually recognize the linear sequence of amino acids. So it's likeyou can specify a zip code or an address on a protein surface.

And we can then engineer a SeqR thatspecifically detects and binds to that sequence on the target protein. Okay.Okay. For people who don't know, protein is basically made of amino acids,right? So it's a sequence of amino acids. So what you're saying is that theseare like unstructured regions that the protein has where like you have thesequence of amino acids. So the SeqR proteins are going to go and bind there,right? And then affect the function of the GPCR. You're correct. So

the common model is think of beads on achain. Most of the time, those beads don't have any shape on their own. Butoccasionally, you can fold them. Let's say some of them have magnets. Some ofthem have charges. They can fold into a three-dimensional shape. So there willbe portions of this chain that have a structure and portions that don't have astructure. And they are highly flexible. They can wobble around. And AikiumSeqRs can actually bind

to those structureless regions through thespecial SeqR protein. Great, great. So are you now with these, are youtargeting the GPCRs that have already been targeted or are you looking at thenew GPCR? Like how are you choosing the targets? The target selection isactually a much bigger deal than we realized when we started Aikium. So whenyou think about targets, we would love to be able to say,

you know, the new technology has enabled usto go after GPCRs that have never been targeted before. And so we have someGPCRs in that category where we know their therapeutic relevance based on therecent academic work, but there is no clinical trial that has ever been done onthose GPCRs in the context of certain diseases. flip side of choosing thesetargets is we already have modality risk

meaning they're doing something new on thedrug front. But now if you also couple a high risk target that hasn't beenvalidated in the clinic, then you're compounding the risks. So we try tobalance it by picking a set of GPCRs all within the same target class. But onesthat have been validated, meaning there is an approved drug in the clinic,others that are in clinical trials but nothing approved yet. And the thirdcategory where

there has never been any human clinicaltrials conducted on that target, but the evidence is so strong and compelling.So we have tried to spread our bets that way. Gotcha. But then when you'redeveloping something against the GPCRs for which there already are drugs, thenthe bar is going to be much higher, right? Because then your drug needs toprobably beat the traditional drug that's out there, is that correct? That'scorrect.

So what we have done is there are severalindications. So we have tried to pick GPCRs where they are implicated orconnected to more than one disease. So it may be the case that for the -indication is just the word for a disease that a particular drug has beenapproved for. So it may be the case that a certain GPCR has been clinicallyvalidated for disease A. But there is strong evidence for disease B.

But the fact that in a clinical trialsetting in humans, it has been drugged and it has been deemed safe means thatby drugging the target, you're not going to cause any toxicity. So you're justchanging the disease context. What we have tried to do is hedge there as well.We have tried to pick GPCRs that are implicated in more than one indication. Soin one case, we have a GPCR that's implicated in both oncology and anautoimmune disease.

In the other case, we have picked a GPCRthat is implicated in oncology and also in neuroinflammation. OK. OK. And thenso these GPCRs, what are the therapeutic areas that the GPCRs are commonlytargeted against in general? So that's a good question. But because of thebroad relevance of GPCRs, essentially they're present on every cell servingcertain functions.

Like I said, you know, we have asthma drugsthat target GPCR and they all fall under the broad category of inflammation.Then we have indications like atopic dermatitis. We have indications like,again, all falling under neuroinflammation. Then we have several oncologyprograms where the GPCRs are targeted for two reasons. Either because they aredirectly involved in the disease or simply because certain GPCRs are

overexpressed or present a lot on cancercells than on normal cells. So they are often used as payload delivery targets,let's say GPCRx is present only on cancer cells and then I can take an antibodythat binds to - let's say if I could make an antibody that binds to GPCRx, butalongside I attached a drug molecule that would eventually kill that cancercell.

This field is called antibody drugconjugates. And this principle is often used, is often the reason why someGPCRs are targeted in the context of cancer. Okay, gotcha. And you mentionedabout the obesity as well, right? So that's another therapeutic area. Anotherbig application that has revolutionized our thinking of, expanded the treatmentoptions for a variety of indications in the last two, three years. Absolutely.So now let's talk about the

machine learning a bit. So, Aikium has alarge-scale machine learning platform to screen and optimize this SeqR proteinsand I believe it's called Yotta ML2. So, can you tell us more about that? Sure.So, the big realization that happened about three or four years ago is thatthere were actually two independent sets of breakthroughs that had happened.One in the field of synthetic biology and the other in the field of

artificial intelligence and both had to dowith reasoning about proteins. So in the world of synthetic biology, ratherthan test one protein at a time in a certain experiment, you could now makemillions, billions and what Aikium has done is trillions of proteins at thesame time in a single part and then use that to do experiments. On theartificial intelligence front, two breakthroughs. One is the protein structureprediction that we talked about,

acknowledged with the Nobel, but theequally impactful invention was the use of or development of protein languagemodels. Similarly, just like we all know, chat GPT, similar models have beentrained on protein sequences. So these are called protein language models thatreason in the chemical language of proteins and not in an English languagedescription of the protein.

So basically we are talking because we have20 different amino acids, right? And each is represented by a letter. So yourprotein sequence is a chain of letters like you said, like the beads, right?Each is a bead that has this one of the 20 letters. So basically that's thelanguage that now we're using for the large language models, right? So that tobe trained, is that right? And that's absolutely right. So if we have a 20character alphabet rather than a 26 character alphabet,

And so every protein can be described as asingle sentence in that alphabet. So what Aikium has done, or the bigrealization I was saying was, neither piece on its own can deliver the next bigstep up. So the next big step up in biology is going to come when these twobreakthroughs in orthogonal fields are coupled correctly. So what Aikium hasdone,

is we started out realizing that if youbuild a really good model and you say, hey, my model has 70 billion parameters,you're going to have that advantage maybe for three months. Because at thatvery moment, somebody else is training a larger model or somebody else iscoming up with a different architecture that has something else better aboutit.  In a large language model, you mean?In a large language model or model that's reasoning about proteins or anyaspect of that.

So the model is not a sustainable mode,whereas Aikium is a sustainable mode. So what Aikium invested in is to build aprotein-protein interaction platform where we could get or screen 100 to 1000fold more proteins than anybody else in the world. So that is called the Yottadisplay. So if you look at literature or if you look at

protein engineering companies in general,the principle is the following. You have a protein and you're trying to input,and the protein - let's say is the SeqR protein, and let's say it binds to apiece of the disordered region of the target GPCR. Now you improve it. You wantto make a really, really, really strong binder. So the principle in proteinengineering is if you have a rough starting point that sort of works, then yourandomly make changes to it

and then see if any of those random changesimprove the binding to the target. The question then becomes, can you really dobetter than random? And that's where the AI comes in. The next part of thequestion is, how many can you test? Can you test 100 changes? Can you test athousand changes? How many random changes can I make from my starting point andthen test them experimentally? And you're generating these new sequences usingthe large language model?

That's correct. So Aikium has its ownpatented generative AI approach for designing SeqR proteins given any sequence.So the input sequence is, let's say, the disorder stretch of a GPCR. So we havea chemical specification. And then we feed the input and out comes a SeqRprotein that has been designed. So then we can make that SeqR protein and testit in the lab. Interesting.

So in this approach, the way I just saidit, you take one design at a time, synthesize it one at a time, and test if itworks. Instead, what we have done is we are changing the paradigm from testingtop 10 or top 100 to essentially generating a library, physical library withone trillion different changes to the starting one. So that number, to put itin perspective, the field

has been using a certain wet lab techniquecalled yeast display, where you can make a million different changes to aprotein and test all of them simultaneously. There came another break. Yeah.Sorry. When you say test them, it's basically so in yeast cells, you have allthese different proteins, and then you're putting that GPCR that you want themto bind to

and see how good the binding is. Is thatwhat you mean by testing? That's correct. So let's say you're able to make amillion different versions of a SeqR. And then you have GPCR, let's say, on amagnetic bead. And then all of these million variants are now competing to bindto that target. OK. Only the most potent ones, driven by thermodynamics, willbe able to find to the target.

If you then fish out the target and then dosome clever genomics, you can then go back and say, hey, this must have beenthe protein that stuck to the bead because I was able to establish its identityusing a trick that yeast display allows you to do. You can pull the identity ofthe protein with the actual protein. Okay. Okay. That sounds great. Yeah.Please. Yeast display, you can do a million proteins at the same time. And it'sbeen done for a few decades now.

Then came phage display where they pushedit a thousand fold larger. So in phage display you can make a billion differentvariants of a protein and then all of them will compete simultaneously to bindto your target protein and then you fish out your target protein and check whatis bound to it. You likely pulled out the strongest binders in that billionprotein library. So from there the next technique that came

was called mRNA display. But mRNA displayhas certain limitations. It really you cannot make a... in mRNA display you cango to a trillion but the problem is that the method as described in literature,as practiced by others, is applicable not to large proteins but only to tinypeptides. So in context going back to GLPs those drugs are short peptides andso you can do GLP peptide screening

using mRNA display. By GLP you mean theobesity drugs, right? The drugs are themselves peptides, rather than largeproteins. So peptide is like a tiny piece of a protein. So mRNA display has thelibrary size or diversity advantage, but as described in literature, aspracticed by others, it's not applicable to large proteins. Okay. the Aikiumteam

really spent much of its pre-seed phasebuilding an industrialized version of mRNA display that we are calling Yottadisplay. And this was accomplished by a team with over 125 patents in the spaceof spatial transcriptomics, CRISPR, and protein engineering. So this has beenthe bulk of our effort in the last year, year and a half. And now we have ahighly optimized

and a robust protocol that has cut down thenumber of steps, the wall clock time, the overall efficiency of the process. Sonow we reliably do trillion protein screens against several different targets.Okay, this is pretty much like a wet lab approach you're talking about here,right? Like you have the AI, the large language models are on the one side, butthis is the wet lab inventions that you're doing

on this other side as well. Absolutely. SoAikium is very fortunate to have founders that span both AI and syntheticbiology. And early on, that innovation in just one side will not really lead toa quantum leap in our ability to do protein engineering. So what we are doingis going back to the darts on the target analogy. We can now throw a trilliondarts on the target, but these darts have not been shaped at random.

So AI works ahead of time to shape thosedarts. So the tips of those darts and the other features have been engineeredthrough a computational exploration to design SeqRs that are likely to bind tothe target of interest. So this combination is what allows us to find reallyefficacious SeqRs.

So basically what you're doing is with yourlarge language model, you're generating a trillion variants and then you'retesting all of those in the wet lab using this new methodology to see which arethe SeqR proteins that are best binding to these GPCR. Is that correct? That'scorrect. So to the first approximation, that is correct.

If you really have a generative model andask it to make a trillion designs, a lot of those may be garbage or redundant.So what we do is we have a way of summarizing the kit list, if you will, fromthis model. And then we use an intermediate optimizer to make them better. Andthen we infer the library that collectively spans the ideas that have come outof the computational pipeline so that we have representations of all the ideasthat have come out.

But we also certain permutationcombinations of the suggested mutations that may not have been present in theoriginal list, but this way we are doing a very thorough exploration of thelandscape. Gotcha. Gotcha. Okay. And is that, does that happen in like just onesingle iteration or are there like multiple iterations that happen for this? Socomputationally we have multiple iterations, at the end, multi-daycomputational pass

we have a library recommendation. And then,the library is synthesized. The experimental part also inherently has multiplecycles. So what I previously described was the target, you use like a bait, andthen you're trying to fish out members of your library that bind to the target.So this is called a selection process. But what we do is we do multiple roundsof selection, meaning we make it harder and harder for those who are stuck toremain stuck.

So we really allow those super potentbinders. We also do what is called negative selection, things that bind to thewrong targets, they are discarded. So the library is depleted of members thatbind to the wrong things. So we can get both selectivity and specificity in thesame experimental assay. But this is still one computational pass followed byone experimental pass.

But we have to do one second iteration aswell. So we take the hits that we have found, then generate thesecomputationally searched for, a build a library around the hits that we havefound, then subject them to another round of mRNA display. And very often thatis sufficient. OK, gotcha. And now let's say, you you've done this with the wetlab and you've got some of these potential SeqR proteins that could be gooddrugs. So what have you done to validate these further?

Great question. So when we do theseexperiments, all that we are doing is binding, meaning does my SeqR bind to theintrinsically disordered region of interest? This binding is happening inisolation, in vitro. It really is not happening in the context of an actualcell. So the validation that we have done is we have gone to cells. We havegone to cells where the target GPCR is present in its entirety

under realistic conditions and in thoseconditions we don't just check for binding because binding in that environmentwill be harder than just binding in isolation but after binding it has toachieve the functional modulation that we care about. So in the case of some ofthe GPCRs that we have picked we want to antagonize them, meaning we want themto want to suppress their activity. So what we have measured and validated sofar is that these SeqRs are actually really really potent,

meaning they are able to bind to the targeteven under cellular conditions and actually inhibit the ability of that targetto perform its function. That's the goal in the context of a therapeuticapplication. Okay, okay. And what is the business model around that? Are youseeking to discover promising therapeutic candidates and license them to pharmacompanies? Or you take targets of interest from partners and provide them with

with candidates or are you going to pursueyour own pipeline? Can you tell us a bit more about that? So our core value isgoing to come to our in-house assets. So that is just the nature of the worldwe live in today. That in a platform play, while it sounds exciting, it doesnot command the value of having in-house assets. So what Aikium has chosen todo is to have two in-house programs

that we can take to the developmentcandidate nomination stage and then to IND and beyond through our pipeline.platform we have built is really so efficient and has been worked on so muchthat it would be a shame if we are not able to apply it to other targets. Butit is unrealistic for one entity to have the resources to pursue a hundreddifferent drug discovery or drug development programs.

So what we are simultaneously doing is talkingto pharmaceutical companies and enter into partnerships where the company tellsus the specific target that they care about. It doesn't have to be a G proteincoupled receptor. It can be any protein where they previously employed theirin-house technologies. It could be small molecule based or could be antibodybased. But perhaps they did not get the candidate that they were looking for.

So in those cases, we should be able toapply our technology, the SeqR-based approach and the Yotta ML Squared platform,to be able to give them potent selective binders to that target, which they canthen develop into drugs if they have the other desirable properties that theycare about. So far, we have a very long pipeline of folks we are talking with,and we have one big pharma partner

with whom we've entered the paperwork andnegotiation stage. So we can expect an announcement before the end of the year.Great. And these are early days for Aikium. So you're at the Baker BioEnginuityHub in Berkeley, and you're raising seed funding. So what have been thediscussions like with potential investors? And do you think they are morereceptive to you as an AI-based drug developer compared to a more conventionaltype?

Can you tell us more? Sure. So I think ourfate changed in September because that's when we got the really compellingcell-based validation. So previously, we were highlighting the platformpotential and the fact that if you could extrapolate and go from here, you cansee that it can have therapeutic relevance. But now we have functional activitydata. And in this world, it has to be really, really potent, what are callednanomolar antagonists.

And that's what we do against threedifferent therapeutically relevant GPCRs. So now it's no longer a scienceexperiment. It's no longer, couple of dreamers are building something new thatmay or may not work. It actually works in cells. So now the real challenge istaking it into mice and then eventually to humans. So there, a completelydifferent set of skills are needed. And that's what we're

actively recruiting. We are talking toseveral pharma people, current and former, to bring them on in key roles totake this biologic in its infancy, go all the way to high-end filing. That'swhat we are currently adding to our team. But in the context of fundraising,the AI piece is not going unnoticed. So like I said earlier as well, a pure AIplay, unfortunately, while it sounds glamorous,

with the exception of perhaps Alpha Fold,it's not going to have an immediate clinical impact if you say let somebody doall the validation for me. So you have to come to the realization that a drugdevelopment company has to have AI and has to have protein engineering in ourcase on site. And that's what we have done. And this story is really having alot of takers. So again, we expect a big announcement before the end of theyear.

We are in diligence with a lot of venturecapital firms, but we'd love to have conversations with more folks whounderstand this space. Interestingly for listeners, the field of biotechnology,when it is driven by computational methods, people have flipped biotech and arecalling it tech bio. So Aikium is an example of a tech bio startup that isincorporating AI, but realizes the importance of ultra high throughput experimentation

to generate that label data that the hungryAI craves for. How do you see... So I wanted to ask you about the name Aikium.Is that something to do with AI or what is the origin of the name? Reallyinteresting story there. So Aikium in two different Indian languages, Tamil andSanskrit, means the union or coming together, much like yoga.

But in this specific case we picked thename because three different fields - protein engineering, synthetic biology,and deep learning have to come together to be able to solve this difficultproblem of drugging the intrinsically disordered regions. The 50 % of theprotein real estate that no one has a solution for. If you really even want toattempt to attack that problem you really have to unify forces and Aikium meansthe union are coming together. But

after the name was suggested, we didn'treally immediately like it. But then when we realized we could spell it asAI-KI-UM, so then we had a way of emphasizing that there's an AI-first proteintherapeutics company. So we enjoyed that name and we adopted it immediately.That's interesting. so now looking at the big picture, I mean, these are, it'sa fascinating method you talked about, like in terms of AI,

bringing in really new areas, new type ofdrugs and also applying the new wet biology methods. So how do you see AItransforming the pharmaceutical industry over the next decade or so? So AI hasseveral different applications. Again, I was very fortunate that my careerbetween D.E. Shaw Research, Atomwise and NVIDIA

 

gave me different perspectives on both thesmall molecule discovery, protein engineering, and the broader space. So if youthink about where AI can be used, companies like Recursion and Insitro areusing them to find the targets that need to be targeted. Initially, I presentedit as though it was a solved problem. Somebody does the biology and says, hey,target X,

the GPCRx, go find something that binds toit. But really, the disease etiology is quite complex. And there are oftenmultiple routes to reach the treatment. So what some companies have done is,again, they start with an ultra-high throughput experimental platform. But thenthey employ machine learning to reason. For example, reason over highresolution microscopy images, high resolution protein-protein interaction data,

to figure out the network of differentproteins, how they are talking to each other, and then find those nodes thatneed to be suppressed or activated for the profile to resemble a healthy cellrather than a sick cell. So the first application is AI for target discovery.But now once the target has been discovered, many companies, like Insilico,Atomwise,

what they have done is they've employedmethods to say, hey, small molecule drugs, proven method. Now let's just see ifwe can find a better drug using AI. So this can involve cleverly searchingthrough ultra large compound libraries. Or it can involve generating new smallmolecules that would fit into the previously characterized pockets. Or whatcompanies like D.E. Shaw Research are doing is realizing that proteins are notstatic structures but highly flexible.

Rather than use the one structure in theprotein data bank or the one structure predicted by AlphaFold you can do amolecular dynamic simulation to understand the different shapes that thatbinding pocket looks like and then plug in compounds there. So this is allunder the space of small molecule drug discovery. But the real revolution, inmy opinion, is on the protein side of things because protein language models,protein structure prediction have put us in an era

of reasoning that was not possible eventhree, four years ago. So that's a big revolution where I think within the nextdecade, we will see enzymes, we will see drugs, we will see so many thingsaround us which are all engineered proteins. They may have come from, let'ssay, synthetic biology platforms such as those engineered by Professor GeorgeChurch, or they may come from entirely computational pipelines engineered bypioneers such as Professor David Baker, or it could be some new

tool or application that comes fromDeepMind or other big players in this space. But that won't be in therapeuticsor finding that initial drug. But I think what most people are now reallyconcerned about is what about the rest of the drug development pipeline?Because things that bind on a computer, maybe some fraction of will bind evenin a test tube. But the moment it goes to a cell, already they're not able topredict whether it will bind under cellular conditions. Then it goes to amouse.

Maybe it's bit blind to the target but itcould bind to so many other things and it could cause toxicity related issuesthat we are still struggling to predict. And then ultimately when it goes inhumans, really so far we have not been able to predict which drugs are going tofail in phase 1 or phase 2 or phase 3. Those are the most expensive failuresbecause...

about 10 years and a few billion dollarshave gone in and then you realize you produce something that is unsafe forhumans or has no efficacy in humans even though it worked in mice or otherlarge mammals. That's bad. So what is happening is AI is that we found thetarget. I have now new methods of AI driven target discovery. Then I have AIdriven drug discovery. Then we have AI driven drug development. Then the realquestion is,

How soon can we get to AI driven toxicityprediction and human clinical trial outcome prediction? And so for that tohappen, we need much more data. Like with anything else, we need more lab ratsto be sacrificed to be able to generate the data. And the ethics of all of thisis definitely being brought into question. So there are some people who aretrying to find workarounds -

models called organoids. So even ratherthan use a whole animal, you can take stem cells or other approaches to createsmall living organ mimetics and you can use those to do your drug screen andthat will be much more predictive of subsequent failures. this is also beingenabled by high resolution ways understanding what's happening in a tissue or acell. That's where fields like

spatial transcriptomics, where you look ata slice of a tissue and you can see this cell is trying to do this becausethese proteins are overexpressed here. This other cell that did not respond tomy cancer drug, it looks like it has a completely different proteomic profile.So high-resolution ways, they're combining information about which genes arecurrently being transcribed, what is the cell's morphology, what is theoverall, what else can we learn about the

from the person's genome or blood sample orother components, all of that is feeding. And now we have a richer than we candigest picture of each person. And then we have to tie this to whether thatperson responded positively or negatively in a clinical trial. And as this dataset accumulates, as pharma companies collaborate and share this crucial data,that's when the new kind of reasoning that will all of us will emerge.

I we are maybe 5 to 10 years away fromsaying we can predict some of those things. Absolutely, yeah, there's a longway to go. One more question I have for you is, you mentioned about reasoningmodels. Can you explain, like double click on that, like what do mean byreasoning models? So when we are trying to reason, traditionally it's aboutreading 10 papers and then coming up with a new hypothesis or maybe protein A

works because it talks to protein B andmaybe this is the reasoning that emerged from reading different papers. But thekind of reasoning I'm talking about is multimodal deep learning. So when youthink about a patient or a blood sample or a cell or a protein, you're going tohave, let's keep it simple and just talk about proteins. On one hand, you knowthe chemical composition or the sequence protein. On the other hand, you knowthe three dimensional structure or the

conformation of the protein and then youhave lots and lots of annotations about what are the cellular processes thatthis protein is known to function in. That's the ontology. So if you can seeyou have a text description of the protein, you have its chemical compositionand then you have its three-dimensional shape. So these are three differentkinds of information but yet you have to be able to digest all three differentkinds of information

and then compare two different proteins. Orin the case of, let's say you're engineering a SeqR, have to compare two differentSeqRs and say which one is going to be better for a particular target. So whenI say reasoning, I mean being able to bring in different modes or differentkinds of information and then using to either predict the small molecule drugsthat are likely to work or engineer the proteins that are likely to betherapeutically relevant. Okay, gotcha, gotcha.

Dr. Venkatesh Mysore, CEO of Aikium,Venkatesh, thank you very much for your time today. Thank you so much, Amar,for hosting me. It's really a pleasure talking. hello and thank you to all thelisteners here.

Well, Amar, what did you think? It wasfascinating to really see the company being very active and innovating on bothsides, right? On the AI side, developing the launch language model and also onthe web biology side, what he talks in terms like synthetic biology, right? Socreating these displays of a trillion proteins at the same time -

a very interesting business model in termsof like having both of those in the same company and in a startup company.GPCRs are an important set of biological targets. Venkatesh noted there areabout 150 drugs that target GPCRs today. What's the significance of Aikium'sSeqR proteins in their ability to target what he called disordered regions?Yeah. So one of the things is that GPCRs is a very well validated

target family. I mean, we know that thedrugs against GPCRs work very well because they're sending the signals down inthe cell. So a couple of ways I can think is one is for the targets that arealready there, perhaps these protein bindings, these SeqR bindings, if they canhave better binding than what the traditional drug molecules have, they may bemore effective drugs than what we have on the market for those specific

targets for which there are drugs already.But then also, some of these GPCRs are the ones for which the drugs haven'tbeen there, for whatever reason. Maybe they weren't the great targets, but alsomaybe they were not very well drugable in the sense that it was hard to findthe right small molecule against them. So for those, those now become prettyopen in terms of new targets that we can develop a drug against and again thisnew modality

and having a lot of effect in curing thediseases. So I see it in both ways. It was interesting to hear him characterizehis risk management approach to target selection on how the company's noveltargets for those that are well-validated and others that may be targets of aclinical candidate. What do you think about that? What might it say about theconsideration for the application of a technology that can produce a high levelof novelty?

So I would say that, I mean, that approachseems very reasonable because the new SeqR proteins, which are the drugcandidates that his company is coming up with, those are very novel. So as hementioned, if you have a novel target and a novel drug, there's a lot of riskthat people are taking. On the other hand, if you have a well-validated target,okay, there's no risk there,

but then you're developing the drug itself.And I also mentioned to him, like if for a well validated target, if he'scoming in with a drug, the drug needs to be much better than what's out in themarket for it to be effective, right? So that is an extra challenge that'sthere. But on the other hand, it helps him validate that the methodology isright. The methodology is correct. And so...

And that validation can go a long way interms of even convincing his pharma partners, right? Saying that, okay, theseare the drugs that we can develop these drugs against these well-validatedtargets. We've shown that. So now we can develop those against theseunvalidated targets as well. So overall, as he's trying to push these into drugdevelopment, right, either themselves or...

or through partners, all of that data isgoing to be helping as evidence that the methodology works. They're usinggenerative AI to make SeqR proteins, you know, libraries the size of a trillionproteins. These are weeded out and optimized, but at some point is there adiminishing return? How much of a challenge of using generative AI is knowingwhen you're working with either too small a data center

or one that's too large? When does enoughbecome enough? It's hard to answer that. I can tell you that it's a trillion -may not be enough because what we're trying to do is we're trying to - I mean,there's so many mutations that are possible and right now not much is knownabout these. So at this point at the beginning, he probably

needs to try out a lot of differentcombinations to see what sticks. I think once the models and the researchers inthis company learn more about what kind of sequences, what kind of mutationsare giving better results than the others, then they can guide the models inthose directions. And he said that they have started doing that already, but Ithink that's gonna be the thing. And that's why I also asked him about ...

Are they having multiple rounds? Because alot of times what happens is that computationally you can develop a lot ofstuff, but then once you put that in a wet lab and you get some results, thenyou can improve the computational model with those results and focus in theareas where you're getting the positive results. So in that sense, I thinkthat's kind of like a recursive

way, could be very effective. It's hard tosay right now, like, what is enough because, I mean, enough is when we can geta good drug, right? That's what it's about. Well, it was a great conversationand looking forward to our next one, Take care. Thank you, Danny.

Thanks again to our sponsor, AgilisiumLabs.

For Life Sciences DNA and Dr. Amar Drawid,I'm Daniel Levine. Thanks for joining us.

Our Host

Dr. Amar Drawid, an industry veteran who has worked in data science leadership with top biopharmaceutical companies. He explores the evolving use of AI and data science with innovators working to reshape all aspects of the biopharmaceutical industry from the way new therapeutics are discovered to how they are marketed.

Our Speaker

Dr. Venkatesh Mysore is a co-founder, CTO, and Head of AI at Aikium Inc., a Berkeley-based biotech startup revolutionizing therapeutic protein engineering. With a Ph.D. in Computer Science from NYU, he brings 15+ years of experience in computational drug discovery, AI/ML modeling, and scalable production workflows.