Tackling the New Bottleneck of Drug Discovery

DanielLevine (00:00)

TheLife Sciences DNA podcast is sponsored by Agilisium Labs, a collaborative spacewhere Agilisium works with its clients to co -develop and incubate POCs,products, and solutions. To learn how Agilisium Labs can use the power of itsgenerative AI for life sciences analytics, visit them at labs.agilisium .com.

Amar,we've got Markus Gershater on the show today. For listeners not familiar withhim, who is he? Markus is the chief scientific officer and co -founder ofSynthase. He was a research associate in synthetic biology at UniversityCollege London and a biotransformation scientist at Novacta Biosystems. He hasa PhD in plant biochemistry from Durham University. And what is Synthace? Wheredo they fit into this world of

AIin the Life Sciences. Synthace has created a digital experiment platform forresearch and development. That's a cloud -based platform that allows scientiststo use a natural language interface to use AI to design complex experiments,automate their running, and capture and analyze data all in an automated way.It has integrated AI into this platform to accelerate the discovery anddevelopment of new therapeutics.

Well,before we begin, I want to remind our audience that if they want to stay up onthe latest episodes of the Life Sciences DNA podcast, they should hit thesubscribe button. If you enjoy this content, please hit the like button and letus  your thoughts in the commentssection. If you want to listen to the show on the go, you can find an audioonly version on major podcast platforms. And with that, let's welcome Marcus tothe show.

Markus,thanks for joining us today. We're going to talk today about using AI totransform R &D, Synthace, and its digital experiment platform. But beforewe do that, I wanted to start with a piece you recently wrote that likens AI indrug discovery to the electrification of factories and time it took to realizeproductivity increases.

Sotell us a bit more about it and where do you think we are in the ultimateimpact AI will have on drug discovery and development? Wow, okay. That's quitea lot. Well, first of all, thanks very much for inviting me on. It's a realpleasure to be here. The analogy with electrification is really that whenelectricity first came along to be available to run factories, right, then

whatit was replacing was huge steam engines, which would then have these massivedriveshafts that would run through the factory and everything would be run inpulleys from that, from that driveshaft, right. And so what that would mean iseither the whole factories on or the whole factories off, right, all themachines are all running, or they're not running at all. And similarly, youhave to make sure all your machines are arranged

aroundthat drive shaft, you , you have no other way of getting power to thosemachines. So that's what you have to do. And so when people first, you , hadaccess to electricity, then some factories, they upgraded to, you , electricalengines instead. But they didn't actually see any productivity improvementsfrom it, right? They'd switched power sources, but they hadn't seen any kind ofproductivity improvements. But we think of electrification as being kind of arevolution in the way that manufacturing was done.

Butit was only when people started to actually realize the benefits ofelectrification and reorganize their factories around it that they started toget the benefit. So essentially, it meant that now you can have any machineanywhere, and you can turn them on and off at will. So now you can rearrangeyour factory from being centered around the drive shaft to being centeredaround the production process. And so then

itbecomes like the production line method of actually producing things, which isa of a lot more efficient. And so that's when you end up with the efficiencygains from electrification. So it wasn't actually electrification itself, it'sthe way that it could then change the way of working. But that benefit actuallycame substantially later than the original introduction of electrification. AndI think it's an interesting thing to think about, because we have this newpowerful tool

inartificial intelligence, and it's being applied in various different ways to differentparts of the value chain. But I feel like until it's actually adopted acrossthe value chain, right, then we're not going to see the full benefit of it. Soso that's kind of where that that analogy comes from. And I mean, the otherthing that I sometimes get a bit irritated by is kind of just this generic useof the term AI, right. And I think those of us who've been, you know, talkingabout AI for a while, and maybe

havea bit more experience about where it can be applied. You know, it's similarlyto  sort of exhorting people, you shouldjust use electricity more. It's like, well, actually, you use electricity forall sorts of different things. And in all sorts of different ways. And I thinklike, sometimes, the specificity of what you know computational power can dogets lost in this kind of term of artificial intelligence, which

itcan be helpful in that it can capture some of our aspirations for what it mightdo for the field, but it can also be unhelpful, right, because it  obfuscates everything. Okay. And so whenyou're talking about AI, right, and it's always an interesting one, like, whatdo people mean by AI, right? So it can be like, automation with business rules to machine learning to deep learning togenerative AI, right? So  how do you thinkabout AI? I tend to try and demystify it. There's this,

Ithink she's kind of a sociologist / ethicist here in the UK called HilarySutcliffe. And  she said something whichreally resonated with me once, if whenever the term AI is used, you just replaceit in your head for the word software, right? Then I think  you often get to a much better kind of

intrinsicunderstanding of what it is, because AI sometimes has this sort of mysticism,maybe like this sort of sense of the unknowable around it, whereas, you know,if you say,  software can nowenable,  voice recognition, and,  like text to speech, or,  you know, software can now rapidly generate

condensed bits of information from lots ofdifferent ... it kind of just makes it a little bit more, it makes it moretangible and more real to be somehow. And so, you know, I think essentially,it's building sophisticated software models, right, which can then help us tomake predictions or help us in some way, with the complexity of what we're dealing with. And, you know, I think thatthinking about it that way could then be...

it'smore obvious then that it's incredibly diverse, right? Because software isincredibly diverse, right? So it's just a kind of mental game that I sometimeslike to play to try and reframe the way we think about it a little bit more.Absolutely. And another interesting thing is that like, so electrification is agreat analogy. I've also seen a lot of times like people saying data is the newoil, right? Like the oil of the 21st century. So yeah, what are your thoughtsabout that? Which analogy do you think is better?

Well,I  like the emphasis on data, because Ithink that, you know, again, a lot of people who've been in the space for abit, realize that actually, it's more about the data than maybe the algorithms,right? You know, if you have a high quality

dataset that you're training your models on, then that's a hell of a lot morelikely to give you decent results than if you're... I saw it referred to asdumpster diving the other day, right? You know, you're just basically goingthrough and just trying to find whatever you can, and will shove it all into amodel and expect it to magically come up with something. And again, maybe if wethought  of it as like, okay, we'regiving our software some data, then you kind of, you don't you don't kind ofthink it's going to do something magic somehow.

Right?You know, sometimes I think the perception of AI is that we that it shouldsomehow give us some kind of magical abilities. And to be fair, I mean, when Ithink when all of us started using LLMs, and they feel pretty frequentlycapable, right? And it's only when you start using them more, do you realizetheir limitations and kind of, okay, yes, it is. It's a very powerful andexceptionally compelling piece of software, but it's still software, right?Yeah.

It'syeah, I think  you know, the whole datais the new oil thing. Absolutely. I think fundamentally, it is all about thedata, right? And obviously, how we manage that data, how we bring it together.And  then how we train very purposefulmodels on top of that, right? Again, I think, you know, the data have to be

designedin such a way and the experiments that produce the data have to be designed insuch a way that from the outset, we are thinking of the models that we'retrying to build, right? And the problems that we're trying to solve. That's mypersonal take on it. I think you can look through retrospective data and thatkind of thing, but often there's a real issue that they haven't been built forthe purpose of finding out specific things, right? Yeah, yeah. And that's beena big problem, right? A lot of times,

whendoing data analytics, the data scientists are brought in a bit later when thedata is already generated and were just lying around somewhere, right? And thenpeople are like, okay, well, here's the data. Was it generated for thispurpose? Not really, but yeah, can you use it? Right. And actually it remindsme of a quote from a statistician from like the 1930s, a guy called RonaldFisher. And he said, you know, if you call in a statistician after you've runthe experiment, then the most you could ask of him is to run a post -mortem.

Right?And I think it's a similar kind of thing, you know, and I think what's encouraging these days, though, is thatactually, we're seeing a lot closer interaction between data scientistsand  wet lab scientists, to call itcrudely. And so I think that gap is narrowing. And I think we're all learninghow to speak, you know, each other's languages and understand each other'sdomain more,

whichfor me is part of the fascination of it. I've been very lucky to work on the boundaries of disciplines throughoutmy whole career. And I've always found that the most interesting perspectivesand questions come from, like the different mindsets that you have indifferent, in different disciplines, as well as obviously the differentexpertise.

Absolutely.So where do you think we are in this ultimate impact AI will have on drugdiscovery and development? I think we've made a really exciting start, right?That would be my take. But like another thing that I'm really privileged aboutis that in my role at Synthace I get to go and talk to scientists from allsorts of different companies, I guess, much like, you know, you getting to talkto them on this podcast. And

whatI found that would be interesting to hear if you've seen something similar, isthat the wet lab scientists we're talking to, and they're in kind of earlydiscovery functions, particularly sort of in target validation or assaydevelopment, those kind of areas. And they're saying, look, we're just gettingslammed. Right. And the reason they're getting slammed is because the bitupstream of them has been on bottleneck by AI, right?

Soall this is, this is certainly what they ascribe it to. So targetidentification is no longer the really sort of painstaking academic process oflike, fundamental research into disease, which then as a byproduct, then allowsus to identify targets. And now it's, you know, an industrialized process where absolutely, as I'm sure you'llbe

veryaware of this, of these massive multiomic data sets,  you know, plus other data from other areascoming together such that then these AI models are identifying the patterns inthese data sets in order to then identify new putative targets. And so whatthat means is that, you know, there's this flood of new potential targets. Andat the same time, there's another technological shift, nothing to do with AI,but to do with new modalities coming through, like PROTAC

modalitiesand nucleic acid modalities, which means that you no longer have to look forthe typical features that you previously have looked for to look for a drugabletarget, right, like a nice kind of active site that you can  have a drug go and bind within. You know,these new modalities mean that pretty much anything could be a target. So we'vegot a lot more targets coming in from AI, and the diversity of what could be atarget has also expanded.

Andso what you then find is that  thesefolks who have to then generate the tooling for any new therapeutic program, sothe assays or the cell lines or whatever, you know, these guys now are findingthat their job has just got dramatically harder, but they don't have any moreresource to do it. And they're like, you know, they've got this pressure onthem to get as many new therapeutic programs started as possible. So,  we kind of have

calledthis like the new bottleneck, right? Whereas maybe target ID used to be thebottleneck. And so, you know, there'd be a nature paper published after, youknow, like 10 years of painstaking academic research or whatever. And then, youknow, that'd be the starting gun and every pharma company would then race todrug that target, right? And that's the dynamic that the value chain and theprocess is all set up around. So it's like fewer very high quality targets

whichthen everyone races to be first. And now the dynamic seems to have shiftedquite a lot. And I think,  like there's agood case to be made for  that's happenedbecause of multiomics and AI models built upon those multiomics. And so I thinkthat's one example where, you know, the electrification has occurred, right,the electrification of the factory, but we haven't worked out how to rearrangethe factory to make use of that. right. That that would be my argument. Okay,okay. Great. Great.

Andso as you're thinking about this, right, like the drug discovery and targetidentification, you explained a lot, but like also thinking more about likefrom like the target validation or hit identification, how do you see AIchanging those areas? So target validation is obviously done in a number ofdifferent ways, but I mean, fundamentally, you need to be getting in the lab.

right.And,  for a lot of these things, youcould do a lot of theoretical validation of yes, it makes sense. You know,like, which  could all be done upfront,but then the kind of  CRISPR basedstudies, you know, to look at a knockout of those targets, obviously, is verycommon these days. But also, like making sure that you're working with the mostrelevant cell lines that you can you , for doing those kind of experiments, Ithink is is absolutely vital.

Andsimilarly, you know, kind of in assay development. I think it's, it's somewherewhere I think we need to think very carefully about the way that we'regenerating data going forward, right. And because if we're going to build AImodels on top of these, then any biases in the data, any poor quality in thedata is then going to lead to those AI models either

justbeing ineffective or even worse misleading. Right. So I think that, yeah,there's, there's  an application of AIstill to be done, I think, in the process of generating, say, exceptionallyhigh quality assays, right, for example, because that's the way we produce datais through assay. So ideally, they are a very high quality, high signal tonoise, but also on systems that

youknow, at least some of those assays have to be on systems that are as relevantto the patient, as relevant to the clinic as possible. Right. And so what we'retalking about now is a very high demand on these groups I've already  outlined that are pretty slammed. Right. Andthen I mean, mentioned hit identification as well. And I think alongside targetID, you know, this has to be an area where AI has clearly had an impact, right.But

youknow, I still think that we could be doing still substantially better, were weto increase  the acuity and reliabilityof the assays that we're using throughout drug discovery as well. So,  I've spoken with a number of people in thefield, and those who are looking to up the

qualityof drug discovery in general tend to point to one of two things. So one of themis about  the clinical relevance of themodels that we're using in discovery. It's in phase, we're entirely about labwork, right? And the generation of high dimensional data sets from those labwork, that lab work, that's really what we're focused on. And

so,the way we're thinking about then the application of these technologies is thatwe want to see that lab work, you know, increasing the quality of what can comeout of things. And,  like what I've heardis that actually often the metrics that are used for what is a good assay needto be raised. And the reason for this is because

Ifyou're running a high throughput screen, for example, so these are the datathat you're then going to end up feeding into your AI models about whatchemistry could affect this particular target, right? So it's critical that weget the best data we possibly can in that high throughput screen. And just forthe audience though,  if you can explainwhat's a high throughput screen, would be great. Okay, of course. Yeah. So ahigh throughput screen is essentially where we take, you know, say a milliondifferent molecules.

Andwe screen to see which one of those molecules can interact effectively with thetarget that we believe ... if we that if we can hit that target effectively,then it's going to have some therapeutic effect on a disease, right? So I've been talking about target ID, I realized Ididn't explain that either. But like, yeah, so this is the process by whichthey identify those

moleculeswithin the cell, those proteins within the cell that if you  hit them, then maybe you can, maybe you cantreat a disease. And yeah, the way what we're doing in a screen is we'rerunning like a million assays with a million different compounds, and seeingwhich of those compounds might affect that target. So and what you hope to getfrom that is

ideally,lots of different compounds that might affect that target, because then thatgives you some information about the kind of chemistries which might affectthat target effectively. Now, the issue is that if your assay is not sensitiveenough, then you're going to get a huge amount of false negatives, right?So,  you've got all of your compounds andyou have a threshold that you need  thatany target has to get up,  any compoundhas to get above

soyou can actually see  that compound hashad an effect on your target, right? But if you've got a poor assay, then thethreshold is higher and it means that you've missed that compound. And so thenyou've missed the opportunity to learn, or your AI model ultimately has missedthe opportunity to learn about the chemistries that can affect a particulartarget. And so there are particular kind of metrics that we have againstassays.

Oneof the classic ones is Z prime, right? And Z prime is basically like a verystringent signal to noise measure. And it's measured from zero, well, actually,it could go below zero, but it like the highest Z prime you can possibly haveis one. That's basically where you'd have a perfect assay with beautiful signaland no noise whatsoever. So there's never - you never have an assay of one,right? But in general, it's regarded that an assay with point five or above isa very good assay. But then

I've been having some conversations with achap called Dan Thomas, who spent a large amount of his career in this earlydiscovery space. And he makes the point that actually a Z prime of 0 .5, youstill got this huge amount of false negatives. And the point he makes is that falsepositives we can deal with, right, because you could go back and retest themwith other assays or with the same assay. And you can realize, that's actuallya false positive. But a false negative is lost forever.

Right?You'll never, you'll never be able to go back. Right? Yeah. So it all speaks tothe criticality of having very, very effective data production. Right. And thenwe need to look at those processes, the assays, which make those data and thinkabout how we can make them as effective as possible. Whilst at the same time,trying to alleviate this bottleneck, right? So make it faster, but also better.Sure. And this is where,  so I'veoutlined the problem, right?

Andone thing that we found really exciting at Synthace is that if you runexperiments which are systematically looking at all the variables that mightaffect an assay, but you look at them all simultaneously, so you set up thisexperimental design, which very systematically puts a lot of different designpoints through all of the different combinations about how you might set upthat assay

ina very systematic way. So you generate this exceptionally deliberate multi-dimensional data set, right, which describes the behavior of that assay as youmove through this very high dimensional space of all the different parametersthat you might change to affect the assay. Yeah, so let's do a deeper dive inthat, right? So I'd like to with an example. So first of all, I mean, for thosewho don't know assays,

youknow, like layman terms of experiment, right? So there's an experiment that'sbeing done, right? So in Synthace you build this digital experimental platform,right? So can you like take us through, okay, well, an example where there isan experiment that's being done and what are the challenges right now and howSynthace's digital experiment platform, how that is coming in

andsolving those challenges. Right. So let's take the example of an assay that wewant to develop, right? So this is essentially where we have to set up a systemsuch that when a molecule interacts with a particular protein, then we get avery clear signal that it's done so. And that signal can be measured throughfluorescence. So we have plate readers in the lab that are very good atmeasuring how much light is given out

by a particular well. And this is often howthese assays work, you have to somehow convert this interaction into some kindof fluorescent signal. That's, that's often the job. And the issue isthat,  if you look at that system, therecan be dozens of different things that could affect that system. So  very simple ones would be things liketemperature and pH, but then you also have salt concentration and the salt typethat you have in there, it could be sodium chloride, it could be magnesiumchloride. You have  things

calledreducing agents, right? So these are things that stop oxygen from degrading thesystem too badly. And you might have to choose both which sort you might wantto use, but also the concentration of it. And then sometimes we might want toadd some other protein into the system to make it more similar to the kind ofcrowded protein environment that you have in a cell. There's just, you can goon, right? There's,  there's a hugeamount of factors,

often,especially with new targets, which people aren't used to working with anymore.You know, like, it used to be that actually kinases were so common, right? Soas a target, so then an assay development scientist might, you know, be taskedwith developing their 14, the 14th kinase assay of their career, right? So theyhave some kind of idea about how they're going to go about it. But if you'renow presented with a target, which you just have no idea of, you have noprecedent with working with that kind of target, then ideally, you'd be able to

mapout how all of these different factors, how all these different parametersmight come together to affect that system and so allow you to pick the bestpossible set of conditions for running that assay. So the fundamental problemhere is that because of the biology, right. In physics or chemistry the numberof factors affecting the experiment might be much smaller if you have like thiscontrolled environment.

Butyeah, but in biology is just - in a biological system is so complicated thatthere are just so many of these factors that need to be just right or right.And as you're saying, right, like there are 20 ,000 proteins, right. Sopotentially 20 ,000 targets or even like 30 ,000 targets that we have. We justdon't know how what is the right biological system in which we can get thesignal, right? And that's why you have to like, there's this whole slew offactors that you need to consider and figure out  what is the right level

foreach of these factors, right? Right. Now, but then if we think about how wewere taught to do science at school, right, then you start to be taught aboutlike the scientific method, right? And so, you're taught about the rigor of holding everything constant except forthe thing you're investigating, right? And this is noticed one factor at atime,

right?So you very carefully just vary one thing at a time so that you can isolate theeffect of that thing when compared with all of the other factors that you mightbe looking at, all the other parameters. Now, that you know, that's pretty good for like school level science, but I'm always a bit dismayed that theeducation often doesn't seem to progress beyond that, right? And,  actually, it's perfectly possible to look atmany different parameters at once, so long as you do it in a very structured

experimentaldesign. And it's particularly critical, and you alluded to this, isparticularly critical for biology, because biology is an emergent system,  it has emerged from evolution. And so thatmeans that it has absolutely no obligation to be understandable or simple,right? It hasn't been designed from, you know, individual component parts thatfit together neatly. It's, it's full of unexpected interactions of these partscoming together in unexpected ways to

then generate a particular phenotype, aparticular phenomenon, right. And so, but the problem with just looking at onefactor at a time is you'll never then see how one factor might interact withanother one. Because  you're looking atthem in independent experiments in isolation. Whereas if you look at them allsimultaneously in these very structured designs, then you can actually pickapart the way that these different components are coming together to actuallygenerate the phenomenon that we're looking at. In this case, the

assaythat we're working with, right.  And inour experience, it's incredibly rare to come across a biological system thatdoesn't have these interactions all over the place, right? So  where the optimum level for one parameterdepends on  the level of anotherparameter, the setting of another parameter. And,  there's one kind of obvious way of describingwhat these interactions and that's the synergy.

Right.So synergy is a word that's overused, but a true synergy, like one plus oneequals five or whatever, right? That is basically where you  have two things coming together to give anunexpectedly good result. In experiments design terminology, we'd call that a twofactor interaction, right? So two things coming together. And we see them allover the place. And what we found at Synthace is if you run the experiments inthis much more powerful, multifactorial way, often referred to as design ofexperiments,

thenyou can cut through biological complexity a lot more effectively. So that'slike the first thing. And then if you add automation to that, then the numberof runs you can do and the complexity of the design space that you could lookat goes up substantially, right? So you've got these two things comingtogether. Like, let's look at things in a high dimensional way, but let's lookat it in an automated way as well. But then those experiments get incrediblydifficult to plan and run.

right?And this is where Synthace comes in the digital experiment platform. Okay,because what it allows scientists to do is say, I want to run this experimentaldesign. And then Synthace will then work out the huge amount of intricatedetails that are required to actually then run that experimental design in thelab. Imagine a plate, you know, like of this kind of footprint. So it's, youknow, typical kind of footprint of a plate in the lab, and you have

384 wells in it, or even 1536 wells. And wewant to run an experiment where each one of those wells, it has a differentmixture of components in there where you could have 20 different components,right? And you're varying these all and so no well, no two wells are the same,right? And you're spreading all of these different experimental runs across thehigh dimensional design space. That is an experiment which is to allpracticalities impossible to run by hand

andis also incredibly arduous to plan and program automation to do. Right. Andthat's essentially the problem that Synthace is solving. Then to loop back toAI,  my hope for AI is that then, youknow, you could use the these exceptionally high quality comprehensive multidimensional data sets, right. And because

what we found is that for each individual newassay, to keep on this particular example, each new assay developed, thesemulti dimensional experiments are very powerful for doing that, right? And itmeans that you could get a new assay way faster, so up to 70 % faster. So thisis a process that typically takes up to nine months. And we've got case studiesover and over again, different clients, big pharma clients, where

they'vereduced that time by up to 70%. So both AstraZeneca and GSK are public thatthey  do this kind of work with us. Butalso critically, we see improvements in Z prime, right? Because  you're getting such a comprehensive view ofthat space, of that design space, you can pick those conditions which are fully optimal, right? So you're getting that- you're squeezing the absolute most you can out of that assay system.

Andthen incidentally, it can often be cheaper as well, because you can find theplace where it's most effective, but also uses the least of expensive reagents,because you've mapped out the space so comprehensively. So that's allbrilliant, right? Each new assay, we can make it happen a hell of a lot fasterand make better assays. But then what I then asked myself is, what could we doto properly unbottleneck this using AI? Is there some kind of meta analysis wecould do? Like taking

you know, what could the first 12 very highdimensional comprehensive data sets tell us about the 13th assay development weneed to run? Or the first 50 tell us about the 51st, right? There has to belike, additional information about the very nature of how these kind of systemswork, which I think artificial intelligence could be exceptionally good atdoing that meta analysis, right?  So I

meanthat we can design ever better experiments and get to ever better assays everfaster, because we're starting from a better place each time. Okay. Okay. Andin that way, you know, we've been very successful at applying AI to spaceswhere you can generate very large amounts of data   and you can do that with relatively kind ofroutine processes where  you set up anautomated system that you crank the handle.

Right.So like it's - so things like multiomics, right, you set up your particular,you know,  genomic workflow ortranscriptomic workflow, and then you get a huge number of different patientsamples, and you push it all through, right, and you get a huge data set, whichyou can then apply AI to. Where we haven't applied AI yet, is to kind of  those areas of R &D that are a lot morevariable

whereyou never do the same experiment twice, where you can't just make it high throughputin an easy kind of manner. And I think this is a real opportunity for where wecan go to unbottleneck this new bottleneck of target validation and assaydevelopment, but  also then push throughthe rest of the value chain. OK. So let's say within an assay, let's say thereare five variables and each variable has, let's say, 10 different values.

Andso then you're talking about 10 times, 10 times, 10 times, 10 times, 10, right?We just, yeah, 10 to the five, right? So that's 10 ,000, 100 ,000, right? Sothe Synthace platform, how does it deal with that? And as you said, users, AI,so is it looking at, okay, well, doing experiments, let's say, in a couple oflevels, for a factor, then kind of determining

thismakes sense, then going in that direction. So how is the AI working there?Right. So the process I described for multi -dimensional experimentation isn'tactually AI. This is statistics, right? And it's a branch of statistics calledoptimal experimental design, design of experiments, it's often known as DOE.Well, at least that's what it's all rooted in. And the idea there is that you

setup these systematic frameworks within multi dimensional space. And then, butyou ask very specific questions for each experiment. And so by doing that, youcan strip down the number of experimental runs that you have to do. So I'll tryand give an example. So if we were to look at six factors simultaneously,right, then, and we say, actually,

outputto start off with, we don't know which of these six factors is important ornot. Right. So we're not going to try and find the optimum straight away. Allwe're going to do is find out which factors are important. Then, if you'resetting up an experiment, then to see whether a factor is having an effect, ifyou choose your levels carefully that you set that factor to, then you can getaway with  two levels, right, for thatfactor. So  you can condense the problemin that way. But then also,

wehave to ask ourselves, so we know that there are interactions in biology,right? Like  these factors could interactin unexpected ways. But what we tend to find is that the higher level thatinteraction, the less likely it is to occur. So I've already described what atwo factor interaction is, right? Synergy is an example of it, right? Now, athree factor interaction is where the nature of that two factor interactionchanges when you change a third factor.

So you have the interlinking, thisinterweaving of three factors together. And you can keep going up, right, fourfactor interactions, five factor interactions, and it just gets ever morecomplex. Yeah, but they also get vanishingly more rare. Right. So I mean,  we have run a lot of experimental designsourselves and our clients have as well. And we've never seen anything more thana four factor interaction in the designs that we've done in the kind of

systemsthat we've been exploring. And so then you can make assumptions, you can say,okay, we assume that actually the most likely interaction is going to be a twofactor attraction. And so,  we're goingto do a design where we strip that design down, so that it isn't necessarilylooking for the most complex interaction we might see in this case, a sixfactor attraction, we know that's not going to happen. So you can make thatassumption. But like, if we think about this, like it's essentially like a

asix dimensional hypercube, right? Just like the design that we're talking about here. So if you did everydesign point on the corners of that cube, then it would be two to the six,right? There's a lot of runs. But that would also be looking for six factorinteractions, which we know are unlikely to occur. So what you could do is youcould vary systematically  andsymmetrically take away design points until you get down to say, I don't know,a 32 run design, right?

Andnow 32 run design is perfectly manageable. And the interesting thing about itis if you imagine any one of those factors, we're actually comparing 16 datapoints with another 16 data points. Like there are other factors that arechanging simultaneously, but for each factor taken in isolation, you have a lotof replication  around that particularfactor. And so you actually end up with a lot of statistical power

forthen seeing whether these factors are having an effect, even as you're runningthese very complex experiments. Sorry, it's quite hard to explain all this. No,I think this is... Waving my hands around, but hopefully I painted a reasonablygood picture. After you've seen which factors are important, you could then spendmore experimental power on exploring exactly the effect of those factors as youoptimize them, right? And you iterate your way through the design space, right?

Youwork your way through until you find an optimum. Or at least, that's theclassic way of doing DOE. What's really interesting in sub space like assaydevelopment is they're not restricted to 32 runs. You know,  their final system, they want to end up withsomething that works in 1536 well plates. So that means that, you know, all ofa sudden, you've got 1000s of runs you can work with. And so then that expandsthe scope of these experiments dramatically. And honestly, nobody really knowswhat to do with all those runs.

Itis a really exciting time because  interms of like how we can experiment in the most powerful and effective way toaddress biological complexity, you know, like we're - because automation andsoftware like Synthace has now unlocked a scale and complexity of experiment,which was previously impossible. Like, you know,  the experimental design theorists,  they've not caught up. So  it's really cool. And I think,  this is then where there's going to be thatintersection between, you know,

multidimensional experimental design, automation software, and then AI. And  you know, what a fabulous data set to be trading AI against,  if you imagine,  50 different essays, where you'veinvestigated, you know, 15 factors for each of them and spread, you know, 1536at least design points for all of those 15 factors, you know, you've got thesesort of big kind of multi dimensional data sets in each case.

Andoften, actually, you don't even need to iterate these days, because there issuch a large number of experimental runs, you can often get to the answer justin a single experiment. So it can compress the timelines quite dramatically aswell. So there's lot of huge potential. Yeah, we've seen exciting things.  I'm always, I'm always the sort of personwho's then looking at, okay, so what's next? Yeah. And I think, you know, thereis a lot of

quitejustifiable excitement around AI. It is clearly an exceptionally powerfulmethod. And we all know that it's going to go through more hype cycles, right?Yes. Yeah, like, AI has been through a lot of hype cycles, it's going to gothrough more. But we also know that there is the heart of something very realhere. Right. And so it's very exciting to think about how that could beapplied, you know, when you combine it with these other technologies that canproduce the data sets

thatwe're talking about. OK. And you mentioned about the Synthace system using AIfor the meta analysis, et cetera. That's where we'd like to get to. At themoment, what our customers are using it for is generating these highdimensional data sets, making assays or other biological systems, often celldifferentiation is another area, generating stem cells.

Andthen  differentiating them intoparticular tissue types,  organoid types,or that kind of thing. That's, that's also the very complex areas of, you know,where you're  optimizing these media withall the right growth factors and things to make, you know, to make cells  dance to your tune, right, very, very complexand noisy systems. So our customers are using Synthace for kind ofsupercharging those wet lab experiments. But I guess, my excitement is that,you know, it's all about producing those data sets.

Andas we established before, data is the foundation for AI. And so the ultimatesolution is going to be the production of these really comprehensive highdimensional data sets and then the application of AI to those data sets. That'smy belief. Yes, absolutely. So what is Synthace's business model and how do youwork with the customers? Sure. So we have this very

powerfulsoftware platform that can enable people to do these experiments and, you know,get much more definitive answers from their experiments a lot quicker. And welicense that on a yearly basis, like a lot of software is these days, so it's ayearly license. But what we often find is that it isn't just the software byitself.

Right.So we also have a very deep expertise in these kind of ways of doingexperiments within the company. We have a very long history of applying thesethings to biology and understanding again that interdisciplinary interfacebetween the mathematics and the experimental design and what's theoreticallythe best designs you could run. But also then we really understand what'sactually possible in the lab and you know what's within the constraints ofactually running experiments.

Andso we've got a lot of expertise in that area. And so we work very closely withour clients to support them on this journey of changing the way they're doingtheir science, right? It's actually quite a profound change that it ends up being because you're going froma system where people are doing the science the way that they've been taughtfor, you know, however long is the correct way of doing science to then a muchmore kind of high dimensional

andmuch more powerful way. That takes some change management, it takes, you know,it takes learning about  this new way ofdoing things. And then obviously, the software helps them to actually run theseexperiments. So they see the results. And then that's the moment that peopleget converted, right? It's like, wow, suddenly, we don't have, you know, it's such a normal experience, as abiologist, to get results, which are a bit unclear, which are a bit ambiguous,which we're not quite sure what they mean.

Andthen you'll have lab meetings where everybody's debating about what theseresults mean, and everybody's got their own theory, right? When you have thekind of data sets, these kind of very powerful high dimensional data sets,there's no longer any debate, right? They're incredibly statistically powerful.And you've run pretty much every condition you could imagine running. And solike, you know, you can't have someone coming along very helpfully saying, butdid you try this? It's like, yes, we did.

Weknow we've tried everything and either it worked or it didn't work, but we canmove on, right? We have a definitive outcome. And we can  move forward with it. Anyway, so  yeah, back to the business model. So essentially, it's  software as a service, plus then,  level of consultation and partnership, really, like, we're working with thoseteams and those leaders who really understand that there's

afundamental shift that can happen with their experimental science, theirexperimental biology and the data that come from that, and the insight thatcomes from that. And so, you know, it's not a it's not an overnight change, issomething which has to be entered into very deliberately, but then it paysmassive dividends, right? Because it's assay development, again, back to assay development. It's a criticalpath activity, right? It takes nine months on average

andwe can compress that timeline. So now, even if we did nothing else, then we've moved those drug discoveryprocesses forward faster, right. And, you know, the promises of AI is that thenthat keeps progressing,  compressing evenmore, right? Yes. In future. But we already, you know, a good good way forwardwith that. But then also, we've been talking about assay quality, right? Sowe've made a better quality assay in a faster timeline. So now,

everytime that assay is now producing data, and every decision that has to be madeoff the back of the data that's being produced by those assays are higherquality decisions. Those data sets, those screening data sets, the iterativemedicinal chemistry that goes on, you know, every time that assay is used,those scientists will be getting clearer insight than they would have doneotherwise. And if we start training AI models on those data, those AI modelswill have clearer insight, right?

Absolutely.so, yeah, I mean, that is very exciting to both see, you know, what's alreadypossible, also where it could lead in future, right? So I have, I know a lot ofthese, you know, the AI companies that focus on specific targets or platformsaround that. But then how you describe Synthace, this is very unique in thatsense. Like,  is this kind of a uniquespace, like in this

designor so. Yeah, because I mean,  what are wedoing? I mean, we're combining, you know, a kind of way of doing science, akind of a theoretical framework, if you like, with software, with automation,you know, it's kind of a mixture of expertise and software. And like, youknow,  it's not a simple thing. And we'venot cut - the only reason we got to this place ourselves is because

youknow, we ourselves have been on quite an interesting journey of synthetics. Westarted out as a biology company, we only made software to help us do thesesuper powerful experiments, right? And so yeah, we are pretty unique. Andthat's both a real pro, because you know, we don't have directcompetitors,  at least technologically.And but it's also a con, because, you know, there isn't a pharma company out there with a line item, which says,

weneed to do, you know, automated multi dimensional experimentation, right?So,  in startup world, this is calledcategory creation, right? So you're, you're educating the market about what'spossible, and the problems that they're having that you can solve, and thenshowing them how they can be solved. And that's a much more kind of that's amuch more sort of consultative process. And, where you're chatting to people alot more about science than you are about, like,

youknow, the  specifics of exactly how thesoftware works, right? So, yeah, it's a really interesting journey. It's achallenging journey. But for me, it's just like, I think you could tell whatkind of a person you are offered by what makes a good day for you, right? Youknow, like, when have you when have you gone? You know, when have you gone homehappy?

Andthe thing that never gets old for me is when I see the kind of sheer power ofexperiments that people are running in the lab, right. And when I get thechance to actually see those experimental data and to see the outcomes, that'sjust really exciting to see. Markus Gershater, Chief Scientific Officer and co-founder of Synthace. Markus, thanks for your time today.

It'sbeen a real pleasure. Thanks for the chat.

Well,Amar, what did you think? Yeah, it was a fascinating discussion. This is apretty unique company focused in the design of experiments. And this issomething that was usually when we're talking to alot of the other companies,it's about their developing specific algorithm or AI software to develop aplatform for targets or a specific type of targets. This is a different type ofcompany where

they'reusing data analytics, AI, to get the best out of the experiments and even liketo set up the right type of experiments. One phrase he used that really stoodout to me was, it's more about the data than the algorithms. I see that as amantra on t -shirts. What did you think of that? Absolutely. And we've talkedthat quite a bit, right? It's garbage in, garbage out, right? So if you don'thave the right data,

nomatter how fancy your algorithm is, it's not going to be able to do anythingbecause it needs to be trained on the right data. So that definitely is themantra. Data is the most important thing here. You did talk a lot about theneed to focus on data production and making it not only better but faster. Arepeople thinking about it in those terms with regard to their data collection orare they

justthinking about feeding the beast of AI. I think that has started happeningquite a bit. And I would say, especially in research, the scientists are verymuch aware of that. Because in science, these kind of experiments, especiallythe omics experiments have been going on for a couple of decades now, more thanactually almost three decades now. So the scientists in the pharma companies orin academia are used to setting up these kind of experiments  and then

makingsure that all the data is collected. So I think that cultural transformationhas happened in research. I'm not sure about commercial. In commercial, I thinkthat in the pharma, I believe that's still happening at this point. Somecompanies are more ahead than the others. So I see like in the value chain inpharma,  I would say, research is whereit's most advanced.

Evenin clinical development when it's about clinical trials and running those of coursethe clinicians are very much aware of that but when it comes to getting some some new aspects running, some differentnew types of assays or so, the clinicians they need to be  very well aware of that. So I would sayculturally it's there in research, in some of the other areas. It still needsto be systematized

inthe other areas. He talked about enabling experimentation in a high dimensionaland automated way to address the complexity of biology. How novel is whatSynthace is doing? See, this is something that needs to be done for the rightassays. And we did talk about biological systems being very complex, having somany different factors. So for you to get the right

conditions, you do need to play around with alot of these conditions. And then that's what you talk, you know, this is thedesign of experiments, right? That's the field. And so it needs to be done verysystematic way. I think it varies, like we're different scientists. Somescientists can be very disciplined about it, some others may not be. So whatthey are bringing is they're just bringing like this, they're just...

Andthen they're making it to be very much a science, Like, okay, boom, boom, boom,you have to go through this and you have to do these experiments, you have totry out all of these. So making sure, I mean, what they are enabling is makingsure that the scientists can do the best experiment that's possible. And thatto me, having that kind of a disciplined approach, I believe is pretty interesting.I mean, that should be how all experiments need to be running in the idealworld.

Andthey're trying to do that. Markus, at the end of the day, is talking aboutchanging the way scientists in drug discovery have traditionally operated. Hetalks about being able to win them over with powerful data sets, but how big achange do you think he's advocating? And how much of a barrier is it to changethe behavior of the way drug developers and scientists operate? So it is goingto be how

disciplinedthe scientists are, right. I mean if they if they really want to go through allthe different conditions or so I think those are easy to convert thepeople  but also at the same time youhave to see  that  there are people who are more like - who aremore comprehensive and like more process -oriented versus there are people whoare more like, okay well, you know, I think this is gonna work man, you know,this is, you know, hey I have done, you know, I have a lot of experience inthis area. I know

whatshould be the ideal conditions. Let me just do it my way. I don't want to tryout a lot of different things, right? So there are people who are more like,okay, well, I know what I'm doing. What is this going to add? So I think thoseare the people who are going to be more challenging to convert into this, wherethis is really going into systematic way and looking at the entire space, themulti -dimensional space that he talked about of the factors to really figureout. So it's going to depend on the personalities. Some people are right there,some people are not there.

Well,it was an engaging discussion and I'm looking forward to the next one. Thanks.Yes, absolutely. Thanks, Danny. Thanks again to our sponsor, Agilisium Labs.Life Sciences DNA is a bi -monthly podcast produced by the Levine Media Groupwith production support from Fullview Media. Be sure to follow us on yourpreferred podcast platform. Music for this podcast is provided courtesy of theJonah Levine Collective.

We'dlove to hear from you. Pop us a note at danny @ levinemediagroup .com. For LifeSciences DNA and Dr. Amar Drawid, I'm Daniel Levine. Thanks for joining us.

‍