ML-Enabled Drug Discovery With LabGenius' James Field, Ph.D.

Enabling a community of peers is what the Business of Biotech podcast is all about . It's also what our new Business of Biotech newsletter is all about . Peer driven content , no strings attached , delivered to your inbox once a month . Go to bioprocessonlinecom backslashbob to subscribe .

The Business of Biotech is produced by Bioprocess Online , part of the LifeScience Connect community , with support from Citeva . Citeva also demonstrates its commitment to the leaders of new and emerging biopharma at Citevacom backslash emerging biotech . Check that out .

It's anecdotal , but in my experience , scientists hold a healthy degree of skepticism toward the role of machine learning and biopharma drug discovery , as it should be , given the nature of a business where next level human scrutiny and interrogation are primary to the job .

On the other hand , there's a growing frong of startup biopharma companies to which machine learning is the cornerstone of discovery efforts . My guest on today's show is founder and CEO of a company intent on marrying the best of those perspectives , combining the best of human ingenuity and machine intelligence .

The company is lab genius and it's developing what it calls a smart robotic platform named EVA that's capable of designing , conducting and learning from its own experiments in an effort to discover new therapeutic antibodies . Importantly , it's applying that platform to the development of its own internal pipeline of therapeutic candidates .

The CEO is Dr James Field , and I'm thrilled that he's agreed to join us today for the show . Dr Field welcome .

I'm excited to have this conversation with you , but having a lot more of these AI ML discussions on the podcast with companies that are either jumping in as a platform developer or jumping into AI and ML from the perspective of a therapeutics developer , it seems to me that for the past several years , machine learning was coming It was calming to drug discovery .

People were talking about the potential of its applicability . Then , for some time , some folks seem to be dabbling in it a little bit at the surface level All at once very recently . It's seemingly fundamental to drug discovery for so many young companies . I want your perception on that .

Obviously , i'm going to ask that question , knowing that you may have a bit of a bias . Is this height-fueled perception or is there something real behind this perception that ML has become fundamental to new drug discovery companies ?

The thing that's changed is it feels like there's been this sea change now , where it's now universally acknowledged that the deep integration of machine learning and computational methods into every stage of the drug discovery process is now an inevitability rather than maybe just a possibility .

That being said , i think we're actually just at the very start of the technology adoption curve . It's interesting to think about this when you ask that question . Folks have been , as you say , dabbling in this space for a while . One question to ask here is why are we not further ahead ?

From my perspective , here , it all boils down to the availability of what we call machine learning grade data . Anyone from the computational sphere will have heard the old adage garbage in , garbage out , but I think it really applies to the world of drug discovery here as well .

For the majority of problems that drug hunters need to solve , the real big issue is that there are no readily available datasets of the requisite quality for machine learning . What's the impact of this ?

Well , if you want to ask some of the answers , some of the most interesting and potentially impactful problems in drug discovery , then it means you actually have to generate new data and ensure that from the outset it's collected and that data is structured with machine learning in mind .

So I think the thing that's happened is that there is now this , this acknowledgement that machine learning is going to touch every stage of the drug discovery process , and the race is now on for folks to develop these engines that can generate this machine learning grade data at the right throughput , and that's really , it's really not trivial .

It requires a lot of integration , not only from sort of the data generation perspective and the machine learning perspective , but also around data capture , storage , processing , etc . And I think that that's the reason that we're , that we're not sort of further along in in that process .

I've had discussions with companies who are playing in the space , who are attempting to leverage great big , publicly accessible data in target discovery and so on , and I've had conversations with companies who , on the other hand , are using very proprietary internal data and sort of limiting their application to that .

You know , obviously , different applications , so it's sort of a big question to answer . But what's your perspective on sort of that starting point of data and drug discovery ? is it , is it big data ? is it proprietary ? you know smaller data sets . Does it depend ?

And you know , obviously there's been a lot of work done around answering questions around how proteins fold , using some of these methods and pulling some of those large open data sets there So that those open data sets definitely have sort of utility and they help us to answer certain questions .

From my perspective , some of the most interesting questions The data doesn't actually exist yet , or where it does exist , it doesn't exist in the right quantity or quality on the web .

And so that's where this real challenge now comes in , where , if , if , as drug hunters , we want to answer the most interesting questions , we've got to build the engines that can then generate the requisite , requisite levels of data , and then you can sort of go a layer deeper there and actually say that that look some of the techniques and approaches that you

might use for these kind of you know , large open data sets .

Maybe that's very applicable to those sorts of problems , but actually there's a whole other branch of machine learning which is much more suited to addressing the challenge of how you deal with these smaller data sets and how you search and explore the drug design space using using small data sets and efficiently running that process .

And I guess the point I'm trying to make here is is that we'll see the application of machine learning across every stage of the drug process , answering all sit different types of questions . And we've only just really begun to answer a few of the low hanging fruits there and I think the next few years of for the industry will really be around .

How do we start to grapple with some of the harder questions where you can't just sit down and open your laptop and apply an existing machine learning approach or develop a new machine learning approach ? it'll be around . How do you generate the right type of data sets in order to solve some of these hard problems ?

So the problem space that we're interested in is is obviously in the protein therapeutic space , specifically the antibody therapeutic space , and we're not interested in predicting how a protein folds . We're not interested in predicting how protein binds to another protein .

What we're really interested in is using machine learning to predict and understand how a protein design impacts the way that that molecule will perform in the context of a very disease relevant cell based asset .

And the reason that that's very interesting is is that , historically , running those sorts of experiments is exceptionally low throughput , so you can only gather a small number of data points here , and so this is where we feel that machine learning has the biggest potential to to provide an advantage in the job discovery process , and so the course of application spaces

we're talking about here is you know , how do you develop multi specific , multi valent antibody therapeutics that really have quite complex mechanisms of action ? and what you're trying to understand there is if you pull many different levers in parallel in terms of how that molecules design in terms of its geometry , its valency , its affinity , its topology , etc .

How do those impact on on its performance ?

and we're not just talking about the performance of one property here , we're talking about the performance of several properties , and this is the big challenge in drug discovery , specifically in protein engineering , where , historically , the way in which protein engineers had to engineer these molecules is that you would sequentially optimize one property after another , and the

issue there is you optimize for one thing and then you've actually inadvertently deoptimize for another , and then you've got to try to try to address that , and the beauty of this approach is that , by generating data for each of those features of interest , we can independently create these computational models that are predictive of those , and then we optimize across

them all in parallel . The big challenge there , though , is around how do you generate that data at the right quality , at the right throughput , at the right speed , etc . And so there's been a lot of innovation and engineering around how we solve those particular problems here at Lab Genius .

So Imperial College , which is the university down the road here in London , had a very early PhD program in synthetic biology And I was really excited , signed up to that and got into the lab , and then none of my experiments worked , and in the time between I had experiments failing , i'd sit down and I really had to make progress for my PhD And I thought ,

goodness , the only way I'm going to do this is if I apply some computational methods . And so I actually managed to sort of teach myself to code whilst doing the PhD . But of course now within the business that we are today 50 person company we have much , much better approaching engineers and much better computational people than myself .

You can bring computational folks and you bring traditional kind of wet lab scientists folks together And they speak different languages , They come from completely different walks of education And I know that the education systems generally are starting to get better at marrying those in graduate school and beyond .

But do you , has that been a challenge for you and lab genius at all ? I mean , have you sort of seen that sort of physical human challenge of marrying these two mindsets ?

When I started this company , i thought we were trying to solve an engineering , maybe computational problem , maybe a scientific problem .

Actually , we're a major part of this is an organizational engineering challenge And , you know , i think actually being a startup gives you some really nice advantages over maybe some of the larger incumbents , in the sense that , rather than having different teams that are siloed , you know , in different parts of the world , you can get everybody in one room , in one

office , and it's only through that really really close collaboration , in that cross functional way , can you really start bridging the divide between some of these different domains .

And the reason that's so important for a company like LabGenius is , as I say , you can think of this technology stack that we're building a little bit like a pyramid , where right at the top you have , you know , the data analysis and the machine learning And the real foundation that that sits on is the ability to generate data , to capture and store it , to

process it . And again , if you can't do any of those steps , then the machine learning isn't going to perform as you need it to , and at each of those steps you have to have really really tight integration between the domain experts and the domain experts . For the data generation size side are not the domain experts , obviously on the machine learning side .

And so , from something just as simple as how do you structure and design an experiment , if you were to do that in the conventional way , the conventional way in which maybe like a trained biologist would do it , you wouldn't necessarily include all of the controls that are absolutely essential for allowing the normalization and the noise reduction that you require from

a machine learning approach . So it's definitely non trivial . It's very hard to re-engineer existing processes and , say , a big pharma company to achieve this kind of to this marriage of the two domains , and so I really think that small startups , biotechs , are actually at an advantage here and really kind of need to lean into that .

Because the field is moving so quickly And there is so much competition , you absolutely have to make sure that you're pulling every lever that you can stay ahead .

And those people are fairly rare and unique . And so I would say , when starting a business , you absolutely have to make sure that you have these folks , these multidisciplinary folks who have experience not only with the industry of drug discovery itself but with those deep technical domains .

And then as you start to grow the org , people can become a little bit more specialized . But getting the timing of that right is absolutely critical , because if you're not fast enough you fail to go deep enough And obviously if you're too slow you have a similar set of challenges .

And specifically , we're interested in the problems that you can't solve using conventional methods , Because I should say from the outset that there are a lot of problems in drug discovery that you absolutely do not need to apply advanced computational methods and machine learning to .

And often you do see , because there's so much heat and hype around machine learning , I also see a lot of what I would call the misapplication of some of these methods . Applying machine learning to areas in which , to be honest with you , it doesn't provide a huge amount of value .

And that's actually especially the case in biology , because often the knee jerk reaction thing to do , if you're kind of leaning in from a machine learning angle , is to say okay , where are the data sets ? And the data sets are often linked to the assays which you can already run at high throughput .

And , of course , where you can already run an asset at high throughput , often you can brute force the solution itself , and so there is actually there an interesting trap that I think a lot of companies can fall into , where they say , hey look , i want to apply machine learning to drug discovery , where are the data sets ?

And then they get pulled into trying to solve a problem that's already been solved .

So we're really very much interested in solving the challenges of protein engineering , protein engineering challenges that can't be addressed by the conventional methods , and specifically , the one that we're most interested in right now is through the development of sort of novel multi specific antibodies , specifically immune cell engages , where we're really trying to engineer

these molecules by pulling all sorts of levers in parallel that relate to the design of the molecules , to better design them such they're able to distinguish between healthy and disease cells and ultimately address the issue of on target off tumor talks . Yeah .

human engineering , you would have designed yourself . And the reason that's so exciting is , you know , i think drug discovery augmented by machine learning , has the potential to be faster , be cheaper but , most excitingly , it has a potential to be truly enabling .

And it's enabling in instances where the actual output , the actual result , the new molecules that are generated , are those that wouldn't have been obtained through any other method . And the reason that we're able to do this is that , historically , if you're only able to say , use conventional methods to evaluate a small number of antibody designs , you play it safe .

You say , okay , here are ways in which I know I can combine these components And I'm pretty sure I'll be able to make the molecule . I'm pretty sure the molecule will will work as I'm predicting , and maybe you get , i don't know , very , very best case scenario . You can test 500 to 1000 of these molecules in these really complex cell based assays .

The advantage of our approach is that we can evaluate up to 28,000 of these molecules in these very disease relevant cell based assays , and we do that in such a way , in this kind of iterative , cyclical way , with every cycle that the machine learning algorithm is pulling us through the design space and how to better recombine the parts .

So what you end up , what ends up happening , is is that you get pulled into these regions of design space that is a rational human , patient , engineer you would have never even thought of going into because it wouldn't , it wouldn't make sort of rational sense , although we knew precedent for it .

And that's the thing that I think is is is most exciting about this sort of approach .

You know , i guess maybe it's human nature to push back on that a little bit . You experience any , any , any pushback or healthy skepticism .

And then and then the algorithm says here are your first set of designs that you should test , and the human protein engineers look at those and they say Why on earth are you wasting like very precious resource in the lab testing ? and they tell you for reasons x , y and z that those molecules are never going to work . And you know what ?

sometimes they're right and sometimes they're wrong . But the difference here is that the algorithm , the process , learns from every experiment that it does And that knowledge is institutionalized and in a way that when you run your next program , your next cycle , it doesn't make the same mistakes .

And I think , i think that's the , that's the really interesting , interesting thing here , which is when you're taking this data driven algorithmic approach , the algorithm specifically sometimes requests you to make designs that a human protein engineer may say you know , i don't think this is going to work . But the algorithm what , what , what it's doing is .

It says here are areas of design space that I myself am uncertain about . I want to test these so I gain more information , more predictive ability there .

And again , this is the thing that is ultimately the , the lever that we can pull that pulls us into these areas of design space where , yeah , the human protein engineers would say , you know , maybe I actually have no intuition around what's going to happen there , and we often , often see that what I would call these rules of thumb , for sometimes they work , but

in other programs , actually the if you apply the rule of thumb in terms of engineering these molecules , it breaks them , and I think that's where this very data driven algorithmic approach has a potential to really outperform a human level protein engineering .

The pot is brought to you in collaboration with Citeva , a global provider of technologies and services that advance and accelerate the development , manufacture and delivery of therapeutics , from idea to injection . Check out their resources at Citevacom . Backslash emerging biotech . That's CYT , ivacom . Backslash emerging biotech .

How far down the path you mentioned , like you know , the thought process might , process , might be well , that'll never work . You know the human thought process and then machine learning can , can contribute to proving that human thought process wrong , wrong and right . They , wrong and right , experience different degrees of relativity along the drug development continuum .

Right , like it's a constant , constant experiment . So how far along would you say lab genius is with some of its more promising molecules to , to proving them truly , truly right ? And how do you assess that Is ? that question makes sense .

and one of the ways in which we can limit risk is by reducing target risk . So we don't necessarily want to go after targets where there's a lot of , you know , novel biology that may or may not be proven out in the clinic .

So we're actually privileged in the sense that we're going after very well validated targets , often where there are clinical benchmarks that we can draw upon and run head to head with our molecules . And that's typically what we'll use in any program where the molecular product profile that we're shooting for will have these clinical benchmarks in mind .

So we can certainly say for the molecules that are coming out of our discovery process , they're beating those clinical benchmarks with respect to the parameters that we think will make those molecules certainly valuable and useful for patients .

it's so capital intensive , time intensive to onboard any , any assay , any set of experiments so that they can generate data at the requisite level for machine learning . You have to be really sure that this data generation pipeline is engine , that you've created maps to multiple different potential programs .

you don't want to be setting up one of these engines where maybe it just addresses one particular problem . so so we've had to think very carefully around what are the right spaces within the world of drug discovery where you can set up this pipeline and continuously run multiple programs through it without having to refactor the platform every every time you do ?

And that's why we were pulled into this area of thinking around how do you equip these molecules to differentiate between healthy and disease cells ? because ultimately , that boils down to the , to the molecule being able to sample multiple receptors on a cell surface and effectively make a decision .

And so the area that made most sense for us was in the immune cell engage space , and so the indication area we're most focused on is , of course , oncology , and then within oncology , obviously I think you could apply this technique to address multiple different types of cancers , but we're most particularly focused on solid tumors .

So within that area , we're developing T cell engagers that have been engineered so that they are much more selective in terms of their killing profile for cancer cells over disease cells .

So the show we focus on new and emerging biopharmas and these stories about how companies form up are super instructive , right and inspirational to our audience . So tell us a little bit about that experience for you , your sort of academic into industry overlap and how that sort of directed the beginnings of Lab Genius .

And then I took part in this competition called the IGM Competition or the International Genetically-Engineered Machine Competition , which was basically this really unique opportunity where as an undergraduate , you're given a summer to do any research program that you want and then you all go out to MIT to present your work And it's this fantastic , as I say , international

meeting of the minds for anyone interested in synthetic biology . And that really sort of set me on this path of hey , actually there's something really exciting here .

I had gone from looking at biology as a descriptive subject to actually something where you could start to engineer the substrate of life And that book that was really sort of a lit a fire up inside me .

And then I had this incredible mentor and professor at Imperial College who said looks like you had fun at that IGM competition , why don't you come and do a research master's ? You can work on any problem you like , you can do whatever you like in the lab , and that led into a PhD .

So I was given these four amazing years of just sort of messing around doing experiments in the lab , at the time when all of these breakthroughs were happening , like the cost of sequencing and synthesis was coming down , all of this robotic automation was being installed in the labs , and it just felt like this very opportune moment where it was very clear to

everyone that protein engineering was currently very inefficient but it had the potential to be transformed , and so really that was the impetus for me to kind of start the business . And , as I say , we just had an idea at that point . But it was just at the right time because the UK government made this investment in the translation of synthetic biology ideas .

So they set up this incubator at Imperial And again I've gone through having the run of the labs for my PhD to be able to set up a free lab space in this incubator , and it was really only sort of the confluence of those factors allowed me to get the business off the ground .

And probably one of the characters to really draw out and recognize is our chairman , edwin Moses , who built and sold a very successful business , abelings , which he sold to Sinovia And that's one of Europe's greatest recent exit stories But has subsequently mentored a huge number of founders and CEOs across the European biotech ecosystem .

And I would say that's probably one of the relationships that has been most valuable to my own personal growth journey . And I would add that one of the really interesting things about building a business like this and I say this to the team is that every three months it feels like a materially different company .

And that might be because someone new has joined and they bring in incredible sets of skills and experience . It might be because the market's just changed from being a bull market to a bear market , or vice versa . Or it might be because one of your programs has advanced and suddenly you need to be thinking of a whole new set of considerations and challenges .

So I think it's one of the most exciting industries to be working in for that very reason . And really then , in terms of , I guess , equipping yourself to deal with that , it all comes down to adaptability . Can you continually learn and grow every three months as you're effectively getting a new job ? or at least that's what it feels like .

That can be tough for employees to embrace , Like you may be willing to embrace it , but it might be tough for your organization to embrace a constant rate of change and feeling like it's a new company all the time . How do you manage that , Like how do you keep the team paddling in the same direction when it seems like the currents are always changing ?

So folks who have worked for big tech companies experiencing one type of working environment , to other folks straight from the academic labs and other folks from biotechs And every one of those people comes with a very different history and different expectations .

And I think the reality here is the best way to provide some stability and consistency to a team is to be really , really clear about what your corporate goals are . That's only in terms of from a delivery perspective . Everybody comes in and your corporate goals don't change . That's your North Star .

But probably more importantly , there are the cultural aspects as well , and I think you have an opportunity there to define not only the values but also the mindsets that you expect from your team and what each member of the team should expect from each other as well .

And again , if you can get that right and set those expectations , then again that's the ballast that gives your team confidence in an ever-changing world , especially now where there's so much uncertainty around what's the space doing , the financing , environment , inflation , etc . It's really important to try and create that kind of consistency .

I know if you go to Silicon Valley there might be a great , big , giant appetite for machine learning startups In biopharma . Maybe it's a little more tepid , i don't know . You tell me what have you found the appetite for machine learning based startups in the space and do you see maybe some change ?

So I think actually it's changed very much for the better , because a lot of those investors who made their initial bets a couple of years ago have now got either were really smart to start with , have got really smart or they've left . So you're now pretty much the investors who are still investing in the space . They absolutely know what they're talking about .

A lot of them have the scars from where they've made bets that haven't worked And they've equally had the pattern recognition for backing the businesses that do really work And enables you to have a much more informed conversation about the type of business that you're building , because the reality here that you and I know that a lot of the value in this space is ,

of course , locked up in the molecules And it's not the case that you can sort of march straight in and say , hey , we're going to build this whizzy new machine learning tool and scale it in the cloud And then we're all going to be millionaires . It's the case that this is a long-term process .

We're going to ultimately find new molecules , and the process of finding and developing new molecules takes time , it takes capital , and I think the investors that are remaining in the space now are very attuned to that .

And then that gives you as a business the opportunity to have a much more informed discussion with any prospective investor about the journey that you're going on as well .

Those firms are either starting in , certainly in our backyard London-based offices , or you have homegrown firms that are popping up in the space as well . So I would certainly say there's less of a geographical restriction than maybe you would have seen sort of five or six years ago , Specifically within the , I guess , the biotech space .

What's really interesting is that some of the most sophisticated investors that I've seen are actually coming from the pharma companies themselves , in the sense that every pharma company now has its own machine learning , has its own AI strategy .

They've been working on this for some time , making in many cases , very , very large internal bets on their own capabilities And , as a consequence , their venture teams have been able to kind of really get in deep there , understand again what works and what doesn't work , And you have some exceptionally smart folks in some of those teams as well .

So I would say the background to that really is that building a company is a very , very painful experience and it's very challenging , and the main reason for wanting to invest in companies is really where I feel like I have a good connection with the founding team and hope to be able to at least share some pain and hopefully some learned experiences there as

well . Typically , the types of folks that I invest in are folks who are operating in the therapeutic space and typically who I've known for several years And it's been really interesting , after making a few investments over the years , watching which of those companies really accelerates and which are the ones that maybe are moving a little bit slower .

And I would certainly say I can take some of those learnings myself back to my work at Lab Genius And I think that makes me a better CEO as well .

And I would say probably the most important or consistent sign of success is that the companies and the founders who are most consistent in the way that they tell their company story and their focus are the ones that have done the best , and I think really that's a sign of there was this exuberance in the capital markets a few years back where investors were

pushing companies to do more . So they would say , right , grow the opportunity here by , rather than doing one thing in a very focused way , do two or three things , et cetera .

And those companies , those founders who remain very , very focused , saying , okay , we're just going to ignore all of that noise , we're going to focus down and solve this one problem really well , those are the ones that have done terrifically well .

And , again , i think it's that continual reminder of focus which really is brought into sharp focus by the current economic environment . Now That is what is required , not only in a small startup , but across every organization .

I mean , you can certainly elaborate on that , but it's intriguing , yeah .

But what's really interesting in their company is it fits into this sort of broader thesis of the fact that biotechnology , our ability to engineer living matter , is not just obviously limited to the therapeutic space .

That's a space that everybody obviously really focuses on and it's where the most established value is being created but ultimately will touch every component of our lives . And I find the kind of the food industry kind of interesting , right , because it's this other absolute extreme where maybe in the farmer space , you know , cost of goods can be exceptionally high .

In the food space your cost of goods has to be exceptionally low . And so from a kind of technical engineering challenge , that is a really fascinating challenge of how do you solve it , how do you make really high quality food grade substance from some of these engineered cell lines ?

And really that requires innovation , both on the computational side , the bioprocessing side and of course , within synthetic biology as well . So again , really fantastic team doing really excellent work .

And again , this is , the applications are very broad , not only the food chain , but other areas as well .

But again it's an interesting analogous story of somebody who again got really excited by the IGM competition , has founded their own business and is now doing amazing , you know , really transformative work And that I see is a really consistent narrative across a lot of the synthetic biology startups that are in existence today .

That obviously then feeds into what are the things that we need to develop on the healthcare side , And so , if we really want to make inroads into some of humanity's biggest challenges , I think you have to take an integrated approach and think about all of these things in parallel .

So , you know , I think the next phase for lab genius is keep our heads down , keep pushing , keep working , progress those molecules towards the clinic .

And again , that all takes time and money , but I feel like we've got the right team here to do this And I'm personally very excited about that journey because , as I say , every three months feels like you're in a different company , And so the learning journey there is is going to be phenomenal .

I mean what you guys are doing compared to a lot of the conversations I'm having with therapeutic developers or , as I say , dabbling in machine learning , what you guys are doing is pretty advanced And I really appreciate the opportunity to talk with you about it .

If you like listening in on conversations like the one you just heard with Dr Field , subscribe to the Business of Biotech podcast . Sign up for our newsletter at bioprocessonlinecom backslash B-O-B . Also , be sure to leave us a review , let us know how we're doing And , as always , thanks for listening . Music .