Talking Biotech 404 with Dr. Kevin Folta

Guest: Dr. Daniel Veres, Turbine

Kevin Folta: [00:00:00] Hi everybody, and welcome to this Week's Talking Biotech podcast by Colabra. A study of biology in action has some important limitations. How do you look at a biological process in a diversity of patients or understand what's happening in a specific cell type or developmental condition? How do you understand something complex like genetic changes associated with cancers?

Well, one way is to explore a reductionist hypothesis and try to dissect the nuts and bolts of that biological process. But even then, it can be tricky, if not impossible. Plus it takes money and careers and ultimately has a high chance of failing or, or not translating to actual clinical scenarios. But what if someone took the opposite approach?

What if analysis of massive data sets could be used to generate a synthetic cell? A computational output of merge data that could not only mirror a biological process, but could also predict missing [00:01:00] biochemical players or better yet point to vulnerabilities that may serve as new drug targets. The This is what's happening at turbine.

Turbine synthetic cell is a confluence of big data machine learning and using AI to make predictions, and it's leading to the discovery of new potential therapies. Today we're speaking with Dr. Daniel Veris. He's the c e o and Co-founder of Turbine. Welcome to the podcast Dr. Veris. 

Daniel Veres: Thank you very much, Kevin.

Nice to be 

Kevin Folta: here. Yeah, it's very exciting to have you aboard because I learned about your company and had so many questions because I see this as such an awesome technology to be able to actually be able to make predictions in silico using basically a simulated cell. But before we start talking about that, what are some of the major traditional methods of drug discovery and, and maybe some of their limitations.

Daniel Veres: Yeah, so usually how you run drug scale project is of course you have a hypothesis [00:02:00] from the get-go, and that hypothesis can come from literature or it can come from some previous experimental work. And from that you formally that, okay, I'm looking for a new target or a new biomarker, combination, et cetera.

And you start a traditional. Process of why they did that. Hypothesis and validation is usually starting, especially in oncology and turbine operates in oncology field exclusively. You start with an in vitro phase where you are looking at for example, cancer cell lines, potentially primary cultures of human tissue, and you use that.

And either pharmacologic or genetic report or be to recapitulate some of the biologic Omahas you had in your hypothesis. And that's already a limitation in one on one hand because you have a very limited set of models you can use even in the in vitro phase. So you can use roughly, let's say 1000 cancer cell lines if you have access to those which are representing.

A big chunk of the patient population, but many are missing. So for example, [00:03:00] a good example is prostate cancer. There are maybe six cell lines, which are prostate cancer cell lines, and none of them are really representing, for example, the castration resistant metastatic prostate cancer patients. So you are missing out on many.

Potential contacts, magical contacts, which are representing patient populations. That's already a problem in the in vitro phase. Once you see a signal in vitro, then you go to the in vivo, which means that you are taking a higher level animal like a rat or a mouse. And you are doing an, an experiment in that certain model where you try to ate your hypothesis, which.

Had a signal in vitro. So here, one potential bias is that you only validate what already worked in vitro, so you will not get an answer, which is only working in vivo because you didn't really. Run a process, which can just start with an in vivo hypothesis. You need to have the in vitro first and then the in vivo.

And once you are done with the in vivo proof of concept, then you're [00:04:00] moving to patients. And many of the drugs are failing in the clinical trial is mostly phase one, phase two, even in phase three. And the reason of that is that the drugs are not working in the patients, which can be on one hand because you can't dose high enough.

So basically you don't reach a dose in which you see efficacy because of the. Toxicity related to the dose of the drug, or you are, you, you are able to dose high enough to hit the target, but the target is not doing the same biology as it was supposed to do based on the in vitro and inv one analysis.

And because of this high failure rate in the clinic and because of the low above translatability of the in vitro and InVivo models, everybody is basically focusing on. Getting as fast as possible to patients and failure there, because that's the ultimate proof or failure of a program. And this of course means that we are on one hand dealing with a lot of failure, but also on the other hand, we are dosing patients with drugs, which we are not work.

Or if they work, they have a very limited benefit. [00:05:00] This is something we call a translatability gap. So basically there is an issue of what are the cancer models you use and how predictive are they for clinical response? And the idea behind what turbine is doing, for example, is to try to limit this gap and basically have more patient relevant behaviors predicted much earlier in the process so you are able to get to patients level behavior earlier.

And with that, increase the success in the clinic. 

Kevin Folta: Yeah, and from what I've understood about turbine's approach is that you're able to use machine learning and AI to be able to make predictions about particular targets using an in silico model of discovery. So could you tell me a little bit more about that?

Daniel Veres: Yes, indeed. So our technology is called SIMULATE itself, and it's basically at the current development stage, it's a representation of a population of SA cancer sas which is based on their signaling. So how they are deciding about. [00:06:00] Certain internal external stimuli, how they are deciding about sulfate, how they are responding to certain drugs or other perturbations.

And we are basically building up these simulated cell models based on the signaling model, adding omics data like genomics, transcriptomics, proteomics, et cetera. And with that, we are able to build. Kind of inci avatars of hundreds of cell lines, hundreds or thousands of InVivo models and also patients.

So the possibility here is that you are able to build, for example, a model of an InVivo cancer model or an in silico model of a patient in the basically first steps of the discovery process. And with that, you are able to. Guide the experimentation. So we are running a much bigger source space with a lot of inci patient relevant models and finding the right models in which we then run the validation process.

So in, in some way or form, we run the same process, [00:07:00] which I've described in the first answer, but we are guiding it with the simulations and it's important that it's not just AI and ml. But it's basically a combination of simulations and ai. So with that, we don't just get scale from the AI part, but we also get interoperability, which comes from the simulations, which is in biology important.

Because the predictivity rate, even for the best models, is not yet close to, let's say 90 plus percent. So you need to understand what, which prediction is actually making sense. And for that, you need interpretation. For that, you need magical understanding. 

Kevin Folta: And it seems that your approach really as a first step really just limits the commitment to these longer term processes.

So is it really just a question of speeding that first step and taking a better educated guess at what these targets may be? 

Daniel Veres: That's a very good question. Not exactly. So basically we are applying the technology throughout the whole drug discovery [00:08:00] process. It's not just a point solution happening in the very beginning, but we are informing basically most of the badge related questions throughout the process.

So it, it can be. You can start with identifying a target using the platform. Then you can continue identifying biomarkers, and this is positioning for that certain target later a drug. You can also use it to select which lead series is the better if you have those information to run simulations. You can design combination strategy, you can design clinical trial exclusion, inclusion, cri exclusion criteria.

So it's basically an iterative process where we are using the data, which is available, and of course the data is accumulating throughout the process, inform the model, have a better prediction, and then test. In vitro inpatients if that response is really there, what we have predicted. Well, as a 

Kevin Folta: scientist, I, I find this super exciting because I've thought about questions in molecular biology and protein, protein interaction and drug targets for a [00:09:00] long time.

But one side of me is really skeptical because I think that cancer cells, as we discuss frequently on the podcast, evolve rapidly and have many differences between them that dictate their instability. But, Does your model, your simulated cell kind of group cancers according to commonalities between them that really offer the, the targets and similarities that all of them, or at least a part of transform cells may share?

Daniel Veres: Yes. So basically you can imagine this as an integrative framework, framework of protein interaction network signaling network bid. Cell type specific omic dataset. So basically a simulated cell model is not a general model of a cancer cell, but it's specific to that certain, let's say, sample you have.

So it, it can be a certain. Cancer cell line, or it can be a certain in vivo model. It, it can be a patient [00:10:00] sample based model or even it can be a single cell resolution set of models if you have that data. So with that, you are able to capture hundreds or, or thousands or even more different contexts of cancer cells, or not just cells.

These are basically populations of cells and about heterogeneity. This on one hand gives you a, an intro, two more heterogeneity. So basically between the different tumor, more types and indications, you have representation of all of them with different let's say cell lines, models or patients. And this is for example, important here that I mentioned, prostate cancer before.

In vivo. And of course in patients you have a much better representation of prostate cancer, so you are able to actually run experiments in iCal in those samples and ask questions which you would otherwise not be able to do in vitro. The other layer of heterogeneity is basically intra to more heterogeneity, and that's basically what is driving evolution.

There we are trying to capture some level of heterogeneity. In [00:11:00] the technological level, but that's, for example, something which we are still working on. And that would require further development to really capture that. And of course it would require data from more like a time series experiments where you actually can model the adaptation of the tumor to the treatment.

Currently what we are doing is that we are able to predict what are the potential escape route or resistance mechanisms. So you can, for example, apply a drug on the system. And then perturb certain pathways or certain proteins and see if they are changing the response. And we have shown and validated experimentally that in many cases we are predicting.

Escape route and mechanisms, which if you can block, then you have a better response and you can block that by having that as a biomarker and intrinsically basically not having that escape route. Or you can block it by doing a combination of two drugs. And what we are working on now is also to basically build that, let's say, [00:12:00] polypharmacology into one drug.

So for example, having a kinase inhibitor where you're not just. Targeting one pathway, but two pathways with the dual inhibition. And that is informed with the simulations about these potential escape route. 

Kevin Folta: This is really cool stuff. Super, super powerful stuff. So how do you build a simulated cell?

I mean, where do these data come from? Are these just a public resource or are these substantial amount of proprietary in-house data that are going into building the models? 

Daniel Veres: Yeah, it's a combination of three origins. The first is public data. So we are using publicly available signaling databases, protein interaction databases, omic data sets, also drug response data sets, generative perturbation data sets.

So it's a very integrative framework in which we can put in many of the commercially usable public data sets. But of course these are going through very rigorous quality control and, and filtering to really work with the data, which is as much possible, reliable, but [00:13:00] also because these data sets are also like inherently biased because of the hard super nature or because they are around a certain let's say assay type.

We are building more and more proprietary data around the system. So we have started roughly I think three years ago to extensively run validation and. Audation capital data generation around our work. So we are generating prep post treatment, RNA C transcriptomic data. We are generating our own genomics data liabilities.

We are using different assay types to capture best the bio mechanisms we are exploring. So the, the internal turbine property data is something which is growing extensively and we are opening the turbine lab, hopefully. Q3 this year. And we, that will enable us to scale this further and run specific data generation to feed that into the model.

And the third aspect is prietary data, which is coming from partners. So we have partnerships, for example, with Cancer Research [00:14:00] Horizons on the seven program. And we receive data from Cancer Research Horizon on what they have done already. With that asset. So they have drug response data, they have in vivo data, et cetera.

And we use that to feed that into the simulations and have the best possible training and, and output of behavior. So it's always a combination of public. Proprietary turbine and proprietary partner data, but obviously the partner data is something which is after to the legal constraints. So in many cases, we can't use that data in other projects, but still for that certain partnership, we can deliver the most using the proprietary data.

What 

Kevin Folta: about the exact data themselves? You mentioned signaling data. You mentioned interaction data. So are you looking mostly at protein, protein interactions? Are these all modifications say ubiquitination status of a given regulator? Because we know in cancers that's such a big deal, so what are the major data that are included in these model?

Or is it all of the above?[00:15:00] 

Daniel Veres: It's a bit all of the above. So the, the, but the focus is indeed signaling. And we, the original idea behind building the simul itself was that we are, we would like to predicts decision and that's why we have been starting in cancer because it's a very strong phenomenon of cancer cells that they would like to leave and proliferate.

And of course that's something that dependence your proliferation, which is basically an output, like viability output you can simulate. So it's a bit easier than, for example, going for neurodegenerative diseases or other disease types where you don't have a clear phenotype. Here you have a phenotype, which is a perforating, immortal cancer cell, and you would like to kill it.

That's the, that's the starting point. So for this, we are using. Data to characterize the models, which is genomic, trans omi, proto data and also build the signaling network further and further. And the signaling contains post sensational [00:16:00] modifications like cubic ation, but it also contains genetic, more like transcription, regulation aspects as well.

And we are constantly expanding the mahas in which the simulated cell is working well. And for example, to your point around visibility. Of course it can do everything. We have limitations in certain areas of biology by and other areas. We are very well suited. So for example, we have been focusing many of our work around DNA DME repair.

So our model is very well defined there. We have a lot of property data in that field, et cetera. But for example, we are just moving into immuno-oncology and modeling immuno cells as well, because that's another level of modeling. Yeah, that 

Kevin Folta: was really my big question. You know, as a biologist who studies genomics, we still say that we don't know what 30% of e Coli genes do.

So, so really you're working from what we know, and especially in a cancer cell environment, we know certain signaling pathways are required in order to, or are, are actually [00:17:00] useful for a cell to be able to continue its transformation and, and proliferation. So, Let me just go back to how this is working as a business model.

Is this something that you outsource the software to different companies that maybe they license from you, or is this something that you work as a partnership where Turbine is a long term partner, eventually in the trials and all the way through development and that, that way is the way in which turbine is involved in the process.

Daniel Veres: Yeah. It's more the second. So basically we have been starting out, so turbines eight years old now, and we started with validating the platform. And in the first couple of years we have been working with big pharma companies. For example, Bayer in a, in a more like fee for service basis. So they were.

Contracting us for certain questions on biomarkers, combination, et strategy, et cetera. And we were running the platform, developing the platform for that given purpose and providing them outputs which at that [00:18:00] stage was predicted in silico data packages. And then we realized that there are several drawbacks of this both scientifically and from commercial perspective.

So we switch basically to build our own pipeline. To capture most of the value both scientifically and commercially. The pipeline building is still something we are doing and that's one of the main areas of the business. So that's one part of my answer, that we are building our own proprietary pipeline.

And that's something which is of course, a huge value. And we can demonstrate the platform benefits throughout the pipeline in a more controlled way. And the other, which we started to do in the past year basically is to engage in more strategic partnerships where we are sharing the risk, but also the downstream economics of these partnerships.

And those are the partnerships which you were describing. So there we would like to be there early in the process, ideally from target identification, but, but it's okay to join the earlier stages of. Let's say did optimization or hit finding[00:19:00] and then be able to support a certain program throughout this life cycle because we think that that would deliver the most value to the partners, but also turbines the most of these programs.

If we can be there, and we have demonstrated in all of these collaborations in the past that we are able to support programs throughout the process from Target ID to actually clinical trial stage assets. And in the new collaborations we are launching or launched we are exactly doing that.

And that's the, that's the goal basically, to have this as a more integrative partnership with the partners. 

Kevin Folta: Well, we're speaking with Dr. Daniel Vera. He's the co-founder and C s O of Turbine, and we're speaking about how simulated cells are using large data sets with machine learning and AI to make predictions about drug targets.

This is a Talking Biotech podcast by Collibra, and we'll be back in just a moment. And now we're back on the Talking Biotech Podcast by Collibra, and we're speaking with Dr. [00:20:00] Daniel Veres. He's the Chief Science Officer and co-founder of Turbine. And we're talking about how simulated cells created from huge data sets are informing drug discovery, especially in the area of oncology.

So let's talk about process. Do you start with a, a bunch of AI predictions and then move those predictions to the wet lab and actually perform those in vitro and eventually in vivo experiments? 

Daniel Veres: From a high level, that's totally correct. Yes. And but if you zoom into the details, it's a bit more general than that.

So ideally you would run a closed loop of predicting with ai, putting in the data sets. Productivity, AI measuring what comes out of it, learning from it, and then around the whole process again. We are not yet there and I think the whole field is not yet there. And that's mostly because productivity in biology is, is much lower than it would be, for example, in computational chemistry.

Finding keys for certain targets. So [00:21:00] what we are doing is we are collecting the data sets. Training does simulate, it says, and then running the predictions. But then those predictions are basically filtered on one hand through algorithm processes, also involving machine learning, but also like traditional bioinformatics which gives us a short list of ideas.

And the short list is something which you can already bring to validation. But in many cases we are not validating all of these, but doing basically translational research to understand the relevance of these, and that is bringing in many aspects which are either not captured by the simulations or ai.

Or you need to build confidence in the meha. So we are going back to the simulations, understanding what happens on the bih level, but we are also looking into independent validation in bio data sets. We are looking into commercial availability. By validation feasibility. So there are many as drug ability if you are talking about the drug target.

So there are many aspects we [00:22:00] are building on top of the simulation predictions to make a very comprehensive case that yes, this is a target or this is a combination and a biomarker, et cetera, which versus investment, because validation, if you want, would like to run it right. It's investment it's much more expensive than running a simulation.

So we are, we are building a process where our aim is not to do highest throughput validation. Our aim is to do the highest throughput part, INO in the computer, and then do the right experiments in a more low throughput fashion. And basically the funnel, which, what which I was describing is resulting in those hypotheses which we are bringing into violation.

Then we run the validation which for example, just to mention what the, what right experiment means in DNA damage repair. If you run a certain drug perturbation or a genetic knockout in an s preferential, s a four. Three days, you'll not really see an effect. Even in five days or seven, you are potentially not capturing what is happening because you just need time[00:23:00] for accumulating damage to see the phenotypic change in relation with that damage.

So you actually need potentially a week or two weeks or three weeks to get to that effect. And that's why, for example, par. Cell sensitivity data is not even close to who are the patients who will actually respond in the clinic. So when we validate the DNA damage or repair hypothesis, we are violating that in the right assay with the right timeframe.

And with that we are capturing the phenomenon of it a much higher chance. So the validation is actually meaningful. And if it's it was vibrated, then it's a good. Point, and then you can take it further to pipeline building, take to the partner, et cetera. If it's not validated, then it's still valuable because we can learn from it.

And there comes the, the active learning learning loop part where we are basically feeding back all the data which we have generated to have a better prediction next time. So if I was saying that, okay, this is a good biomarker of PARP inhibition and I measure it and it didn't, didn't pan out. [00:24:00] I can put that back to the system and next time it should not predict that certain biomarker or if it pres it, it should be in another context, which that we haven't ed yet.

And if you're doing many of these cycles then you will have more and more high quality loss throughput data around MAs. That's very complimentary to the highest throughput data, which is going in in the beginning. And of course, one aim of us is to basically make this loop as automated as possible.

So with that, you can eliminate many of the steps which are currently requiring, for example, translational scientists, manual inference. And also you are able to reach a productivity rate where you can just put that into a low throughput experiment and, you know, don't need to decide what to measure and what not to measure.

And with the lab, which I mentioned earlier, our, our aim is to basically have this as a much more efficient process causing this loop and running it again and again, and, and improving on productivity. Well, 

Kevin Folta: your idea of a, of a, of a, [00:25:00] the idea of a simulated cell isn't being born in a completely naive environment here.

That we know quite a bit about different compounds that inhibit cancer. You know, you mentioned PARP inhibitors, those kinds of things. We know cell cycle regulators, we know quite a few different compounds that have intergress into into cancer. And so has this simulated cell actually predicted things that we already know.

Daniel Veres: Yes, and that's basically one of our, so on one hand we are training on that data. So once we are setting up the same models and we are exploring, let's say, PARP inventory resistance, we would like to see in the first step that the model is able to recapitulate PARP inventory response. So for example, BRCA mutant or r d patients should respond to the treatment with a much higher.

Chance than the ones which are known r d or non BRCA mutant. So we are basically training the system to recapitulate known biology and then you can start asking novel questions. The second [00:26:00] is that we are looking at these known phenomena as positive controls. So for example if I'm predicting that BRCA is a good biomarker for response, That's a positive control.

Even if I was not training on it, I would like to see that if I introduce a BRCA mutation to the system, it, that model will be more sensitive to PARP inhibition or the other way around. If I reactivate that BRCA gene in the system, then it should be less sensitive. So basically these are the. The results, which we use is positive control.

And many times that's very important for us internally, but also for partners to see that, okay, we have a reliable model, we are recapitulating, non-biologic, and then you have a higher trust in the novel. Questions than the novel answers, and I think you had a, a good question before around how much we can basic, well, how much do we know about biology and how much is.

Unknown and how does that limit our technology? And I think the main [00:27:00] implication of that is how you can discriminate artifacts from novelty. So many times. It's a very, very narrow. Border between something being just an artifact of AI or simulations or being the most exciting research result in the past decade.

And you need to have that judgment built into the system that you can decide what is artifact and what is novel. And that's something, for example, we are constantly refining how to make those decisions. 

Kevin Folta: Yeah, I feel so much better about this now. I, I entered into the conversation feeling a little bit skeptical, like, how could you possibly be able to do this?

But I understand that there's so many things. We do know that if you're seeing those coming out of your model, that it really does validate that it works. And then the novelties can be discerned from artifacts to really guide you to good targets. I think that's pretty cool. So let's talk about some of the potential compounds or strategies that have been identified in the pipeline.

Most of them seem to be associated with D n A repair. [00:28:00] And so is that because that is a very well-described inter acto, maybe with some already well-defined players? 

Daniel Veres: Partially, yes. But it's also basically an evolution result of evolution of the company. So when we started out working with Bayer and others Basically opportunistically, we have been working on many of the DTR inhibitors.

So that meant we had experience with PARP inhibitors, V one, et cetera. So we, we had a very good understanding already of the biology, and we also basically developed a network of advisors who are really topnotch in the field, and we understood that in ddr. Despite the fact that it's a very surged area of biology and we know a lot, there is no integrative model of D DDR R.

So we fed that. Our approach of having an integrated decision framework basically built into the simulated cell around how to react to what kind of DNA damage, it's an advantage [00:29:00] over other methodologies. And in this case, it's really a complex, intertwined model of. Thousands of different DDR pathways, which are resulting in the decision on the SA cycle checkpoint level, if that certain cell should further proliferate or not.

So it's a very complex set of mechanisms. And for that, our modeling is an ideal tool. So there is that aspect. There is the accumulated knowledge around ddr and also we felt that commercially this area, especially with immuno-oncology as a combination opportunity is really evolving. And basically all of these aspects, Resulted in us focusing on idr and I almost forgot the unmet need of patients, which was one of the major components, for example, behind going into PARP inhibitor resistance as one of our, our focus areas because PARP inhibitors are wonderful drugs and they are happy millions of patients.

But still both the response rates and medical response time is, Something which we can improve. And we think that [00:30:00] understanding better the biology behind the resistance and finding novel modulators of that resistance, which are potential novel targets and eventually drugs, that's one way to do it.

So for example, in that field, we have identified Two novel targets. One of them is in vivo, ated the other is in vitro and vr. We have started drug discovery on those targets, but these are fairly novel targets from the DDR space. Well, one particular 

Kevin Folta: target looks to be PARP inhibitors and so could you explain what PARP is and its role in BRCA mediated cancers?

Daniel Veres: Yeah, so PARP is a very important enzyme in DNA damage repair, especially related to double and break repair and homo recombinant homologous recombination. And the biomarker and the synthetic lethal marker of it is BRCA what you, what you mentioned already, and this is a very fine example of synthetic fatality where you have a certain.

Impairment if a ge, a gene, for example, BRCA one or BRCA two, and that [00:31:00] certain cell with that impairment will respond very well to a PARP inhibitor. And basically the idea here is that you, with that you are able to. Force the SAS into accumulating DNA damage and they are not able to repair it because they already have an impaired repair mechanism.

And with that, they are driving into SAS psycho arrest and eventually SAS through this accumulation of DNA damage. The, the, the main aspect of, of PARPs beside the synthetic fatality and their advantage in many cancers where they have been approved like pancreatic cancer or prostate cancer, is also that there are certain resistance mechanisms which can develop against them.

And those are usually around reactivating some of the DNA damage it repair s So basically that's an opportunity. To interrogate further the DDR pass phase, but also you can combine them outside of the DDR space and it's really evolving. So there are many inhibitors already in the clinic and the PARP one cell active [00:32:00] inhibitors are really promising, which are now entering already enter to clinical phase, and they have a much better therapeutic window.

So there is a much bigger space for potential combination therapies. 

Kevin Folta: We've also identified so many players in cell cycle and their interaction partners, and you've mentioned kinases, which are a huge part of that cell cycle. How, no. How have novel compounds been predicted that affect cell cycle? 

Daniel Veres: So basically in, in our model, for example, cell cycle is an inherent part of all the simulations.

In the end of the day, you need to somehow interfere with cell cycle. It can be either that you are putting the cell into an unresolved checkpoint, and with that you cause arrest and that will lead to status or you are. Directly activating, for example, apoptosis or other sada s and stopping cell cycle activating apoptosis leads to SDAs.

So this is something which is an inherent part of predicting viability of, of cancer in the simulations. And there are many drugs [00:33:00] which are interfering either indirectly or directly with that process. Of course, indirectly. Basically every cancer drug will. Somehow cause Sasac related effect more directly.

We can talk about, for example, the CDK 46 inhibitors, which are recently showing very good data in the clinic. Not so recently. And there are many other Drugs which are interfering with cell cycle. Many opportunities to combine these cycle related drugs with, for example, DDR related drugs.

So, for example, the program which we have with CDC seven with crh ca cetera, horizon, that's also on the edge of. Replication, SA cycle ddr. And that's something where we see opportunity for combinations, but also if you think about V1 or some other aspects these are very, very close related mechanisms.

And basically all the SAS are reliant on proliferation. So in the end, you are interfering with SaaS cycle. 

Kevin Folta: Yeah, it really does seem like, like a lot of the d n a [00:34:00] damage approaches would be good adjuncts to traditional therapies just to kind of add an additional level of control to slow the evolution and changes in proliferating cells.

So it's really cool. It, it seems like an appropriate approach. I guess the last question I would have is, you know, you mentioned before that you focus on cancer because we understand. The aspects of proliferation in that this is where you're likely to find some potential commonalities between different mechanisms.

But if you were to think of the next type of disorder going out, what might that be in terms of the next type of simulated cells that may prove valuable in identifying new therapeutic approaches? 

Daniel Veres: Yeah. So we have started out in cancer because it's better understood and there is data. And I think that's one of the main aspects of where we would like to move further is where the data brings us.

And in the past couple of years, there is not just an evolution, but more like a revolution around immunology, especially immuno-oncology. [00:35:00] The data generated in that field is immense and we are now. Able to tap into that data and start simulating the interaction between cancer cells and immune cells.

And we see this as an opportunity to move out from cancer as well. So once we could prove that interactions between tumor cells and cancer and immune cells is possible and we can simulate some aspects of immunology then we have the possibility and the path forward. To go into immunology, and I think that's a very good place for a model like ours because it's a very complex network of different Cell types of immune cells.

Also very, it's, it's basically it's a network similar to protein infection network, but it's a network of sas. So it's really showing the next level of hierarchy. But you can still apply the same principles, what we are using on the cell cell level. So if you ask me with that combination of data and with the basically similarities in, [00:36:00] in terms of basic principles, I would move with the technology.

Through immuno-oncology to immunology, and with that you can tackle diseases around autoimmune, but also potential infections, et cetera. So there are many potential use cases there. 

Kevin Folta: Well, this is really exciting and a place where I would encourage my students to think about a potential career. So if you were hiring a turbine and you were going to advise.

A next generation of scientists to start to prepare for a job at Turbine, what kind of classes would you want them to take? Would you want them to be computational people or molecular biology people or combination of both. And what level of training do you think is necessary to be really impactful in this field?

Daniel Veres: Yeah, I think the combination of those is the idea. So if it's, it's not. I would not say it's not about just biology or, or data science anymore because you can be an expert in those fields and that can be [00:37:00] also very useful. But especially in our space. Biologists should talk with AI scientists and they need to each other understand each other and they have a very different language.

So basically the, the barrier which we are constantly working on to, to not have in turbine is this language barrier between AI and biology. And if you just have ai, you'll be not able to speed the language of drug discovery, because that's mostly around. Biology and chemistry. If you are only speaking biology, you are not able to talk with AI scientists and you are not able to give them advice on how to go further.

Just think about Chad g pt. If you are not giving the right prompts it are not. Give you the right answers. And in simple terms, badges are giving prompts to the AI scientists. I would like to see this or that, and then they can help making that happen. But if they don't understand each other, they are not able to work together.

So I think My advice would be that for a biologist, try to [00:38:00] explore the space of data science, bioinformatics, talk with those people, try to be familiar with the language and also vice versa. AI scientists, data scientists should look into biology and understand it better. So once they are in the same room, in the same team, they're able to communicate with each other and they are able to deliver research, which we would not expect if they would be not able to work together and understand each other.

Kevin Folta: So Dr. Daniel Veris, I was really excited about this topic. I learned a lot from it. Mostly just expanding my my beliefs. I mean, I, I, I really always hope that something like this would be possible. And it's great to see the technology is actually getting there. So, if listeners really wanna learn more about Turbine, where would they look on either websites or social 

Daniel Veres: media?

Yeah, I think these are the good first steps. So I would suggest on one hand a website where you can learn more about the technology, but you can also look at the news and what is happening around turbine. You can also contact us if you have questions or you would like to partner, [00:39:00] et cetera. And of course on social media, especially on LinkedIn.

We are sharing many of the, not just job opportunities, but also updates around where is the team, where you can meet us, where are potential interfaces to have some discussions. And we are also on Facebook, which is more about, let's say, team building and related cultural activities. So I see these are the three main forums where you can most easily find us.

Kevin Folta: Well, very good. Thank you very much for joining me today. I really appreciate it and best wishes to you going forward. And please, please, please reach out next time you have a big discovery that you'd like to talk about with a, with this particular audience because I think this is super cool. So thank you very much for joining me today.

Thank 

Daniel Veres: you very much for the opportunity and very nice conversation. 

Kevin Folta: And to the usual listeners, thank you very much for listening to The Talking Biotech podcast by Collibra. This is just another great example of how we're using computational tools and big data to be able to hone the starting points for drug discovery [00:40:00] and identify new candidates that could potentially work to solve some of our most pressing medical problems.

So get excited about this. I think this is really cool stuff. Thank you for listening to The Talking Biotech podcast, and we'll talk to you again next week.