Simulations Plus, Inc. (SLP) Earnings Call Transcript & Summary

Welcome. We're allowing everyone in a few more seconds to get connected. But while we do that, I wanted to say we're excited to kick off a follow-up to last year's webinar introducing the ADMET Predictor HTPK Simulation Module and how it can be used to effectively provide early PK assessment in discovery programs. Today, we'll talk about workflows and strategies, and the speakers joining me are Andrés Olivares-Morales from the Roche Innovation Center, all the way from Basel, Switzerland; and Eric Jamois, Director for Key Accounts and Strategic Alliances for Simulations Plus. Before I turn things over to today's presenters, a few housekeeping notes. We take your privacy rights seriously. And by attending this event or participating in the Q&A session, you are allowing us to contact you for follow-up. Immediately following today's presentation, there will be an opportunity to ask questions. [Operator Instructions] Now again, as we all get settled in and we're approaching about 200 wonderful peers joining us today, I'm going to ask a quick get to know you question. Which sector best describes you? We'll give you just a few seconds to answer, and then I'm going to turn it right over to Eric Jamois. [Voting]

Just a few more seconds left. It's wonderful to see all of our colleagues joining. It looks like we're going to have a full house today. So this is fantastic. All right. We're going to go ahead and close the poll, and it's wonderful to see a large part of our commercial and industry family joining us today. So without further ado, allow me to introduce you once again, Eric Jamois. Eric take it away.

All right. Thank you, Arlene. So good morning, good afternoon, and good evening, everyone. Thank you for joining us for this very special webinar featuring the deployment of the HTPK technology at Roche. As you can imagine, the ultimate reward for us is to see our technology deployed and successfully used at client sites. We've been privileged to work with Roche towards furthering the HTPK technology. And Andrés will be sharing with us the details of this effort. We have hosted prior webinars detailing the technological background of HTPK and notably the mechanistic modeling foundation. You can still find these resources for replay on our website. And actually, I think at the end of this webinar, you will also get a link to these replays. So you can capture the full background behind what HTPK actually does. We won't be focusing so much on the background here, but much more on the application and deployment of the HTPK technology at Roche. So without any further ado, it's my pleasure to introduce Dr. Andrés Olivares-Morales, Principal Scientist and Project Leader at DMPK/PD Modeling and Simulation and Clinical Pharmacology at Roche. Andrés has extensive experience in PBPK modeling, and he and his team have been working closely with us to bring the technology to the Roche pipeline. Andrés, the floor is yours.

Thank you, Eric. I hope you can hear me. Thanks for this very nice and kind introduction, and thanks to all the people that actually joined the webinar today. I hope you can also see my screen. As Eric mentioned, what we would like to share today with you is the use of this technology in our pipeline and how we are actually implementing it as part of our workflows in the early discovery space. So I will cover that and also a bit of introduction and why we're motivated to include this in our day-to-day work. Before actually I move on, I'd like to say that this is -- I have the privilege to be here and present on behalf of the team. But this has been a large team effort for several colleagues -- of several colleagues from different functions, from DMPK Modeling and Simulation, from pRED Informatics to enable this technology and also for colleagues that work in Small Molecule Research on these chemistry and computational chemistry. So before I go, I'd like to briefly introduce PBPK, although in this audience, I probably -- I don't need to do that. But as you know, we had PBPK models. They originated actually from a theoretical framework in the 1930s by the work of Teorell. But in the recent years, in the last 20 years, we see more and more use of PBPK both in industry and academia. And PBPK models obviously provide a significant advantage to us because we can separate what is drug-specific from system-specific. And this allow us to play around with a lot of scenarios where we investigate what are the potential effects of compound properties in the pharmacokinetics of a molecule. And recently has been very well established and it's gained also regulatory acceptance. And we have commercial softwares available that allow us, obviously, to use this technique in mainstream, in the industry and also in academia or other settings. In our pipeline, actually, we use PBPK from -- in different stages, in different flavors. We can go from very early lead identification to help us design, experiment or to rank compounds or to understand how the ADME and PhysChem properties integrate to each other. And as I mentioned in the beginning, also the scenario assessments, which is really important to know, where do we need to optimize and what do we need to find in the early space. But as we transition towards the clinical candidate selection and to the phase -- the clinical phases, PBPK takes a more prominent role into defining the entry into human doses and the projections for a new molecules. Also can help us to define tox studies or also to design tox studies used in animal models. But also, later -- in late development or early late development, we have a strong focus on the prediction of drug-drug interactions or special populations or biopharmaceutics, which is -- it's a growing area in the PBPK field. We have this also post-marketing and virtual bioequivalence. And as you know, there are different -- many implications. And PBPK now inform drugs labels. So here are 3 examples from us of COTELLIC, ALECENSA and ROZLYTREK, the 3 compounds -- the 3 molecules that have -- they have information that is based on PBPK modeling in our label. And this is a relatively new paper where -- this is from the FDA in terms of the uses of PBPK in regulatory submissions. And we see, again, there is an increase in number of applications with PBPK analysis over time, so very -- growing exponentially almost from -- in the last 3 years and probably there is a more updated figure for this. But as you see, the main applicability is in the drug-drug interaction field and followed by absorption and also special populations such as pediatrics. PBPK is something that is not new for us here at Roche. And Neil Parrott and other colleagues, Thierry Lave and Hannah Jones, very early on saw the potential of PBPK for the drug discovery space. And their ambition strategy, which we use fairly regularly in our pipeline, which is basically verify our PBPK model in preclinical species, understand our in vitro assays. And once we gain confidence then use that knowledge and translate it for our human dose predictions. And this is, as I mentioned, where mostly our PBPK work is, so in the translational space, in the early clinical development from Phase 0 to Phase I. And this strategy has paid off and PBPK is systematically applied since 2003 in all our small molecule projects. And this is a figure from Neil and other colleagues where they checked on the predictive -- predictivity of PBPK on entry into humans. And in general, for around 33 projects, we have a twofold error in the prediction. And so 70% of our predictions are within twofold and an average error of twofold, which is fairly good and is consistent with this strategy. Now if we look at the pipeline and the different stages of drug development and discovery, we see that PBPK, at least for us -- and I don't know how it's for other companies. It's fairly limited to this stage. So the transition from nonclinical to clinical. And the uses in the early space are fairly limited. And the reasons are manifold, why we're not using it in the early space, but most of them have to do with obviously availability of different software and the fact that in the early space there are several thousands or hundreds of compounds and, therefore, prioritization of a dedicated strategy for modeling a single compound or a few of them -- it's a bit tricky in the early space. And therefore, the -- in the early space from the lead optimization and lead identification, there is -- PBPK plays a role, but it's a minor role compared to other approaches, which are mechanistically based, but are simple to implement such as equations, such as the early dose to man or Efficacy Index or LIpMET and LiPE, lipophilic efficiency. And we wonder why we don't have -- we don't use PBPK in the early space, and we identify a few barriers. And one of them has to do with, as I mentioned, the number of compounds and the limited time that the project leaders or the modeling and simulation colleagues or the MedChem teams -- they have in order to perform the simulations and also the need for many different data sources and software. So we have different commercial platforms and in-house softwares and lengthy and complex data transfer processes from -- data from in vivo studies or data from in vitro or from compound properties. So that's very -- those limitations, as I said, actually we don't use it that much. And also we don't have the use of people that are not non-experts in the field. So most of the PBPK simulations are conducted by our specialists, people that really have the knowledge and training to understand the models. And therefore, we embarked in a project with the colleagues that I mentioned at the beginning to bring PBPK early on. And for that, we're using the technology that Eric mentioned that is available in ADMET Predictor. So we have a project internally where we want to implement faster, simple and easier PBPK simulations in the small molecule teams and bring this to the early teams as soon as possible. And we hope that this will help us change the way in -- the way we perform drug discovery by bringing this expertise early in our project teams and even early to the design stage. And one of the ideas that we had is try to remove this burden of manual data transfers and also try to remove these reliance on simplistic equations that sometimes they don't incorporate all the ADME processes. For instance, absorption is generally overlooked because the questions are focused on potency and clearance. So we wanted to overcome that by bringing PBPK. And at the same time, we wanted to reduce animal experimentation. This is one of our goals to also as part of this approach is to reduce animal experimentation and to have a more meaningful drug design cycle by using human relevant simulations. But also, which is actually a trend nowadays in the industry, to make use of more of sparse data, i.e., machine learning, so make use of more machine learning properties and try to figure out which properties are well predicted and which are not and, therefore, focus on those. And as I mentioned before, the only way to implement this and to achieve these goals -- it's by collaborating. And we created this collaboration between these 3 different elements internally and also we work with -- closely with our colleagues at Simulations Plus for their guidance on how to implement this technology and also for the scientific assessment of certain parameters that came from the HTPK module. And for that -- that's the overall view of the project. And I just wanted to show you a case study of -- one of our internal case studies and a very simple and early proof of concept. So we have a small molecule program. And this is the way we expect, obviously, to look this in the future and actually what we're working towards, is that we're trying to -- we have a small molecule program where we know what is the dose target that we want to achieve. So we want to achieve at least a predictive human dose that is less than 200 mg and a half-life for this compound in a range of 12 to 48 hours. Now with the equations that I mentioned before such as the early dose to man, we can easily achieve this by combining our microsomal stability or hepatocyte data with our potency data and get an idea of what is the actual projected early human dose, and that can help guide the teams to rank compounds. But for the half-life it's a bit more complicated. And normally, there are equations where these SKUs are based, but we prefer to use the PBPK approach. And we use these high-throughput PBPK simulations to have insights in terms of design and to prioritize our compounds and to help us find the right candidate basically or candidates that could beat this criteria. And this is an example of how this could look like and actually is looking for a few projects like this, where we have visualizations that are already made, are the output of these high-throughput PBPK simulations where we can gain insights in terms of what we're looking for. So here, on the right-hand side, you can see projected human doses for different candidates. And also on the y-axis, you see the projections of the half life. And easily, we can find that actually, if we only focus on those, we might miss the compounds that actually are the right ones for meeting our criteria. While we have compounds with shorter half lives that we want or even longer half lives, while we have the same dose. So these level of insights were not available before we implemented the PBPK approach. And what is nicer is that also we can get a simulated PK profile, so we can get an understanding. So all these compounds, they have the same level of target engagement, though -- although they have different pharmacokinetics. So this has to do with the potency and the compound properties. So we can get more insights. We can get ideas of the Cmax or the AUC or the Ctrough and look further into other properties such as toxicity and so on. So all this can be achieved with this approach, which actually we find it really informative. And what's more we can look into? From a design space, we can look into properties and see whether -- which properties are relevant for the compound. In this case, we are focusing on half life, and we can see here that if we have the right combination of clearance and lipophilicity, we end up in the right space in terms of half life, depicted here by these green colors. And these are kind of insights that the team really appreciate and can help admission of chemistry to define better molecules. Now we can also look at the impact of other properties into basic fundamental pharmacokinetic properties, for instance, solubility and permeability, which normally you think you need to optimize for dose. But given the space and the PK/PD properties of this compound, the impact, for instance, in the absorption of these properties is relatively limited, even at the doses at reducing relatively low. So in this space, we are not focusing that much for this example, but, obviously, for other examples or other projects could be different. And this can provide the team guidance of where we need to focus and what is actually relevant for the projects and what's not. And all of this is generated and all these insights come from the systematic PBPK assessments. But these are obviously prospective simulations. We also want to look into retrospective simulations and see actually whether we can rely on this. And we have the luxury of having some in vivo data available for some of these compounds, and we can look into the IV predictions for these projects, for instance, in terms of volume of distribution clearance and half life, and we can see that most of the predictions are really twofold, which give us a good confidence that we can extrapolate our in vitro assays. But also for the oral pharmacokinetics, we see that actually there is a really nice correlation between observed and predicted, although there is an overestimation of certain parameters. But for ranking purposes, this is ideal because we want to prioritize the best possible compounds that we can rely on this ranking -- for this series and these compounds. And that's the whole idea of being able to implement PBPK early on in this stage -- in the discovery stages that we can focus more on predictivity and also move faster with predictive models. So we gain these insights and we have optimized. We actually -- we are prioritizing compounds based on these insights. And obviously, one key take-home message from this as well is that we have to have good and predictive assays for each project. So in this case, we knew that our in vitro assays were predictive because we checked the idea before, so hepatocytes, for instance. So we gain confidence in moving forward. But you may ask, well, is this actually an one-off and what if -- how are the other compounds or other projects and new projects will be predicted? And this is very fresh. This is a work from our PhD student, Doha Naga and [ Neo ] and myself, where we actually looked into certain compounds for which we have the in vivo data. And we have every single in vitro parameter that actually was necessary to predict the pharmacokinetics after oral administration. And so we're looking to 200 (sic) [ 250 ] structurally differentiated compounds portion. You can see here the different properties, the distribution of logD, of molecular weight, permeability, fraction unbound, solubility and clearance. And we looked into this by asking several questions. So first, can we predict the IV clearance of these compounds, again, by just taking the in vitro data and moving forward? Secondly, can we use this -- how is the predictions in terms of -- when using machine learning model? So for this, we're using the ADMET Predictor software to use -- to guide us in terms of intrinsic clearance and properties that are needed for the prediction of clearance. And how does this work in terms of oral predictions? So if you know the clearance or how is the absorption model and how reliable are these predictions. So these are the questions that we ask to these data sets. And here, you can see the results. And here from -- I have 6 finals here. I hope you can see that. So we have different weights of scale in clearance. So here, this clearance is scaled by using hepatocytes, and this is rat data. We use direct scaling dilution method or we assume that everything is unbound or we use machine learning model or the [ OSIM ] method in ADMET Predictor. And here, you can see that our predictions of clearance are within threefold for 63% to 76% of the simulations. For high clearance compounds, direct scaling works better. For low clearance compounds, the dilution method seems to be better. The machine learning model actually doesn't do that rat, although the correlation is not as great as when we have in vitro data. And if you assume that everything that you have is unbound, then you have this underprediction, which is reported in the literature. Here, on the left-hand side on the bottom left, this is for reference only, and this is a back calculated intrinsic clearance that was fed into the ADMET -- the gastro process in this case, actually, to be precise and to simulate, just to check that the software does what it's supposed to be doing, and which was the case. So what was surprising for us is actually the machine learnings predictions for this set that is actually completely new for ADMET Predictor in the sense of they actually hasn't seen our structures, is that the success rate was within 36% to 60% within two to threefold, which is not, obviously, the best but it's actually interesting in the sense that it allow us to get some idea of the ranking. Now if we look into the oral predictions like similar plots, but from the oral space, then, of course, the prediction success is a bit reduced, from in vitro data only. So between 40% to 60% of our predictions were within twofold for using the direct scaling and the dilution method a little bit less. The bias was around threefold or fourfold with the dilution method. But the correlation, actually, in terms of ranking -- and remember, these are naive prediction. So actually, there are no refinements or not learn and confirm cycles here. It's good for what we want in the early space. So we basically were able to rank our compound. So this suggest that this strategy actually to shifting into a more predictive approach in the early space is actually valuable. Now what we looked into in addition to that is actually -- is we wanted to understand what is the source of arrow in these predictions. So here on the bottom what you see are predictions that are made while knowing the clearance. So basically, we use that intrinsic clearance value that was determined from in vivo as a plug in. And then we use all the in vitro properties to predict the oral predictions. And this is where you see in the case of this rat you see. And we see that actually the predictions improved massively within twofold. So which tells that we really need to understand our clearance model and our clearance predictions. While it seems that the absorption model predictions are relatively good, especially for these early stages that we're looking at ranking. And here on the bottom, you see our predictions using the machine learning models. So what we learn here and what we want to say is that, although you might not have the best prediction, the directionality and the ranking is on the right direction when you use in vitro data. And machine learning actually gives you an idea of where we should -- it also give you relatively decent predictions. So you can see a scenario where you can work with sparse data and still get a meaningful information out of the PBPK modeling. So therefore, we were happy with this. And I mentioned -- I think I should have mentioned that all these simulations were done in GastroPlus, the previous ones. And then we wanted to compare whether the GastroPlus and actually the simulation technology that is in ADMET Predictor and the HTPBPK can also maintain that performance. Actually, we did an assessment. And for those who don't remember, so a traditional PBPK model is a larger distribution PBPK model. Plus the absorption model that is fairly complex. The high throughput model is a simplified distribution model and the complex absorption model. But for what we want -- again, it's good for what we -- the purpose that we're looking into the early space. And we looked into this with -- actually with Eric and other colleagues on Simulations Plus. And actually, we're picking press of the consistency, at least using our data and internal data between the 2 software. So we compare ADMET Predictor and GastroPlus when using in vitro inputs here for Cmax and AUC in this rat data set. And then we use -- we also repeat the same exercise by using the machine learning input parameters that come from ADMET Predictor and GastroPlus. And we got a really good consistency. And why I'm showing this? Because it's important to mention that the difference in speed of the simulations is dramatical. So with HTPBPK, we can simulate thousands of compounds or hundreds of compounds in minutes or seconds, whereas in GastroPlus or other platforms, it takes a bit longer than that for one compound, basically. So for the discovery space, that's what we need. We need speed, and we need obviously reliability as well. So that's why we're focusing on bringing this technology to our pipeline. And what can we get out of a high throughput PBPK simulations? Well, we can simulate a rat and a human PK. We can simulate an oral dose or an IV bolus, and -- which is something that was added recently to the software, which is something that really happy to see. But we get all the standard PBPK parameters, AUCs, Cmax, half-life, fraction episode, fraction bioavailable, which is really what we need and volume of distribution, which I also mentioned is here calculated using mechanistic equations, which is something that the design teams really want to know. But in addition, there is actually a feature that it give us a guidance of what could be the potential early doses, which can guide -- and take all these properties, both in the clearance, solubility, et cetera, and combine this by defining a concentration, which I want to achieve, and this can be done by compound or it can be done by the whole series. Depends on the data we have available, and we can look for a Cmin concentration, a Cmax or Caverage, which give us a lot of flexibility in terms of guiding our teams. And yes, in all this, we were very happy, and that's why we contacted the colleagues on Simulations Plus and we enter into these discussions on how we can implement this. So going back to my initial slide, so that -- where I mentioned about the strategy that we have, this learn and confirm cycle, which is something that -- it has proven to be successful, and therefore, we're not going to change it for -- the entering into the clinics. But in the early space, we didn't have much of the PBPK. So now taking a bit of the analogy that Simulations Plus does with their software, we want to focus on speed, ranking and prioritization at the early stages. And therefore, we will bring this technology or actually the technologies are right there in our early space and try to reduce this learn and confirm cycles as much as possible, but try to make sure that we develop, we design the best compounds early. And then when we go into the clinical candidate selection, we will transition between these 2 approaches into a more dedicated and tailor-made PBPK approach. Here, for instance, in Basel, our colleagues -- they have a really nice Tesla car. So we can really customize our software and customize our modeling approaches when we move in the late stages. But in the early space, we don't really need that level of customization. We want to make sure that we bring the insights and we help the design of better comps. So how do we implement this? And one of the things that I mentioned was this data transfer and data transfer processes. So the way we actually work on this is actually we look into our databases, and we created an app. This app actually brings all the information necessary to run the simulations, retrieve the company information, seamlessly -- in a seamless manner translate, sends this data into the software, in this case, ADMET Predictor, and sends the results back directly to the project teams via web portal, for instance. For this we use Spotfire to send the visualizations, and we have created all the IT infrastructure to do that. In addition, we have access to our PK databases, so we can actually validate models, as you saw for the previous -- for the example that I gave. And for the early compounds that have not been synthesized, we use this in -- by just using the machine learning models and combine with a service. So this makes the integration of this PBPK and the complicated steps that to develop a PBPK model -- they are basically gone from the user -- the end user. And now we have this in -- well, we have this for most of our projects very soon by -- in a few weeks and months. So here's how it looks like. So we have an app that just, as I mentioned, where we can retrieve the data and actually generate the input data and send it to the software. And the user just log into a landing. They select the project that they want to see and then they click a few times and they get the data up and then simulations later. So you see that we can transfer this to our -- from our servers to ADMET Predictor. And then in the end, we get this kind of visualization similar to the one I show you before, where we get the compound space, the we develop the case simulations, all these things are done in a seamless manner, fully automated, just a few clicks from the users. So that means that all these different data source connections are actually now done automatically for all the users, and this is the way we are implementing this in this project. So with that, I think that's my last slide. So we are able to implement these high throughput PBPK simulations in our small molecule project teams using our in-house data inputs or machine learning models or any other models. It's a seamless process. We have created internal workflows that allow us to do that. And the users -- the end users, actually, they can set up this very easily with minimum integration, which bring us to the point that the data consumers, basically, which could be the project team, so medicinal chemists, DMPK colleagues or modeling and simulation colleagues, they can even run this themselves. So we don't really need knowing actually the software in too much details because this is very narrowed to the specific task, so running simulations and creating ranking based on PBPK insights. And also looking into other space like in the design space and looking into the early properties with these visualizations. And obviously, the most Important thing is that it's all integrated with our workflows in terms of PK, so that we can validate the models retrospectively and not just for a single compound, but for a whole series of compounds. which is the most attractive approach. And that's why we're very excited about this and very excited that we're finally there. And that by working with our colleagues, again, in Simulations Plus, we are where we're supposed to be and looking forward to make sure that this is continued into our project pipeline. So with that, again, I want to acknowledge all the team members that have been really collaborating from a scientific point, from a technical point, making sure that we can -- we are at this stage of our project where we can share something with you. And we look forward for the implementation. And I'm happy now to take questions, and this is the end of my presentation.

Thank you, Andrés. Great presentation. Two quick questions before we launch the Q&A session. One, are you currently using the HTPK Module within ADMET Predictor? We'll give you just a few seconds to go ahead and answer. If you haven't done so already, again, you may send your written questions using the questions pane on your control panel. [Voting]

And the last question, does your organization plan to deploy a Discovery PBPK program within the next 3 to 6 months? Again, we'll give you just a few seconds here. The questions are coming in. So I'm going to be bringing Eric back into the conversation. And Eric, go ahead and take it away.

All right. Well, first, I would like to thank Andrés again for a really nice presentation. Thank you so much. Really interesting -- very interesting stuff. There are a few questions that have come over the wire. And 1 that I think is probably actually on a lot of different minds. And it has to do with the use -- I mean, when you are predicting -- when you're doing dose predictions, and so you're using clearance and volume of distribution information in order to input -- I mean, to provide clearance and volume of distribution information or you're using Roche proprietary models? Or are you using the models that are prebuilt in the software? And I think it's a really good question because -- I mean, actually, the software allows you to do both. So I think you answered this maybe partially on Slide 23, but it's a very interesting question.

Yes, maybe I can activate my video as well then. So yes. So we use both. And for -- when we have the in vitro data, we use the in vitro data to extrapolate and to -- provided that we have a good in vitro to in vivo extrapolation. So we check on this. This is always a must in our project. So we check on the microsomes. We check on the hepatocytes. We try to understand their in vitro systems. But when it comes to volume of distribution, we generally rely again on mechanistic equations, so PBPK modeling. I mentioned this Roche's Lukacova equation. And this is the -- this is, I would say, is our preferred approach. So we focus a lot on our mechanistic understanding. Now when it comes -- when we don't have the data, if we have a machine learning model that can predict us in vitro properties, we might use that one. If we don't have that one, we will use the 1 that is built in ADMET Predictor. And I think I mentioned in our discussions, for instance, pKa is 1 of those things that we generally go with the ADMET Predictor because, obviously, it's predictive. And again, it depends also on the sensitivity, right? We need to check -- for PBPK models, it's really important to understand what parameters are sensitive. And sometimes the dose might not be sensitive to certain input parameters. And for those we can have limited data. I hope that answered the question.

Yes, yes. And there was actually a related question on how does the automated system decide when to use the in vitro data versus the in silico data in the HTPK simulation?

That's a great question, and normally it's the users and the project teams. So the project teams define the rules. They can define what is the primary, what is the secondary and tertiary and so on and so forth. So the automated system -- what it does is it doesn't run the simulations for you. There will be an user that will click next, next, next in an app, but they will be prompted to make decisions. If the user doesn't want to make any decision, so there is a naive project, something that just started and there is not much knowledge, then we will use a default approach. We will be always in vitro, followed by in silico. So that's kind of the rules. But if the users, they know that they might have an in silico model that will be better than the in vitro for whatever reason, then they could focus on that. And again, it's customizable.

Yes. I'm reading another 1 here. Have you explored how the predictions work for different classes of compounds or different ECCS classifications?

Yes. Actually, the work that they had -- it -- we -- I didn't actually -- if I go back to those slides, we actually look into the ECCS. And I have -- I apologize that the labels here are different ECCS classes. So obviously, metabolism driven clearance. It's the majority of the compounds in this data set, as you can see here, from here on the Class II. So -- and -- but we do have compounds that are predicted at least to be renally eliminated or by other mechanisms. Unfortunately, the number is fairly limited to really make a conclusion on those, but the ones that are -- we focus -- we also look into this as well.

Okay. I had another question. How do you integrate the machine learning models with the PBPK model? And -- well, that's kind of part of the ADMET Predictor environment, but I will let you answer the question.

Yes. So as you mentioned, thanks for answering partially. So we do -- again, if we do have all the in vitro data, all the input parameters, we will use that as a primary. But if we don't have, for instance, the fraction unbound in plasma, then if we have a good machine learning model for that, then we can simulate our compounds or our doses, predictions using, again, a combination of both in vitro to in silico data. And in the early stages where there is nothing, then, of course, we will rely on the machine learning models. But again, this is fairly early, and we haven't really gone into that process given that actually the technology -- the HTPK technology is fairly early, right? So the latest release we got was last year, the version of the software we have today. And I think we are finalizing this. So we don't have that much experience in the early space, like in the really design space, but we're trying to get there, so to understand what is the best approach for that. But for the compounds that already exist, then a combination of, again, in silico, in vitro is the way to go, at least in our minds. When there is no in vitro data, other is sparse data.

Okay. And Arlene, are we still pretty good for a couple more questions?

Yes, we are doing very well. So lots of good comments and good questions in the panel there for you to choose from there.

So another 1 relates to how to find the impact of variations of the -- I mean, how the PhysChem parameters are going to impact the HTPK simulation? I mean I know that for some of them, like PEF, you can see if compounds or permeability limited or solubility limited, but maybe you have further insight about other PhysChem parameters influence on the whole HTPK simulations.

That's actually an excellent question. And again, it's something that we need to look on a project-specific level. It will depend. So if you have a project where all your compounds are extremely potent, okay, and low clearance and your doses predictors are relatively small, then permeability and solubility will have very little impact on the dose. They have obviously an impact on the pharmacokinetics, but on the dose it will not be limited by that. And you can do these assessments by project. So that's the only way to assess what is relevant, right? So you need to check on the overall on the combined approach. Now there is a secondary maybe that is related to that question, and it has to do with the uncertainty on the measurements or the uncertainty on the machine learning related properties. And this is an area of actually extensive research in the recent maybe 2 to 3 years. And the approach for that, in my opinion, is sensitivity assessments, so sensitivity analysis. At the moment, I think ADMET Predictor includes some sensitivity analysis maybe for 2 properties, right?

Yes, yes, yes.

And I think an expansion on that, it could be interesting to check on other properties. And I think that's the key because all the base models, right, the system of differential equations like PBPK models, they have no linear. So the propagation of error is not linear, right? And therefore, some properties might have an impact and some other properties might not. And our -- actually, we have a PhD student, [ Doha Naga ], that actually looked into this from a disposition point of view. And she published recently a paper when you see, depending on the parameter space you are, you might not have an impact at all on the volume of distribution because you are like in a complete and sensitive space. And I think my recommendation will be that in terms of PBPK you need to do sensitivity analysis. It's not -- there is no other way.

Yes, because you could try to see something wrong with a compound whereas that property has literally no effect whatsoever. So...

Exactly.

And you can also find situation where if that's a real problem, you know the compound probably cannot be safe just because there has no influence on it. So I mean in terms of rescue, I don't know if it has -- just it has any implications in terms of sort of compound rescue potential that -- if it makes -- I mean, if it doesn't affect that property, then potentially you wouldn't be able to be rescued, right?

Yes. And 1 of the things that -- actually, 1 of the limitations for global sensitivity analysis or the local sensitivity analysis is sometimes the speed of the simulations, right? That's 1 of the limitations. But having a software or a script or anything that can run thousands or hundreds of thousands of simulations in a relatively short amount of time will allow us to actually to check on different uncertainty scenarios for properties. So for instance, you can take a distribution of fraction unbound in plasma. And then stimulate all of them and see whether they have or not an impact in a certain endpoint, and that's something that we can do with softwares like this. Of course, you need to find a workaround, but it's something that can be done.

Yes. I agree with you. Actually, speed actually changes the way we think about scientific problems in many, many respects. That if something is going to take half a day or a few days to actually do, you're going to kind of think twice about investing that time to actually do it. But if it takes just a few minutes, I'm just going to run it and see what the results are. So really, that couple orders of magnitude in terms of speed actually does change, I think, quite a bit how we react to it as scientist than if it's fast enough, which is we have this nature to just try it. So I think that's really pretty critical. And maybe one last one. Can HTPBPK rank order predict BBB permeability or maybe incorporate that in the design in early discovery compounds?

I don't know. That's the answer...

Yes. We have a blood-brain barrier model within the software. So if you wanted to incorporate that as 1 of your parameters, you definitely could. I don't think it's 1 of the parameters that play in the HTPK simulation. But if it was sort of a filtering criteria, you could definitely do that.

Yes. Because at the moment, the distribution is just a central compartment basically. We don't have -- there are no organs at the moment. So okay. Thanks for that because I wasn't aware of the BBB model.

All right. Well, I think we're going to gather all the questions. We're not able to answer them all, but we think -- definitely thankful to you for answering a large number of them. So -- and also very thankful about some really great content that you were able to share with us. And if everything is good, I think we're gearing up to close the session. Arlene, is there anything else you'd like to say before we close?

Yes. Thank you, again, Andrés and Eric. We appreciate you sharing with us your approaches for deployment of the HTPK Module. And we invite our audience to learn more about the scientific accuracy and simplicity the HTPK model -- our module provides for early drug design and optimization. Please visit our website at www.simulations-plus.com. The webinar has been recorded for playback and will be available on our website and YouTube channel. And again, as Eric said, we will get back to anyone who had a question in a follow-up e-mail. Again, thank you. Have a great day. Bye-bye.

Thank you so much again, Andrés. Have a great rest of the day.

Thank you. Bye.

Thanks, everyone, for joining. Bye-bye.