Webinar Transcript: Accelerating Clinical Translation of Antibody-Drug Conjugates with Hybrid AI

The following transcript was taken from a recorded conversation between Dr. Jo Varshney, CEO of VeriSIM Life, and Dr. Annick Menetrey, Clinical Pharmacology Lead at Debiopharm. Read on to understand how VeriSIM Life and Debiopharm partnered to scalably investigate first-in-human (FIH) dosing strategies for antibody-drug conjugates to reduce tumor burden safely and effectively. (Edited for clarity and length.) 

Dr. Jo Varshney

Thank you xTalks, and hello everyone. Welcome to this amazing and exciting webinar with Dr. Annick Menetrey, as well as VeriSIM Life. We are thrilled to be here and really appreciate the opportunity to sit down with Annick and discuss all the elements that are taken into account when it comes to clinical trial design as well as drug development.

A little bit about us: with VeriSIM Life, our one mission, which is very aligned with Debiopharm, is to bring novel therapies to patients everywhere, faster and effectively. We were founded in 2017. We are backed by leading life science investors, as well as deep tech investors. We are trusted by leading global pharmaceutical and biotech customers, and we have seven different patents around the technology, as well as drug assets that were developed based on this technology.

With that said, I'd like Annick to give a brief overview of Debiopharm, please.  

Dr. Annick Menetrey

Okay. Thanks, Jo. First, I would like to thank you for the opportunity to participate in this webinar. It's my pleasure to be here and indeed, I will give a few words about Debiopharm. So Debiopharm is a privately owned pharmaceutical company. It is based in Switzerland and it was founded in 1979 with a staff of only two people. And now we are over 400 employees. We have quite a strong pipeline with 15 products in development, and this pipeline is focused on oncology and infectious diseases. It's a diverse pipeline, as we develop small molecules to biologics, and also because our programs are at different stages of clinical development, from non-clinical stage to late stage, phase three trials.

Debiopharm’s influence extends beyond our borders, as we have strong collaboration with pharmaceutical companies in Europe, U.S and Asia, but also, partnerships with academia. And as you said, this is important for our mission in order to bring new innovative medicines to patients. And there's more - there is also that we have an activity of investment in startups dedicated to digital health. 

And, smart data - I think that's worth a mention because thanks to this activity, we have the opportunity to be aware of new AI platforms and innovative technology, which may enhance the efficiency of our drug development. So all that is to say that indeed, our firm is oriented towards innovation and believes in the use of digital tools and AI to do so. So my pleasure to be here again.

Dr. Jo Varshney

Thank you so much, Annick. So why are we here? This is a question every pharmaceutical company is concerned to address, and yet also may not be able to address. Currently, over $300 billion is spent annually in R&D by pharmaceutical companies. And although this investment is getting significantly higher each year, the output or the number of drugs that are being approved by the FDA is still only 50, which is a really challenging reality for human health.

It’s because drug development is complex, and there are different elements within the drug development lifecycle that impact this throughput challenge. Some of them being generating targets that are not translatable, basically (challenges) getting the chemistry to work in animal models and ultimately into the patients is a significant challenge. And then there are different silos of knowledge, data transfer, as well as having the incorrect animal model selections. And ultimately in the clinical trials, there are issues with safety-toxicity that impacts the clinical trial success, as well as not having enough information to even go into the clinical trials.

So why do all these challenges exist? We know that biology is extremely complex, and I don't think anybody would contest that. You know, if we knew everything about biology, we would be in a very different time. And there has been remarkable progress made in medicine as well as understanding drug behavior and diseases. But with all the complexity, the data quantifying our understanding of biology is still behind, and we believe that AI alone cannot solve this problem.

We need deep expertise in biology, physiology as well as the different interdependencies between different elements of chemistry, how it interacts with patient variability, animal model variability, and then being able to identify those hidden patterns to reduce the inefficiencies and the silos that come with different expertise across different pharmaceutical companies. With that said, VeriSIM Life has developed a computational platform called BIOiSIM.

And we made an attempt - full disclosure, we don't believe that we can solve everything - but we have made an attempt to bring all these different critical aspects in the computational platform together to mitigate risks in clinical trials before the drug enters the clinical trials. And the way we do this is we combine different aspects of chemistry, the physiology, the physiological difference between different animal models, different target engagement behaviors, and clinical patient variability and condense this into a scoring system. This enables us to discuss the reasons why our AI system with the knowledge integrations, which we call hybrid AI, is informing or giving that score, and that enables us to have better conversations with customers and more importantly, with the experts who know more about their drug programs than we do. It also ultimately helps our platform understand the different knowledge domains that are required to solve these translational challenges to create a more successful clinical program. Of course, there's a lot to unpack here, and this is why we have an expert such as Annick.

So Annick, you being a clinical pharma expert, please give us a little bit more information. What does that really mean for the audience? What are the key responsibilities you have for your department as well as how you are thinking about AI, and how you're involving and evolving integrations of it into your own programs.

Dr. Annick Menetrey

I will start with my role as a clinical pharmacologist in a pharmaceutical company. So indeed, my role includes a diverse range of responsibilities. I think the key applications here - I can give two main ones - are really about supervising the pharmacokinetics and exposure response analysis in our programs to ensure safety and efficacy for the patient.

And I'm also tasked with designing and supervising dedicated trials, which we call the clinical pharmacology trial. So you may have heard about human mass balance study where we give a radio labeled version of our drug, to elucidate the route and elimination of the drug. We also have a dedicated trial for drug-drug interaction or food effect, where we investigate the impact either of food or a drug on the pharmacokinetics of our drug or another drug.

I think all these activities are linked to the critical part of my job, which lies in selecting and optimizing the dosing regimen. But for sure, I'm not doing that alone. We are carrying that out in close collaboration with a full team of different people, from non-clinical stage first to discuss the prediction of the human dose, and also with clinicians, biostatisticians, and other functions.

So the teamwork is crucial. It is crucial because we need to integrate different parameters beyond pharmacokinetics, to elaborate on necessary clinical pharmacology strategy. We need to consider the indications, the populations, the biomarker, safety efficacy data, and I think this is where AI can play an instrumental role in supporting clinical pharmacologist like me.

Yet I have to also clarify that we did not wait for AI to try to integrate these parameters into models. We've been leveraging this in silico modeling for some time now. This allows us to improve our PKPD and analyze the PK and PD outcomes and at the end, to support our dosing rationale.

So indeed, these traditional methods, like PBPK physiologically-based pharmacokinetic modeling, have gained recognition and validation by authorities worldwide for their effectiveness. So they have proven to be valuable and reliable on a case by case basis, depending on the data quality and depending on the use. This is why, while we are assessing the AI advancement in our field, the traditional methods still continue playing a critical part in our job because we still have these uncertainties around the acceptability of AI-driven prediction by FDA, even if FDA published discussion papers around AI, and even if we know they encourage model-informed drug development, we are still not confident in this part. But, we still need to gain confidence here.

Dr. Jo Varshney

I think this is a very critical aspect that you mention, and thank you so much for sharing your collaborative approach at Debiopharm, as well as how you think about combining different elements with traditional methods. One of the things that we as an AI-enabled company want to understand is: how do you see an AI-based approach like ours being better - or not better - compared to what currently is being utilized (traditional methods)?

And then we'll come back to the regulatory aspects after this question.

Dr. Annick Menetrey

Okay. So here I may introduce some of the details of the modeling and simulation approaches that we usually use. So one of the uses of these different approaches is to consider factors that may influence the pharmacokinetics of our drugs, that may introduce variability in the exposure and at the end, in the efficacy and safety of the drug.

So this factor may be due to the individual characteristics, such as genetics, or may be due to the medical conditions of the patient, or to concomitant treatment that the patient may need. There are all these multiple parameters that we tend to still address sometimes on a case-by-case basis, one at a time, so that we may, for example, look at the impact of genetics on the PK, and maybe we don't see anything relevant. Maybe then we look at the medical condition of the patient. And again, there may not be anything relevant, and the same for concomitant treatment. But we can wonder, what about all these in the same patient? So here we go more towards the personalized medicine and optimize the treatment to know what would be the outcome.

So I think that's one of the advantages with AI right now. And also as humans, we tend to specify which structure we want to analyze. So we may define what we call the covariates or such factors that we may investigate during our modeling and simulation activities, while with AI we could just leave it open and maybe we could discover new factors that we may not have investigated at all without AI.

Dr. Jo Varshney

Very, very interesting. So is it safe to say that one of the big differences that you see is AI creating a more parallel approach of different parameters coming together, instead of the traditional methods where it's a step by step approach? So having that approach all in one element will help create more understanding of the patterns that were missed because of the step-by-step approach.

Dr. Annick Menetrey

Yeah, I think that’s a nice summary indeed. One of the key advantages that I can see in AI is the ability to bring all this information together and then consider a real world population and maybe, some virtual pre-selected populations.

Dr. Jo Varshney

Definitely. And that's something we believe is an important aspect of artificial intelligence, no matter who does it - to really reveal unhidden patterns or even identify patterns from previous learnings that are very much embedded in the way AI methods are being implemented. 

So, yes, the regulatory aspect is definitely a big challenge.

And, you know, one of the things from our perspective that we have seen, and we are also learning directly from the regulatory agency (FDA) itself, is that there is a significant interest in application and integration of AI. Actually yesterday, they announced draft guidance on platform technology, which is another step forward towards implementing these newer technologies. It means that if your technology is using existing or previous data, they would consider looking into that information to create faster response cycles, which I think greatly demonstrates their spirit to embrace the new technologies.

But of course, the elements come down to the details. And one of the ways that we try to work with companies, and also with the regulatory agencies, is by creating specific case studies where we can show very specifically how a platform technology like ours can be helpful in addressing that specific challenge, because that does two things. One, it helps explain how and why and what that case study will be solving within the drug development. And the second is it embraces knowledge sharing. So that when a company like yours works with a company like ours, the agency already has an idea, like “Okay, there is this application we could look into because we understand what the technology’s doing and what it’s about.”

And I think it's almost like a wait and see scenario. But knowing how much the agency is really embracing, even within their own teams, the implementation of AI in different elements, I do feel very optimistic when it comes to acceptance and change. The other aspect is that in regards to model-informed drug development, they have released guidance within the same framework of use of AI and use of alternatives to animal testing aspects. Now they are embracing this further into the Modernization Act 2.0, as well as several discussions that are happening at the U.S. Congressional level to really do the right animal studies that actually translate into human studies.

There is a saying that you can pretty much cure anything in an animal model; you can cure cancer to diabetes, to obesity, to depression, in animal models. But we still have to see those things translate into humans. So what we tell the clients and the customers that we are working with is: can we really, together, convince the agency why they should consider this in silico approach versus the animal approach? Along with how much should they rely on a particular animal model - is that animal model really going to be sensitive to and help us design the clinical trials? And those are the kinds of key conversations we are having at the agency level. 

With that said, Annick, we have worked together and I’m sure you’re going to go into how we’ve worked together, what your experience was - but before that, what are Debiopharm’s thoughts, within your team and also the entire company, on embracing these changes and working with platform technology companies and working with regulatory agencies together to create those changes internally as well?

Dr. Annick Menetrey

So that's really an interesting question. When I answered about my view of AI, I was quite focused also on the clinical pharmacology part and the model clinical development. But we have several case studies and key applications using AI during the drug development with you and companies such as yours. So we’ve had several opportunities at the non-clinical stage to investigate the use of AI to, for example, select drug candidates, to predict toxicity and the efficacy of these candidates and select winners, or to identify new indications or new potential combinations of treatments.

And then there's also opportunity at other stages, like in the clinical stage, to boost the efficiency of our clinical trials. And also, finally, in the application of AI it was critical for us to identify a biomarker able to predict patient response to the treatment. It’s amazing to imagine all these capacities of AI to identify biomarkers that then help to select a patient population that will respond, that will benefit from a treatment.

And, so I think we are gaining internal confidence. We still need to work with that at different levels so that everyone can have access. First we need the access. So that's also a nice thing about this partnership, that indeed we now have accessibility to AI. That’s, I think, the first thing because we are all quite busy people doing a lot of things. It’s easier when we can have simple access to it, and (beneficial) to have a partnership with people that know the language of AI so that we can rely upon you to perform the simulation. And then we need to gain confidence. So it's not only that we have the regulatory acceptance of the data, but also needing to see that we can use the predictions to support internal decisions. And once we observe that the experimental data match with the AI driven predictions, the more we see the validations of AI, the more we will feel confident to use it. 

Dr. Jo Varshney

Absolutely. And I think validation is super key, right? Because AI doesn't differentiate from apples and oranges because it's all code. Thank you for sharing about working with the AI partners. From the AI partner’s perspective, we believe that we have to work with experts like yourself who have been doing this for decades and understand challenges that are probably not written on paper and are perhaps being shared internally. The more we work with companies like yours, the more we not only learn how decision-making happens, but also how we can make that process more efficient.  

And we really enjoyed working with you. I don't think there was any discrepancy. We really learned a lot from your expertise and your team's expertise. So, I wanted to ask you, how was your experience working with our team, and were you able to utilize some of our predictions for internal decision-making and what that looked like?  

Dr. Annick Menetrey

So first you asked about the way we collaborated together, and indeed the collaboration was highly interactive and productive. And as you said, we really needed to exchange information, and we did that. We started by really aligning our expectations for the work - the objectives, and the context of this work as well. The team at VeriSIM created a lot of opportunities for us to discuss the data, the relevance of the data, and the process that was implemented.

And in the end, it's critical in our role to try to understand the process as much as possible and get as much transparency as we can, so that we are then able to better interpret the results and communicate the results internally. So that was really nice, and they were really proactive in answering and addressing any arising questions. So, fantastic project management again, and thanks a lot for all the exchanges. 

And then if you wanted to know a bit more about the project and the extent - so, we started this project at an early stage, and selected the drug candidate super early. So we did not have extensive PK data, nor GLP tox data. This is where it was really great to partner with you. With BIOiSIM, you were able to provide for the estimates and incorporate them into the PBPK model approach. So then to have the prediction between animals and humans, that part was so great.

And in addition to that, there was the other aspect of leveraging AI for pharmacodynamic predictions. Indeed, the translation for efficacy from in vitro or animal models to humans is sometimes challenging. And here, your team really addressed this issue by proposing different external sources for our review. So that’s something I had to discuss with my colleagues internally, about the relevance of this external data, and then with your team. Then based on our insight, they decided upon the best option to select to have the prediction of efficacy. And we did the same for toxicity. 

And, one important thing that we did not mention yet is about the speed of having these predictions. We were quite early on, with not a lot of data. And then we had the predictions quickly, and this was really a support for our CMC activities and clinical activities. The project, as mentioned by the title of a webinar, was about predicting the therapeutic dose in humans.

And also, your team supported the selection of the starting dose in the initial clinical trial. Having that dosage prediction super early then supported the ability to start discussing the clinical design, which in turn supported the selection of the dosing regimen. You can also use this data to define the number, of course. At the end, we also now have this piece of information in our package to support the selection of a starting dose and the predicted therapeutic dosage.

Dr. Jo Varshney

So we did a lot in a short period of time. Are you able to share how much time we took to provide all these predictions to you?

Dr. Annick Menetrey

Yes. So from the initial discussion to the results, it was less than three months.

‍Dr. Jo Varshney

Okay. Great. So in less than a quarter, you got the information in a limited data set. That makes me very happy. Another big thing also that we didn't discuss and is in the title is, it would be great to hear from you about the differences between ADCs and small molecules. Even when we are good experts in these aspects, we sometimes forget that different challenges exist in large molecules, especially like an ADC molecule, which is complex in every angle.

But I'd love to hear your thoughts on how you look at ADC challenges? Or, how easy this was would be helpful for our audience to understand a little bit.

Dr. Annick Menetrey

As you said, it's completely correct to say that these ADCs are inherently complex. They are built with an antibody, and this is the biologic part that is able to recognize an antigen to be directed against a target. And this antibody is linked to a small molecule, the payload or the cytotoxic.

And this payload may also be transforming to metabolites. So you have to consider all these entities when you want to perform pharmacokinetic or exposure response analyses. So it's complex. It's really interesting because these ADCs are designed to deliver the payload intracellular in the tumor. And so we may be able to avoid the circulation to healthy organs and to avoid side effects.

So they really have to have highly potent molecules, administered to humans. That's one of the big advantages. But then in terms of knowing what's going on with all these entities in the body, that's a challenge, because usually all these antibodies don't run in parallel in the body, and so you may have different impacts for toxicity and efficacy.

And I think this is where AI could further support us in managing all these entities. One additional thing is that, as I said, we tend to have these cytotoxic drugs delivered intracellularly to limit the circulation of this small molecule. So when we need to monitor these small molecules, we need highly sensitive, analytical methods. And it can be a huge challenge, and we may not be able to do it. So here again, we may leverage AI to have the information about the elimination of this small molecule. What does it do in the body, how it is eliminated. This is also where AI could help.

Dr. Jo Varshney

Very interesting. And definitely a big, big challenge because it really comes down to the chemistry, the linker chemistry, how the molecule intracellularly behaves in the different organs. So we'll do our best to see how we can leverage some of the existing knowledge to eliminate some of these challenges, especially when it comes to sensitivity.

It’s an incredible time to be alive, I'm sure, for all of us, because AI has been around for so long, but it's just changed recently. And, as you mentioned, Debiopharm has been embracing AI from different elements, I'm sure from ChatGPT usage to working with companies such as ours. 

Annick, please share - and this is more of a personal question to you - how did you get comfortable embracing something so new and so different than what you’re used to, from the science and the learnings that you have over several decades of your experience. And this could be something that other experts within pharma could be utilizing internally since they are also probably interested in applying AI, but often as humans, when the learning is so different, it can be scary. And if there are ways that you embraced it that helped, it could be really great for our listeners to know how they could apply that internally as well.

Dr. Annick Menetrey

Yeah. So maybe in order to better understand my answer to that, I need first to clarify that I'm not a modeler. So I don't look at AI at the algorithm level. I usually relied upon pharmacometrician. I ask a colleague or service provider to build a model and to execute the simulations. And this doesn't mean that my involvement was peripheral or passive; rather, it is my critical role to have this active oversight. 

And indeed, I am used to discussing the plans, the objectives, to challenge the assumptions for modeling and simulation and also to interpret the results. Another critical part is to communicate these results to external or internal stakeholders, because if you cannot provide clear data, then we cannot make an informed decision based on this data and it is not useful. Already from a modeling and simulation perspective, it's sometimes challenging to do so. I have to digest these results, then I have to give digestible information to my stakeholders so that altogether, we can make these decisions. So with this context in mind, I think I was open to do the same with AI.

For sure, it's been a learning journey for me, because even if I'm used to discussing the data for modeling and simulation activity, understanding the different options for AI and the impact of the decision on the results was new. This is why it was fortunate that we had all these meetings and discussions with your team. So I think I kept my position of asking a lot of questions to challenge you and try to understand, asking for as much transparency as possible so we could align each time on the results, on the data, and on the next steps. And this is how I gained this additional confidence in the data.

Dr. Jo Varshney

Right. So is it fair to say that explainability - the transparency on the approach - is critical for AI partners who work with companies like yours? Additionally, if you could explain how that transparency helps communicate internally what was done and how it helps your decision making.

Dr. Annick Menetrey

I think it's crucial to communicate, it's crucial to encourage questions, and it’s crucial to have a comprehensive understanding of the predictions and to get as much transparency as we can. And there is always a more gray zone, yes - but then we are at the interface between the resource and the stakeholder, and we need to ensure that we give them the message, and this is critical.

Dr. Jo Varshney

Yes. That's very helpful for our listeners to know that AI is not going to solve that problem internally until it's explainable and it's transparent enough for experts to be able to really take that and incorporate that into their regulatory package, as well as to showcase that to the senior leadership, to really have that kind of embracing involvement and evolution within the organization.

I'm hoping that our listeners will really take that as a big message, as it’s one of the big things from our perspective that we emphasize - the question of how do we simplify something to get to the expertise as quickly as possible? With that concept, we came up with this Translational Index score - can you share some of your feedback on it and all the other learnings around it?

Dr. Annick Menetrey

The Translational Index is comprised of the prediction of efficacy and of toxicity, so it’s one number that represents the predictions for the dosing. In my case, it was for the dosing regimen. So it was easy to understand that the higher the scores, the more optimal the dosing regimen.

So that's the best, and it is a great way to communicate the data. I was still also discussing a lot with your team how to understand the absolute value and relevance of the differences between two scores. Your team shared their insight and interpreted the results, I shared mine, and we came up with the overall interpretation of the results. So that was also a nice part. I think this is also because I'm more of a scientist, that I wanted to know a bit more than just the number. But for communication, it’s a great tool to support the decision.

Dr. Jo Varshney

Right, our objective is more on the communication. We also shared how the score was generated and the different models - was that helpful?  

Dr. Annick Menetrey

Indeed, because then I was also able to go back to whatever the pharmacokinetics predictions were and also the efficacy metrics and the toxicity metrics, and this really helped me to better apprehend the data.

Dr. Jo Varshney

Great. So I think that serves our purpose, being able to communicate internally from the AI perspective and then explaining what went into that score and how those individual models performed and what kind of outcomes were there, which is the data you mentioned that really gives us that explainable aspect. And more importantly, when you took this all into account, did you also use that to change the method and see, for example, how it would impact a score if you emphasized more on the toxicity outcomes?

Because ultimately we believe that you know more about your program than we do, and you may have seen certain elements experimentally that we may not have incorporated into our approach or the predictions. And we hope that we can utilize that information back to learn more about how this could be more beneficial into furthering the studies.

‍Dr. Annick Menetrey

Yes, models are not set in stone. So if we have new data, if fresh questions arise, we may need to update the model to incorporate this data, or to reevaluate the model for new questions. That's also quite frequent with traditional modeling approaches, and really we should keep that in mind as well. A model is not a binary thing; usually it's not good or bad. It's just that you have to delve into the uncertainties of your model and consider the past, where you're more confident to make a decision, and the part where maybe you have more uncertainties and how you will mitigate these uncertainties. Then we have to discuss together within the team to ensure that everyone is aware of that.

Dr. Jo Varshney

I think that's a phenomenal aspect, and we hope the field of AI starts having these sort of disclosures that nothing is perfect, we just want to improve the inefficiency. Then it would be a step towards change, and it would be a collaborative aspect from both sides of the world, which is the experts in drug development as well as from AI.

My one last question, and I ask this to everyone who I get to interview, is if you had a magic wand and you really wanted to see, like, “Hey, this is something in drug development or in my team I really want to solve, and I hope I can solve it.” What would that be?

Dr. Annick Menetrey

Okay, let's dream and let's think. What if we could just select a disease and ask AI to give us the structure of a compound that could cure this disease? I think we may not be there yet, so maybe we could start from a structure to predict the efficacy and toxicity in the world population. That would be a great tool to add.

Dr. Jo Varshney

Yeah, we can dream, right? And you never know, the field is evolving at a speed like nothing else. So we are very excited and we believe that we are on the right path. And with folks like yourself who are our partners, we are going to always be better than the last time.