NEXT-GEN INDICATION STRATEGY WITH AI AND REAL-WORLD DATA — FAST, VALIDATED, SCALABLE

Welcome

Eric Mertz: Good morning and good afternoon, everyone. I am Eric Mertz, and I am pleased to welcome you to our webinar, Next-Gen Indication Strategy with AI and Real-World Data — Fast, Validated, Scalable. Today we will explore together how combining artificial intelligence with real-world data can help accelerate indication strategy, cutting development times by months, increasing asset value, and supporting faster, clinician-validated decision making for R&D teams.

Eric Mertz: Over today's session you will hear how to build a scalable, evidence-based indication roadmap using AI, real-world data, and phenotypic modelling; how to quickly identify high-impact, patentable opportunities; and how this work can progress in weeks rather than months. We will close with a real-world case study from Tridek-One's CEO on reshaping their asset-development strategy.

Eric Mertz: Our three speakers are Dr Bernard Hamelin, Chief Medical Officer at chAIron; Gilles Paubert, Senior Vice President and Global Head of Cegedim Health Data; and Laurence de Schoulepnikoff, Chief Executive Officer of Tridek-One. Each brings a distinct perspective — from data-driven clinical strategy, to real-world-data infrastructure, to applying this approach in biotech asset development.

Eric Mertz: Here is a quick look at today's flow. We will start with an overview from Bernard on why this represents a shift in how indications are found. Next, Gilles will share insight on the importance of data quality and reliability. Then Laurence will present a real-world biotech use case from Tridek-One. Finally, we will open the floor for your questions. Please do not hesitate to post your questions during the presentation, and we will address them at the end. With that, I will hand over to Dr Hamelin to begin.

Why indication selection is the bottleneck

Bernard Hamelin: Thank you, Eric. Good morning and good afternoon, everybody. My name is Bernard Hamelin, and it is my privilege to introduce an indication-finding approach developed by a group of scientists, based on the latest advances in AI applied to electronic health records.

Bernard Hamelin: Clinical research is central to evaluating the safety and efficacy of health products; it is the transition from scientific discovery to an innovation validated for medical use. However, this process remains slow, carries a high and well-documented risk of late-stage failure, and is increasingly expensive. Tools such as electronic health records and claims data, combined with analytical platforms, have already transformed several aspects of clinical research, particularly in clinical operations. Yet other important bottlenecks have not been addressed.

Bernard Hamelin: Prioritising which indications to pursue is a crucial decision. It applies both to treatments already on the market and to novel mechanisms of action. The objective of this presentation is to explain how an analytical approach can significantly improve the benchmark for indication selection.

A data-driven approach

Bernard Hamelin: The traditional development path — target validation and indication selection — is hypothesis-driven. It is based on the literature, on phenotypic modelling such as cell-based or animal models, and on insights from biomarkers or key-opinion-leader experts. It is a complex, sometimes uncertain, and always slow process.

Bernard Hamelin: Our approach is a data-driven one. Drug targets are often functionally pleiotropic and may be relevant to multiple indications. Rather than looking up diagnosis codes and hoping they capture the biology, we reason from a mechanism toward the patient populations it predicts and corroborate those candidates against real-world evidence — pairing mechanism-grounded reasoning with data-driven discovery, reconciled inside one analytics environment.

Bernard Hamelin: The concept of reverse translation is relevant here: starting from the clinical effect of a drug, or a combination of drugs, on a clinical outcome and working backwards to understand the biological pathways involved in the disease. In practical terms, the workflow reasons from a drug's mechanism of action toward the patient populations it is predicted to benefit, corroborates those candidates against real-world evidence, and ranks them on development and business considerations. It also opens opportunities to partner or license new applications of an asset under development, which can add significant value.

Bernard Hamelin: The output is a list of indications — qualified, quantified, and ranked on development and business considerations — and that is the final result of the work.

Bernard Hamelin: So what is the benefit of this approach? Let me start with the traditional method. It is usually slow — 6 to 12 months to generate a list of indications per pathway. It is qualitative: you cannot readily compare the probability of success across different indications. It tends to surface the most obvious indications, driven by the key-opinion-leader experts you happen to consult, with the attendant risk of bias — you only see the indications those experts already know about — and its robustness is difficult to assess.

Bernard Hamelin: The data-driven approach operates in a much quicker model. A validated, ranked, and testable indication list can be produced in a focused, phased engagement of roughly five to eight weeks. It often surfaces new, unexpected indications. Because it is based on a real-world dataset, it can also be used to support patents and the value of the asset. It is a validated and robust method that can subsequently be confirmed with randomised controlled trials.

Bernard Hamelin: In conclusion, indications can be visualised in relation to the relevant biological rationale. We believe this method offers patentable evidence relevant to a client's composition of matter, a prioritised list of indications, and new opportunities for partnering and licensing. I will now let Gilles introduce the data we use and explain how we work together. Gilles, over to you.

The real-world-data foundation

Gilles Paubert: Thank you, Bernard. Good morning and good afternoon to everyone. We have just explored how real-world data and artificial intelligence can fundamentally transform how we approach indication strategy, bringing more objectivity, speed, and predictive value to early-stage decisions. But this data-driven strategy relies on one critical foundation: the quality and reliability of the data itself. Without robust, granular, representative, and scientifically validated data, even the best AI model will produce limited or misleading results.

Gilles Paubert: That is why I would like to introduce the real-world-data infrastructure that Cegedim Health Data brings to this work: a pan-European database designed to meet the highest standards of scientific, regulatory, and AI excellence. It stands out thanks to five key strengths. First, a truly pan-European scale with exceptional longitudinal depth — a single integrated clinical database launched in 1994, covering over 50 million patients across seven countries, with on average around 15 years of historical data per patient. Our physician network in primary care captures a maximum of patients with rich electronic medical records spanning primary and hospital care, giving the possibility to understand the full patient lifeline.

Gilles Paubert: Second, the clinical richness is unmatched. The size of the network and data portfolio allows us to capture structured, coded data on patient characteristics, diagnoses, treatments, procedures, and claims, across more than 30 medical specialties. Where our proprietary databases do not fully cover certain diseases — for example, highly hospitalised conditions such as haematology — we have built partnerships that enable additional data-collection registries or linkage with existing hospital databases, such as in the UK or the PMSI medical-administrative database in France.

Gilles Paubert: Third, the database is natively built for science and AI. It is structured around a common data model and harmonised terminologies, which makes it inherently ready for research workflows and machine-learning applications. Fourth, its scientific and regulatory credibility is well established: it has been used in over 2,000 peer-reviewed publications and regulatory submissions, and has been approved by authorities. Finally, the database is representative — prevalence, including for rare diseases, is aligned with population-level epidemiology across Europe, which ensures robust signal detection and meaningful clinical insight. It is this combination of scale, granularity, structure, credibility, and representativeness that makes our data infrastructure a strategic asset for any data-driven R&D initiative.

Gilles Paubert: Beyond the intrinsic value of the data, a critical success factor is our ability to operationalise this resource quickly, precisely, and in full alignment with project needs. Our commitment is to go from a very large patient base to a high-quality, AI-ready cohort in a matter of weeks. This is enabled by several factors. The first is scalable, flexible patient selection: the infrastructure allows fast, flexible access to longitudinal profiles, enabling efficient phenotyping and reverse translation. The second is operational excellence and client support: a dedicated team supports cohort design, testing, and delivery, ensuring scientific rigour and reproducibility while keeping timelines short. That timeline is not set in stone — anticipation and close cooperation with the client can reduce it further.

Gilles Paubert: Third, targeted cohort extraction is performed once the cohort is defined, based on clinical history, treatment exposure, and comorbidities, tailored to your project needs. Fourth, AI-ready formatting follows: the datasets are formatted to feed the modelling work, which significantly boosts efficiency and saves time for the downstream analysis. This capability substantially reduces time to insight while maintaining the scientific integrity needed for clinical decision making. High-quality data combined with robust operational delivery can rapidly turn strategic vision into actionable insight. To make this tangible, I am pleased to hand over to Laurence from Tridek-One, who will work through a real-world use case where this approach was applied to an active development programme.

The Tridek-One use case

Laurence de Schoulepnikoff: Thank you, Gilles, and thank you, Bernard, for your kind introduction. I am very pleased to introduce our use case — how, at Tridek-One, we leverage this approach to identify potential indications as early as preclinical development. Very briefly, Tridek-One is developing first-in-class, CD31-modulating bispecific antibodies — a platform-based approach for inflammation and immunology.

Laurence de Schoulepnikoff: Why, and when, can we start searching for indications? Based on the mechanism of action and the target cells, we can begin very early to look at potential therapeutic indications. What I want to share is why you should start early. The potential indications serve to prepare the preclinical development plan, to define the in-vivo models, and to inform the translational work. They are also instrumental when you present programmes to investors and potential partners. If you approach the external world, the first question they ask is which disease you are targeting — even more than your mechanism of action. Saying you are working in inflammation and immunology is not enough.

Laurence de Schoulepnikoff: At Tridek-One we started this work almost a year ago and looked at various approaches. Today there are many approaches combining AI with different data sources, algorithms, and expert assessment. In a nutshell, the classical approach takes about 6 to 12 months: reviewing public data and literature, databases such as clinicaltrials.gov, and epidemiology, then moving to expert judgement and interviews with key opinion leaders — which, to Bernard's point, can give a somewhat biased view, because two key opinion leaders may hold completely different views. That is the classical way: you arrive at a number of clinically relevant diseases and then prioritise them against fairly standard criteria.

Laurence de Schoulepnikoff: What indication would you look at first in a biotech start-up? The objective differs from the pharma world. In a start-up, a clinical proof of concept is a pivotal value-inflection point for the company. You look for an indication with a very high probability of success for clinical proof of concept, based on preclinical and translational studies — the golden nugget. You also want an indication that allows expansion within the same pathophysiological pathway. As a start-up, you cannot afford very large, expensive Phase 2a studies; you want a Phase 2a study that gives you proof of concept in patients and that you can then expand. Some start-ups exit on a single indication, with the larger pharma partner taking it forward; it is the proof of concept that delivers your key value-inflection point.

Laurence de Schoulepnikoff: In terms of scoring criteria, these are very classical. You look at science and feasibility — it has to be relevant for the mechanism — and at developability, which is really important. Developability screens out a number of indications: patient recruitment may be too low, endpoints may be unclear, or you may have to run on top of standard of care, where it becomes extremely difficult to differentiate. You also want unmet need, treatment value, and a sense of the addressable patient population, which gives you an idea of the potential. You might also consider orphan diseases, though that is a strategic decision. Then, depending on your therapeutic area and mechanism of action, you assess what other pathophysiological factors should be considered for selection.

Laurence de Schoulepnikoff: Here is an illustrative example of the kind of mechanism-to-population output you obtain, where a mechanism's relevance to a number of candidate diseases is surfaced and then corroborated against real-world evidence. Take asthma as an example: macrophages are very much involved in the disease from a biological perspective, and that is why asthma surfaces as a potential indication. This is a clearly illustrative example, not a specific result.

Laurence de Schoulepnikoff: Now to the more concrete example from the collaboration we carried out with chAIron. We ran this work on our lead programme in thromboinflammatory diseases. First, it is important to note that this is very new: we are first-in-class on both targets — a bispecific, first-in-class on both arms. So it was important to confirm that the mechanism is relevant for the diseases we are looking at, and that came out of the work. The analysis also distinguished primary and secondary thrombotic diseases; we were more inclined toward the secondary thrombotic diseases, and those were the indications that ranked highly.

Laurence de Schoulepnikoff: Most importantly — and as Bernard mentioned — some inflammation-related indications are obvious to everyone. What matters more is which novel candidates make sense from a mechanistic perspective. The analysis surfaced both expected inflammation-related indications and several novel, mechanistically-grounded candidates — including conditions such as systemic sclerosis and sickle-cell disease — that we are now evaluating further for developability. Another candidate, peripheral arterial disease, was identified and corroborated, through the analysis and our other work, as an indication of relevance. One further indication, end-stage renal disease, came out as more challenging to develop, but it is always interesting to see what else surfaces.

Laurence de Schoulepnikoff: As for next steps: once you reach this preselected list, it is extremely important to validate it and identify the right key opinion leaders to confirm developability. That is what we are doing now — conducting interviews with key opinion leaders and preparing early clinical-development plans based on these indications. Depending on the indication, you can foresee Phase 1b in patients, or translational studies and disease models. Then, for the selected indications, you want to do a deep-dive analysis on scientific validation, the possible setting, the development strategy including timelines and estimated budget, the competitive environment, and the targetable patient population. As we progress our current financing, I can testify that these are exactly what investors and pharma partners look for. So my message is that you should leverage all these new insights to increase your value proposition. With that, thank you so much for your time.

Audience Q&A

Eric Mertz: Thank you very much, Laurence. Bernard, Gilles — we have time for questions, and we have received a few. The first asks about human validation by a key opinion leader or other expert at the end of the process. Perhaps a question for you, Bernard.

Bernard Hamelin: You should see this as a tool that makes the evidence available to experts to evaluate the potential. The approach learns from the available data — clinical trials or real-world data — to understand the pathways and mechanisms of action, and we need the key opinion leaders to validate what the data are showing. So it is not a question of not using key opinion leaders; it is an additional tool to accelerate the process — indication finding through artificial intelligence and human intelligence together.

Eric Mertz: Thank you, Bernard. The next question is for you, Gilles: of the patients in the database, what is the prevalence of those who also have laboratory values, on top of diagnosis codes?

Gilles Paubert: A very good question. It is not 100 per cent; it depends on the disease we are working on, with a range of roughly 20 to 60 per cent. As a medical-software provider, we designed the data for two main uses — care and research — so we are very close to the day-to-day practice of the physician, and the data carry real value of use and medical intent. So, depending on the therapeutic class and disease, it can be between 20 and 60 per cent, higher for some long-term conditions. With a very large patient base, we have enough patients with all the information we need to run the AI model.

Eric Mertz: Thank you, Gilles. The next question asks whether we can share a case study of linking a mechanism of action to a list of indications. I think that is what Laurence presented at a high level, and we can go more granular or share additional use cases on request. The following question: are clinical-trial data — for example from clinicaltrials.gov — also incorporated into the model, to learn whether an identified indication has already been targeted? Perhaps one for you, Bernard.

Bernard Hamelin: The principle is that you learn from the effects observed. Those effects can be observed in real life — hence the diversity of situations in the data — but also from clinical trials. So the answer is yes: we learn from clinical-trial results. Clinicaltrials.gov provides some degree of results, the literature sometimes provides more detailed results, and these are combined with the effects observed across the diversity of real-world settings.

Eric Mertz: Thank you. The next question is for you, Laurence. You presented your recommendations and next steps; where are you now in your journey in implementing those findings, and what reaction or engagement have you seen from stakeholders?

Laurence de Schoulepnikoff: We are right in the middle of it — interviewing key opinion leaders. What really helps is that, once you have this list of indications, you can identify the relevant key opinion leaders. Because there are no key opinion leaders specifically in thromboinflammation, you have to go deeper: are you in primary thrombosis, secondary thrombosis, and which type of indication? So where we are today is interviewing key opinion leaders and defining preclinical and translational studies and disease models for the indications, and we hope to reach a final selection in the next quarter. As we progress our current financing, the first question you always get is what the indications are, their relevance, and their competitive landscape. You have to be clear that the indications you propose make sense, at least from a pathophysiological perspective, and that you know which criteria you have to evaluate.

Eric Mertz: Thank you, Laurence. Back to you, Bernard. Two related questions on the approach: does it generate new indications every time you use it, and do different drugs come up with the same result?

Bernard Hamelin: The diversity of the data allows us to find indications systematically — though not every time; it depends very much on the use case being tested. There will generally be indications in the result. As for whether different drugs come up with the same results: in my experience, no. I would not claim we have an exhaustive set of models to settle the question, but it is very unlikely, and so far we have not seen different drugs return the same results.

Eric Mertz: Thank you, Bernard. A further question, on the example Laurence provided around the CD31 bispecific antibody: how does the approach handle indications for one of the two targets versus indications where both targets are relevant?

Laurence de Schoulepnikoff: Without going too far into the biology, our bispecific targets different types of cells depending on the second arm. So you have to start the indication search again: the mechanism is similar, but the target cells are different, so you begin again.

Eric Mertz: A last question on the data: when you already have generated clinical data, how does that work in the approach — do you still use it and incorporate it with the real-world data? Perhaps for you, Gilles.

Gilles Paubert: Yes. In Europe we start with our own data, and we can also add other information in different ways, because we are expert in secure data environments and trusted-research environments. The legal process can take longer than expected, but we can link and integrate different datasets. Anticipation matters here, because we have to work through the legal process and the relevant scientific-committee approvals — but the answer to the question is yes, we can.

Eric Mertz: One last question. If I understand correctly, the bispecific antibody was designed before this exercise, so you likely had indications in mind. Did the approach return those highly ranked, and what was the order of prioritisation?

Laurence de Schoulepnikoff: I would rather not share too much about the ranking, because that is what we are working on now. The point I want to make is that, once you have your mechanism and your target cells, it is important to start this work — because, although the output comes quickly, it still takes time to process, prioritise, and engage the key opinion leaders. That subsequent work, after you have the preselected indications, is very important, because once you select your lead indication it is difficult to go back. What I can say is that the approach surfaced a meaningful proportion of genuinely new, mechanistically-grounded candidates alongside the expected ones.

Eric Mertz: Excellent. I think we are almost on time. If there are no further questions, I will invite any final comments from our speakers; otherwise I will thank everybody. We are happy to follow up, and I hope you enjoyed the session. I wish you a good rest of the day and a very good weekend. Thank you very much.

Laurence de Schoulepnikoff: Thank you for having me.

Bernard Hamelin: Thank you. Goodbye.