← 返回

2026.09.23 | TB | Life Science | 10x Genomics Atera & AI

原标题：10x Genomics – Atera Assessment & AI Applications Analysis

#medtech

Third Bridge · 2026-09-23

Analyst-led • 14 Sep 2026

Companies: TXG US, BRKR US

Vassili Soumelis (VS) Chief Medical Officer at ArcaScience SAS

Anjelique Soriano (AS) — Third Bridge Sector Analyst

Agenda

Key Insights

Introduction

AS: Welcome to Third Bridge's Interview entitled 10x Genomics – Atera Assessment & AI Applications Analysis. I'm Anjelique Soriano and I'll be facilitating today's Interview with Mr Vassili Soumelis, current Chief Medical Officer at ArcaScience SAS.

Vassili, before we start today's Interview, please state I agree or I disagree to the following statement: You understand the definition of material non-public information and agree not to disclose any such information, or any other information which is confidential, during this Interview.

VS [00:00:34]: I agree.

AS [00:00:35]: Could you begin with a brief introduction to your background?

VS [00:00:41]: Yes, I'm a Medical Doctor, trained in clinical haematology, I have a PhD in immunology. I got appointed as a full Professor at the Paris Cité University. I'm currently on leave of my academic position and after working 3.5 years as a Chief Medical Officer at Owkin, I am now currently a Chief Medical Officer at ArcaScience. My expertise combines medicine, biology and data science, AI and all related fields.

Questions

[00:01:16] How do biopharma customers justify securing budget for large-scale spatial transcriptomics projects?

VS [00:01:33]: Of course, I cannot speak for all biopharma in general, but in my experience and my view is it's really a strategic decision. It comes usually from very high in the executive and the boards of the companies because it's a very strong investment in money, time and human resources, and it has to be fully integrated into the strategy of the company. That means generating those data needs to come with a strategy of what the company wants to do with the data and what they aim to accomplish and how they're going to make value out of the data. For example, if it's a target discovery programme that is in drug developing company, there needs to be a very clear path before even starting any data generation programme on how the data will serve the different objectives, the broader objectives of the company.

Only then, the budget will actually be unlocked. The budget is really conditional to the objectives of the project and how well the project or the programme fits within the broader goals of the company. This is because we speak of very, very large budgets, it's tens of millions of dollars. It may go up to tens of millions depending on the breadth of the initiatives. The higher the budget, of course, the higher the decision will be made. If it's a tiny 10 sample or a couple of modalities project, of course it's very different. What I said is for large projects that engage a lot of resources from the company.

[00:03:31] How does spending behaviour differ between SMID [small- and mid-cap] biotechs and large pharma customers, recognising that it differs from company to company?

VS [00:03:50]: Yes. The big pharmas obviously have more potential because they have a bigger budget. They also have a longer and more tedious and complex decision-making process. The thing is that these fields evolve very fast. If you start thinking of something and it takes one or two years to actually decide and build the project and fund it, sometimes your choice is not optimal because the technology has evolved. In smaller structures like biotechs, the decision-making process is faster so you can decide faster, but you have less resources. That means you have less potential to sequence or to analyse a large amount of samples. That's the key differences I would say.

[00:04:41] Are biopharma and discovery teams and groups actively moving budgets away from bulk and single-cell sequencing towards spatial platforms? Depending on whether they are or aren't, what are your thoughts surrounding these shifts?

VS [00:04:58]: Yes. I think there are many shifts that occur. Number one, many companies, they shift first early discovery programmes, they shift budgets away to later programmes. What I'm seeing is that the budgets dedicated to early discovery, which is what we are discussing in a large extent, are going down. Then in this early discovery, there's how many companies actually use omics data for their early discovery, either for target identification or for target validation, both. These today are the majority, but they're not all equal. Some of them invest a lot. For example, Genentech is known for investing a lot of money into omics data analysis and they've recruited people who are excellent experts in the field. Other pharmas like… I cannot name many names, but Servier for example, invest way less. They don't have the expertise that Genentech has. There is a difference, discrepancy, the level of resources already invested.

The companies that invest a lot, they would go to either shifting from one modality to another or combining the different modalities. I would say the ones that want really to have the deeper knowledge out of a sample, they will combine. Even bulk transcriptomics has a value because it explores, it represents a larger volume of the tumour. If you go and analyse spatial, you have a high resolution on the section you're analysing, but we know there is a lot of intratumoural heterogeneity, this is really very well-established since 10-15 years and there are large studies that show that if you cut your tumour at different sections, the one section will be different from the other.

The spatial is not necessarily better on everything and in all areas. It is much higher resolution, but it represents a much smaller amount of the of the tumour volume. The bulk is a larger amount, but much lower resolution. The single cell can be done also on more cells, so larger volume, high resolution but without the spatial, the distribution which is obviously, it's super important. The spatial, it's high resolution, especially if it's single-cell spatial distribution, but much lower amount of volume.

AS [00:08:06]: What is the volume differences in proportions?

VS [00:08:15]: In proportions, it's lots of magnitude because one section is about, let's say, five microns. It depends on the technology we would use. You can generate a bulk transcriptomics on at least 10 sections. 10 sections, you generate the bulk. One section, you generate spatial. It's at least 1-10 difference, if not more.

[00:08:48] Does spatial transcriptomics expand overall market demand, or is it partially taking budget away from single-cell?

VS [00:08:58]: It depends on the limiting factors in patient sample analysis. There are basically two main factors or three. The first is the budget because the more modalities you add, the more you have to pay. The second one is the availability of the sample because if you want to generate many modalities, basically, you need a larger sample, so it's more difficult to get when you contact clinical centres because many samples have already been used up by studies or you don't have enough material. The third one is the computational part, which in some pharmas and in general, because of AI and advances in bioinformatics pipelines, it becomes easier but it's still very, very challenging and I don't think AI today is able to solve the issue of multi-modal integration and how you're going to integrate and make sense of the different data layers which may be spatial, bulk, single-cell and cistromics, on all these multiple data layers. Ultimately, if you want to generate many modalities, you need to overcome all these challenges. Budget, tissue availability and computational analysis.

[00:10:18] How do the all-in sample costs and data quality of spatial transcriptomics compare to traditional bulk RNA sequencing?

VS [00:10:31]: I don't know if we can say how it compares because both of them, there are different technologies. They have different quality checks to perform, to run. The spatial has a little bit more because you need to also to check the quality of the section itself, you need the pathology check. And then you need to go and check the quality of the RNA before you start sequencing. I would say it's a little bit more complicated, but the bottom line is both of them have very strict quality criteria. If the sample doesn't pass those QC, it should be discarded. This makes it more difficult in one… Maybe spatial transcriptomic studies, you may have up to 50% of your samples being rejected for quality reasons.

It's just because of the way the samples are stored, because of one section may not be optimal, because of the RNA quality and all these things. If the low QC samples are not discarded, this can have dramatic consequences on the data analysis step because it adds noise and it can also lead to false discoveries in the data. A little bit more complex to perform for spatial. Now, if you work with teams that are acquainted with spatial technologies, they are also used to performing all the QCs. It's never a bottleneck in terms of operationally applying those QCs. Just like even more than bulk, it can lead to the exclusion of samples.

[00:12:22] You said that up to about 50% of samples tend to be rejected because of various factors, such as how samples are stored. Are there ways to decrease that percentage? Is AI coming into the mix here, or is that a common threshold assumed to persist and be part of the whole journey?

VS [00:12:55]: AI is not very helpful in that matter. It's really due to the quality of the sample itself. The tissue, the way it's been preserved, it's been fixed and the way it's been cut, and how it's been processed and manipulated by the technicians. It's a technical factors. What AI can do is to fix things. Basically if you have a bad surprise, many low or very average or just above threshold or close to threshold samples, the AI can help fix that in a way that it will try to normalise to bring all samples to a little bit higher quality and to remove what the AI will see as a noise, as compared to the real signals. It's very difficult to sort out because you are in a disease situation where something that AI can see as maybe it's really (audio distorts 14.06) of the (audio distorts 14.09) because it's a…

I was saying that AI can try to fix things to some extent in terms of data normalisation and removing some noise, but it has limitations because we are in a disease situation where something that AI interprets as noise may be a real biological signal that is associated to the pathological tumoural process. Personally, I wouldn't trust AI 100% in removing noise out of a sample if it's of low quality. I would push towards excluding the low-quality samples rather than keeping them in a study and using AI to fix things. It's also my clinical bias. I'm a Medical Doctor, so I want to be sure that the sample is as close as possible to the real tumour. The sample goes through (audio distorts 15.25) that you necessarily induce biases. The bits (ph 15.31) of the sample is the key to anything that comes downstream.

[00:15:37] Could you discuss how much Scale Biosciences' barcoding technology is able to lower single-cell costs? Does this cost reduction open up high-throughput screening contracts within SMID biotechs vs large pharma?

VS [00:15:57]: Yes, for sure. The barcoding technologies have become very efficient and I think it does reduce costs a lot. I cannot give you numbers because it's not my area I'm not dealing much with. I'm validating budget, but I'm not really the one actually designing the operational aspects of the study anymore. For sure, you can manage higher volumes and higher throughputs.

AS [00:16:29]: Just to confirm, it is significantly lowering single-cell costs to then open up more high-throughput screening contracts, or am I misinterpreting?

VS [00:16:38]: I don't have the numbers in mind, so I cannot tell you because I'm not doing the budget in my company.

[00:16:51] How are biopharma customers balancing upfront instrument costs against ongoing reagent spending when planning multi-year projects?

VS [00:17:05]: Yes, that's a good question. Again, it's really a case-by-case situation because it all depends on first, the size of the biopharma and how much they want to invest in the long run. I guess biopharma would buy equipment and generate the data themselves only if they have sufficient budget (audio distorts 17.36) and data generation will be sustained in time. Other companies like the ones have been, and it's public information, we were outsourcing to service providers. Big platforms that actually generate the data for you. I cannot tell you how many have internal data generation capabilities and how many outsource. My feeling is the majority would outsource because the sequencing facilities today are very, very efficient. They are very skilled. They have very reproducible processes and they're also very pretty cost-effective.

To your question of the investments, it really concerns only the few companies that decide to invest in sales and then they will have to manage the budgets that come with reagent costs, with maintenance of the equipment. Personally, I've always worked with service providers that manage these costs and they actually charge you a cost per sample cost that includes everything that is associated to running the sample.

[00:19:17] What would you say are the main bottlenecks, especially cost-wise, when scaling spatial projects from small pilot studies to cohorts of hundreds of patient samples?

VS [00:19:35]: Again, so my experience is the bottleneck including cost-wise, it's the sample access. It's not necessarily the data generation, it's once you have secured the samples, the data generation, if the sample is good quality, it's becoming fairly easy because of the technological advances, because of the protocols that are very precise and because of the SOPs, the standard operating procedures. What increases the cost is the sample access. If you contact a hospital, academic centre, a biobank, they would actually provide the samples to you. Sample access is associated with a lot of costs. The costs are overheads to the academic institutions. It's the pathologists' time to actually access the sample and disarchive the sample. It's the pathologists' time for quality check, it's the technicians' time to actually process the sample. Everything that happens before the actual data generation effort, per se.

Basically starting using a platform or a technology like spatial transcriptomics or single-cell to generate the data. If the data generation for example, is 200 per sample, everything that happens before can be up to 1,000 per sample, 2,000-3,000. It depends which samples we're speaking of. It depends if it's a frequent or rare tumour, for example. If it's a tumour, it depends on the institution, it depends on the costs on how much. By the way, in France, we call it the environment costs. It's basically the costs that are associated for a given institution are associated to sampling. It includes the surgeon's time, pathologist's time and so on, it includes all that. It's in the negotiation, in the contract negotiation to access the samples. All of that comes into play. In my experience, that's the biggest part of the budget, it's not the data generation itself.

Also, the data generation is impacted by the technology and as you mentioned, some technologies they now allow to lower per sample cost. All the steps that are before, they're not sensitive to technological advances because it's all manual human work. That's the part of the expense that will not change in the next 10, 20, 30, maybe more years because it's related to getting the sample out of a patient and analysing through processes that have not changed in the past 50 or 100 years. Meaning, you fixed your sample, you cut it, you put it under a microscope and you just look at it.

[00:23:03] Is there a general price per sample that you think spatial platforms ultimately need to hit before biopharma customers are routinely willing to approve high numbers of sample clinical trials, perhaps 500-plus?

VS [00:23:24]: Again, it's not the largest part of the budget. Today, the sequencing, what the cost of the platforms, the difference would be what? EUR 200 instead of EUR 300 or EUR 400, it's a couple of hundred euros, but it's not even a limiting factor and it's usually not the decisive factor on a project. Even for of course, you have 500 samples, you're going to pay attention to 200 vs 300, but the sample access is the key and the quality is the key. Personally, I've been in R&D academia and industry for the past 20 years and I've learned and I'm still applying the rule of quality comes first. If I have a facility or a platform or technology gives me high quality for EUR 300 per sample and the other one is a bit lower quality for EUR 200, I go for EUR 300.

What I would ask the providers, the platforms, whatever, the service providers that use a certain technology is first, the quality. I'm willing to pay a little bit more to get that quality because if I don't have that quality, it can be 10 years of downstream research or target discovery or mechanistic studies that will be jeopardised because I didn't invest enough in that quality. The value is in the quality. Today, I've never I've never seen a big project blocked because of the sequencing cost today. It was an issue maybe 10-15 years ago.

[00:25:20] How do biopharma customers evaluate raw 10x Genomics spatial data to determine if it's clean and standardised enough for AI target discovery models?

VS [00:25:41]: Again, it's something that today is very well-standardised. There are procedures that will tell you whether the data are AI-ready in terms of quality control, in terms of normalisation, in terms of removing as many biases as possible before feeding the AI model that then would actually will have no way to actually to discriminate. I cannot give you a list that will be too technical to drive you through all these steps. It's usually AI teams that actually do that, but with the help of biomedical teams like the one I had at Owkin, that actually ran those processes and we have quality criteria that tells us whether the data are AI-ready.

[00:26:36] Would you be able to discuss any specific quality metrics that are used to determine whether a dataset is accepted for AI model training? Are there any specific 1-2 criteria here?

VS [00:27:01]: One is the quality of the section because if the section is not of good quality, basically any spatial information will be distorted. It's the quality of the RNA that is captured on the pixels or each of the spots. It's the position of the sample within a larger cohort. There are many bioinformatics, actually pipelines, that try to detect outliers. The outliers may be technical outliers that actually are falling very much apart or outside of a cohort, of a patient or healthy sample of the cohort. All the objective of the pipelines is to detect and eventually exclude those outliers, if the fact that they're outliers comes from a technical issue. What you would do is hierarchical clustering, it's principal component analysis or related classification methods that would actually look at the structure of the data and tell you which data are really outside of the structure. This is classical. What I'm telling you is really nothing fancy. Any bioinformatics or data science team working at international standards will actually go through those steps.

[00:28:44] Are there any specific run-to-run batch variabilities that customers tend to run into when combining 10x's spatial datasets across different experiments?

VS [00:28:59]: Yes, of course. There are always the possibility of batch effects if you work with multi-centres. Multiple clinical centres for example, and one centre has one fixation protocol, and the other centre has a different fixation protocol, or it can be the age of the samples. Some samples are very recent. Some samples are 10 years old, others are 20 years old. Every time you have some differences in any of the steps of, what I call, the data life cycle. The data life cycle starts from the patient, not even from the sample itself. It starts from the patient from which the sample was taken. Every time you have differences, the differences actually can affect the data and create batch effects. For example, if the treatment habits, the treatment protocols in one centre are different and in one cohort, the patients are not pretreated, in another cohort, the patients were even slightly pre-treated and received some low-dose corticosteroids for example, it cannot directly affect the samples. All these need to be checked at all of the levels.

You need to check the homogeneity of your cohorts at the clinical level. The patients, the disease and co-morbidities and treatments. You need to assess at the pathological level, you need to assess at the cellular and molecular level. If you do that, you actually can get a grab of all the possible sources of biases and batch effects. Once you've identified those possible sources and when you observe first, there is a decision to make, whether you combine or not, the cohorts. You may decide to leave the cohorts separate. If really they are too different, you cannot integrate them. If you decide to combine them and there is a manageable batch effect, then there are computational methods to actually to correct for the batch effect. Those methods are limited. They cannot correct for huge batch effects, and so that's why it's very important to identify all possible sources of biases and really stop, think and decide whether or not to combine.

There is a tendency in the literature and also in the industry and biopharmas that more is better. More samples, more data is always better, I don't agree. More is not necessarily better. It can really introduce noise. Sometimes you're better off analysing smaller cohorts but are more homogenous and you can combine them in different ways without analysing them as one cohort. You can use one for discoveries, for example and you can use the second one to validate a specific finding. This is very different. Yes, you can do it because the validation of a few pre-identified hypothesis is not affected by noise, the same way exploratory analysis will be affected by noise. The noise is impacting more when you explore thousands of data points, that's where the noise will actually be impactful. If you want to validate a couple of hypothesis, the impact of noise is really reduced, if not absent. This is all about the data analysis strategy that is specific to a project and to what you want to learn.

[00:33:14] How materially do data issues impact the accuracy and reliability of downstream AI models?

VS [00:33:26]: It can be huge and that's why it's super important to work in an interdisciplinary manner. Basically, you shouldn't give the data to data scientists who are expert computer guys or mathematicians and are super strong in AI without working with medical doctors and biologists. Whether they are in the same team, whether they are collaborating, whether they are in the clinical centres. Those people should be also acquainted with issues of data generation and omics data analysis. AI expert isn't necessarily familiar with all these aspects and will not think of maybe… Some patients are different from the others because they come from different centres and because the centres have different habits, they may not even know that maybe you can pre-treat a patient or a disease with these glucocorticoids and maybe not the other one. It's really the teamwork that is required to actually identify all these batch effects.

Of course once they're identified, there will be a discussion. When I told you the teams need to stop and decide whether they're going to merge or not, this is a joint decision. I'm really supporting, I'm always favouring the collective intelligence and collective decisions because ultimately, this can impact the whole research and discovery programme, and so everyone should take accountability in a way. Everyone should be aligned before actually deciding something. When I say everyone, it's medical doctors, biologists and data scientists. Not data scientists alone, everyone.

[00:35:31] In large multi-centre studies spanning independent clinical labs, what SOPs [standard operating procedures] are needed to guarantee data reproducibility?

VS [00:35:45]: It's so many SOPs, so I don't know if we have time to discuss that because actually, every step needs the associated SOPs. In a study, in a multi-omics data generation, you may have up to 20-25 different procedures that has to be applied and you need to have experienced people actually writing these SOPs and distributing them. You need to train the centres, the teams to apply the SOPs because the fact that you have an SOP, it's not the key to everything. It's not the end of the story. It's like having a recipe and you give the recipe to five different people, a cooking recipe I mean. Some people will apply the recipe very correctly because they are experienced in preparing a certain dish or a cooking. Others with the same recipe, they will just not do it and the food will be crappy because they are not experienced and some steps they will not pay attention to. You need to train or you need to work with the clinical centres that are also experienced in each of those steps. I'll tell you, it's about 20-25 SOPs. You have an SOP for how to cut the section, you have an SOP for how to check the quality of the section. It's very detailed.

[00:37:06] How would you size the TAM and timeline for selling high-plex spatial data sets directly to AI drug discovery companies and model developers?

VS [00:37:20]: No, I cannot because there's too much variability on the early discovery strategies. It's also dynamic. Even if I would tell you something today, you can name a few companies that use actually omics, maybe in a couple of years they will stop using omics because they will realise that it's not efficient in discovering new targets. Also because the budgets in big pharmas, they are really cutting budgets from early discovery. There is a renewal of interest in these multi-omics technologies and AI is contributing to that because maybe people think that AI will actually be able to do things or see things that the humans cannot.

I'm not necessarily very optimistic or confident that suddenly AI will solve all the puzzles because the difficulty and the bottleneck is not in the data analysis, it's in the biology that is associated to the data and the data interpretation. The data are so complex that it's very easy to bump into association or something that is predicted by the model to be important, to be reproducible, but there is no biology associated to that and it doesn't make sense because you manipulate hundreds of thousands of data points. AI cannot sort out what is meaningful vs what is not meaningful.

[00:39:05] How would you say virtual cell foundation models are using 10x's spatial data to speed up target identification and drug discovery?

VS [00:39:14]: Eventually they can help, but it will take a lot of time to build them and then to also provide the proof of principle because everything I tell you, which is a personal opinion for example, that multi-omics and AI will not solve the biology puzzle. Maybe they will. Maybe I'm wrong and maybe they will, but we don't have the proof points as of today. If I ask you to cite me one drug, one even interesting target that has come out of multi-omic studies, I don't know any of such targets that actually is a validated successful therapeutic target. Someone may say it's too early. Yes, it's too early. That's why it's speculations and some people may think multi-omics will do better than single omics, but we don't have the proof points.

It's a little bit the same, even more with foundation models being cellular foundation models or whole organism foundation models because today, basically, there are many teams and companies that build foundation models for almost everything and it's completely obscure. It's not obscure the foundation model itself, we know how to build it, we know how it works. It's very hypothetical how much these foundation models will actually help to find better targets. Also, the other difficulty is if you want to build a foundation model of anything that is representative of pathology, you need to include a lot of pathological states in the model because the pathological states will not behave like the physiological state. Today, the foundation models are based on static data in health and a lot in cancer.

They don't represent inflammatory states. They don't represent autoimmune or allergic states. They don't represent neurodegenerative state. I think there is a hype and overstatement of how much these foundation models are universal. I know AI companies that have very strong claims about the universality of their foundation models, but when you look at which data they have used actually to build the models that will actually determine the relationships and how their model is structured, is very limited data in terms of the diversity. They have a huge quantity of data because without the quantity, you cannot build the models but it's very limited in diversity of cellular states. That's a big limitation. Again, it's speculative because no one has today, any direct evidence of how much they can help or not.

The people who tell you, cellular foundation models will transform target discovery, it's speculation. I'm telling you, it's my expert opinion that they will not transform target discovery, it's also a speculation. We don't have the evidence. The evidence will come much later. To actually demonstrate that the target is successful, you need to bring it to the clinic and you need to show efficiency in patients. It's at least 5-7 years before we actually know whether all these foundation models are actually bringing something that is actionable and useful in the clinics.

[00:43:25] Do you think 10x is better positioned than competitors such as Bruker or Vizgen to capture the AI data market based on customer trust and installed base?

VS [00:43:40]: Yes, I think they're very well-positioned and for different reasons, because of the quality and the reliability of their technologies and also because they were among the first in the different areas. They've been able to actually to catch a lot of technological turns and remain at the forefront. They started with one of the first single-cell platforms, but also very cost-effective in data generation, and then they were one of the first to take the spatial transcriptomic turn. Now, they are at the single-cell spatial transcriptomics turn. Why this is important? There is a huge amount of data that has been accumulated worldwide using 10x technologies.

If you want to build the foundation model and continue to fuel it with data, it's easier to continue using the same technology that has been used to generate the already available data, because the first data these companies use are publicly available data. I guess if they want to continue to generate data to fuel their foundation models, they will probably use the same technology that has already been used for the data they already have. This gives a very strong competitive advantage to 10x. At least if I was at the heading a foundation model company, I would try to limit diversity of technologies that are being used. There are so many biases that you cannot control in terms of the sample quality, the patient cohorts and so on. I wouldn't want to add an additional bias that can be significant and that would be related to the platform or the technology to generate the data.

[00:45:50] How much does whole transcriptome spatial profiling – thinking of Atera's 18,000-plus gene panels – improve AI target discovery compared to smaller, targeted gene panels?

VS [00:46:10]: It's almost impossible to address because it all depends on the strategy that is used for target discovery, whether it's more hypothesis-driven and it's really focused on certain processes. If you know those processes are well covered by the smaller panels, you're good. If your smaller panels allow you to diminish the noise and to increase statistical significance, you're good. If your target discovery is purely unbiased without any a priori hypothesis and you're looking for fishing the complete, new, out-of-the-blue type of mechanism and target, then of course you need a broader panel, but you also need the quality of the data and you need the number of samples that will provide you with a statistical significance. It's always a trade-off between the numbers and the quality.

[00:47:14] How would you say customers are currently viewing 10x's overall data quality when compared directly to players such as Bruker with its CosMx and Vizgen with Merscope?

VS [00:47:33]: I think they still view it very favourably, I mean favour of 10x, because of the many reasons we've discussed, because of also the ease and reproducibility of their processes, it's because they've been around for so many years that now, it's really important to build confidence. In parallel to that, any team or institution or biopharma that really wants to start a huge data generation initiative from scratch like it's a new project, we want to sequence thousands or hundreds of samples in one disease, it's never been done. My guess would be that they will compare directly the different technologies that are available at the time they actually start the project. This is what we've done at Owkin and we've published part of that comparison. Because it's published, I can tell you that the comparison was between NanoString, AtoMx and 10x.

We really compared side-by-side, based on different metrics, on adjacent sections and that's part of how we made the decision of going for 10x for our big MOSAIC initiative. If I was to launch a similar, large initiative today in any disease area, I think even if I trust 10x Genomics a lot, I would still go through a pilot study and a side-by-side comparison comparing basically the three that you named. Vizgen, 10x and CosMx. Based on my criteria, my objectives, my hypothesis or not, or exploratory, basically my strategy. Based also on the samples, the tissue I want to analyse, maybe some technologies perform better than others in that tissue or another or the fixation and so on, I would really go through that side-by-side comparison and try to select the technology on a direct evidence base, rather than by reputation or even past publications and so on because all the technologies evolve.

I think 10x is evolving in the right direction and in a very dynamic and efficient manner, but the competition is also evolving. I would want to check whether Vizgen today, how they are performing relative to 10x. Even if the cost, I told you my priority is always quality, but the cost comes second. If I have two technologies giving me similar quality, then I would also negotiate the cost as a number two decision factor. There may be other factors like contracts, different things. If Vizgen for example, is similar in quality, but offers me a much better bargain, maybe I work with Vizgen. I don't know.

[00:51:08] Does 10x's software and informatics infrastructure create a strong moat against lower-cost hardware competitors? How would you say the company's Xenium Explorer and cloud compare to Bruker's AtoMx and Vizgen's pipelines?

VS [00:51:28]: I've not used directly the pipelines because I'm not a data scientist myself but I can tell you, I've had computational biologists in my team for 10-15 years. The pipelines that come from a company, they're useful, but they are more useful to non-expert teams. They're useful to teams that don't have a lot of data scientists and experience in computational data analysis and multi-omics data integration. If you go to teams, companies or academic teams that are really highly skilled, they would actually run their own workflows. They would have their own. At Owkin, we had our own workflows. We've worked with the CHUV in Lausanne, Raphael Gottardo.

They're developing their own and they're publishing them. They are leaders in publishing computational workflows to analyse, to normalise, to remove batch effects to de-convolute the signal. All these things are better done if you know exactly which algorithm you run and what exactly it will provide you. I think it's certainly it has a value, the fact to have all these already available, computational pipelines. It is facilitating the work of non-experts or young bioinformaticians. If you go to teams that are AI experts, they would probably use internal pipelines.

[00:53:25] What's your view on 10x's new Atera programme and its launch? Would you say a customer purchasing pause ahead of Atera shipments is a sufficient explanation for the company's recent instrument revenue decline?

VS [00:53:47]: I cannot tell you because I've not used Atera. I've seen some communications, I've read about the platform and it looks very interesting. It looks very having a lot of potential and capabilities. Obviously, to go to subcellular level in the analysis, it's super interesting but I've not seen data out of the Atera platform and I have not worked with teams that actually have used the platform. I cannot provide a first-hand information. Then about the purchasing equipment, it goes back to the first questions, and it's about the internal strategies of each biopharma, whether they want to be equipped internally with those equipments or whether they want to work with service providers.

Personally, I've worked most of the time with sequencing platforms and service providers, and maybe that's part of the explanation. I don't know. We need to do a market analysis. When technologies evolve very fast, it's not promoting the purchase of equipment because then, you're afraid that if you purchase an equipment two years later, your equipment will be worn out because of a competitor or a new version will come and you will need to switch again. If it's just for one or a couple of projects or very few projects, it's not worth it. That may be part of the explanation.

AS [00:55:50]: Have you heard whether or not biopharma labs are holding off on legacy Visium or Xenium orders while awaiting to receive Atera shipments?

VS [00:56:06]: I've not heard of that, which doesn't mean it doesn't happen, but I've not heard of that, no.

[00:56:14] For an installed Xenium lab, do you think moving to Atera requires buying a whole new instrument or have you heard of any upgrade options? What are your thoughts with regards to Atera's commercial viability and scalability?

VS [00:56:37]: I have not heard anything and this is a very technical question, it's actually a question for 10x people because it's internally. I don't know what they offer to potential customers, if they offer the possibility to upgrade an existing platform or if they only offer to buy a new platform. This is super technical.

[00:56:57] What specific 10x platforms are you currently using or are familiar with, given that you haven't used Atera yet?

VS [00:57:08]: I'm familiar with Visium because that's the technology that we used for the MOSAIC project when I was at Owkin. We sequenced more than 2,000 samples with Visium, so I can say I know the Visium very well. Besides that, no. I cannot compare with the Atera because we only used Visium. We did some smaller scale studies with Xenium. We did a comparison from Visium to Xenium, which is single-cell, and we used Xenium more for data validation, for hypothesis validation. The Visium was our discovery cohort and we used single-cell resolution spatial transcriptomics for validation of some hypothesis. Of course the same programme, if we had to start it today, we would probably start with a single-cell resolution technology. When we started MOSAIC, the single-cell resolution were either too expensive or not sufficiently mature, in terms of the data quality and reproducibility.

When you sequence thousands of samples, you want something that is super reproducible. That makes a big difference. If I was in a research lab and we had a 10 sample or 10 plus 10 plus 10 sample project, I would have gone to single-cell resolution immediately. When you work with five different centres in different countries and you sequence 2,500 samples across six modalities, you want to go to something that is robust, reproducible and you have all the benchmarks already established and you know what is a good quality image, what is not a good quality image or sequence and so on. You also have already some basic workflows to analyse the data that are available that you will build on.

[00:59:35] Over the next 3-5 years, how do you see the spatial biology landscape evolving across whole transcriptome profiling, spatial proteomics and full multi-omics integration?

VS [00:59:54]: Again, as I said in the beginning, I think there is still a complementarity between all these approaches. Even if the spatial goes more and more detailed and granular, it will never make it whole tumour. You still analyse only a very tiny section and a part of the tumour, and so you need to analyse larger volumes in order to get a better grab of the intratumour heterogeneity. Unless the spatial technologies become so easy and cheap because here, there's also a cost factor. That instead of generating a bulk transcriptome, you decide to generate 20 spatial. Imagine you have the tumour, you cut 20 sections that actually represent different layers, different floors like a building, of the tumour. You pile up those 20 stacks to give you an idea of what happens in the whole tumour.

The day the spatial transcriptomics becomes for the same price almost or a little bit more expensive but not too much of a bulk, you can get 20 spatial and with good reliability, then I would anticipate everyone switches to only spatial transcriptomics. The volume, you got to reconstruct it by multiplying the number of sections and the number of speciality per sample. You have 20 spatial transcriptomics per sample. Maybe a direction that the field will go, but it depends on the vendors. Also, it depends on how much they want to make the technology accessible because if you want to generate 20 sections per sample, you imagine the costs go sky-high for now.

AS [01:02:12]: Thank you for that. This was really helpful. I'd just like to close out by saying thank you so much for hopping on the line and for all of your insights. Hopefully, we can work together again in the future.

VS [01:02:23]: My pleasure. Thank you.

Transcription ends at 01:02:30 of the recorded material