And welcome, my name is Naima Spohun, and I will serve as your moderator today. Thank you for joining us for today's webinar, Artificial intelligence based screening and selection of Co formers for Co crystallization. As your moderator, it is my role to ensure that we make the most of your time with us. I'm here today with Mitali Baghrat. Mitali has a master's degree in pharmaceutical science with several years of work experience in the pharma industry on generic product development and filling. With a degree in intellectual property, she previously worked on developing IP strategies. After completing her MBA in International Management and consulting in the pharma industry, she now works at Merck Kgaa, Darmstadt, Germany as a strategic product manager. Her responsibility lies in the field of excipients used for solid formulation drug development, wherein she is leading the digital strategy with its flagship product and Predict. Before I turn things over to our presenter, I'd like to cover a few housekeeping items. At the bottom of your screen are multiple application widgets you can use. There you can also find our reaction button, indicated by the Sums up energy that allows you to give immediate feedback on the presentations, topics, or anything that stands out. All the widgets are resizable and movable, so feel free to move them around. To get the most out of your desktop space, you can expand your slide area or maximize it to full screen by clicking on the arrows in the top right corner. If you have any questions during the VB Now, you can submit them through the Q&A widget. We will try to answer these during the VB now, but if a more detailed answer is needed or if we run out of time, it will be answered later via e-mail. Please know we do capture all questions. You will also have the opportunity to participate in a couple of quick Paul questions throughout the session. I encourage you to take part in these surveys if you're watching this webinar on demand. You can still submit call responses and questions via the Q&A widget. The webinar is being streamed through your computer, so there's no dial in number. For the best audio quality, please make sure your computer, speakers or headset are turned on and the volume is up so you can hear the presenter. An on demand version of the VBNR will be available after and can be accessed using the same link that was sent to you earlier. Lastly, attendees who wish to receive a webinar certification will need to fulfil the criteria of a minimum 30 minutes viewing time and completing 2 Pol questions within the duration of this webinar, and the certificate will be sent to you automatically per e-mail. And please know that right after the VB, now every attendee can download today's presentation. So that's it from my side. It's my pleasure to turn things over to Mitali. Thank you very much Naima for the warm introduction and a very warm welcome to all of you. From my side as well, I'm very excited to walk you through artificial intelligence based screening and selection of Co formers for Co crystallization. Please note that the life science business of more KG Darmstadt, Germany operates as Malapur Sigma in the US and Canada. With that, I'm going to spend the better, better time of the one hour that we spent together introducing you to solubility and Co crystals, giving you a deep technical overview of M predict Co crystal prediction service, which is the artificial intelligence based service. Finally, I'll introduce to you our offering and with an executive summary slide, we will move on to questions introducing you to solubility and Co crystals. Now most of you must be aware or very much would have gone through this kind of situation in the laboratory. Many scientists have got a promising molecule which needs to be formulated, but oftentimes we don't have enough time or even enough drug substance to do so. And plus the fact that it's often very hard and time consuming to find the optimum formulation. The expected cost to develop a new drug is constantly going through the roof, and 30 to 35% of that cost is related to product development. This is mainly due to Lindy experiments being a development bottleneck, substantial API material requirements which drive costs, and the fact that optimum formulations are not always realized due to resource constraints. Now solubility adds a problem to this. Low solubility increases the risk of compounds failing during formulation development. And if you drive your attention to the future distribution of drug substances based on pipelines, you can see that 60 to 70% of drugs are are actually falling into the BCS Class 2, which means poor solubility plus high throughput screening and target oriented drug discovery often result in challenging and poorly water soluble APIs. So 70 to 90% of drugs currently under development are poorly water soluble and thus addressing this challenge of poor water solubility is of extreme importance. Now we have a plethora of products from our company which can be used to enhance solubility. For example, if you look at the purple side of the the flow chart, we have products like Partec CCS, Partec M, several dissolution enhancers, hydrophilic lubricants, etcetera. Plus we have some technologies that we can also offer. Today however, we'll be focusing on the pink side of things and namely Co crystallization in the enabling formulations. Co crystallization is useful for substances without an ionizable functional group as opposed to salt formation. Now what exactly are Co crystals? Co crystals are a combination of an API with at least one different compound combined together with a non chemical solid-state bond. For example, in industry, usually what is happening is if there is a drug candidate, let's assume it's metformin. Metformin has been on the market for many, many years, but let's imagine that you want to create a Co crystal out of it. So there'll be screening experiments back and forth. And finally, after these experiments, you'll be able to figure out that tartaric acid is indeed the Co farmer that you could use for metformin to form a Co crystal. Now coal crystals have several benefits like physics. They have better, they can enhance physical properties namely drug solubility, dissolution rate, etcetera. They also offer better control over physical and chemical stability and may show advantages during manufacturing. They may also help extend the life cycles of drugs using it for cleaning or they may also have several benefits for regulatory in terms of genetics or for intellectual property in terms of genetics for early market access or even in having a new IP protection as well. Add to this that's Co crystals and solid-state form can alter utility and properties of the API and hence the drug product. Several properties are linked to the crystal form of an API and hence the solid-state form. Solid-state screening is already conducted in this preclinical phase or even phase one clinical phase. And this is where most of the pharma companies want to find out as much information about the drug substance as possible. And This is why pharma companies are investing a lot into R&D and to understanding the solid-state properties and potentially to find the optimum crystal structure of an API. These screens should be carried out as early as possible in the early development phases to mitigate and manage risks. We've also seen quite a lot of examples of marketed Co crystal drugs which you see here on the on the table at the top. At the bottom, you also see certain common Co crystal Co farmers and here you'll see a graph which basically just is a reference. It's for Co crystals in the formulas database. The key take away though on this slide is that marketed API Co crystals are steadily expected to grow because companies know the know the benefits that Co crystals offer. With that, I'd hand over to Naima for the poll question. Thank you, Matali. And our first part question is, have you used coke crystal crystallization as an approach to take our solubility challenges? And we have two options. I said no, and we look forward to your responses and we give you some time so that everyone can participate. Have you used Co crystallization as an approach to tackle solubility challenges? Let's have a look. I'm going to give some more time. Still some submissions coming in, so let me give it a few more seconds. Thank you very much for all your responses. That's very interesting. And Mitali, with this I hand over back to you. Thank you, Naima. So moving on from the poll question, we see that selection of Co farmers can be super challenging because there are several experimental screening techniques that can be used for Co crystal formation, example solvent evaporation, solvent assisted grinding, sublimation, slurrying etcetera. After the fact that Co former selection is dependent on API purity, structural features and physical properties. Additionally, comprehensive experimental screening is rarely possible due to limited API availability, lack of resources and know how and in general the time needed. Now to solve these challenges. We actually developed the M predict Co crystal tool which selects optimum Co farmers for your API to form a Co crystal. And we've developed this tool fully in house and trained it using a high quality and extensive experimental data set. It's unbiased and screens a broad chemical space and it's machine and data trained to understand interactions between drugs and Co farmers. Additionally, it's built on a unique combination of computational chemistry related structural and interaction features using artificial intelligence based algorithm and it predicts optimum Co formers based on the likelihood of Co crystal formation. And we'll learn more about this as we dive into the technical overview of how we made this tool. But if you just focus your attention a little bit on the figure that I have here at the bottom of the slide, again, let's consider with that we are working with metformin and UV actually put the structure of metformin in our tool. Then in just a matter of days, you will be able to, or the tool will be able to predict that tartaric acid is the right Co former in this case for your drug or metformin to form a Co crystal. Now, how did we actually build this tool? This is where we entered the technical overview part. Now building this tool, let's say, was a very complicated process, but in order to simplify this process, I've tried to break this down into four steps. The first step is understanding and defining the problem. This is where we identified what exactly do we want the tool to predict. With that we identified what chemical parameters should be included in the model development and we selected the prediction targets and defined outcomes and expectation. The outcome was that we wanted our tool to predict does this combination form a Co crystal, yes or no? And what is the likelihood if it does? So this leads us to the second step, which is literature search and data gathering. We conducted data mining using CCDC and several research articles. However, we quickly identified several issues with literature data and This is why we planned and implemented our own data generation campaign, which was then fed into the tool. A key point that we had here was that data quality is important in order to build a superior tool. Finally, we move on to experimental design and algorithm fitting. This is where we used data science approaches like clustering and active learning, which is an iterative process which guided us to do the appropriate experiments and the model gained new knowledge as well. We started model development already with the first data campaign results and we ensured that a broad chemical space was covered and experiments were designed to maximize knowledge gained. And finally which most of our customers are interested in, we did performance validation of the tool. So not only did we compute the standard statistical metrics, but we also compared the models performance with a test with using a test set. And we did some better tests and the conclusion was that model shows good metrics for a difficult problem and that better tests recommended different Co formers. So I'd like to hand over to Naima for the next poll question. Thanks Mitali. Coming to the next one, which step do you think took the longest time? Is it A understanding and defining the problem, B literature search and data gathering, C experimental design and algorithm fitting or deep performance validation? Again, looking forward to your replies and while I give you some time to do so, I would like to also remind you that if you have questions by now or upcoming questions in the next part of our Baby Now, please feel free to use AQ and a widget to enter your questions. And we try our best to reply to most of them during the Baby Now, but if not, we will do so by e-mail afterwards. We still have responses coming in. I give it some more seconds. It's still very busy. That's nice. Absolutely. We have an engaging audience. Yes. Awesome. OK. Thank you so much. That's a good sign. I think that's very interesting. Thank you. This is an interesting outcome. So let's look into the details of those four steps that we first discussed, right. So the first step was that prediction targets were identified based on the target question for Co crystal prediction. And let's say that we have an API and a Co format which you see here. And what we wanted our tool was a tool to predict was, is the Co crystal likely yes or no? And if yes, how likely is it to form a Co crystal? So in in in other words, the tool was basically able to solve a classification problem. Now based on this, we started conducting literature search and we identified that there are quite a lot of publications that are coming up about in silico solutions for Co crystal screening. This is where we based model development initially on literature data. So we applied an algorithm, we built a model and we tested it using external data set. But the model failed the performance test. Now we knew that our model architecture was good, but the literature data was sort of an issue because it includes hundreds of labs and experimenters. It covers a significant span of history, different sources, different equipment vendors, so on and so forth. The key outcome was that literature data is inconsistent, has a narrow focus and a skewed distribution towards positive results. You can also look at this peer reviewed paper that was published by my colleagues. It's linked on the slide or you can just Google it. So as I mentioned, the key outcome was that good data quality is the key challenge and requirement for a predictive tool. Hence we worked on building a high quality data set. How did we do this? Our aim was to cover a broad chemical space of APIs and Co format structures with specific parameters. That is include all possible diverse chemicals. Now this slide has a lot of text, so allow me to walk you through this. Let's look at this yellow figure. We use something called as clustering. With clustering, we wanted to cover the API and the Co formal space separately, evenly and systematically. Now the this left little, this yellow Oval that you see this model has no information about these APIs which are not covered in the white dots. However, if you check the other Oval where the white dots are clearly nicely distributed, it ensures that a good selection with high coverage is intended in terms of chemistry, let's assume. That the model which has no information about APIs probably only has information about alkenes or alkynes, whereas this one actually has information about alkenes, alkynes, ethers, esters, ketones and so on and so forth. Right. Once we have narrowed down this chemical space to ensure high coverage, we applied a common protocol for the experimental data generation. We use solvent assisted solid grinding because that's what we identified as the best way to find a hit and the Co farmers that were used are included were non-toxic affordable, included in the CCDC database, grass EFS and the FDA in active ingredient list. With this we had 14121 potential Co farmers. We use the same experimental technique and quality concentrations. With this we ensured data robustness because we conducted all these experiments in a controlled setting. Now that we had a lot of data, even after clustering and narrowing down, there still was a lot of processing of data that we needed to do. We further needed a science backed approach to guide our experimental design. So to give you an example, stage one, we started with 30 APIs, 40 Co formers, probable combinations being 1200 data points. With this, our total experimental space was 3256. The large number. This is a large number of probable combinations and it was not feasible to do all of these experimentally. That's why we adopted something called as active learning, which is like a sophisticated DOE, and this reduced up to 80% of our experimental effort, but it also ensured adequate experimental data, amount consistency and model performance. So how does active learning actually look like? Let's say the first yellow square is all your possible combinations. That is 3256. We select 10 points of that. We train the model ensemble on selected data points and then get prediction for these ten data points by running the model. Now, sometimes there'll be a large ensemble disagreement, and sometimes there'll be a small ensemble disagreement. Small ensemble disagreement means that the model is doing well, but a large ensemble disagreement means that the model could learn more. And this is what we did. We took all those data points or combinations which showed large ensemble disagreement and we constantly continued these iterations until model met certain exit criteria which were predefined. So we conducted approximately more than 1000 experiments. And with this, we covered 30% of the experimental space and the model kept on working on the results where the predictions were weak to ensure active learning. And this is where this is where it all comes together. Let's imagine this slide as a workflow of our tool. So we input the chemical structure of an API computational chemistry program converts it into 3D structure and conformers. Now based on these first principle calculations are carried out and cosmologic generates Sigma profiles and statistical thermodynamics generates features. Now features is very important. They are also known as parameters and they contain all information that is deduced from the chemical structure of the API. So we have the melting point, we have the molecular weight, the log P solubility and several other things. Now we have the Co former features separately and the API features separately. What our tool also includes is the interaction features. So how is a Co former actually interacting or behaving with a particular API, right. And this is one of the technical differentiators of our tools. Another important point about features is that once a chemical structure is converted into features, you cannot reverse engineer it back into the chemical structure. So essentially our tool does not really hold the chemical structure of your potentially confidential API. And this is one of the key important points that our customers also ask. Now we have features. Now auto Gluon is used to automate machine learning based tasks and algorithms are fitted and model improvements are carried out. And finally, you have a dossier of optimal unbiased predicted Co formers as your product. With that, I move on to the next poll question, Naima. Thank you Tony. How many potential Co formers has our tool learned from? What do you think? Is it around 500-800-1000 or 1400? Again, very curious about the outcome. And another reminder, it's still time to submit your questions in the Q&A widgets and I see a lot of responses coming in it's. Great to see. Yeah, still very active. I give it a few more seconds to give everyone the possibility to reply and have their thoughts. Very exciting topic yes, Yeah, AI is just so much to learn at the moment and in our industry I think that's very, very special. Thank you very much. Let's. Have a look, OK? Very good. That's yes, that's that's encouraging. So moving on to one of the most important parts of tool development, which is performance validation. And as I mentioned, most of our customers are very much interested in doing this. So we actually checked the performance of our tool and here on the left side you see the graph which is the ROCAUC. Now ROCAUC is a graphical representation of our model's ability to differentiate between classes. That is what is a Co crystal and what is not a Co crystal. We, if you see it at the plot, you see that the industry standard is at 0.61 and we are at 0.72. 0.5 essentially indicates that it's random. That is the tool probably does not have any predictive power. Now you might ask me whether one is possible and yes, 1 might be possible, but for such a difficult problem like Co crystallization, our model is actually showing, showing quite, quite nice metrics. How did we assess the model performance? We assess the model performance by sampling challenging API and Co formal combinations repeatedly from the whole data set. We randomly picked 10 APIs each time from the data set and tested 60 to 70 APIs 50 times. And Please note that the test set was always split and kept separate from the training set. So this graph is essentially an outcome that the tool has predicted from the data that it had never seen before. So the key take away was that our tool shows good ROCAUC for a difficult problem. Based on the graph earlier, you can see that we are actually twice as better versus industry standard and versus alternative solution with an AUC of 0.94. We are we are also better than the alternative solution versus random selection. We are three times as fast. And finally, this slide actually shows certain metrics like standard classical statistical metrics. For example, 83% of predictions are correct even when there are more non Co crystals than Co crystals. And this is of utmost importance because in the real world, there are always going to be more non Co crystals than Co crystals. 73% of predicted Co crystals are actually Co crystals, which means there are less false positives. And 90, they're actually 96% faster than experimental screening. And you'll learn a little bit more into how we reach this number when I walk you through the value case that we built. Finally, an update on the beta test that we did. We screen a commonly used anti diabetic API using our tool and the dossier or the tool recommended 7 unbiased Co formers and we did some experimental validation as well of what the tool recommended. So we did XRD, we did NMR and the API actually formed Co crystals with six out of seven Co formers predicted by our tool. With this it's easier to find a hit with favourable properties for further drug development. On the next slide you'll see some of the Co formers that are too recommended and also conclusion of the insilical outcome being confirmed. Using XRD and NMR you'll get these slides. If you have questions, please feel free to reach out to us later as well. Moving on to how do we actually offer this particular service now we are. This service is basically positioned at 2 levels with optional experimental validation and this is how we aim to bring value to your drug development. At the first level, we have an unbiased Empradico crystal tool screening wherein customer provides a SMILES code and then as a product receives a dossier containing optimum Co formers in a ranked order with conclusion. Now let's say that the customer needs more information and opts for the optional consultative experimental service. In this case, we also do an experimental validation of prediction which has several analytics that we do in our central analytics laboratory. And then you get an experimental dossier conforming a predicted optimum Co farmers with analytics. Additionally, Please remember that our tool is agnostic to the supplier of the Co farmer. But if our product is recommended, we also provide a wide range of Co farmer chemicals that are already in our portfolio. Moving on to show you how we actually provide value to your drug development. So we tried to map of how our, we tried to do a map of how our customers will go through our production service versus the manual formulation testing that potentially customers are doing right now. And I'll not go through the map here. But basically this helped us, this helped us go into some value drivers which have been mapped here. Some of them are like we have a three times hit rate, we have enough labour savings, we are faster to production and the entire experimental space is tested. We quantified some of these value drivers and some of these value drivers that remain non quantified because then that's just a fact, right? Based on these value drivers, we did some assumptions for each screening cycle. These assumptions were generated from in house testing as well as literature review. Primarily important of these would potentially be the chance of hit. As you can see current industry practice is zero to 5% and we are actually 10 to 20% total hits found. Also we are a lot higher and the combinations tested we are also more than 1400 here now. This helped us come to the conclusion that we are actually 60% now we'll actually help you save 60% of your COGS. Now this is primarily related to the labor and cycle time and primary cost drivers include reduced FT time. No need to reduce cleaning or characterization because we are so confident with our tool that in case our tool does not recommend a Co former for your particular API, then Co crystallization is probably not the right API processing technique for your API going forward. We also did some non cost of goods related metrics. Some of them are like the cycle time and this is where the 96% faster comes from. The number comes from with 20 days, let's say we need three to five days for screening with additional experimental or analytical testing we are at 20 days and current industry step practice is a lot higher. So with this we are actually you can actually get your ideal Co forma formulation 96% faster and also see a 99% savings in the cost per combination tested because we screen an entire like a broad chemical space which we already discussed in the past couple of slides. So how does this how does this screening look like? The customer provides us a SMILES code via interlinks or SharePoint, which is a secure and confidential confidential transfer with the CDA. Our tool, which is an unbiased computational screening in about three to five working days is going to provide you with a dossier with a ranked list of optimum Co farmers and it also has conclusion and recommendation. If you want a sample dossier, please feel free to reach out to us to see how our product looks like. We'll be more than happy to send you a sample dossier as well. Now again, let's say that after you receive the sample dose here, you would like to go forward and also get the analytics conducted in our central analytics laboratory here in Darmstadt, Germany. With this, with our experiments, you'll be able to confirm the outcome predicted by the tool. You'll be able to characterize the Co crystal with common analytical techniques and also demonstrate the performance of the Co crystal. And finally, let's imagine that one of our portfolio, one of our products is actually recommended by the tool and then the central, the analytics also indicate that that product is actually one of the good Co formers for your particular API. Then we already have a wide portfolio of Co formal chemicals which you can use as well going forward. Having said that, a key point that our tool is agnostic to the supplier of the Co farmer, but just a point that you have an option also to buy it from us. So how does this business process look look like? We have a customer enquiry about screening an API. There'll be a technical brief discussion if needed. We'll also involve our subject matter experts. There'll be an alignment on the project plan deliverables, the confidentiality agreement and pricing and so on. And once you share the smile score or the API structure with us, we kick the project off. The technical team executes the request and within three to five days, you're actually getting the report that we previously spoke about. And then if you would request certain analytics, this is what we conduct in two to five weeks depending on the kind of experiments that you'd want us to conduct. And finally, you will receive our dossier containing all that information that you actually provide, that you that we actually will provide you. One of the key points which we actually get as questions is the confidentiality and security aspect of things. We divide this into three different three different pillars. The first one is the API structure. We use interlinks ideally as a secure platform to share the structure. The API structures is entered into directly into our tool and then you can delete the structure or the SMILES code via interlinks. Once the API structure enters our tool, it's converted into features and features is not possible to be reverse engineered. I mean, the structure cannot be reverse engineered using the feature. So your confidential API structure will remain confidential. Our tool we come to systems. Our tool is hosted on a secure system in a private environment. Our optimized data and analytics ecosystem is managed by gold standard data governance and regular audits are conducted by our internal audit and compliance team. We make sure that all relevant internal SOP's are applied and system access is limited to core technical user group. Finally, with regards to documents and data, all data is labeled as confidential and the CDA will be signed. So the key take away is that your API structure is going to remain safe, secure and strictly confidential, which kind of takes us to the last couple of slides for the executive summary. So as a summary, our prediction service is an AI based computational screening to predict optimum Co farmers with an optional experimental validation of prediction. With the analytics results via our central analytics laboratory, we help you save time, which means we find optimum Co farmers three times faster than random digital screening and 96% faster than experimental screening. And we help you save resources as well. We are two times more reliable and accurate than industry standard and alternative solutions. Our business process is shows a fast turn around and has a straightforward execution and we are confident that our tool is reliable which steers you towards promising API processing strategies. Additionally, we have discussed key points including data quality. The model has been trained on consistent high quality experimental data set which we generated and discovers a broad chemical space using smart iterative approaches. We followed all standard data science practices and we look extremely. We took a keen look at the security and confidentiality and created a confidential digital environment adhering to all relevant internal Sops with the business process. As you can see here, the business process in short, we have an API structure uploaded to the secure platform. At the end of day five, you receive a dose here with recommended Co formers. And if you want more experimental validation, you receive a dose here with experimental validation of prediction. And then you sort of continue with Co crystal development independently. So get in touch with us if you have any further questions. I'll hand over to Naima. Thank you, Mitali for this great presentation. Now it's time to answer a few questions that have come from in from our audience. But before we do, I would like to remind you that it is not too late to send us your questions now using the Q&A widget. This also applies to on demand viewers. We will try to get through all of them, but if we run out of time, we will respond to you individually. As a reminder, this bibinar will be available on your on our website soon. All participants will receive an e-mail notification when it is available for you. Now back to Mitali, who will start answering questions that have come in. Thank you Naima. The first question that I'd like to answer here is that is the API structure saved on the tool and no it is not. As I mentioned, we have the features which the tool works with, so the API structure will be fully deleted or will not be saved on the tool. You will however need to share the API structure with us or the SMILES code with us. After it's included in the tool, we sort of cannot reverse engineer into reverse engineer back it into the structure. The next question is what is the size of the APIs the tool predicts? The tool can predict anywhere up to 1000 Daltons of the APIs. The third question is how AI based screening is important to predict selection of screening for Co crystallization. Well this is a question that has several sort of parameters to it. The first parameter is that This helps you narrow down the design space because with AI based screening you are getting the outcome already very quickly. As I mentioned before. And let me just bring you back to this slide, it's quick. It's three times faster than random digital screening and 96% faster than experimental screening. So with this, you are actually saving quite a lot on time and resources. And additionally, we have confidence that our tool is actually going to steer you towards promising API processing strategies. It may even predict certain additional Co farmers that you may not have considered before. So I hope that answers that question. And the next question that I have, can you detail the assumptions for the COGS analysis that you presented? I think we would like to go into details of this. Maybe we can reach out to you directly because it's quite a complicated case study and we'll sort of reach out to you directly, if that's OK with you. The next question that we have, let me just does your tool identify also the preferred technique or technology to be used for the production of predicted Co crystal? If yes, does the tool suggest experimental prediction protocol? The first answer is holistically. We do not really consider a technique or a technology that could be used for a production because we just help you narrow down the design space in those initial stages. And then this helps you to continue Co crystal development independently. And the tool does not also suggest experimental production protocols because we are non GXP and we don't really look at the at the later stage of the cycle. Do you predict also the characteristics of the Co crystal such as kinetic solubility and melting point? We may be able to give you certain answers to this depending on the experimental validation or the analytics that we do. But beyond that I think it'll have to be, we'll have to have a detailed conversation with our subject matter experts in the laboratory. Coming to our last question, can your model handle Co crystal prediction of multi components solids such as a salt Co crystal? This is a good question. It's already difficult to predict the Co crystal while we are working on the model also being able to sort of predict or handle salts. Currently our model cannot do this. Having said that, our analytics laboratory has subject matter experts and we can sort of have a consultation with them and they should be able to guide you further based on the Salt Co crystal depending on also the experiments or analytics that they perform. Thank you very much for all the questions. If you have any additional questions, please feel free to e-mail our presenter directly. To register for future webinars or to access our archive webinar library, please visit our website. You can also download the presentation slides in the Take Action field that will pop up on your screen once the webcast has finished. I would like to thank Mitali for today's presentation and thank you to our audience for joining us. Have a great day. _1733307575291

          _1733307575507