NVIDIA Corporation (NVDA) Earnings Call Transcript & Summary

Hi, everyone, and thanks for joining NVIDIA and OneAngstrom today for our webinar entitled Transforming Molecular Design. Before we begin, I wanted to cover just a few housekeeping items. More information can be found in the upper right corner of the window. All the windows open on the screen are resizable and movable. If you have any questions during the webcast, you can submit them through the Q&A window, and we'll try to answer these at the end of the event. A copy of today's slide deck will be made available to you via e-mail after the call. Here are a few tips that can make this event as best as it can be. To maximize the quality of the audio stream, please close any open applications aside from your browser window. Also, a browser refresh can solve any -- most any glitches if they arise. So if your audio sputters or the slides seem to be lagging, give that a try first. You can also try opening this event in a different browser if you have any issues. If you encounter any other technical issues today, let us know through the Q&A box, and we'll do our best to help you troubleshoot. Now I'd like to turn the event over to NVIDIA's David Ruau, he's Head of Strategic Alliances for Drug Discovery Artificial Intelligence, to begin the presentation. Go ahead, David.

Thank you, [ Gail ]. And welcome, everybody, to this webinar. It's a real pleasure to be able to host it today. So today, I'm joined with my colleagues, Stephane Redon, who is the CEO and Co-Founder of OneAngstrom. OneAngstrom is the developer of the platform we will see today. It's a computer-assisted drug discovery platform. And also with our colleague, Carmine Talarico, who is Head of Structural Biology in the Exscalate group of Dompé Pharma. So it's a real pleasure to have them both with me today. I will start with an introduction of what is BioNeMo so that I set a bit the scene and it will be followed up by the presentation from, let's say, the fancy part and the fancy presentation from Stephane. So let me introduce a bit what we do at NVIDIA. So NVIDIA in health care has a suite of tools and an offering called NVIDIA Clara. NVIDIA Clara is proposing solutions for the entire drug discovery pipeline, solution for developers, let's say. And that's a bit thematic, we are developing tools for developers at NVIDIA. You see here on the screen, something which is fairly known. It's a bit of a representation of the drug discovery pipeline with the different activities. I'd like to say it's absolutely not linear as a development system, developing a new drug. Myself, I've been 10 years in pharma, it's absolutely not linear. Below, you see also the frameworks. That's what we call it at NVIDIA. Basically, it's a library tool solution, which are there to help drug developers. In this case, we have a framework for genomics. We have the BioNeMo framework, which is around biology, generative AI and large language model, which we'll talk about today. MONAI for medical imaging, an entire suite, open source. FLARE for federated learning, also open source and freely available. And of course, very trendy nowadays, the NeMo, generative AI suite, NeMo being more on the textual NLP side of the house compared to BioNeMo. They are linked. So both are linked. In fact, BioNeMo is built on NeMo. The slides would be made available to -- and you will receive them at the end of this webinar. So let me press next. So as you all know, artificial intelligence has revolutionized drug discovery research, and it has -- I mean, it had a tremendous impact across basically the different vertical and different research discipline inside the drug discovery area. You certainly know some of the algorithms, which are listed on the right. You see the number of paper exploding daily. There is new paper which are groundbreaking. It's very hard to follow what is new, what is the top-notch algorithm you need to use next. And it's a bit why we are trying to build basically a solution around that, to propose basically a framework to use and, let's say, easily leverage the power of this algorithm. So without going into too much detail, you see basically the language of biology explained on the right. So people work with amino acid sequence of protein. They work with DNA sequence. They work with language model. They work with basically structural models, and AI has an impact across the entire [indiscernible] even inside basically as a quantum mechanic approach. So the generative AI approach in drug discovery has an impact on the way we generate new leads. It's a bit what we have seen. And here is an example that is crafted based on feedback from our customers, early adopters, where they build what we call that NVIDIA lab in the loop, it's a bit of a coined term, which allow you to generate a new hypothesis using AI model, let's say, a generative AI model. And you go -- after having selected the best candidate, you go to the lab, you generate your new compound, for example, or do a test of your new compound onto -- as a 96-well plate or a compound assay. You analyze that. And those data feed back into the database and are used by the algorithm to improve itself. And that's particularly the virtuous loop. That's what we call lab in the loop, which is a way of progressing and improving your generative AI assets. What we built at NVIDIA is a suite of cloud APIs, which are there to simplify the access for the computer-assisted drug discovery platform, in short CADD platform. And we display inside the cloud -- API BioNeMo will display different type of models, which are ready to be used, fairly simple or documented, optimize, of course, I mean, that's our core basically, knowledge at NVIDIA. We do accelerated computing after all. So you have model for generating new molecules such as MoFlow, MegaMolBART, MolMIM. You have -- and we'll see a demo of DiffDock for molecular pose prediction, 3D structure generation, like AlphaFold, OpenFold, ESMFold, but also property prediction for protein with the ESM family, ESM1, ESM2. You can reuse some of the molecular generative model inside property predictions. Some are suitable for that such as MoFlow and MegaMolBART. You can even -- we put also into the mix ProtGPT-2 for protein generation. All of that is available through different cloud platform and ready to be integrated in various CADD platform. So it's included in fact into what we call the BioNeMo framework. So BioNeMo framework is the underlying machinery, which allows those API to work. And it's a framework which is enabling users to bring their data, redo the training of the algorithm that they want. The algorithm we provide inside BioNeMo, we provide a bunch of them to start from, but you can also bring your own model inside BioNeMo and start from that. The advantage of the BioNeMo framework, which is a PyTorch library distributed freely as a container, is that any type of job that you do for retraining basically your model will scale. And that's really the biggest advantage of using BioNeMo is that the job will scale across multiple GPU and across multiple nodes, nodes being servers, okay, with GPU inside. There is a bit of a dark art of training, basically doing distributed training, which is like one you see in the middle. When you have your customized foundational model, you can switch into the AI factory and part of the work being fine-tuning your model for a specific use case, generating your predictive model and using it, for example, into your lab in the loop with intelligent experiment, generating your proprietary data asset, coming back into your fine-tune model for being retrained. So you see the offering in front of you there. So you have 2 offerings: you have the API and you have the framework. The BioNeMo framework is for training biomolecular foundational model. It present, as I said, high performance for training at scale. And you see a bunch of comparison here. You see the tensor parallel basically on the left, you have the total number of petaflops you're able to scale to. And on the right, you see the training time in days for the different basically type of model. So you have, for example, here in dark green, you have the 650 million parameter model, you have a 3 billion parameter model and you compare that using BioNeMo compared to using that, let's say, the original training time from ESM from the publication. And you see the comparison time that you're able to achieve. And of course, at scale, that's a bit the principle. You can go to 512 H100, and it go really fast in this thing. So where can you access BioNeMo? So BioNeMo is accessible online. You can request access for it. We are in general accessibility for the cloud API at the moment. You have basically the NVIDIA training time as just as a reference guideline, and I wanted really to show this particular slide at the end, where you can take the BioNeMo framework, which is already available in, for example, AWS, that's for the training approach, and start from there. And that's basically a beautiful way to interact with us. So that's the last slide I had for today for an introduction, really a rush a bit through the intro of BioNeMo into the whole world. And I would be very happy to take any further questions at the end inside the Q&A box. And of course, feel free to contact us if you want more information. And now I will pass the presentation to my colleague, Stephane.

Thank you very much. So I'm super excited to demonstrate the integration of BioNeMo inside the SAMSON platform. So I'm going to share my screen so that I can do just a few slides and then a live demo. So entire screen now, share. So now you could see my first slide. So we are OneAngstrom. And our mission is to help scientists create drug and materials by essentially democratizing access to knowledge and computation. So what do we mean by that? The thing is that molecular design is super hard. So even though lots of people need it to create many different things like classical things in a sense like drugs, vaccines, treatment materials. But even in the future, more and more complex systems, DNA, nanotechnology, complex systems. The thing is that even though many people need it, it is very complex. You need a lot of different expertise to be able to do molecular design, to understand molecular design, to develop molecular design tools, like chemistry, biology, physics, now machine learning, AI, large language models, generative AI. So it's super complex. And as a result, it's a very big mess to work and design molecular systems because of the vast array of tools that you need to be able to combine. And so sometimes you see angry users who suffer because they need to have very complex workflows to do things. So what we're trying to do with SAMSON is help people integrate everything in a unified platform that makes it very easy to do molecular design. So SAMSON is the platform that we developed. It's an integrative platform. What we mean by that is that you can very easily integrate very different tools, data models, data sources, AI models, services, apps, editors, force fields, visualizations. Everything you want, everything you've been developing internally or using from outside, you can integrate in a single easy-to-use platform. So what we mean by platform is really the platform model. You connect the users -- we connect the users and the developers for molecular design. And it's the same internal data model for biomolecules for material science, for anything at the atomic scale. And so it's future proof. Like some people use SAMSON from [ project ] design, some people for material science, some people for nanotechnology for creating nanosystems, it's unified. And the 6 -- maybe 6 unique characteristics of SAMSON are the following. So first, you have an embedded AI that I'm going to demonstrate quickly based on GPT-4 to help you with molecular design. Then since SAMSON is really open, we integrated very advanced rendering algorithms. Then you can also scale -- so SAMSON is a desktop application because it's developed in C++ to make it super-efficient. But you can also, in a few clicks, and we'll demonstrate that in particular with DiffDock, integrate cloud services. Then the thing is that -- another thing that's unique is that you build your solution on demand. You only install, you add only the extensions that you need, so that makes it much easier to control the cost and develop your own custom solution. But the 2 key things are really this, the part that thanks to the SAMSON SDK, you can integrate everything you want, basically, so external tools, internal tools, make them very easy to use. And what's unique also is that you can create executable documents. So I'm going to show you how you create an app, and this app can be anything, again, AI model or something doing complex calculations in the cloud. So we're seeing some traction in many places. Now there are more and more extensions being installed, users and countries. We have clients in 30 countries now. So I'm going to start the live demo. It's always a bit risky to do a live demo. We'll see, but we'll take the risk. So the first thing -- I'm going to go briefly over several things. The first thing I want to show is what you can do about visualization. So you have many different ways to -- so hopefully, the screen does not lag because of the transmission. We have many different ways to represent things. But what's unique is that we've observed that many people tend to use external software like 3D software, like Blender, for example, which are very difficult to use when you come from biology, computational biology, chemistry, when you're a medicinal chemist. So we integrated Blender cycles, the renderer of Blender, inside SAMSON to make it very easy to render nice images and animations. So what I can do is click -- sorry, I click here, and this activates the Blender cycles renderer and then can have the possibility to edit materials like show metallic objects, glassy objects, any colors in just one click. You don't need to export. You don't need to convert your models. You just use the path tracing of Blender inside SAMSON that we integrated. So that's just -- again, I'm going to go very briefly because I really want to focus, of course, on the BioNeMo integration. That's one thing. Another thing that you can do is create animations. So again, a bit like in advanced animation software, you can create animations with different tracks and then control these animations. So you can make molecules move. You can make cameras move. You can make objects appear, disappear, and you control this by having key frames in your animations. And then, of course, you can combine this with the renderer to create movies with data field effects, glassy effect, et cetera, et cetera. So again, that's for visualization and animation. Then we -- another thing that we integrated, thanks to this open architecture -- so what we mean by open architecture precisely is the fact that the core is proprietary, but then you have software development kits in C++ and SAMSON -- and Python, sorry, that you can use to develop extensions. And for these extensions, you decide whether you make them public or not and whether you make them open source or not. And to help you in general with SAMSON, you can use the SAMSON AI assistant, which is based on GPT-4. So I'm going to use the mic here to say what I want. Hi, I would like to predict protein structures using NVIDIA services. How can I do that? So first, I see the transcription. And then SAMSON AI is actually augmented GPT-4. What we mean by augmented is that, first, it uses all the documentation of SAMSON and all the documentation of all extensions to build the answer and give me direct access to documentation on the web and to commands. So for example, here, I have directly the link to the extension. And you see that it mentioned BioNeMo services. And I have also the link to the predict command that I can use to use this prediction service. So I'm going to run that. So to run the prediction service -- so there are several by NVIDIA, AlphaFold, ESMFold, OpenFold. What I'm going to do is just demo ESMFold since it's the fastest. For that, I'm going to select a file that contains 10 sequences in the FASTA format. And I'm going to click on start prediction. And here, I have a pop-up. That's some type of security that tells me this extension wants to use your computing credits, which you get on SAMSON Connect. And so I'm going to say, yes, this is fine. I'm going to describe my job. I have an overview of the pricing. That's temporary pricing. It's still evolving, and this is still being organized. So I click okay because I accept the start of this job. And the job manager appears here, and there is this new task -- let me sort them by date. There is this new task that started. So it's initialized, meaning the input files have been sent to a private secure bucket in the cloud. And now the job is running. So it's going to take a few minutes. I ran it earlier. It took 10 minutes. It took 0.25 credits. That's again temporary, but it's going to be very cheap anyway because ESMFold is very fast. And so I can run this job. And when a job is complete, I can download everything. So here, I have downloaded the files already. And I select this one, I right-click, I import. And I have the imported structure and then I can combine that with, of course, all other extensions. For example, I could want to simulate it using GROMACS, either locally in this computer or in the cloud. So that's the GROMACS interface where we prepare the system, minimize and collaborate, simulate. And then in the cloud -- so I'm just going to download the result of the simulation, let's say, this one -- I'll make it fast. So that's a 1 nanosecond simulation done with GROMACS. I copy the path to the results. And I import them. So here, I have the result of the simulation that ran in the cloud, I could visualize the secondary structure, hide the solvent and play the trajectory. I don't know how smooth it is in the webinar window. So that's what's nice with SAMSON is that you can combine -- since everything has some unified data model, once something is integrated inside SAMSON, you can pipeline your work, you can combine extensions very easy. Let me close this document. So you see that the job is running. Again, it should take a few minutes more. So I'm going to move to the next step, which is demonstrating the embedded apps. So something that is, we think, completely unique is the fact that you can put inside a SAMSON document anything you want, not only molecular models, but basically any file. And this can be research notes. This can be PDF, it can be other SAMSON file or other structural files, really anything you want. And this includes Python scripts that you can edit and run, let me hide this for now, from inside SAMSON. So here, we integrated the Monaco Editor, which is the editor that's part of Visual Studio Code, so that you can develop in Python directly from SAMSON. And here -- so if you read Python, you see that I'm creating a little app that creates a GUI, graphical user interface, with 2 buttons, one to add bonds and one to erase bonds. And these 2 buttons are connected to 2 slots. This one adds bonds to the documents, and this one erases all the bonds. And this, of course, is just a very simple example, but you can imagine that you run advanced analysis protocols on simulations or structures. You can imagine that you can integrate web services. And this is great when you have internally some people who develop and some people who use tools, so that you can send them executable documents. So here, for example, I can just double-click this app and have a security warning that tells me that I shouldn't run everything that I receive. So imagine that this file has been sent -- either put online on SAMSON Connect website or sent by e-mail or put in a drive that's accessible to colleagues. So I say, okay, I want to run this file. And the Python console appears and the interface of the app appears and I now can add bonds, erase bonds. So could just open any structure and modify it. So this means that very easily, you can create an app, send it to colleagues to test. Let's say you're developing a machine learning model and you want to send that to medicinal chemists so that they can see. They can improve the model by checking the outputs of the model. You can make that very easily in SAMSON. So that's the final thing I wanted to show before doing the demo of DiffDock. So this job is still running, but it's fine. It's in the cloud. So anyway, I'm going to launch another job in parallel. So here, I'm going to create. So I click on the DiffDock extended interface, and I create a new docking project. And so what's nice is that we can very easily create batch jobs by selecting folders which contain potentially multiple receptors and, of course, multiple ligands. So I have selected here already folder. So this folder -- just to make it fast, this folder contains 2 receptors, 2 different related structures of enzymes, and this folder contains 5 ligands. And I can preprocess the receptors. So here, I could say that I prepare them, meaning remove alternate locations, ligand, waters, et cetera, et cetera. They have been prepared already, so I don't need to check the box. And I could also minimize the ligands. Potentially, you have downloaded some 2D ligands. You want to minimize them first. So I'm keeping this box checked because I just took the files from the ZINC database. And I'm going to click dock. And again, here, I have this pop-up that tells me the planned cost of the job. So again, this is going to vary. This is just an internal temporary. I can annotate this task. I'm going to say, okay. And a new job is going to appear here, DiffDock initializing. So again, it's the same unified process. You input files. The receptor structures and the ligand structures are being sent to a private secure bucket in the cloud, and it asked me if I want to start the job, I say, yes. And now it started. And you see I had run that before and it took 2.5 minutes. So this varies because it's in the cloud. But anyway, what's billed is known in advance here in this case because it's depending on the number of tests. So I can actually show you -- let's not wait, I can show you the results of a DiffDock job. Basically, it's going to cross-dock all receptors against all ligands. So this is how it looks like, the organization. I'm going to -- so I have pre-downloaded the files already. I'm going to copy the path to the job. And here, I'm going to say load results. And I have the cross-docking with different rankings of all the receptors against all ligands -- so the ESMFold job has completed. So for example -- and they're all ranked by confidence. So if I double-click this, it's going to show me the best score, the best ligand against the best protein. I could change the confirmation. So these are the different poses, and it's something I didn't say. I asked for 10 poses, 10 modes per docking job, per docking task. So I have the 10 here. And of course, I can load other ligands, compare structure. So then I'm going to use other extensions of SAMSON to visualize interactions between the ligand and the protein, do more analysis, potentially use MDAnalysis in Python to develop my own custom scripts, et cetera, et cetera. Here, I can check the jobs of ESMFold. So we see that they have not been downloaded yet. I can click download all, it loads all the files. And I could say that I import this one. Probably I would download it to another document. So the next DiffDock job has completed. So again, I'm going to open all the files and download all the results. So that's the idea. You have a general interface that makes it very easy to use potentially very different services, apps, external, public or proprietary, and that allows you to communicate between people who are in the different places and the spectrum between programming and more scientific like from medicinal chemistry to computational biology, machine learning data scientists. So that covers the spectrum that makes it very easy for people to communicate and work together. So of course, we are just beginning to integrate services by NVIDIA. We want to integrate everything BioNeMo, all the generative AI algorithms and of course, also the different ways to train models -- foundation models to train -- fine-tune models because you can make -- you can imagine that we're going to build more and more complex workflows. So I'll just close SAMSON now. Maybe something that I didn't say is one nice thing with the cloud jobs is that with your account, you can install SAMSON in multiple places, let's say, your laptop, your desktop, workstations. And you can sync your jobs, meaning you could create a job while you're on the move on your laptop in a coffee place and then you go back to your office. And on the desktop computer, you download the result of the cloud jobs, and you can download these. And when you're sure that you don't want to have these files saved in the cloud, you just delete the job, and they're gone permanently. So I'm just going to show the last slide because what we're doing is if you -- we just have this offer for the next 2 weeks. If you create an account -- so it's free to create an account and many, many things are free. It's a freemium model. So they have even a free plan. Many of the things I showed are actually free. But if you want to test the BioNeMo services, we get -- we'll give you 100 free credits if you sign up before -- with a professional address before the end of this month. So I'm going to pass -- I'm going to stop sharing. I'm going to pass the mic to Carmine Talarico, who's the Head of Structural Biology at Dompé. And -- so Dompé is -- we're allowed to say that Dompé is actually integrating many of the tools that are developing as part of the Exscalate project inside SAMSON. But I will let Carmine speak about that.

Thank you so much, Stephane. I'm very happy to be here with you all. Yes, just a few words on the Exscalate platform. The Exscalate platform is a drug discovery platform by Dompé farmaceutici and is basically based on supercomputing and AI to try to enhance and accelerate the drug discovery process. The platform is mainly composed by some complementary modules that combine virtual chemical libraries, protein targets and a structure-based docking software that is called LiGen, and an AI module step allow us to predict toxicity and reliability of the molecule. So the idea of the Exscalate platform is to support the -- in that case, Dompé Pharma to boost the identification of new drug candidates. And as Stephane already mentioned, we are approaching some of these tools and modules that integrates this kind of AI technologies also in our platform to support the drug discovery process. So that's all from my side. Open to any questions.

Yes, exactly. Thank you, Carmine. And we are going into the discussion. So we -- I was trying to answer a bunch of Q&A questions that were popping up in the chat at the same time as Stephane was presenting. But I would start with also the discussion part. So -- and my first question would be for you, Carmine, in fact. Would you see the BioNeMo tool as they are integrated in OneAngstrom platform useful and for your work at Dompé Pharma?

Yes. Well, here, the main point is try to use this kind of technology that BioNeMo integrates mainly for those parts that can be very hard to face. I mean the problem related to the protein folding, for example, can be one of the main keystone in case of drug discovery process that can have not any kind of targets on which you can have, I would say, structure-based drug discovery process. So using this kind of tool that are already integrated in BioNeMo will be [ done ] in the process. So this can be the first aspect that came in my mind.

Thank you, Carmine. It's very interesting. We see -- at NVIDIA, we see a lot of interest on BioNeMo. We populated BioNeMo, as you saw, with a bunch of models like DiffDock, AlphaFold, OpenFold, ESM, but we see also great -- strong interest, in fact, from the community to contribute model inside the BioNeMo tooling. And we released at JPMorgan this year in January the Phenom-Beta model from Recursion, which is a model of a different type, also accelerated on BioNeMo, taking an image of a cell culture and generating and embedding, aiming at the end -- so kind of producing in the good old days, particularly demonstrating the reduction of an image, in this case, and being able to compare vectors embedding -- 2 other vectors generated by the same algorithm, enabling you, in fact, to relate your subcultured image to annotated subcultured image from cell culture, let's say, treated with specific drugs or different type of annotation. And the tool kit is exploding. So we see also a lot of new algorithm popping up for docking. Improvement on the type of model similar to DiffDock is definitely in the news these days. And the interest is strong to use a platform like BioNeMo in this respect. And we see that as democratizing releases, the use of tooling such as that by the wider community. It's not -- I mean, I wanted to specify that. It's not that those tools will replace everything. There are additional tools that are good to have. I mean who could live nowadays without AlphaFold2 or ESMFold or OpenFold to generate additional structure of your own protein quickly and, in fact, almost for free compared to being able to wait or crystallize your protein or to [indiscernible]. So it's additional tool that you have in your belt for achieving your goal. And I was quite happy -- so Stephane, my next question for you, I was quite happy to see the slide, but maybe I think we might not have come to it. There was a slide about the worldwide engagement with SAMSON. And I was pleasantly surprised also to see that there is a wide community. And you mentioned to me, there is also a wide community of developer because SAMSON is having a programmatic interface. You show it into your demonstration. How easy is it to program and develop in SAMSON AI assistant?

Thank you for the question. Yes, it's actually -- I only show very briefly the SAMSON AI. But -- and I said that it's augmented in 2 different ways. The first way being that it knows about SAMSON and it creates in all extensions and it creates custom answers based on that. The second way is that we're trying to make it an agent. And there are various AI commands that I didn't show that you can use to increase the knowledge of the AI, so that it can work specifically on the research papers that you are interested in, but also program for you with a slash script command, so that you can ask for a script and then run it directly. There is also the slash do command. So you can say, do select everything, apply a [indiscernible] model, make everything blue. And this will convert your prompt into a series of commands that will be applied automatically, so you don't have to do them. But at the same time, it will show you which command it has used so that you can learn and reuse them. So that's for the Python side. And on the C++ side, we also have extensions that write new extensions. So what you do is you say, I want to create a new extension that will contain an app, a force field, a new visualization model. And this creates C++ codes with lots of commands that can immediately be compiled, and then you fill in the blanks to do your custom extension. So yes, we did a lot of efforts to make it super open and super extensive.

Excellent. Excellent. And I see a few comments and questions in the Q&A, which we could try to answer. One of it -- I mean, reading through it, there is a bit of a thematic, and you touched on that, Stephane. It's the thematic of being able to retrain model. So it's possible through BioNeMo framework. It's -- you can be trained in a different way or on your own, if you want. The topic is that those models are the model from the open-source community. They have been developed for research purposes and not for, let's say, production purposes. So really, it's a gain. And here, I would amplify the -- I would repeat, not amplify, but repeat the message is that AlphaFold is performing [ so very ] well. But model like DiffDock, model like ESM2 need to be potentially retrained on your own private crystal structure or, let's say, genomic protein structure, so that DiffDock can learn new performers, new distribution. And that can increase tremendously basically the power of such an application. And as I said, those were the first models. There are new models popping up, which have been trained on larger dataset and are made to be -- meant to be in production. And we see definitely NVIDIA in the future into the ability of users to bring their data, retrain their model and create what we would call, let's say, AI asset, for the sake of having a name here. And it's a bit what I think you're creating. You have your own platform coming at Dompé Pharma. You are creating your own model, you have your proprietary asset like every pharma. How much time do you spend on retraining models?

Well, thank you, David. Yes, this is exactly the point. We spent a lot of time to, well, first of all, get high-quality experimental data that can -- we can combine. And we train models in terms of models able to predict toxicity, liability on certain molecules. And so we are in a sort of continuum, in a sort of loop that can allow us to retrain every day the models. They are basically based on the small molecules, not mainly on the protein structure in that case. But I think that you've got the main point that the idea here is now to have this technology as a support and known as a ground truth to make stuff. So this is our perspective. Another point is that given the high quality in terms of new algorithm, new technique, we can test in the next future to retrain our model by using directly, not experimental data, but also in silico, high-quality data. And this can be another very interesting aspect. And this is the direction we are following.

That has been -- I mean, so that would be -- that is something we see appearing in the field that people start having -- retraining their model with, let's say, simulated data or in silico data. That has been, in fact, the trend which started into the imaging world since a very long time. We all knew the original technique of basically rotating the image, skewing the image, modifying, inverting the image, which were the early technique. Nowadays, there are models which are -- I mentioned MONAI, which is our medical imaging platform, which are able to learn a particular type of image and generate new image. Those images are meant to be basically used inside the training phase, not as ground truth, but they are enabling basically the development of the algorithm to great speed. And now it's coming, yes, indeed, into various fields, different from the imaging world, and it's quite interesting. We see a lot also of development into the genomic large language model, interpreting DNA, being able to predict from a particular string of DNA with mutation, what will be the impact on expression for example? And it's quite interesting to see those developments. It's a beginning, but a lot of people start doing research into this particular space, yes. Very good. I think -- I mean, we do have a lot of questions, which are relating to SAMSON, the interface, but also to BioNeMo. Is there a particular comment you would like to make, Stephane or Carmine, based on the comment or the question?

So I've been answering your questions. Thank you very much for all the questions. So there is one actually that just appeared. SAMSON, can we run multiple MD simulations in parallel? Any sign and time scale limitations? What other MD software are integrated into SAMSON other than GROMACS? So yes -- so you have the choice of running either on your desktop or in the cloud. When you run in the cloud, yes, you can run many simulations in parallel. That's an obvious benefit. That's really nice to test different conditions, different [ box sides ], different systems and do all that in parallel. At the moment, we integrated GROMACS. We are seeing demand for material for NOMs, more for material science side. Again, it's open so you can take any existing package, say, quantum mechanics, classical mechanics and integrate them, like quickly make a GUI for that and integrate it inside SAMSON? Or if you don't want to code, we can do it to you. A similar question, do you provide any API to enhance the modules on your platform and/or add customized features? Yes. So this is -- so when you create an account, you go to the download page and you see both SAMSON and the SAMSON C++ SDK. And with this SDK, you can develop any extension, so new apps, services, force fields, visualizations, editors. So an editor is something that reacts to mouse events and keyboard events. So if you want to create an app that creates a DNA's trend based on a mouse motion, you could do that using the SAMSON SDK, any machine learning model, yes. So some of the tools are open source, so you can also directly access the source and modify them and make a custom version for you.

Yes. There was a question also, can BioNeMo be set up to work with [ VIP, Palmer ] and [ Kymera ]? So BioNeMo is very, if you want, very compact. It's not meant to be -- so we will not at NVIDIA basically start integrating, building kind of CAD platform, absolutely not. So within SAMSON, you can have some of the feature. I'm not sure you work with [ VIP, Palmer ] or [ Kymera ], but you have visualization tools inside SAMSON. And that's where you would do -- basically, within those CAD platforms, that's where you would basically start changing those -- basically, step into your development. It's -- any comment from you, Carmine or Stephane, on that?

So yes, I answered this question. I said probably yes, but I don't know how you would do it. I guess -- of course, you could -- the thing with SAMSON is that's redone for integration like with all of these tools to make it easy to integrate. But BioNeMo -- I guess, you could develop a plug-in that would use the API of BioNeMo, but then you would have to do all these things like for visualization of the results, preparation of the inputs, the receptors, the ligands and stuff. I don't know if you have some tools for preparation, for example, in PyMO -- I'm sorry, I don't know enough the latest version.

No, no, with respect to PyMO, I think that given it's very easy to integrate in terms of Python script, that 2 things can be quiet easily incorporate.

Yes. There is a question that just popped up, maybe one of the last ones. We are reaching the end. Any direction towards RNA structure prediction? Is it possible? I guess it's possible in SAMSON to display RNA structure, definitely. But prediction of RNA structure is an open -- is an active research field. I can -- I know that.

Yes. I know that there are some tools that we have not yet integrated inside SAMSON for RNA prediction. Yes, we can visualize them. I don't know in BioNeMo, if you have some...

No, in BioNeMo, we don't have tool for -- so you cannot -- in fact, I don't know if there is a model that exists where you would give an amino acid, so an RNA, basically, string and would compute a structure of an RNA. I would need to dig in little. But we don't have such tool in BioNeMo, that's for sure. BioNeMo will be suitable to train such thing. But yes, we don't [indiscernible] [ domain ]. But yes, it's a hot topic, RNA, since the pandemic. That's really popped up as a target therapeutic area of choice, yes. I think there is a question for you also, the last one that popped up.

Do you provide discounted prices for nonprofit academic research? Yes, we do. Actually, when you register with an academic e-mail address, like a dot edu, you essentially have 90% off on basically everything, and even some things are free when you're an academic. So yes. For computing credits, it's a bit different. In general, there are no discounts on that because like we need to pay the same price to run the service for you, whether you're in academia or in industry. But you can contact us anyway.

Very good. So I think we are reaching the end. We have 4 minutes left. There is a few words from [ Gail ] at the end what [indiscernible]. Great questions.

Great questions.

Yes. Hi, everyone. I just want to thank everyone for joining again. And a reminder, we're going to make a copy of the presentation available. It will be sent to all attendees via e-mail following the call. And also an on-demand version of the webcast will be available in about an hour, and you'll be able to access that using the same link you did for the webinar. So keep that in mind, and thanks, everyone, for joining. Have a great day.

Thank you. Buh-bye, everybody.

Thank you very much.

Thank you. Bye.