 And yes, we are back, and we are back for the last talk of this big thingsie conference at the Attic, so don't go because this is your last chance to enjoy some of the time with us and our two next speakers. So allow me to introduce them. We'll have our COVID stories to tell, who knows when any of the latest released vaccines will be available to all, if ever. In the meantime, it is vital that we tackle the virus as best as we can with the tools we have, and that means repurposing existing drugs. Traditionally, this takes a long time and relies on heavy and expensive computation. Well, our next speakers are taking a different approach. Intriguing, right? To tell us more about it, let's welcome Albert Mercadal and Borja Medende from Fujitsu. Hello. Hi, hi. Albert, Borja, you have the responsibility of being the last speakers of this conference at least in the Attic, so no pressure. And no pressure. You're going to be El Broche de Oro, you know, you are the last touch. That's what everybody will remember of this conference until next year. No pressure. So whenever you're ready, we'll be looking forward to listening to you. And remember at the end, if this time will allow the audience to ask questions. So guys, your last chance to ask our speakers, use the chat. Send your questions as soon as possible because sometimes they arrive when they're gone. Don't wait until the last minute, okay? Let's enjoy our last chat and last talk at this Attic today. Come on. Yeah, thank you. So let's start. So yeah, so my name is Albert Mercadal and I'm heading the Global Center of Excellence in Advanced Analytics for Fujitsu. And with me, we also have Borja. So Borja, can you introduce yourself? Yes, sure. My name is Borja Medende and I'm leading the Digital Annual Team here in Spain under the Center of Excellence of Advanced Analytics in Fujitsu. So yeah, so let's start. So today we would like to talk about how quantum-inspired computing or quantum computing can optimize and may help in accelerating the current process of drug discovery and more specifically in how we can improve the molecule screening process during a drug reproposing for COVID-19 fight. So let's start. In order to start, I would like to give you a bit more context about the problem. So here in the slide, you can see a traditional timeline and drug discovery process which normally lasts for in between 12 to 15 years, as you can see. So starting from the search of new drug molecules, ending by the preclinical and clinical trials until launching the new product to the market. So having said that, specifically all the tasks in the preclinical drug discovery market normally are applied by inefficiencies and anything we can do in order to optimize this timeline, it's great for the industry. Just to give you a figure or a bit of a view of the kind of expenditure that we have nowadays, it's more or less forecasted that by 2024, and this is a forecast before the COVID-19 appears, it's about the levels of investment will be around $200 billion per year from the pharmaceutical industry without, of course, taking into account other research, public institutions and so on. So the discovery of new therapeutic drugs is slow. It can take a lot of time and it's really important and this is not something that is just affecting the industry, the pharmaceutical industry, but it's also affecting the medical profession and, of course, the health of millions of people across the world. And I would say that the need for this disruption is clear or was clear and, of course, with the current COVID-19 situation worldwide, it's even more clear. So here the question is not that it can be, if it can be or not, it can not be, or it cannot be improved, but how it can be improved. So if we jump in the next slide, today we will talk about specifically of the phase that we call design and optimization and during this time we will focus on the design on the screening of molecules. For me more specific, normally this task or this kind of task takes around 15 months and we think that we can lower that or down that into more or less two months. So if we go a bit more in the of the disease-causing pathogens, normally often contain specific proteins that are responsible for the infection. So blocking these kind of proteins is key to stop the any disease. So early stages of the discovery normally involve the identification of small molecules that normally in the industry are called HITs, which normally are capable of blocking these proteins by hiding them or binding them. So the problem here is that there is an almost infinite number of molecules to access with a potential, with the potential of one only and only one molecule will bind the protein that we want to bind. So finding the right one, it's like I would say like finding a needle in a high stack. So Korean libraries for searching this kind of solution are only normally capable of reviewing up to few tens of millions of molecules and existing processes take, as we said, like two years or so. So here we're going to talk about how quantum-inspired computing enables the screening of a lot more molecules, which reduce the HIT molecule search timeline to just eight weeks. So this, at the end, increased the speed of the analysis, means the range of targets can be progressively reduced, leaving just a core of high value candidates significantly lowering the risk of failure, which is also costing a lot of money to the industry. So if we jump on the next slide, you can maybe ask why we're talking here about quantum computing or quantum-inspired computing. So in optimizing the design of neutral for COVID-19 or for other diseases, the pharma industry as a whole, normally, let's say, have to do two tasks, depending on the proposed. So first of all, there is the task of finding new drugs from zero, from scratch, or from existing molecules, which normally can be solved as a constraint-certification problem through a combinatorial enumeration of the chemical space. And then, and it's the case that we are going to go through later on for drug repurposing, we also need to analyze the similarity among molecules, and this is also can be translated somehow as an optimization problem. So having this in mind, anyone could say that quantum computing could be significantly important to face the challenge. So here the question is, by what? So if we type it in Google, we Google it and we Google quantum computing will. The most common answers are quantum computing will never work, or quantum computing will change the world. So of course, as Fujitsu, as technology providers, we really think that we are in the second one, so that quantum computing will change the world. But if you see the Garner curve of the maturity, the typical Garner curve on the maturity of the technologies, someone would say that quantum computing is still in the research phase, even if some providers, as you may know, Google, IBM, and so on, they already have some quantum computers ready for research, but not for industrial processes. So here Fujitsu, as well as other providers, we have created what we call quantum-inspired computing, and at the end it's a digital circuit that runs in a normal temperature, so in a normal data center, but allow us to reproduce some capabilities of the quantum computers today, and of course solve some industrial, some really big problems in the real world. So here the question is why we are talking about quantum, if it's not really, let's say, the physical, the physics quantum that a quantum physicist would mention. So mainly because the circuit or the device that we are talking about, that in our case, or the case of Fujitsu is digital and illar, it's mainly inspired by the three of the main principles of the quantum physics. So first, it's inspired by superposition, sorry, because it has parallel speedup, which means that we can deal with several states in the same time. Second, it's inspired by quantum tuning, which is the capability of escaping from a local minimum and cross a higher point of energy to reach the global minimum. So what we are doing in digital leader, it's what we call an healing process, that at the end it's escaping from a local minimum in trying to reach a global minimum of the function. And finally, because it's inspired by a danger in the sense that how the circuit is created or is designed, all the nodes are interconnected and that allow us or the mathematicians to easy map the problem into the device. So here the question is how difficult it is to work with this kind of technologies. At the end, as you see in the slide, it's a quite straightforward, let's say, process. It's a three-step process. The first one is identifying the problem as any data scientist of the world. And at the end, what means identifying a problem, in this case for a combinatorial optimization problem, means to define the target function, so what we want to minimize or maximize. And secondly, decide or understand which are the constraints that impact this target function. Just to put an example, for example, for the travel salesman, typical problem on combinatorial optimization, that the travel that the salesman can go through the same city twice. So this is a typical, let's say, constraint in a combinatorial problem. Secondly, at the time that we have defined the problem and contextualized the problem, what we do is mapping the problem into a function. So these functions are called like both Isaac models or Qo, and Qo, for the ones that I don't know, stands for quadratic and constraint binaritimization problem. And look likes the one that you see in the screen. And finally, at the time that we have mapped the problem to that kind of functions, the only thing that we have to do is query the system in the cloud, like as a software as a service, and just receive the answer normally in a few seconds. Today, just to give you an overview, for example, in our case in Fujitsu, we have a capacity of 100 k bits in terms of power of computation plus 64 bits for precision of the models that we can run over already. So, and finally, before we turn to the yes case that we want to go through today, and we will take it from there, and we just wanted to give you an overview of other examples in the industry that we are working right now, and we have been working for the last year in order to give you an overview of what this kind of technology can do for different industries or for different practicals. So, just to give you some examples, for example, we have worked together with BBA, the Spanish Bank for portfolio optimization, in this case with the asset management in trying to demonstrate that cube models and this kind of technologies enabled the portfolio managers to build more diversified portfolios, which is quite key for the industry. Then we have other examples, for example, in the logistics of the supply chain world, we have worked both for Toyota and the Japan Post in trying to optimize some parts of the supply chain, most of times reducing the cost by two to five percent, which is a lot of money for this kind of processes, and we also have worked, for example, with several automotive OEMs, so car manufacturers, specifically in Germany, where we have also, let's say, optimized some processes specifically on the production lines, and we also have worked, for example, for designing new parts of the car. And finally, also with the pharma industry, and of course with the King's College of London, that's the case that Borja will start explaining in a few seconds, we have worked, for example, with the Rai Industries, which is a Japan-based pharmaceutical company in trying to find more stable structures for identifying new protein candidates. So that's all for me, and Borja, I let you... All right, Albert, thank you very much. Yeah, let me show the screen. Okay, so tell me when you see my screen, the presentation, right now. Is it okay? Can you see my screen? Yeah, yeah. Ah, okay, okay, perfect. So, well, we have identified two scenarios where quantum-inspired computing may help in drug discovery within the novel drug design problem and in drug repurposing. So, starting with the first one, we have partnered with Polaris QB, a company that joined quantum computing with artificial intelligence to build a platform that relies on digital annealer to accelerate drug discovery. So, this platform delivers solutions in five steps. So, the first one, the platform identifies lead molecule candidates from a diverse virtual library of several billion molecules and assesses their quality. Second, leads are evaluated using the structural information of the pharmaceutical target and a set of physical chemical constraints. Third, the platform leverages an annealing-based molecule filter. Fourth, the output is then refined and ranked with Polaris QB's machine learning algorithms for physical chemical properties and quantum mechanics simulation for binding affinities. And finally, the final output is a short list of high-quality molecules that are prioritized for synthesis and in vitro testing. On the other hand, we also have partnered, we have the drug repurposing problem, sorry, in which we work together with Thorai Industries. Thorai is well-known for its textile, plastic and carbon fiber businesses, but the company is also developing a life science business. In this endeavor, we are leveraging now how in advanced materials for research and development of pharmaceutical products and medical equipment. At the Pharmaceutical Research Labs, they developed three new drugs using digital annealer to predict the most stable structure for protein sidechains. Proteins have various important functions in our bodies, including in the transport, synthesis and breaking down of substances, as well as information transmission. Many drugs are designed to bind to proteins in our bodies in order to control their functions. Depending on the arrangement, they have specific structural forms and perform certain functions. The structural arrangement information of proteins is especially useful in designing drugs that bind to proteins and control their functions. And much research focuses on the structural arrangements. However, it is very difficult to experimentally determine the structural arrangements with measuring equipment such as x-ray or electron microscopes. In order to solve this, we are making efforts to predict structural arrangements through a method in which we calculate the lowest energy state required for protein to achieve a stable structure. In this experiment, we limited the target to sidechains and we used digital annealer to predict the combination that has the least amount of energy out of all flexible sidechains in relation to the corresponding main chain structure. Specifically, we used another method to exclude some sidechain combinations that would definitely not become stable structures. And then we calculated the optimal sidechain out of a number of combinations about 10 to the power of 100. So here we can see that we make use of graph theory because graph theory gives us a 3D structural method that is richer than the conventional methods that relies on fingerprint. We will see more later. As a result, a process that previously took more than four hours was completed in just 20 seconds. And in addition, we could obtain calculation results for proteins that we could not in the past. So given the drastic impacts from the recent COVID-19 pandemic, the search for a treatment is of course on across the globe as you may know. And the School of Immunology and Microbiology Sciences at Kings College London is doing a lot of effort trying to fight the COVID-19. This school is a multidisciplinary facility focusing on innovation and knowledge development in the areas of immunobiology, inflammation and infectious diseases, and is carrying out research into understanding the immunobiology of disease, the host response and SARS-CoV-2 diagnostics. In order to accelerate the research, we started a collaboration with the Department of Infectious Diseases within Sims at the very beginning of the pandemic in order to help them finding a cure for COVID-19. You can see in this picture, part of the team with the principal investigator and PhD in molecular biology, Rocío, at the right. There are several ways of tackling the virus that range from blocking its entry into cells to inhibiting its replication. Either way, treatment is urgently needed. Considering the length of time required for a new drug to be approved that we saw with Albert, repurposing approved drugs is a valuable option to accelerate the drug discovery process. Thus, we are using digital annealer to find similarities among already approved molecules and decide properties for COVID-19 treatments. So as we have seen in the terrarium industry's example, most of the well-known methods for measuring the similarity among molecules used to the molecular fingerprints to encode the structural information, which are efficient in terms of execution times, of course, but lack the consideration of relevant aspects of molecular structures. Thus, considering 3D structural properties of molecules increases the accuracy of results at the expense, of course, of higher computing times. So how do we do that? Well, we first make use of a bioinformatics library named RDKIT, which enables us to easily get the information regarding the molecule. We can gather information about every element, including how are they connected to each other. That is the 3D extractor of the molecule that we have just seen, as well as its properties. So this is easy to see as a graph, like here in the picture, in which the vertices are the elements, the atoms, or cycles of atoms, while the edges are the connections between them. Once we have the graph extractor of each molecule, we need to create a conflict graph. This conflict graph will have information about the two molecules and will have some properties that make it useful to solve an optimization problem called the maximum independent set problem, or in a more general way, the cocaplex problem, in which we have to select the independent vertices among them that maximize the weight of that subset. Solving that problem will lead us to obtain which elements have in common both molecules. So we can measure the similarity between them very easily. So going into more detail, each of the elements in a molecule have some information, like the element itself in the periodic table of course, whether it has explicit or implicit hydrogens or the formal charge of it. And this information is stored in a node of the graph, like for instance this one. Since a cycle in a molecule is very stable and we do not lose information compressing it, the graph representation consider a cycle as just one vertex, as we can see here in these cases. So this also simplifies the graph information. When considering this compression of the cycles of the rings, we can see here that we have three rings together. And then we need to create one node for each of the rings and a new artificial bond between them. Okay. In a second step, we need to create the conflict graph, considering the information stored in each vertex. So for instance, if we look at vertex V1 from graph G1 and VA from graph G2, we can see that the labels are exactly the same. They have exactly the same features. So we can see that having those same labels, we can put together that information into one vertex of the conflict graph. In this case, these yellow vertex, okay, that has information about the vertex V1 and VA. In the case, we don't have exactly the same labels, but they are similar enough, as we can see with V2 and VB, that they share one of the features in the labels. We can also put them in the same vertex of the conflict graph. In this case, in these blue vertex here, okay. The difference between those vertices is that in this case, in this new case, the weight associated with that new vertex will be lower in the mathematical model, in the cube model that we need to send to digital annealer. And we need to take into account that we need to maximize the total weight, okay. In order to create the edges like this red edge or this other green edge here, we need to think about the connections between the vertices in the original graphs, okay. So for instance, considering the vertex V1, VA here and V2, VB, like in here, we don't need to add any edge between them, since we know that V1 is connected to V2, as well as VA is connected to VB, and they are very similar as we have seen before. So we would like to give them in the final solution, that's what we would like to have, okay. In the same way, we add an edge between V1, VA and V1, VB, because if V1 is similar to VA, then at the same time, V1 cannot be similar to VA. So we want to avoid these kind of things, okay. We also add edges in the case of vertex V1, VA and V2, VC with this green edge, because V1 is connected to V2, but VA is not directly connected to VC. So it's kind of strange having those nodes in the final solution, okay. So we wouldn't like to have those vertices in the final solution. All these edges, the red ones and also the green ones, will add a negative weight to the model in such a way that it's not worthwhile to add two vertices that are connected by an edge, getting somehow like a penalty term when they are put together in the solution, okay. So in the end, what we have is an objective function, like in any other optimization problem, that will be to maximize the weight of the selected subset of vertices and some constraints that will tell us which vertices cannot be in the final solution at the same time. So finally, we construct the mathematical model, this cube model in this way, okay. So here we will have the objective function that is just adding every vertex with its own weight, okay. And then we add the penalty terms that will be those edges between the vertices in the graph. Here we put a minus because we don't want two vertices that are connected to be in the final solution, okay. But in the specific case of digital annealer, we need to consider that digital annealer only solves minimization problems, okay. So we need to change the signs and instead of having a maximization problem, we will have a minimization one with a minus in this position and a plus here, okay. So after having this mathematical model, we convert it into Python with a library and then make a call to arrest API, okay. After digital annealer solves the problem, we want to have a similarity measure that will depend on this delta value, okay. Which controls whether we want to weight more the maximum value of the similarity or the minimum one, okay. We have two different values here since we are measuring the similarity between two molecules, so it depends on which side you look at. Okay, so let's play with a demo. Let's see how it works. So here we have a web application that we built for our collaboration with the King's College, okay. This is just a simplified version of the web application, okay. And here we will select several target molecules like Grand Deciby, Phonine inhibitor or Fabi Piraville. We can put it here wherever we want and add it, of course, to the web application. And we select these target molecules because we think that maybe they have some properties that could be good in order to fight COVID-19, okay. So then we have a dataset that is comprised of more than 11,000 compounds. These compounds are those ones that are already approved by the FDA, the Fat and Drug Administration, okay. So here we can put several of them like this one or maybe the Acetyl X-Raptide or maybe the other one is Sophos-Buby, okay. Here in a real case scenario, what we would do would be comparing whatever target molecules that we want with the entire dataset, okay, with an entire set of compounds that are already approved by the FDA. Of course, we cannot do it right now because it takes a lot of time and it's not possible for the demonstration, but it's okay just showing these comparisons, okay. Then we can select a value for the Delta. We usually use this 0.5 and that's because we have two different values of similarity, the maximum and the minimum and we would take the average in this case, okay. If we put a higher value of Delta, we will take more part of the maximum value of the similarity and vice versa, okay. Also, we can filter molecules depending on its similarity. So let's say that we only want to have molecules that are more or less similar, at least 50% of similarity, okay. And then after pushing this button, the start comparing button, what we do is get in all the pairs that we want to compare it. Here we have three target molecules and three molecules from the dataset, so we will have nine pairs of molecules, okay. And for each of the pairs, what we do is get in the graph representation of the molecule, okay, the 3D structure of the molecule. And then we create the conflict graph, okay. After creating this conflict graph, what we do is build the mathematical model, the cube model that digital annealer will solve, okay. And then we send it through the net, through a REST API as we have seen before. And digital annealer solves the problem, okay. After solving the problem, what we have is a set of elements of every molecule that will be similar, okay. And then what we do is calculating the similarity not only with this value but also in a picture way, okay. So for instance, for the first one, having the render severe and the GS620, we have a similarity of 87% that is the average between the maximum and the minimum similarity, okay. So it's kind of similar. And we also show here which parts of the compounds, which parts of the molecules are similar, okay, in this magenta color. So with this magenta color, what we see is the structure that is similar between the molecules. We show this 2D representation but also we make use of a library in order to show this 3D representation of the molecule, okay. So these representations, these pictures of the molecules with the parts highlighted that are similar, okay. This help us a lot, help the researchers at the lab a lot in order to know what are the real candidates, the real potential candidates, because maybe sometimes we can find that we have a pair of molecules that have a similarity of 60% but this is not enough, okay. The number is not enough. In the case of render severe and the acetylhexapetite, we can see that the similarity is not about the 50%. So it makes no sense to show the results here, okay. And then we have the render severe and so forth we will see a comparison that we can see is a 64% similarity and there is some structure that is similar between those molecules, okay. So maybe they are rotated or maybe we can see in the 3D representation that maybe the angles are not exactly the same but at least we know that some parts are very, very similar, okay. In the case of the protein inhibitor and the gS6620 happens the same as before, we do not have a good value for similarity but maybe here like in the protein inhibitor and the acetylhexapetite 3, we have a similarity, a value of similarity that is above the 60%, so it's more or less good but if we see the parts that are common in the molecules, they are not together like in the other ones and maybe this is not a real good potential candidate, right. And for the rest of the comparisons that we would like to have, we can see that there are no comparisons that are above the 50% that we put in the filter, okay. Okay, so having this said, what have we achieved so far with this collaboration, okay. Well, we have validated that this approach of molecule comparison with a quantum inspired computer is feasible, okay, solving the problem with digital annealer. We have very short execution times, thanks to digital annealer and the precision of the results are improved thanks to using a graph model that considers more information than a traditional fingerprint methods as we have seen before, okay. And well, I think that the key part is that we do not only show the similarity between the molecules, we also show which specific parts of the molecules are similar and this is key for the revision of experts in order to know what are the real potential candidates in order to perform some experiments at the lab. So well, let me share with you a short video that summarizes what we have seen in this presentation. The pharmaceutical industry has never needed to apply disruptive innovation faster to find new ways of reducing the $2 billion and 12 to 15 years average process costs. This urgent need for the cures to life threatening diseases comes as globalization has spread diseases at alarming speed. It means any acceleration in pharmaceutical research and value processes could save lives. Imagine being able to cut the early stages of drug discovery process from 48 months down to just seven weeks. Potential molecules can be analyzed and filtered up to 10,000 times faster, making it a reality to go from searching millions of molecular structures to trillions before selecting those that will deliver the greatest chance of success pushing boundaries to be always optimal. Okay, so well, thank you very much for being here and we will be happy to answer all the questions that you have. Borja, thank you so much Borja, Albert. That was fantastic, super interesting. Wow, you know, the future is here. I cannot listen to anything. I don't know if there's a problem. Hang on, hang on. Yes, okay now. I was saying thank you so much for that fascinating presentation. The future is here. Sure. This is it. They're actually asking you a question where you need the crystal ball for the future. Will quantum computing replace classical computing in drug discovery? Well, of course, there is a promise. If you tell us when? There is a promise of quantum computing to solve maybe one of the most interesting problems these days that this is like the drug discovery problem, but we don't exactly know when this will happen. I know. Yeah, because it's just still in development. Just in case you knew, I mean this is, you are obviously early adopters. Sure, we have to follow suit. In this sense, they also ask you about talent because they mentioned there's no really people with the development of the training that you guys have. You cannot just put a bunch of traditional computer analysts into a room and transform them into a quantum computing experts. Well, I would say that no one gets. I would say that no one gets. So I get that anyone that has a research or a PhD, anyone, I mean there is a few people in the world, but there is people that knows about optimization can solve these kind of problems. At the end from a technical point of view, what you need to know is the language that is used in this kind of technology. So we have normally we go to Python, so typical coding method for anything related to data science or engineering, so that's not new. And then if you go to ABM or to other providers, they do have their own language, but normally that's not the tricky part. So I would say the tricky part is the technology itself, so that the whole quantum gates have to evolve more in order to be able to for industrial use and for quantum inspired technology than us, it's ready. I would say that it's more important to have the business knowledge and find out which is the problem that we want to solve, that the technology itself. So I wouldn't say it's easy, but I wouldn't say it's the most complex problem to solve. So it's more about the capability and having the talent, but the talent is there and it's just about there's a lot of potential then for a lot of future engineers and experts. And in terms of scaling, because they say molecular dynamics are hard to replicate, how do you cope, what's going to happen to the scaling of this model? Well, in fact, this is a different approach because we are not looking at the the mechanical part of the molecules. We are only comparing the structural part of the molecule, so it's more or less easy or an easier task than that you are mentioning. So what here is like, as I've said before, it's just comparing some graphs and anyone that has working with the computer science field and know about graphs. So it's not that complicated, it's not the quantum mechanics part. Okay, so there's a lot to do, a lot to learn. The future is here, guys. We couldn't have a much better talk to finish these three days of Big Things Conference 2020 in the attic. You've opened the door to a lot of things to come. We can't wait to see what's happening, what's going to happen. So you better come next year to keep us posted on your studies, on your work. Thank you so much for being with us, both of you, to be the, as I said, the final touch on this fantastic three-day conference. And I wish you all the luck and let's keep us posted. Stay tuned on what Fujitsu does with all this quantum computing new frontier. So all the best. Thank you. Thank you very much.