good morning everyone uh so welcome today for the block four module 21 uh focus on Healthcare and life science so um today's presentation will be hosted by three speakers three longtime collaborators to the executive master in LA in AI um first we have Mr joanni Bugan so he's the chair of AI in digital medicine at University of M he's also a lecturer at digital Health at University of Le and a lecturer at the University Le brel he leads a ai4 health and at ai4 Belgium and consult on health innovation in AI he um Mr bugti completed a postdoctoral fellowship at Harvard University in 2020 and 2021 under Professor Richard mcnell focusing on Ai and statistical methodology uh then after we have Mr Roman Aldo so he's a senior lawyer at CMS the backer um he's a lawyer specialized in EU and burgon competition law specifically on abuse of dominant uh dominant position cars and Mer merger control with a focus on electronic communication media and farmer sectors he has acquired a deep and concrete knowledge under competition in regulat regulatory law of questions such as dominance assessment price and cost modeling price quiz test rebates and network effects and then the third speaker is Mr tibo Hut he's the founder and CEO at DNA analytics so um he's a Computing science engineer and holds an PhD in MA machine learning from us UCA he also um holds a master of management uh science from the Len School of Management he is the co-founder and CEO of DNA analytics company offering data science technology platform in health care or [Music] R&D bom manufacturing and public health applications so um the first segment of today's presentation will be presented by Mr Joan banti who will be um giving the his lecture U uh from abroad I think and with no further ad du um Mr brigantti the floor is yours okay I was on mute sorry about that um thank you for inviting me once again to give this lecture um my colleagues will have much more complete uh and um overviewing points of view on many many as aspects uh I am a clinician um I do Psychiatry uh as a specialty and today I will mostly focus on the clinical uh aspects and challenges of artificial intelligence so um I was struck uh last year when uh there were problems arising online on the use on chatbots for romantic relationships uh this slides is even more relevant now after uh the appearance of new uh large language models that can uh well um talk and discuss with a human being without virtually no latency uh and uh real time um with a um with a voice that is quite humanlike um well last year uh we had several scandals surrounding different romantic chat Bots um that uh were becoming too romantic uh making um several uh several claims and several um saying several things to uh the uh um the adults were talking to um it is important to to note that last year uh we lost a um we lost one of one of our citizens in Belgium um death by chatbot um after six weeks of the discussion with a chatbot surrounding Eco anxiety that that mental disorder that arises when we focus and stress about the ecological situation of the earth so um because of that scandle I'm uh quite um convinced that as Citizens whatever uh are uh job uh medical doctors or lawyers uh or researchers or Engineers uh our job is to inform people around us and to um be the Defenders of our society I will come back to that so um AI medicine is of course uh the application of artificial intelligence that comes to mind when you ask people uh around the street uh what do you think the best application of artificial intelligence is in our society they will answer definitely medicine uh but why and where are we right now so uh artificial intelligence is becomes necessary when we as doctors have a limited time on Earth uh to um to to do our profession to do our job and to learn things about our job um because our learning is limited by the time that we can give to medicine and the amount of information that we can acquire and we as humans uh do definitely have uh an impact on that AI does not have the same problem that we as humans encounter because it can exploit large amounts of data and gain experience um well the high level expert group on AI defines artificial intelligence as system that exhibit intelligent Behavior Uh this intelligent word has been replaced in layer works with the word rationality rational Behavior because you can understand an artificial model would answer X or why answer uh when prompted to answer that um because we could never uh Define as intelligent A system that can push a citizen uh to Suicide uh that is plainly clear and um there are very few scientists uh that spoke very loudly about artificial intelligence and there are quite nuanced enough to make people understand that what we are living through are certainly amazing times but we we we are never uh in the face of true human uh intelligence um many many times in the Press we we see uh papers surrounding AI versus doctors in which we say AI is uh smarter than Radiologists or AI as modern and Pathologists and uh these are press papers that uh well make things more difficult for us than if they weren't published um first let me reassure you AI is indeed smart but not that smart um in the uh ladder of causation written by judia Pearl um AI stands on the first level of intelligence that is uh Association AI is able to associate X and Y but cannot uh do uh well uh more difficult and more advanced stuff like estimating the effect of an intervention on the world this is something that a toddler can do the toddler knows that if he PO on the side of the toilet and not in the toilet when their parents will get mad and uh that kind of intelligence is uh well um pretty nurturing and pretty Advanced but it is by the time of adolescence that we as humans uh develop something that AI cannot do um and it will take in my opinion uh still um many years to achieve that it is the contrafactual reasoning um it is our ability that we as humans have uh to imagine things that do not exist to create and to imagine parallel universe in which our different actions that we do in one day have different as different effects that is we are able to Lan cause and effects it is the way we as humans do uh and you as lawyers or um experts in in in in in law uh jurists do uh well most of your time establishing cause and effects and and and the impact of that cause and that effect well uh to get back to L uh many many um efforts are now being done to achieve the level three of intelligence uh and to be able to um well do that counterfactual thinking well in in medicine artificial intelligence is um is divided into two main parts um and it is a simplistic separation uh you will see there are many many definitions of AI and Medicine uh I will broadly class uh the AI that can help with the clinical tasks uh into the field of augmented medicine that surrounds the medical doctor and the patient or the nurse and the patient the kinesiologist and the patient the physiotherapist and the p and the patient psychologist and the patient uh to do the job better somehow and to do things that we were never able to do before and the parac clinical AI all the AI that works in medicine but that does all other stuff AI for research AI for uh the the the the bioscience industry AI for administrative tasks for organizations for Logistics in hospitals Etc so that will be broadly in the second domain um well from the clinical point of view um well the domains uh are definitely a lot but these These are eight macro domains that you can if you want remember uh well to Define what is AI and how it is applied in healthcare uh monitoring um can definitely be done um uh well from a distance we can monitor patient to predict uh things that can happen uh based on their records based on their parameters based on what is happening on the moment and this is something that we were not able to do before um because um we um we we we weren't uh able to well achieve uh knowing what a patient is is going through at the given point in time without being there and in belgi specifically uh tele medicine remote monitoring weren't possible until covid um we were doing it before of course but it was uh not remers it was not within the scope of uh our practice but now we can do and it it is definitely a good Advance we can monitor heart diseases we can monitor epilepsy diabetes is a perfect example because it is a high stigma disease um and type 1 diabetes right now is being treated by gold standard treat treatment that involves the use of artificial intelligence um in that case in that sentence um uh Clos Loop Technologies uh in the field of artificial pancreas are now reimbursed and allow patient to Auto regulate the the level of uh of glucose in the blood uh without the intervention uh of external care providers that is amazing uh the diagnostic field is definitely um a very developed um in Radiology in pathology in laboratory medicine uh these are the really uh strong uh domains one learning that we get from Radiology which is the medical imaging field that has so many papers around the ey we learn that almost the uh entirety of literature published on diagnostic medicine and AI does not replicate uh or replication efforts are not being done so for those of you uh who are not in contact with clinical research I think it is most of you uh we as uh medical doctors do not reason uh like the people that come from the industry uh in in in techn in technology and AI uh when uh Engineers develop systems they uh develop something they validate it internally and then they compare it with the gold standard specifically in the field of radiology um and uh in in that sentence in that regard once the algorithm as surpassed the gold standard in the test that the the the engineer does well then they will prefer the gold the new gold standard that is the algorithm to the old gold standard well in in the healthcare uh department and this is why specialized uh firms um that do AI for healthcare are necessary for industry is that uh you need to validate multiple times uh things before you are sure uh that it works in clinical setting and this is what we are lacking nowadays is the lack of clinical validation that impedes progress in artificial intelligence adoption in our Healthcare uh system so what do we learn about this is that definitely we need to start um studying the true impact and true validation of Technology uh what's staggers me is that uh many systems are adopted in in hospitals without clinical Val without beforehand clinical validation and so that this means that your parents or your children uh are being nowadays maybe treated in part with systems that are not validated in clinical practice of which we do not know um similar uh very simple um numbers like specificity sensitivity F1 score etc for uh for tasks that artificial intelligence systems do in healthcare and this is a truly important thing um prevention um we are not good at prevention we need to get better and um the use of AI system can definitely help in to induce lifestyle change of course in the context of the AI act and I believe my colleagues will talk about that um it is becoming more difficult to induce change with an artificial intelligence based system because the risk can fastly become unacceptable and uh of course inducing meaningful change is the only way to do prevention um the wh as released a couple of weeks ago a model called Sarah that is a large language model that allows for prevention in the general population uh by providing information that can induce lifestyle change so I will leave you to to that um well many uh many Technologies exist one that I give extremely importance to is uh the one that reduce the administrative burden um on the healthcare provider 60 to 80% of uh a healthcare provider a healthcare professionals work uh is focused on administrative tasks we did not do medicine for this we did not do nursing for this and so um many Physicians many nurses do experience burnout from administrative tasks it is the leading cause of burnout and uh one of the major causes of burnout is the early uh idetification informatization of uh Healthcare specifically with electronic health records the softwares that we use that um uh that are similar to what you use to to deal with your clients uh well to follow up on patients and that course uh are leading cause of burnout so the question the second question that I want to ask today is um where will will be in five years so I get asked this this question uh at each conference I give at each talk I give and uh since last year the answer is I don't know maybe I'm not expert enough maybe I'm not knowledgeable enough in the field uh but I do not know what will be in 5 years I'm nowadays incapable of predicting uh what will happen the next day I will wake up up uh surrounding my field uh it is definitely um a staggering feeling uh an une easing feeling but one that I have to deal with is that the the the world of AI in healthcare in world of AI in general is advancing so fast that I'm not capable anymore of predicting Trend but what I what I can predict that if I were in implicated in in most AI development in healthcare there will be three trends that I want to talk about the first trend is specialized Health AI so uh large language models are definitely taking Healthcare um by uh and and are holding Healthcare in their hands many many people are um hoping from large language models that it can improve Healthcare what I what I'm sure of is that in the next five years we'll see and we are starting to see it right now by the way uh Sy systems that are able to integrate multiple source of information and helping the clinician uh take decisions in ways that we couldn't do before um and this is kind of a um and this is like autopilot on steroids for for healthcare um this will allow us to do many many things that we were weren't able to do uh before or to do things that we already do but much much better this is the first Trend that I want to talk about the second trend is the augmented patient um I talked about Sarah chatbot um I'm afraid that as patients will get access to more and more technology that will let them have a better health that it will become less and less accepted by our society in which many many things are becoming less accepted that one can have a bad health and one one is destined to have a bad health condition throughout their lives I'm I'm pretty pretty sto pretty um um well uh pretty stoked about the potential of artificial intelligence but quite afraid of that Evolution as will get more um potent it will get more strong it will get more uh performant we will get systems that basically will make individuals uh more likely to be in good health so what happens with those that are already having difficult lives and they will not be able to comply with the new augmented patient error that is an open question that I leave to you and the third trend is old rest and uh tiut is is doing amazing work with the industry uh um with spearheading really specific Healthcare U uh Healthcare AI approaches uh with the industry uh well uh we are now being able to discover new uh medication in fields that were abandoned from a pharmacological point of view um antibiotics that will be needed in the future Health crisis uh are being discovered new candidates um as antibiotics I hope from my own field Psychiatry psychopharmacology has been left the same quite for some decades now I'm hoping uh that the industry will reconsider their abandonment over our field um to ReDiscover new medication that can help our psychiatric patients get better and so in the field such as mental health infectious diseases uh but also uh old age medicine in which we can do new things uh well AI will help us discover new things in research AI can help us discover new research hypothesis at a break neet speed uh I use them myself to generate new hypothesis surrounding surrounding mental disorders to modelize mental disorders in uh in patients but in any case in any discipline AI can help us uh achieve new things in research but also Logistics and organization AI can help us definitely definitely achieve uh better outcomes in in the management of healthare organizations um dealing with human resources dealing with um patient flows this these are all uh truths today they exist very very well a lot of solutions exist I believe that as Citizens uh and especially healthare professionals our role is to defend Healthcare in the face of the unknown we are faced uh with the unknown and I will ask my colleagues of course afterwards tibo and if if if they are maybe less lost than I am in in this field but I'm definitely lost but what I can do as a healcare professional as an academic is being the voice that defends Healthcare uh in this world in which we do not know what tomorrow will bring especially in the field of AI I feel that that is our mission and people like the speakers that will follow me are definitely well placed to be that Defenders and to inform our politicians to inform the general population about what AI brings and what AI does not bring and what we need to keep defending our system I'm also uh feeling that the AI field is going to head towards a replication crisis I I I did it I I said it last year uh in the same course I I'm I'm feeling it right now I'm seeing uh I'm seeing the number of papers um in the in medical resure becoming increasingly uh uh well um retracted we are we have never been retracting more scientific papers than we are today an increasing number of Buddy of knowledge is becoming retracted because the results do not replicate and in the field of AI and Healthcare this has a direct impact on patient care this has a direct impact on patient outcomes so we need to be more careful about what what we're doing with these systems we need to be more accountable we need to be uh more scientifically thinking when we try to adopt uh these these systems into our Healthcare environment the third question I want to answer today is are doctors endangered species because I get asked this all the time of course not of course doctors are not an endangered species of course we will not be able to replace nurses I was reading an announcement a couple of days ago uh or maybe it was a week ago I don't know this field grows so fast and is impossible to locate this news um an Hospital group in America wanted to replace a number of nurse practitioners with AI assistant-based practitioners this is nuts we cannot replace Healthcare professionals with artificial intelligence and the people the managers that think that we can replace healthare providers e professionals with artificial intelligence based systems with our little human that alongside artificial intelligence will be the one that validates what happens do not know anything neither about artificial intelligence or about Healthcare and this staggers me it is it's becoming an increasingly debatable question will AI replace doctors we are losing our minds with this question we should not be asking that question it is not a question that is worth to be asked but the doctor's role will change their General Practitioners will all the new uh technologies that we have will have much more um uh bandwidth to be less insecure and to be less uncertain about what the patient in face of them has definitely our way of providing augmented care will make the general practitioner the leading figure in in healthcare and I believe that will get that many much deserved uh role while the specialist we as Specialists will have to broaden our Horizon and being able to face the highly Specialized Care that this shift will entail some points of attention future prospects uh for us to discuss of course uh I will make my presentation uh a bit shorter because I want that discussion with you about what what uh what worries you what what questions you about that uh because unfortunately I w't won't be able to stay the whole day and listen to my colleagues um uh because I I would much want to uh yeah let's stimulate the discussion in in in Belgium we we have the potential to be the best in the world when it comes to Ai and Healthcare and those who do not think so are not enough in touch with the amazing ecosystem that we have in Belgium we were the first to make uh a federal uh so Secretary of State Matthew Michelle was the first to propose uh a um an adoption plan a a a a federal plan for artificial intelligence this was this slides kind of uh resume sums up the the healthcare thing that we we've been doing for a couple of years now in AI for health improving uh the health of the population increasing the skills and quality of life at work for professionals um sustaining Health Care Organization stimulate research and development and um and and pilot uh Public Health policies so Belgium has a unique ecosystem in the field of AI and health I will not stress that enough it is the best country to do Ai and Healthcare without a lot of financial support because we do not devolve a lot of support uh towards AI development in this country but the ecosystem is amazing I love working with these people we have hospitals that are champions of clinical studies champions of that worldwide number of clinical trials done in this country is phenomenal for a country of 11 million people we have the strongest farmer representation in Europe we have unities only from the French speaking side more than 700 researchers do artificial intelligence and I estimate that at least a third of them does um healthare this is a lot for a country of 11 million people we have more than 300es uh apart from large companies uh Giants and Consulting three 300 and pluses that do Health AI healthare Technologies or digital Health this is amazing and we need to support our companies uh in many ways many many ways we have um the mutuality that uh are really interested uh into the field coalitions that are up and coming in any uh kind of level of governance in this country they do uh interest in AI and we have clusters and competitiveness polls that are investing massive amounts of public money uh into AI projects this is amazing so uh I will skip the regulation because um um my colleagues who talk about it just a comment on the uh the Medical Board in in Belgium the current advises are not enough to company our physicians and so uh Works uh that have pred see that your class uh and of study works for the for the same class that you're following uh tackle that advises they are clearly not enough for the current framework and they need to be reinforced given the more recent advances in the field but I will let my colleague speak about that uh we were the first in the world to do two national Nationwide studies on the adoption of AI uh in our Healthcare System uh we found out that there was a lack of Education and a lack of involvement in healthare professionals and the priorities were uh quite um well um evident and we use that priorities to form the national plan for artificial intelligence and this is why also uh since 2021 we have been very focused in uh offering education to healthcare professionals uh on the field of artificial intelligence we are leading we are leading the sector in that regard there is no another country there is no other country in the world that is doing more efforts in education of he professionals than Belgium we are the Pioneers in this field um doctors want to improve the the the the how quick how fast and how reliable decisions we make are they want to free up time for other value tasks they want to improve uh the way we interpret results for uh Imaging or for biology studies and they are afraid that AI will dehumanize H care now you will have understood by my presentation that with minimal amount of Education we understand that AI does not dehumanize Healthcare it rehumanize it if using correctly that was my uh last message for today and uh my uh my closing remarks are that the Belgian ecosystem is solving every problem that we have with artificial intelligence in Care One by one without a lot of money without a lot of support but we have amazing people and the two speakers that are following me today are two of those people that are making the field advancing in in Belgian um I also do believe that every one of you has an important role to play into the adoption with our values with our Belgian values uh of artificial intelligence and Healthcare and of course um um everyone has a role to play and I'm counting on you and I'm available to you if you want to discuss this further and now with the questions thank you so I not know if there are any questions uh I did not access the chat uh okay I do not see any questions um and of course uh you are able to uh contact me at any point in time if you have yes there is a question from the audience okay I don't think he can see us but anyway um no it's not really a question um um I I'm one of the students and I am probably the only healthc care professional in the team I am Valen from the European commission and I would like to well to to Deep dive in in AI in healthcare is our main role now in in D connect and I will talk to your your colleagues now but I would like just to let you know that probably we will be in touch on the clinical field because everything that you said is quite relevant um we already have other much more uh clinical examples on what where AI is going now but you are right that we need before they they they get into the clinical setting they need to be very well um tested and it's not the case so we probably I will contact you for that and H just to let you know also we are leading the European cancer imag initiative so Radiology of course as you know is one of the main areas that where II is going so fast and oftalmology in in and dermatology when it comes to Imaging um probably you didn't mention the virtual human twins but um we are also following that very closely in my unit and the third one is the genomic where I will focus my dissertation um especially in pharmacogenom and uh as you said the in the Pharma field there is a lot to do but as you also know pharmaceutical companies are quite interested in the AI settings especially developing new drugs so than thanks for all your your inputs are very relevant and especially I'm happy that you come from the clinical side and you have the the view of the of well of the doctor and of the clinical community so thank you very much and we will be in touch with pleasure thank you for your comments I definitely agree um any additional questions whether from the oh yes uh Mr Roba has something to say thank you jaavan and thank you for this enthusiastic presentation of the the Belgian landscape it's uh it's it's good to hear some good news sometimes so it's really good uh but I was more a bit touched by what you say about the augment get patient is it something that you feel that you already see some because as you already pointed out there are some people who are sick I would say by Design people r a disease and so there's nothing to do uh and so do you already see this kind of you perceive this kind of of gap between patient and augmented patient I I I see well the Gap already exists now I have the the the younger patients that come to my to my consultation uh that have say a bipolar or psychotic or depressive disorder uh and that have um digital Health apps that they use to they monitor their their things and and those patients are definitely more educated to their disease and they are more able to predict when they will go bad uh compared to older patients so the the digital Gap is uh impacting directly patient care right now and and I see it with my own patients now what I'm afraid and uh I know that you work with genetics and um rare diseases these this field uh there will I I I'm afraid there will come a time in which a choice will be done do we invest a lot of money to treat a few people that have rare genetic disorders and let us not Advance the research and uh to try to treat them and or focus on the good side of the population because what we will exclude if we lose our values of altruism and social accountability for healthcare um if we lose them we will lose patients with rare diseases uh patients with so difficult socioeconomic conditions uh patients with mental health disorders we we will lose them we'll not be able to get them back but uh we are seeing some countries uh adopting political measures of inducing lifestyle change through digital L apps and my fear is that it starts that way and afterwards we'll start to measure those who do not improve their lifestyle and I'm afraid with the current political landscape that the Far Away We Go the more likely we are to adopt extreme position that will not benefit the health of the population and we will exclude the P we will forget those who are more easily forgotten that is my fear but that is a fear right now uh but I do observe the digital Gap it already exists there are augmented patients and there aren't augmented patients and like doctors the augmented patients Fair much much well that uh the non-augmented patients thank you J any additional remarks or question you would like to address to Mr dvan okay well um well thank you very much Joan for joining us today online um well if any students have further questions obviously um they can contact you and I think you'll be quite happy to answer their their queries as always thank you so much for having me and uh good luck to my colleagues uh I'm sure that their presentation will be very insightful have a good day everyone bye-bye yes thank you very much javani bye um okay maybe let's take a short break of um five minutes before we can enter the second segment thank you e e e e e e e e e e e e e e e e e e e thank you very much and thank you first of before everything uh to all of the people who are present in the room uh thank you for the four of you because it's kind of a weird experience to speak in front of an empty room and people are electronically far away from here but thank you very much to all of you I hope it will be quite interactive and that you will have a question and that we will have a kind of of dialogue so um let's begin so thank you for the the first introduction indeed I'm a lawyer indeed I'm focused in competition law indeed I I have a specific focus in telecommunication and Pharma but also and it's why I'm here today I for some personal reason I develop a kind of knowledge intensive knowledge in the field of genomics and with that a member of something called the vas which is the European reference Network for card cardiovascular disease and I'm a member of within the AR of the Les group that is a group that is supposed to to help the r to interact with other stakeholders and I'm a member of the Belgian mirror group for the 1 million plus genome initiative which is a brilliant initiative at the European level and here in Brussels I'm member of something called The genome for Brussels Consortium and here today I will show you some of the achievement that we that we did with this the Fantastic Team of the genome for Brussels and uh last but not least I'm the founder with my wife of something called the 101 genome Foundation that um aims to um use the progress of genomics to help people with our disease to have access to better treatment so um quite early in my career I've been exposed to artificial intelligence in the field of competition law where you know that there's a complete discussion about possible collision between operator using artificial intelligence but it's a question that we will not uh deal with today today we will focus on genomics and perhaps all of you are specialist in the field of genomics but I'm not and so uh I will come begin with the beginning and some biological explanation and for those of you that will that are not specialized in this field and so the first thing that you discover when you discover genomics is that the complete Human Genome has more than 20,000 gen 20,000 genes that are you know 20 23 pairs of chromosome and there you can find the complete DNA the complete message of uh the genome you know it you know the difference since the co time everybody is a specialist in this field the difference between the DNA and the RNA you know that within the cell the core of the cell we have the complete copy of the complete DNA and after we have reproduction copies of it that are produced to help to produce and code the protein that we use and this copy are called the RNA we know that the RNA only uh represent 3% of the complete genome you know that the complete DNA it's written in nucleotid actg you have these four letters and with these four letters you can write the complete sequence of human being and this complete sequence it correspond to an alignment of three billion nucleotid and this sequence this alignment it's at 99% identical for every human being she really we have this idea that we are brothers and sisters huge family when you study the genome it's what you see you see that we at really high percentage 99% of the genome we are really similar but this little difference of 1% it make us what we are not only the genome but this difference is specific to uh uh all of us so this is what you see and in this kind of ocean of nucleotid it's quite difficult to isolate this little difference of 1% and uh this however it's really important because within this difference of 1% It can help you to understand why some people are sick and some other people are not and also it can it can have a crucial role when you uh when you have to define a diagnosis but also in the in the field of disase but also in the field of cancer where to have to you have to be able to identify what are the modifier the modification the mutation in the in the cells that are at the origin of a cancer but this classification of the pathogenicity of a variation of a mutation in the gene it's really not easy because sometimes you have some kind of mutation where the message of the gene that code for the protein will not not be lost for example in the first sentence you have o instead of a u and you can still get the message meaning that the message that the gene translate will be correctly communicate and the protein will be communic where we should be um properly uh produced on the other uh another circumstance you have variation mutation that will stop the production of the protein or where the message will be totally lost you have one letter that change and and the you have change in the sequence and the message of the pro then the message it it's lost and you have the third situation where you have what we call the variant of unknown signification you don't know what will be the consequence of a variant on the message that it it's carried we have here we have D is all L can ask for you don't know what will be the quality of the um the protein that is produced and what's difficult to understand is that not all mutation are deuse are pathogenic all all of us we are carrier of thousands of mutation that are specific to us but are that are not affecting our health and so it's why this question of the classification of the pathogenicity of the mutation is crucial in this field and it's crucial you know it when we speak about breast cancer you know that some women they have and some men actually they have some specific um exposure to some kind some some form of breast cancer due due to some specific mutation pathogenic mutation in specific Gene so it's really crucial to isolate and be able to characterize the pathogenicity of the mutation to POS the diagnosis and the researcher and the clinician they can use these kind of tools where they can try to identify if a variant is pathogenic or not pathogenic and after when it's done it also it helps you to pose the diagnosis but it also helps you to define the therapeutic approach that you will use with a patient for example still in the concept in the breast cancer if the tumor it's positive to pathogenic mutation you will select specific threatment you will select specific drugs that will be specifically Target this kind of tumor so it's crucial for the diagnosis but it's crucial also for the therapeutic indication that will be used and so when you have a lot of disease there are a lot of diseases are what we call monogenic meaning that they are caused by one mutation on one specific Gene and we believe that the impact of this mutation is so strong that it at the beginning that it it create generate the disease but more and more we discover that disease they are not only monogenic but they can be oligo polygenic meaning there are the combination of variant of pathogenic variant not in only in one gene but in two or more Gene and begin more and more difficult to identify those different uh those different I would say network of genes network of combination of pathogenic variant and it's something that's well known and here you have a um quote from somebody called Nim TB that I quite like where he explained that you have this kind of curse of demential that when you are studying the genomics with this first monogenic approach you can have some kind of and it's already an extreme Challenge but as as soon as you pass by above the the monogenic level and you read oigo the polygenic level you have this dis Humanity that that makes things even more complex to identify the networks of genes and it's where the human beings and the traditional way to work it's limited and it's where uh and it's why a lot of disease even if we have a deep knowledge in monogenic approach there are still so many things that we don't understand that's why when you have this communication about the P the the the kind of Future Vision of what we will be able to do with the selection of Gene and the the choosing for you know the building a child this kind of thing it's really really really far too early to speak about this thing because at this time we are still in front of a galaxy that we begin to discover and we are we are not so so far to to be at this level so what will help us is what we call the genomic Revolution and the genomic Revolution it's before everything a technological Revolution it's that before in the past traditionally uh when people were trying to identify the the pathogenic variant the research they were working in the labs and they were working in on Amplified genes so they were working on one or two genes and so meaning that they were it was impossible for them to have a kind of vision of the networks and the interaction between the genes and thanks to this Evolution we are passing from the genetic approach which is which is a study Gene by Gene to a genomic approach which is which is a a study at a whole level of the whole genome and for that we use what we call sequencing machine and those sequencing machine they're able to translate the biological DNA the biological RNA into a Digital Data okay this is what the sequencing machine are doing and so you get it when you are talking about uh RNA we are talking about the exome and so we are sequencing 3% of the complete genome and when we are talking about the genome we are talking about the Ws and we are sequencing the entire genome so doing that uh it produce uh what we call the VCF file and VCF file and on different for format you have the fastq you have the BAM you have other new format but at the end of the day the researcher they're working on a VCF file and for each individual individual person each person the genomic information it represent around 300 gigabyte of data for each and every single person so meaning that the translation the translation traduction of your biological DNA into digital DNA it correspond to around 300 gigabytes of data and so a lot of people they get quite early that the way you gather and you you conserve this data that you get it at 99% it's identical to every other human beings so you need to store it to keep it for a long time for some part but there's another part that you need to be able to explore and uh to work on to have a kind of cold and hot storage and it will be exponential the idea is that we will gather more and more and more genome so more and more people and we can begin with Dr geraline Vora who a Belgian scientist who was working at the the broad Institute of the MIT in Harvard she came with this idea really early that the genomics revolution will go with the cloud gen Cloud Revolution so it's why she she wrote this book Gen genics in the cloud and she anticipate what we can see today is that more and more scientists studying genomics are using Cloud solution and we will see more and more genomes available also because there is a price decris at the beginning the first genome sequence that were really really expensive and as of today we are 100 we are around 700 EUR something like that there was some announcement of genomes at 200 euro per genome sequence but it's really optimal situation and are really not in the kind of average scenario so we can say that today we are around at 700,000 € 700 Euro for the sequencing of one human genome and why is it important to sequence The genome it's important to sequence The genome because each genome sequence set with phenotypical data will allow us to kind to decrypt this Rosetta Stone which is the DNA which is the genome because as of today we still don't understand this language we understand some part of it we see some part of this galaxy but we don't have a complete view of it and so each time we have a new genomic genome data a new Digital Data we can help to have a better understanding of the Genome of the complete genome and and we can go with this in the direction of the real personalized medicine that you you heard a lot about it so the question it's really to understand the interaction between the genes within the Gene room and why is it so important here you have an example of a study which is called the resilience project and it's a study that I I like a lot it's a study uh that was conducted on uh let's say 600,000 genomes that there were already gathered and sequenced in other context and a group of researcher they say okay we have access to this data and we will try to reanalyze those data and see if we can identify in this data set the pathogenic mutation of disease that appears during the infancy and they are deadly meaning that there are pathogenic variant known to be causal causative for a disease and for lethal disease in infancy and in this set of 600,000 adults in reasonably good health they identify 30 people harboring pathogenic variant people that should be dead and are not dead meaning that this this is something big there is something in their genome that say if we have only the monogenic approach we say those people they should be sick and they are not sick so why it is working this way it's working this way because you have this network of genes and that we see more and more that within the human being within the human genome you have the action of what we called modifier genes what we call epistasis genes there are genes that can have a protective effect if even on a Del use pathogenic mutation on the genes they can compensate this effect and make the human in the case of the resent project you were in front of those um person who are not even presenting the characteristic of the of the disease they should be sick affected so there are three uh different scenario with this resilience situation so perhaps there was a they are not sick and there was a mistake with the class ification of the varant perhaps they are sick and they don't know that they are sick but it was highly probable in the in the the situation or problem they are not sick and they should be sick and it's because there are those modifier the problem with the resilience project study it was conduct on the court of people who accept to share their genomics information but the Rec contact was not possible meaning that you have a complete set of data and if you complete and you gather genomics information without being able to have a dynamic interaction with the members of the court it's useless like it was in a residence project you just know that there is something but you can't go back to the participant and ask them specific question so this is not the way to go and with this example of the residance project it really unes what we call the Quest for modifying genes this Quest that allows to identify in the human genomes of some specific individuals some specific variation that protects them against disease and actually perhaps some of you you know the story of CC s Delta 32 of marav and the story of this person this person is stepan lion cron and um due to his lifestyle he was supposed to to have ads and he had not and he was exposed to a lot of behavior and compartment and people around him dying and he was not and the researcher they discovered that this person had a specific variation on the gene CCF CS5 Delta 22 that closed the door the entry door to the cells to the virus the Eds virus and so this person was immune against most form of the Ed virus and with this discovery see this discovery that helped to create all the new therapy that's used in the field of ad so you have this human variability that that you see more and more we are more and more conern about that that you have some people who are sick when they should not be sick and you are some people who should be sick and are not and with those people we need to find them we need to identify them and why they are sick when they should not be and why they are not when they should be and this is what the genome can help what the study of the genome what Gathering the genome information can help and here we have somebody which qu amazing it's uh professor kazanova is really using this approach in the field of several diseases and identify many different genes that have what is this amplifier or protective effect here you have somebody that's that's really amazing is Professor G Cano and he did this kind of Discovery on the gene P3 C CA and being able to identify the data use effect of this specific mutation he came with a he proposed a drug repurposing of an existing drugs that could help it was in the context of a rare disease nearby the the clove syndrome the clove syndrome everybody knows it it's the elephant M disease which is really an awful terrible disease with really short life expectancy and uh Professor Cano he realized with studying the genomes that people are Bing this disease at impact on these specific genes and he knew that in the field of oncology there was a specific drug that was targeting inhibiting this specific Gene and he obtained a kind of drug reposing not the complete one but he obtain an extension of the drugs able to reuse a drug that was used in the general population for cure not would it for cure people with this clo disase this is incredible you have you have to to to see what he does it's it's really incredible and so you have with Prof Ken you have the example of a Drugs That's created in the field of the general population used in the field of the real disease and here we will touch with What U Jan say earlier here we have exactly the contrary we have people with a rare disease called hper cheria and those people they have a specific variant in the Gene pcs9 and Professor Katherine Boo which is an amazing person that you see with Queen um she identify this variation in thanks to this this family of for Pati and thanks to that Mo today most of the people who are um followed for problem with cholesterol they are cured with the drugs that have been developed thanks to this rare people so when joanni explained that there is the risk that you are focusing only on the general population missing the ra disease we are really miss missing something big because people with the r disease they have a lot of things to to to give to us okay of course the humanity and the the the ethic and and everything that we have to do uh for include them in the society but in return they can help us to understand better the genome understand better the interaction between the genome and can be profitable for them also for the general population like we have here with this example with Professor Katherine W which is quite amazing and so it's why when you have all this idea of human selection of elimination of some genes it's really a wrong direction to go because this is this extreme variability of the human being of this Humanity that we are of all this variability of the genome that we have that allow us to have some key or some people who are carrying the cure to the other people this is something really important and to do that to identify those variant what we need know nowadays we need artificial intelligence you get it artificial intelligence is as good as the quality of the data that it has to feed you know it after all the session that you had before I will not explain it garbing garbage you know that but here what we need what we have in the field of genomics it's clear data it's a digital sequence of 300 gigaby of alignment of rctg so you have it's huge it's a volume which is which is difficult to cope for people like like Tio explain that later but we have clear data and we can explore those data and we can't explore anymore those data only working with uh one dimensionality that the human being is doing so it's why we need to use artificial intelligence and it wise the couple artificial intelligence and genomics it's even difficult today to to to to separ so here I will give you an example of what we do and what we did with the Geno for bristel initiative uh we create a specific tool at a monogenic level for the gfb one we created that with Professor G Smith that you can see a and we created a te an algorithm machine learning tool artificial intelligence tool that helps to solve this difficult question of the variant of unknown signification and to help to categorize the variant in when we were in front ofi and I will go fast on it but I would explain you what we did the idea at the beginning was to create this kind of chessboard trying to use the FB one Gene and to use all the possible situation all the possible variant that could appear in the FB one gen and so we created we generated its first keyboard we limited some the reference position that of course were not deluse uh we we eliminate all the synonymous uh variation that we knew that they will not impact the production of the the the protein and after we use the machine learning to try to and use Glade which is fantastic Tool uh to uh this is f with this idea like like we're using a padlock we try with each and every combination to see if the um the mutation will be pathogenic or not and we manag to create this with it which is gvap live that it's live it's on live on you can use the link and you can go um I can see see it on the web uh which is a tool that allows you to check each and every possible mutation on the Gen geneb one the one that exists and even the theoretical one that have never been met in the in the reality so it's what we we created and we were really proud of this this tool and last year something else appeared it's for us it's bigger than chpt and andms it's Alpha misons and Alpha missons is something that was launched by Google and Google did exactly the same thing that we did for the FB one gene but they did it for the complete genome for the 20,000 Gene of the genome which is amazing this is a really important tool this is something that will really help and you know joavan was explaining that we not never get rid of the doctors but artificial intelligence will help the doctors to have faster diagnosis and I can tell you that in the field of disease it's really and not only field ofono so having uh timing of a diagnosis is crucial having tools like that it will help to save life okay so not only this tool you only needs the expertise of the doctor and the capacity to do conduct the the the the additional research but this one it will help to go faster so we integrate Alpha Miss in our gvap live proposition and you can see that in some situation or or prediction perhaps a bit better uh but it it says that it's really a kind of combination of source combination of approach that will help you to uh to yeah that will help us to improve our ordinary Health by the way yesterday or two days ago emble announced that they also embodied uh Alpha missons within their own predictive tools which is we discussed it a bit earlier which which make alpha instance even more usable that know is available on the top of that Google created something really amazing which is called Alpha fold I don't know if if I can no I can't if I do that what happens this is a video there's an animation uh you have the link here when you have the SL you can go see the see the animation and you will see this is something that we created with the help of tibo we use the gemap and we selected some of the pathogenic variation on the G fbn1 and we wanted thanks to Alpha fold to visualize how concretely is the impact of the mutant variation on the sequence we know we can see it on a text and know we want it and we can do do this thanks to Alpha fold we can see in three dimension what is the impact on the production of the protein and this is amazing this is something that okay it's specific in the alpha fold world I really I regret that I can't show the the video because you can see it in 3D and it's moving and you can see really the what we did the alpha fold provide the reference structure of the protein and what tibo did is that it allowed to change the code of arold to show us the impact of a mutation and it's really amazing to compare the two images but you can do you can do it on your side by the way TBO at the beginning of this month there was a release new release of of f know there's a version three which is which is available so I told you about monogenic approach and now let's speak about exom approach so the use of artificial intelligence not at the level of the mon of one gene but at the level of the complete exome the 3% of the genome and here as Joan say it really enthusiastic it's true in Belgium we have really champions of artificial intelligence in the field of healthcare and one of those Champion it was a firm called deployed Moon that created an amazing tool that allows to do the characterization of the pathogenicity of the variant within the whole exome in hours and this tool was bought by invit few years ago uh I can tell you that no the deal it's closed with INF and they are back on the game in Belgium and they will come with with new new surprise and thanks to the utilization and the creation of this deployed Moon tool they were able to provide a diagnosis within of rare disease within 48 hours which is amazing it's game changer for the life of the the patient in this this in this publication it's a publication from the liage University and it's really it's really fantastic and things that you can do only only thanks to artificial intelligence okay and no oogenic uh in the genomes and this is I'm really really proud to work with Professor Tom Lenard within the genome for Brussels initiative where he works on this quite unique source which is called Orval that allows to identify the networks of genes involved in the declement of a disease uh I will not tell more about that is better place with paint but the IDE is that thanks to artificial intelligence in the past we have the traditional Ino study we study mice and Mo and fishes and and different animals models but really too often we have drugs that works in animal models but they are not they are not working as well in the human beings so we have the invivo approach we have more and more what we we call the invitro approach which is what we call the IPC which is cells that are reprogrammed to mimic human cells of a human individual it's better because you are going in the direction of the human it's biological so we are going in a good direction but now what we can add with artificial intelligence is the inico approach so it's not it will not withdraw the two approach it will add add a third path in Vivo in vitro in silico and this is the primaries of artificial intelligence in the field of genomic so regulation the gdpr it was fantastic for a lot of Reason in this field it was difficult to cope with and let's say it's it was a bit of break okay no thanksfully you're at the European Union level we came with the European air data space the European data Government Act and all those different tool that were there to help to facilitate a bit the access to the data and I think it's it's really going in the good direction and also we took the time to tame the gdpr and no researcher different University uh they are they are they are more used to it and some people they use it quite well uh so it's going in a good direction no the AI regulation the AI act um Char told me not to speak too much about the AI act because you already have a lot of introduction and discussion about the a act so I will not say a lot about that just know I think it's a new question mark because the a AR it was targeting Terminator and know it's touching this guy tibo so um jooan was talking about all the Innovation that's put in Belgium and all the people while investigating into using I Ai and Healthcare in Belgium here you have somebody Jonathan dupy is's in charge of the research and development in Brussels Regio investment in artificial intelligence and the day the AI Act was adopted he had the reaction that was not let's say not so as positive as as others could could have uh we understand the merits and the benefit of having an act for sure but on the same time we realize that it will be a new burden to Innovation and to research but it's so it is and we will cope with it and we will succeed so we will discuss first about some different actor of the genomics Revolution and we will begin with the DNA sequencing industry so you have different players in the field so different producer of the machine that are used to translate the biological DNA into digital DNA and you have Pacific bio for bgi and most well known that everybody knows ilumina which was for a long time the uh the leader on the field at one point he even tried to acquire Pacific bioscience why Lumina wanted to acquire Pacific bio science because elumina is using what's called a short read technology meaning that they use the genome and when they do the digitalization of the biological information they cut genome in min mini part that are Rec combinated this What's called the short technology and PBO came with another approach which was called the longer technology they're using not short part of the genome but long part of the genomes and so the idea of elumina was by acquiring Pacific bio they will be able to use both technology the short and the long read and that they will be able to produce data that will be even more informative and it will be that will help to identify more easily the patterns the FR regulator in competition law in field of antitrust and competition uh we're not so enthusiastic about the acquisition of P bio byina because they see it as the destruction of a nent competitor why not but in the same time Illumina was was very too confident about its own position they were really in a pos position where most of the of the the scientific group around the world are using its technology and he he had few competition at the moment and at the moment there was something really I always think that's quite amazing the the former CEO of imina he he say at one point that they were able to charge only 100 100 EUR per genome but at the same time they were Charing 1,000 EUR per genome I think which was really not so smart from from this guy to to say it and he say it exactly in the time when they were purchasing P by you so it was really meaning that they were not afraid of the competition and that they have control on the price which was really not going in a good in a good way so it's why competitor on both sides of the Atlantic decid that was not such a good idea to allow Illumina the region to allow Illumina to acquire P bio so determinate the deal so is it good for the consumer is it good and who is the consumer is it good for the patient is it good for the science this is a question that we can ask when they decide to to renounce to the acquisition of P bio they decide to acquire a company called Grail and gr is a company that does machine learning in the context of diagnosis of early diagnosis of cancer and so it's really here with gra we are really at the the intersection between artificial intelligence and genomics and the ID you can go on their website and you will see it on the idea is that when you have cells when they die when they affected by by by cancer they have a specific signature in their DNA and when the DNA when the the cell dies you have the the DNA that goes in the flow of the blood and you can find this what they call the cell-free DNA fragments in Blood and when if you identify the cell- fre Diamond uh cell-free DNA fragment in the blood quite early Grail create an algorithm on machine learning algorithm that they believe it's able to identify some pattern identify that if you find this kind of cellfree DNA fragment in the blood you can identify that there could be somewhere else in the body a Cancer and they have another tool that allows them to detect the origin of the cancer okay okay so the idea to have a faster um preventive uh diagnosis of cancer and they came with this this solution so they wanted to to buy Grain to acquire grain to allow researcher who are using their technology their sequencing technology to use immediately The Grail Al algorithm on the data that they produced and on this side um I will not go too deep in it but you know that have this competition law background um the commission under the request of the the the French competition authority decid to bring a new interpretation of the article 22 to allow the commission to investigate into the merger even if this merger was actually regarding the threshold not in the threshold so they change the interpretation of the law and they appears for the first time in the context of this specific matter so here we are really at the intersection between genomics artificial intelligence and competition law and what Illumina did it's quite crazy elumina even if they knew that there was some concern at the level of the European commission they say we don't care about the concern of the the European commission we strongly believe that our position it's grounded legally grounded and that the commission aired when it interfere some power extensive power from article 22 and they decide to continue with the integration of Grail immediately the commission jumped at them and so explain that no we are in we are launching an investigation for gun jumping which was of course what they the commission has to do actually and so uh the okay we will go fast on that but you know it if you follow the case ilumina went against this position they lost in the F first instance and so uh it went to a kind of Revolt from the actioner of ilumina and a group that challenged the position of the CEO and the CEO of uh it went it conduct the CEO of fumina the Souza to resignate to quit its position at at cumina and uh even when he left he was came with this this belief that this strong belief that the addition of gril will have been something really important and huge uh for Consumer actually uh of course one month later uh the European commission finded Illumina for 432,000 Millions Euro for having done this decision to to implement the acquisition of Grail without the authorization with the clearance of the commission of course and uh imina has only 12 month to devest from from gra what's new in the story is that in March of this year the general Advocate unored the position of Elina he came with the constat that the position of the commission according to him it's not legally grounded and so we are in a position where ilumina withdraw the acquisition of Grail lost e CEO pay more than 400,000 million fines and it's in front of a court where the decision of the commission is still so let's see what what will happen uh it's it's really it create a really difficult situation for the European Commission because no a lot of decision have been also adopted on the same ground there kind of Cas law generated from this I use the cas law to bracket from this decision and so in the same time we can see that the competition in the field of sequencing is still Rising you can see we have this Oxford nort technology which is really promising technology which is a new way to study long read sequencing that is more and more present on the market and this week actually two days ago the the the the global pharmaceutical company Rush uh announce its intention to penetrate to enter the segment uh the market of sequencing technology with its own nanopore sequencing Technology Long technology so we are on the market still evolving and this week you have the editorial in natur medicine when they explain how important the cell-free DNA fragments circulating in DNA in the blood detection it's important and how it could save millions of life and all immature the technology is still and still a lot of money need to be involved in this technology to have something that would be really useful and so now the question about the gr question is really why and for who benefit this decision came because it's not really good for the AI developers is it good for the consumer but who are the consumer is it good for the cancer patient is it good for Pharma company who are selling the drugs is it good for health insurance company is it good for science is it good for legal science we can discuss that and is it good for lawyers it's really an open question and I really wonder what will happen with this with imin gr decision I'm not involved at all in this case to be really clear okay uh you have the story about the direct to Consumer DN test industry because this is the one that attract a lot of attention a lot of people see it uh it's it's a real industry mainly in the US uh you have different actors the main one is the 22 andme uh if you want to have one I suggest you to have a look to nebula genomics um but most of the time it's used only to have ancestry research and sometimes some food and allergic test but U more and more you can see it even used for Elf question and T2 me was the Pioneer was the first to have the authorization to use its uh platform in the context of Elf but once again they are not using most of the case they're not using the the whole genome technology they're using specific sneak net and so they have view on something but it's really incomplete and let's say that's um it's interesting but you should not trust only this kind of of test and I will go fast on that the idea is that in sometimes it was also useful to solve some what we call C case when people share the DNA that they have they had sequence thanks to technology ELR TR me they use a specific platform when they share the DNA and some people from the FBI the ID to check with the DNA that they are in their archive with the DNA that is present on this platform of sharing of DNA and it helps to re identify many case but what they explain they explain all the technique how they did with you get match but what they explain is that they really with only the DNA they would have been they will not have been in position to identify by the people they were seeing for they were only able to do it because they crossed the information with Facebook information LinkedIn information and following the people in the streets it's way it's why they they did it but what interesting with the this DNA test uh Market is that we can see that it's a two-way access two Side Market industry it means that on one side 202 is selling the test to the general population but on the other side T me is selling the result of the test to the farma companies so it's really there are two stream of revenue for 2020 and this is the kind of two-way access industry that you have to to get in mind when you are selecting studying this kind of company on the field of Pharma industry I will let Tio will explain better how artificial intelligence can really fit in a different process I can just tell you that once again Germany is really right in Belgium it's moving fast there are a lot of things that's done here there's a really nice initiative that's conducted by Florence Bosco that that you know really well with B AU that's try to emulate and construct and generate some revenue and new therapy therapeutic syndication thanks to to artificial intelligence uh unfortunately we need to be also really cautious because we have the example of the anen situation where perhaps you know about that but in Iceland they decided quite early to sequence a vast party of vast majority of their population so there was a specific program called decode that gather the Genome of the Icelandic population and for a lot of reason they put that in in a commercial company and at one point they were bankrupt and am both the company and so the Chinese Pharma company AMJ owns all the genome from the Icelandic population which is quite quite crazy and uh so they can use it they can test it they can conduct their research on it and when when code was alone the code was fueling the research because a lot of researchers were asking to have access to decode now it's really propietary to Ament so it really Chang them so it's why we need to have what we call institutional project meaning project where you gather the genomes of people that will you will be able to to allow other group of research to use and investigate I told you about the importance of the people with r disease in the understanding of the genome this is something that the UK understood really really well and really early it's why they create the 100,000 Genome Project which is a fantastic project uh which was created by David Cameron for in a specific context and know this this uh this one th000 Genome Project in the UK it's amazing the the the result is amazing in few years they gather their target they have more than 100,000 genome now and they're really fueling a lot of research they really help the research in the UK to go for F and you will see that it filled also the industry the economical industry uh that helps the NS they are working obviously know with a cloud provider which is I they are this is the point they're using illumin technology so they 100,000 genomes that they have they have been sequenced in in in 30X technology so if you want to have interoperability with the others group data set I think the other one we have no no no much choice to other technology you have it fued also Innovation and you can see in all the innovators in the for that you have them all of them they have artificial intelligence Dimension so there are no planning to sequence five Millions genomes within the UK I can tell you that off record some people involved in the in the project told me that UK wants to be the master of the genome like they were want the masters of the oceans this when she told me that wow so but they're going the way they're going in this direction in Europe we have what we call the 1 million plus genome initiative which is a wonderful initiative that was launched in 2018 at the beginning with the impulse strong impulse of the UK uh the IDE was to reach a target of 1 million genome by 2022 unfortunately we are not there of course the brexit happened the co happened uh but we we missed the the mid L we is the the meting from what I see but perhaps people in the room have more intelligence from what I see as of today there's around five uh 75,000 uh genomes available but actually most of them there are uh in Finland uh in a specific project called finen so the 1 million plus genome initiative has to a bit readapt it its ambition and to have a new road map and know they have a new road map until 2027 and decided to create what they call genome for Europe with this idea that they will build this network of national initiative that will be inter aable and allow to share the different datas of the different nation which is brilliant and really a good approach and uh you see everything is on the website I really recommend you to to have a look on that if you are interested and so the idea will be to have National initiative like the finen uh explain too much about finen and and to build on it in Belgium the current situation is that we have eight centers of genetic six uh which is really a chance in comparison with other other countries and that the Belgium joined quite early the 1 million plus genome initiative with a lot of expectation and hopes and so I'm part of this mirror group I'm working with the in the work package to uh LC on legal and ethic and social issues I'm also involved in the creation of what is called the Belgian genome Bank where the idea is to gather the Genome of a part of the benan population and on the of people with diseases uh for example with the work of the LC group we recently had to comment a proposition from the the World Health Organization about the sharing of the genomics data which is really key which is really under the core of of all the research that we want to do and what I found really interesting in this proposal from the the World Health Organization it's the the point 94 when the the on that we promote in Europe and everywhere when we share genomes Dynamic concept meaning that when somebody share his genome you don't have only a one way contact dialogue with the person who share his genome and it makes sense you see it with the residance project if it's only a one way you can sequence the data put it ins the side it's useless you really need to have this two way stream to make something from the data you get it and something that the Australian genomics project understood really really well and so they create a dynamic consent solution that allows the people who are joining u a Genome Project to have this this interaction between the two in Europe because at this point people want things to go fast they want to use the kind of the exception of the gdpr for the secondary usage of data in the context of research and they want to get rid of the conent approach they want to work with a system of opt out and people can just decide to not join a research but not actively decide to join one research and they say they go in this direction because it's too complex too burdensome to acquire the consent of a person yeah it's also complex and burdensome to do elections uh it's not why we are not doing them and here we are talking about something really specific to a human being which is genome it's really information which is protected and it's quite paradoxal to see how we protect the information we share on on Facebook and other social media and where on the other side when we come to our core information The genome we should share it without return I think it's a bit particular and specifically because in the field of genome for Brussels we created a specific Dynamic consent management solution and it's not rocket science and you can do it and we can see in the context of genome for bruels we have more than one uh we have 125 participant who share their genomes with us and they did it using this damic consent solution and it's something that I think should should increase so we will face a notion of genomic data there will be more and more genomic data available there will be more and more solution to H the data and so algorithm and genomics are designed together to transform medicine as we know it for the diagnosis for the the the development of new therapetic the therapeutic approach for for the drug repurposing in every way the this couple genomics in artificial intelligence we change everything and it's already started and we can't we have to address this issue we have to be in the game and in Europe I think we need to join the race now and fast well that's it for me today if you have question I would be really happy to reply I'm sorry uh yeah I don't really have a question it's a a lot of information and I just want to say that it's very impressive um it's uh hard to say where it will lead us but it's yeah it's very impressive so thanks for the presentation and thank you for being here because it was really nice to have you in the room and follow you that you were for me thank you very much hello uh yeah thanks for the presentation very interesting uh I'm not a scientist uh so it's a yeah a lot a lot to take in uh but from a from a Regulatory Affairs perspective are there any sort of specific regulatory interventions that you foresee for the next political mandate for example in relation to uh the protection of genomic data and so on there'll be like a review of the gdpr expected this type of thing uh is there a way that European polic might react to try and make it easier to to progress on this okay um the easy answer is yes they should change everything and make it happen easily uh the other answer the lawyer answer is that uh please let us work with what we have let us understand what we have what we deal with let us try to construct and find solution and because each time you came with a new solution actually it came with a new complexity and we we see we have the gdpr and all the hospitals that were already terrified with the gdpr they put a lot of hopes in the Newan Health Data space and now the C say that's it now they understand what is it they say okay so it just had a new layer and know you have the data government another layer and so each time we try to solve what we created with the gdpr I'm afraid that the feeling that I have that we had some complexity and uh let us deal with what we have between our hands right now we have more and more tools let us I think it was yesterday that the commission I think it did it well uh decide to to introduce a beginning of a procedure of infringement against Belgium because Belgium didn't comply to adapt translate the data Government Act let's begin with what we have in our pocket and apply what you have to do and let's see for the for the future I'm not such a huge fan of of new new new new Revolution F it's really personal thank you ran you speak on your own name H it's no but you know we will talk about this too much you and me because we will Deep dive on the on this subject I'm very very happy about your presentation very many thanks and I was also just just one point you know the other day I was in a conference on reversed um Technologies and what what China and the US are doing with the with Europe and um what are the interests for them and genomics is one of the main topics for them because of you understand the value of our data which is like the new gold but genomics is like diamond I mean as you said they have uh the data from the Finish was it you can imagine all the European data dyamics data is super relevant for science so in a way we are of course we will we we will try to protect as much as possible regulating this with the new ehds but at the same time we need as you said not to stop the the research and development of our industry so I'm sure this year coming and for the next um Master next year you will you will have even much more information on development in in jomic so it's going to be a very interesting year I'm so sure excellent a lot of people are waiting for you a lot a lot of families and people see okay point on that just really fast two point on that um in China they developed the the being genomic Institute bgi and at one point they came in Europe and they offers crazy price for sequencing meaning they were at 200 300 EUR for whole genome yes when we were still at the time paying 1,000 chome and a lot of research project European research project they use this facility meaning that a lot of European genomes they're already in China and they're already used by the Chinese and it's done so yeah that it's time to wake up it's for sure it's time to look at it for sure but it already begun and began a long time ago and another point when we created our project with genome for Brussels we have genom coming from all over the world and from all over Europe but also from the US and it's quite impressive difficult is it to gather biological samples saliva saliva kit from the US to Europe This is incredible because every day Europe is sending thousands of salivit to 22 me and all the others who are in the US and so the US is already with this kind of solution using has access to a lot of European gen so at least if we could have the reciprocity and be able to import the US uh genome it will be a good thing something in the direction we should explore but thank you for what you do and provides a lot of Hope um I think there was a question written in the group chat that you may address uh let me scroll up okay thank you and for the question it's it's it's always the same uh the question we talk about the military usage of genic information I think there's so many people going on the the dark side and the negative side of the question of the artificial intelligence and the genome that's I'm quite happy that you see me as a a different voice coming with a positive approach of it and I think you should have indeed a vision of the light side but also on the dark side and it's true that they will do it but what will prevent them to do to that as I just explained right now a lot of genomes are already available European genomes are available islandic genomes are just so different from the European genomes I don't know you this this is uh it's why we need to be to be attentive to that and and to find the good balance uh and it's why we need institutional project and and we need this kind of like the UK did this this project is really it's really I think an example to follow and that to go that's that's all I can say um well thank you Roman for your presentation and thank you for the students who uh who actively engage with your presentation um now maybe we can take a short break of let's say 10 minutes and then we will enter the last segment with Mr T head thank you very much okay e e e e e e e e e e e e e e e e e e e e okay everyone please take your seat thank you let's start again so um I will try to adopt a complementary uh perspective to the the two previous talks very inspiring indeed um the the point of view of the the industry the healthcare industry um myself I work at dtic which is a company providing a data science expertise to uh the health care sector um how data science and AI can add value to uh research and develop programs also how it can support the manufacturing of drugs complex drugs uh for example to prevent shortages for the patients and also to increase the competitiveness in our region here um but also how can data science and AI um benefit to the the public health sector uh and um I was interested by the the the question raised by joavan earlier today um wondering where will we be in five years from now and in fact it's very interesting because there are a lot of promises but I think we won't be very far from where we are now in five years um for several reasons I will try to to highlight but but I'm I'm rather optimistic as well there will be developments but it will take uh some time I think first the the hype of language model will uh cool down a bit and I will try to shed light on on why we should be cautious with these kind of approaches uh also because of the the huge load of regulations that are already in place and will come on top um and also because um in fact bringing new healthcare product on the market B for or diagnosis medical devices or new drugs it it takes a huge time to validate and in fact there is already today a very formal need for clinical validation before any product can actually Reach the the market but I understand that doctors feel overwhelmed with um demonstrators new appealing tools but these are not validated yet and shouldn't be used to care uh patients on a on a daily basis um so here in this presentation I will uh try to look at what already exists in terms of Regulation and there is a very interesting paper from the the EMA so the European medicine agency uh listing all aspects of the life cycle of a medicinal product where AI or data science could play a role so it's not a regulation it's a it's a reflection paper on how these techniques may have a role in the development validation and and exploitation of of medicinal products in the broad sense and we can uh roughly split these uh document into different main topics of Interest R&D how do we design new drugs new medical devices uh another um more related to metch so medical devices and and initro Diagnostics the manufacturing of drugs this is very interesting because it's a single paragraph in the in the paper uh whereas actually there is a lot happening um in in that field um related to to vaccines uh cellular therapies genetic therapies um where AI will play probably a huge role uh there are um some parts of the document theing regulations themselves and a quite large part on the the the technology AI uh in fact we can I'm not a lawyer but uh I I see in my my daily practice how different regulations standards um or good practices may impact but we can what we cannot do or how we should do things and so I I just make the link here it's not it's not exhaustive but clearly um when we discuss about R&D and and Roman showed it very very well we need data and data comes from patient for some part and so indeed we we have to wonder about how to make that compatible with the gdpr but also later on uh the the good clinical practice when we have to validate new new products um mettech medical devices and and in vitro diagnostic is probably where um AI will will play its greatest role as Javan alluded to how to to better diagnose disease how to better make informed decision on the relevant uh therapy for the the relevant patients um and in fact we have two very well written and very future proof regulations here the ivdr and the MDR so regulation on indidual Diagnostic and regulation on medical devices um which to me as an AI practitioner I they have been written now a few years ago but to me for this field mettech it they make the AI act useless and and yeah and redundant to some extent and there is a big question mark on how the AI act will coexist with these two very well written regulation um and it we will be careful um to see that the standards that will be associated in the technical documentation that will be required by the implementation of the AI act will not uh enter into into a conflict with what is already asked to the industry through the ivdr and the the MDR so for example clearly these two regulations they already for foresaw the upcoming of software as a medical device for example which clearly encompasses already Ai and need for proof of clinical validity of this product Etc um with respect to to manufacturing um there are the the the good manufacturing practices and also a lot of Regulation also coming from the the the FDA in the US um gam five is a standard that explain how to uh develop software for the the manufacturing industry and the new release dating from uh summer 22 or 23 but very recent with respect to the the first version of the gam five has a complete appendix dedicated to how um integrate AI into the manufacturing practice so the using the pharmaceutical manufacturing practices um on AI itself there is of course the the AI act um just I I will not make a huge discourse on on the AI act but there was this question just a few minutes ago on the likelihood that people may develop weapons based on genomic information and AI actually the AI act does not cover military applications um and there is this very interesting concept uh which is a reflection paper jointly written by the D FDA but also the the British agency and the Canadian Agency on gmlp so the good machine learning practice very short two pages um but with really interesting proposition and and key point of attention for those who want to integrate AI into mettech so just a a few a few links here between where AI could p a role and and existing regulation I will try to to cover in this presentation a few of the the steps of the the life cycle of a medicinal product I will not cover all of them I will uh discuss about the the AI driven drug discovery uh second um more um questioning of how clinical research is conducted now it is and how it could evolve in the in the short short F near future uh and then uh we'll spend also some time um discussing bio manufacturing so the manufacturing of of Biotech products um during this first uh step first stage here I will also make a parenthesis about um large language models and generative AI just to set some misconception right which may be a miscon misconception um so clearly um there is a a willingness and and a need also to uh speed up the the discovery and validation of new drugs a drug typically uh would take 10 to 15 years to to identify develop validate and um more than than less than one among 1,000 drun candidates actually reach the patients so there is a need to to to do this task better and to make it faster and so of course there is a an appeal to use these Ai and data science Technologies to to speed that up so here I will just go through a small uh demonstrator of what can be done it's not perfect just to illustrate some some Concepts um when we uh combine really upto-date data uh with um good logic and generative AI tools so it's a a small application we we made with with my colleagues but mostly reusing existing material um in a a way that we hope is appropriate so the the application lets the the user first select a pathology among all the pathologies that are known and depending on what is selected the application uh will list all the known targets so targets are elements of the metabolism that can be the target of a drug some of these targets are already exploited by um validated drug or drugs under development uh some are thought to be potenti good Target but not exported and once you have selected a disease and a Target um the application will provide you um with three different use cases first it will allow you to list the the drugs that are already known um with and provide a table that that is longer below here with a lot of information about that drug uh some of them being scientific information other are more business information who is the developer of the drug what is the the the status of its validation Etc um so this is more for for landscaping not not redoing what something would what someone would had developed but of course it's it's not enough you can opt for the the second scenario drug reposing so alluded to the the interest of being able to to Repose a drug so for example if you are interested in a given disease and you identified a potential Target for that disease it may very well be the case that someone already developed a drug for the same Target but for another disease then this application will help you identify um candidate Partnerships you can discuss contact the company see if there are possibilities of extending the the indication or concluding a license agreement Etc and um with that respect we also uh implemented uh something something that is called biog GPT biog GPT is like CH GPT you can discuss with it but it is not trained on the same Corpus of information as the general chpt but it has been trained on um only the the few millions of scientific publication in the in the medical so um it's the same concept you can discuss with it but it will not bring the same answer as uh chpt as it is more targeted toward the the medical domain I will explain later why you shouldn't discuss with that kind of chatbot to obtain answers but rather to uh generate hypothesis and and here we just ask biog GPT about its opinion between quotes on why it should be it may be a good idea to reuse an already existing drug a given one for uh given Target on a new indication and and it will and it will reply um and we generate not one hypothesis with biog gbt but a dozen of hypotheses and we let the expert of the disease make up his mind about okay this is already very well known crash this is complete rubish crash all these are interesting fields of research uh also um we there there is a a third scenario of using this app which is more oriented toward drug Discovery so if you do not have an existing drug if you cannot identify a partnership with an existing drug for the same Target but for another disease then you have to invent a new drug and there it's not really visible here but we provide a table of all the the molecules the proteins or the genes interacting with the target you have chosen so Network medicine um and and there you have to uh identify which of these protein for example could be a tool to attack the the given Target of course you will probably have to modify protein for example um and to to help you visualize what you could do you can use tools such as Alpha F to visualize the the normal protein or the modified protein and um to yeah first parenthesis about the the hardware needed to do that so probably all of you are aware that in in most of the computers of the world they are CPU so central processing units uh which are very good at performing some mathematical operations that are satisfactory for most of the the butic applications and and and most uh to make most computer software but when we go for that kind of generative AI model such as Alpha here there is a need to use or to train or just to use um a deep learning model and deep learning models require mathematical operations that are slightly different um and they do not fit very well into a CPU CPUs can perform those mathematical operation but very slowly on the other hand the the gaming industry uh has developed over the years another kind of processing units so the graphical processing units that are not very recent they are used for playing games in computers for years and years but it happens that this particular kind of mathematical operations that are needed to run uh deep Nets deep learning models fit very well into these graphical processing units much better than in the traditional CPUs so you need both um these kind of processing units to be able to to run modern generative AI models uh close parenthesis um so um if you try to mod ify protein you you play in modifying the the genetic code and R explain how modifying the genetic code will end up in some occasions to a modification of the actual protein that is produced based on these um instructions um if you go to the to the public version of alpha you will have um the possibility to uh see the the the reference form of the proteins um by implementing that model on on our Lo local architecture we are able to play with it and to induce mutations and so this is ex exactly the same picture as the one you you you showed and you can then also play through an optimization algorithm um by aligning the reference protein to the mutated one that is what is done here um and you directly see single mutation here what is the impact on the shape of the protein and the shape of a protein is key in its function so you see here there are Parts you see here it's preserved but displaced some parts are really preserved and not displaced quickly cover each other but here you see slight difference here you see bigger differences um so it's it's a a very practical way to play with genetic code and see what could be the impact on the the the 3D Shape of the the proteins um that can be useful in drug Discovery is also useful in understanding the the the nature and the the functioning of some of some disease when you list variants that are appearing in patients with the disease um you can understand why some variants may be more impactful than others based on on that kind of tool among other tools and and human expertise of course and it's here that I will um try to move towards language models I no no first an extra step that can also be performed in silico is suppose you have your disease your target now you have played with genetic code of proteins um you just have to do the production of a new drug you are all aware not that mRNA is a potential way of delivering new drugs um well there are tools also based on on AI and optimization that allow to take as input the genetic code of protein and to optimize the sequence so that the when injected in the the body the protein that will be produced by the the the host so the patient likely uh with the instruction of your modified cin will be uh done with a high efficiency and so there is a need to engineer the genetic code to produce what you intend but with a very good efficiency there are tools also optimization tools able to to do that um and um um I if you remain just one thing about everything I will say this morning um it may be this sentence in AI humans always set the objective it may very well be the case that we forget that we have set an objective but there is always one that is fixed um um with Alpha the algorithm that I discussed to generate 3D representation of proteins the objective that has been s set during the training of the model is how good the prediction of the model so 3D shapes are fitting with the actual 3D shapes for the few proteins for which we know it for sure by other means like crystallography x-ray experience so we have a corpus of prot for which we actually know the 3D shape and we train the model to reproduce that correctly on these known proteins but also to extrapolate on other proteins for which we do not know the the 3D structure but that is the objective minimize the differences between the predictions and the actual shapes for the proteins for which we know the shape you may Wonder okay this would not TR hold for um large language models and and char ppts and all the like they can do anything we ask them to right well no in fact they only do one thing they have been trained for is given a few words that have been written what is the next most likely word not the truest word but the most likely and that is the only thing they are built for so when you ask ask a question to chbt and almost all of the other large language model its objective is to generate a likely sentence not a true one that's why KBT is bad at reasoning because it cannot reason and that also why if you ask chbt to display the sources corresponding to what it just say he will invent the sources because it doesn't care the will be written in a plausible way but they do not exist there is a very funny case of a lawyer in the in the US who who who came and arguing based on on previous cases similar to to the one he was defending and in fact he he used chpt to generate that and he thought that the the model will put actual prior cases but in fact they were just invented um and so I have um a concern um you may have heard that um people are complaining because these models hallucinate it's a recurring word okay chpt has hallucinated this generative AI model has hallucinate in fact they do not hallucinate they just output plausible things we hallucinate when we think it's the true so um you have to to Really to remember about that and so of course people really in the field of developing these algorithm they they know that so they are thinking about other ways to put this large language model at work uh one very interesting uh recent evolution is um the categories of of language models called rag so retrieval augmented generation which is if you want um cross breeding between search engines and chpp the difference is that um the the similarity is that these models are very well able to understand a query in natural language and they will phrase their answer also in natural language so you can discuss with them but what happens behind is different they will map your query into an abstract representation of all possible natural langu queries but before having this system working the the engineers have provided a corpus of text which they selected curated usually on very specific topics um and this Corpus of text have been also mapped in a similar um abstract space of sentences and so once your query has been translated there is a distance that is Ed in this abstract representation between your query and all the text from the Corpus that have been also um projected in that space and the the most similar the the closest elements of the Corpus will be feted corresponding so very precisely to your query and it is only based on these very limited set of documents selected based on natural language uh interactions that the model will phrase an answer that is likely much more relevant to the question you you raised but also this time you can trace the sources where the information comes from because among the thousands or millions of documents from the Corpus maybe it selected only 10 to phrase its answer um and so you will be able to track more closely okay this that is provided where does it come from so it's much better still there is no guarante absolute guarantee that what is output is true it's likely and based on a more targeted uh set of inputs but there is no Guaranty uh even more bold now you have the even more recent AI agents the concept of an AI agent is very old but there are no implemented Das language model um you when you ask a query to these agents the the programmers have put constraints so that these systems will have to break down your request into uh specific substeps um like typically okay um can you based on on on my request can you output one element uh summarizing the objective of the request request one element which is a risk analysis of the request and possibly a plan for breaking down the request into smaller operations and so the large language model will do that so translate a general sentence in a more structured set of text elements and plus the system has access to programming functions in Python in r in C whatever with the documentation of the code and based on the breakdown of the task you asked it will be able to choose which function which data to call and program some things and provide you with the answer so um it's very exciting because you give a lot of actual autonomy not only on of discussion but of action to these agents but again there is no guarantee that it will be uh true and actually reach your goal so here I will show you um a video to see how this is used in the pharmaceutical industry um so don't to share the video I don't know how should I do so it's a video that is publicly available it's from GSK no risk of harm for the patient they use that for their own um research on how AI can play Ro in the e you see here it it breaks down the task into substeps is kept in the loop here to access the next step so you see this is code written by the AI is PR is also is EX on byal okay so you have seen the the essence of it I will return to the the presentation so um it's really interesting to see how natural language can drive he programmation um and generation of results um an important aspect you have witnessed is that there is always a human in the loop to um Grant the authorization to move to the The Next Step um this is of course uh very important um but it it raises some some questions here it's an easy task and second it's a task for which we can validate if the answer is true or not as I highlighted um large language model are not done to say or write things that are true and so the question is will that kind of system first still be able to comprehend much more complex requests how will we be able to verify that it understands it and uh second how can we validate that that this final result is correct so in this simple example we can but if we develop that kind of system is not to answer things that we already know is to make much more advanced queries so not major yet but you see in the in which direction it may evolve very interesting we'll have to to closely follow the the developments um along that that way so this was for the how AI can support the development of new drugs we could say many more things but we we we have not the time to cover all aspects here uh we will now tackle another stage of product development uh which is more related to clinical data science how we validate drugs and this is usually performed in in clinical uh studies and clinical trials um we have um globally in in the world but in Europe and in Belgium in particular um an issue with uh data interoperability for example in Belgium we we were the first country in terms of number of clinical trials per uh habitant we are now um we have dropped at the third place why because um despite of very um promising environment very rich ecosystem uh some other countries have progressed more in terms of digitalization of their Healthcare data um in Belgium and in other EU countries we let a very large freedom to healthcare practitioners to organize the their Healthcare data the way they want it which is good for privacy there is no centralized Silo with the healthare data of every citizen which is also very good for the the the freedom of exercising medicine but it's not it's not that good for the the common good of of research on on new medicines and as I mentioned in in introduction the a specificity of the innovation in metex so medical device and in vual diagnostic is that we need data before we have the product to learn this machine learning algorithm we need the data beforehand in more traditional drugs it's it's more the opposite we we make developments in the lab we have a drug and then we test it through clinical trials before giving it broadly to the to the population so the data comes later but in AI driven innovation in healthare the data should be available before we start a new R&D program so it's really a need um there there is a lot of things happening at the regulatory level the the European health data space um which among other things will impose on the member State uh much higher level of interoperability a citizen from a country treated in one Hospital should be uh properly managed in another hospital in another country and this require more portability of of the data um in in at the Belgian level this has been already partially translated in in some legislative initiatives namely the the PSI Law Public system information law but there are things to to do yet um and also in Belgium it's very complex because the the prevention the the healthcare prevention has been delegated to the regional level and not to the Federal level while the reimbursement of the drugs is still a federal competency so it's very complex um landscape um and um to to to finish this this course on Landscaping on Healthcare data there's a very interesting study made by the the King bodw foundation here in Belgium uh asking to a panel of Citizen a representative panel of Citizen their willingness to share their Healthcare data for different purposes among which even um private industrial uh research and more than half of the citizen would be willing to share their Healthcare data with the private sector so that's very encouraging and this percentage even goes higher for for um public matters the question is okay they they agree to share the data but what data what what data do they have access to I as a citizen what what portion of my medical records have iess access to very little in fact and that is because not entering into the the details but there is no centralized way of managing Healthcare data in many countries including Belgium these are split uh in the the healthcare practitioners in hospitals in in Regional hubs and meta hubs that are in theory all interoperable but they do not easily allow for retrieving data and so this may be a concern for the practice of medicine but it is also a concern for conducting large scale clinical research there is an interesting project that there are many in that direction but here is one um from the the French speaking part of Belgium which is called Ina Institute of analytics for health which has developed um a technical framework so that uh different uh Source data silos from the hospitals but also here at the bottom from the the general medicine to map the data from patients into a common medical data description model which is called omop it's not invented in Belgium it's an international standard favoring clinical research based on real world data the trick is that it also has to be compliant with the gdpr and so there is there are a series of pseudonymization um with the extra objective that after pseudonymization I should be able as a researcher to um still um aggregate data from a single patient not knowing who it is uh in the different care institutions where it has been treated or seen and so there is a complex pseudonymization mechanism with trusted third parties Etc that allows to do that um and now it works and the interesting thing is that um in the previous years this was a research project but now since 2024 there is an actual legal entity created by Federation of hospitals and general practitioners so by Healthcare practitioners not by the state not by the industry operating this platform now since this year um so I will not enter into the the omop framework it is developed by the the ohd SI which is a nonprofit organization operating worldwide with nodes in different countries and we have one node in in Belgium here it allows to represent any useful concept of of medicine so the drug exposure procedure occurrence device exposure lab measurements observations by the the healthcare professional etc etc and of course the the occurrence of healthare conditions um omop is driven by the ohdsi community and um Ina uses that framework which also makes it compatible with International research but it also benefits from the open source development performed by the ohdsi community and um among these developments there are a series of tools that are clinicians um or doctors to exploit their own local data so if I'm work working at hospital a I will be able to consolidate all the these different data cyos from my hospital for a cohort of patients um it allow to to Define cohort and the cohort definitions can be exported to other uh omop compatible systems in Belgium but also abroad it can also be useful to investigate what differentiates different cause so for example patients treated with drug a but with symptom a and with the Sy a but with symptoms B what differs in those two populations very easily and also to um investigate into the the current practice what is the most likely sequence of uh treatments or or or procedures that we impose on these patients with a given status yes yes so the so the the aspect of Big Data um yeah so indeed so first um in in healthare it's true with genome it's true with the manufacturing data that I will show later but is also true here with Clinical Research in general we do not have access to really big data and the million Genome Project is still at its beginning maybe in China they have access to through the the bgi to larger gos but in general we do not have access to um large large data they are large in the sense that we measure a lot of things for a given patient or for a given batch of drugs produced but we do not have a lot of patients at hand to conduct this analyses or a lot of batches to analyze in in biom manufacturing that we will cover after that so it's not big data in the sense that we would have many rows in our data set and a few columns but it's rather the the opposite what I call Fat data we have many different variables collected on a limited set of examples and this is precisely what this kind kind of initiative try to um to to overcome how can we benefit from more massive data sets um on top of the aspect of anonymization of Pon imization um all the the already the procedures already in place more traditionally with Ethics Committee uh data management committees in the hospital all that remains of course Val these are of course important question um and here I didn't say but in fact when a when a research project comes in by the the customers of Ena Ena first analyzes the the phib ility of the request and if it's feasible it pushes the request to every data provider so series of hospitals and and The Syndicate of general practitioners and each of these institutions may have have the the freedom to opt in or not into that specific research project based on on ethical committees consideration or data management data access committees Etc so like in in more traditional clinical research and the second question regarding biog GPT no it's not a tool that we developed um so the the code for biog GPT is available on a repository of uh AI models called hugging faces you can download it the thing is that if you want to be able to run it you have to have an Hardware infrastructure with gpus um either locally or on the cloud so you can and there are not only biog GPT but a lot of models are available for downloading and installing but you need to have the capabilities of managing such deployment from the software and from the hardware perspective uh so I did not get or my colleagues did not get access to the 15 Millions publications of pet the model has been trained on these Publications not by us but by by people collaborating with with bed and only the the model with its weight is published the sources are not published and so it's a it's in open source license you can download it and and use for each model you have to read the the the provided license and see what you can and what you cannot do okay uh let's let's pursue um just to to give an example on how initiative similar to to the Ina project can speed up clinical research both from the academic and the and the industrial part so um they took an actual request coming at an Hospital from pharmaceutical company stating okay with a traditional clinical study protocol these are the inclusion criteria these are the exclusion criteria this is what we intend to measure on the patient can you identify patients in the pool of your regularly visiting visit visiting patients who could match with that project of clinical study and and this is of course again done after the approval of Ethics Committee and they computed um the time it takes by clinical research Associates within the hospital to go through patient file partly on paper partly in different computer system and to check one by one these this CR this is the way it is actually done nowadays for clinical research and they they um obtain the information uh okay for one patient it takes that much minut I remember it's order of magnitude is 400 minutes per patient to be screened and so if you make a projection you have this red line here to include 100 patient I need 600 hours that is the time I have to dedicate a clinical research associate to go through the F of 100 patients if we go um through the more automated way available through the in a platform of course the time for including zero or one patient is not n because you have to code the the request so that takes still a few a few minutes and on top of that not all the requested inclusion or exclusion criteria are already digitalized into this system so it there remains a manual work to be performed but still um with the current version of Ina you divide by three the time required to screen if those 100 patient can or cannot be included in that specific uh clinical trials meaning that either the hospital can uh improve its margin or improve its margin and reducing the price it will um offer to the the pharmaceutical company or it can deal with three times more clinical research programs so huge avenues for improvement and there are um tracks Improvement of the Ina system that could still divide by two this time namely by better integrating the the genomic data and the data from the from the Clinical Laboratory the hospital another example um the the the first actual use case conducted on this platform uh which should be finalized this summer is to check the effectiveness of Blockbuster by the company BMS so they have not requested that analysis it comes from clinicians um to see if K da which is a treatment introduced in 2019 in Belgium um accepted for reimbursement is it as effective as the phase three clinical trial claimed maybe it is maybe it is not but that it is very expensive clinicians are wondering what is the impact at large scale and so in a matter of a few weeks um we will have the answer on a cohort at the level of a region here that is three times larger than the very expensive phase three clinical trial that will be really interesting to to analyze as well so um this is not very advanced in terms of AI but it stresses the need to have good massive data uh very well documented interoperable that is true for the genetic data for drug Discovery it is also true for genetic or clinical data for clinical research and then once we will have good and massive data we'll be able to do more Advan Advanced things and know um the the last step in our journey of the life cycle of of a medical product suppose we have a new drug we have validated it through clinical trials now we have to produce it in fact producing um traditional pharmaceutical drugs small molecules it's chemistry it's rather reproducible you put the the inputs you will most of the time have the same outputs same quality same number of doses there always issues but on on a on a general basis it's it's relatively easy but all the modern therapies these are vaccines Mr therapies um large proteins that are engineered Cell Therapy Gene Therapies all of these Therapies in the process of manufacturing them producing them put a living entity at work a virus a cell a fragment of RNA and the living is not as predictable as small molecule chemistry and so there is a huge variability you try to put every time the same thing in the input of your process you are never sure that you will get what you want in terms of quity but then the badge becomes rejected and will not reach the market or in quantity and this is how there are series of drugs for which there are actual shortages um and patient are prevented from accessing the drug they need because of shortages also from a more economical perspective it's very important in developed countries that we maintain some uh competitiveness in our manufacturing industry well for obvious reasons economical reasons also for Independence reasons and so we can also wonder uh how Ai and data can play a role in the Improvement of the biom manufacturing industry um this is um a view from the FDA about the three main stages of the life cycle of Biotech product you have first to design the process okay I have the work coming from the from the academic partner or the lab that we just have in license we have a few in vitro in Vivo experiments we need to design a process to manufacture it at Large Scale but this is process design and there um we will see how AI can play a ro process qualification it's it's U maybe less interesting from uh the point of view of of AI is just to produce a limited set of batches to show that we have fixed the process and now we are able to produce but then um you will start producing potentially thousands of batches along the the months the years and you have to monitor on a daily basis what happens monitor the quality before U putting the product on the market but also as a company you are interested in um see okay regulatory constraints that's good we we have to produce high quality high efficiency produ but we also are interested in um increasing our margins reducing our environmental impact because this production is also consuming in terms of well electricity disposable material the the tons of CO2 it rejects so there is a multiple set of interest in having a very effective efficient manufactur and of course of good quality so for the um the process design usually the way it it's it's very complex there are hundreds of steps in the manufacturing of one batch of um biotech product from dozens to to hundreds of steps and depending on the product the complete production of a batch can take from a few days to even a couple of months of continuous operation before you have a final product on the market and the way all these steps are designed um is traditionally be is traditionally made by a set of experts per step so one set of expert are really good at making cell grow in a bioreactor so they will focus on the design of that step another set of experts are really good at harvesting which means B based on the the large soup that comes from the bioreactor how to extract just the the protein that we need etc etc but the consequence of that is that people have not an exhaustive view of the process and the design shows they may make for a given operations may not be compatible with the choices made by their colleagues to doors next or be suboptimal Etc Plus in general they do not take benefit from the design of Prior processes that may be similar um and also there are a lot of um know AI models uh designed for specific p of uh this this kind of process for example to predict cellular growth B based on initial conditions um and so the the Nemo project is a collaboration between several companies um a bin project so it's a resar an innovation project in in French speaking Belgium um it also involves my company the analytics but it also involves UCD which is here behind this ion um and the objective of that project is to be able to design a process at once all steps together by taking into account the ex the knowledge of the human experts plus mechanistic models so these are models more based on on biological biochemical equations and all the past data of the company uh through data driven so machine learning based models and so to to design this process uh at point this is one step where AI can can play a role and the aim is to reduce the time of development of the process and so it cost Etc um there are also more possibilities um of of application of AI the people operators uh in these facilities have to deal with very complex uh machines and and Equipment uh they have to be trained so of course all the field of virtual reality augmented reality is really used now to make people capable of entering uh these very clean zones and not wasting useful uh material for the sake of training so they receive training beforeand on in a virtual environment um as I said this this process are very complex so almost nobody in the company knows the the process in in full so building uh objective knowledge and sharing it between colleagues is also an added value of data science and and AI in general you have plenty of other possible applications such as okay I develop a process from in in a factory in the US how can I be sure I can replicate it fastly in another Factory in Europe Etc and of course uh process Improvement and I will show a concrete example uh after there is also um in in cellular therapies no there will be an extra challenge so cellular most cellular therapies begin with a sample taken from a patient suffering from a disease this tissue this sample will be engineered in a specific way and then reinfused in the patient and the reinjected modified material will heal the the patient so that is cellular therapy in that kind of strategy there is no interest in having massive factories such as the the traditional biotech or pharmaceutical industry have like one two three Mega factories per continent to the contrary there will be very small factories here you see actual examples this is the size of a maritim container um where the complete process can be implemented in that kind of factory the the benefit is that you will produce the drug much closer to the patient you so you you would typically have a few dozens or a few hundreds of these units at the scale of a continent close to the patient because it's it's nonsense to ship a sample from a patient at the other hand of the the planet to manufacture something so specialized the Downs side of it is that we we had a question about the amount of data available the down side of it is that for each of these small factories you will have even a smaller data set to monitor the quality of what you are doing the the constants of your operation um and so there is a need to opt for what we call in machine learning Federated learning how can we build models based on local very small data sets and still extract more robust knowledge at a global scale and so this is um a project that we conduct for example with the company or Genesis which which at the same time designed and manufactur this small production units but on top of that actually manufactures cell therapy in different countries they're already operating right so a partnership with University of um no uh let's uh go through maybe I had to I don't know how how much time I still have well I will try to to finish um I will not go into all the details but this is a very now traditional but a very Advanced example of uh what is done uh to improve the yield of a biom manufacturing line uh typically um it's phrased like that okay uh in the future tomorrow and and after I want all the batches that will be produced to have the same yield as the top 20% yields that I observed in the past how can AI tell me what I have to tune on my manufacturing line to reach that objective um a really huge uh obstacle here is that just as we saw for uh Healthcare practice data we have the same siloing of information so raw materials information will be stor in one it system called an Erp uh All Quality measurements will be stored in another system called the limbs uh all the the environmental monitoring of the zones where the produce is made um is St in another system and so there is a a huge challenge to be solved how do we reconcile all all those data to go from 90 and Regulatory perspective on the storage of data to a business way of structuring the data how do we make the data meaningful again for people which business is to manufacture drugs so this is a a really recurring challenges um to describe one batch of vaccines for example you can have up to 10 to 20,000 measures just describing one batch uh so we again have data where we have plenty of descriptors but a limited set of batches here to to build our model so this is a really large challenge I I will skip but the message here is that we need um that the models that are built are validated from a data science perspective with engineer metrics once we are satisfied with that we have to give um the process exper the feeling that they can trust these models and this go through a dialogue with them it's not AI will solve everything it's just AI will provide tools to reason and to move faster we have to provide trust and we also have to um make useful models if all models recommend that I move the uh I don't know the the pH within a bio reactor I cannot there is no button to adjust pH I have to act on other reagents to adjust the we have to build models that are validated trustful and useful and then it's not sufficient to make models the model should be put at use to make recommendation on how to reach that go so that goes through an optimization um yeah and and it can pay off in in a recent examples we shared with a with a customer we reach Plus 15% yield on a two billion uh Euros a year product it's it's not nothing uh yeah already talk about the difficulty to assemble data um I I invite you to read the gmlp position paper it's very accessible two pages very a list of good principle that should be taken into account for implementing AI in Medtech but it's also Val for the the other uh stages that I discussed today um and the this position paper but also the the guidance the gam five on how to develop software code uh integrating AI in manufacturing industry and also the AI act they all have in common that they put the stress on um informing and educating the users the human in all these uh regulations standards good practice they all put the stress on keeping the human in the loop educating users uh advertising users and I will I will finish by that um so AI in medicinal product life cycles uh has a huge potential at all steps of this of this life cycle the Gams may be financial and environmental social um a lot of regulations already in place or upcoming apply or will apply and we have to to keep that in mind and the the main roadblock is the data availability and uh structuring and contextualization and this is more General than for healthcare but when a user is facing a digital application Poss encompassing AI aspects um the user should always ask himself am I facing an information that is generated by a human or by a machine and if it's by a machine what is the goal of the AI system that generated it for what has been what has it been trained for and is it good at reaching that goal I can trade a very poor model it's not useful um and um yeah so be critical the potential is huge but there is a need to know where it can lead us um and not overestimate the the promisees be reasonable with it thank you well thank you very much tibo for uh conclud concluding this very interesting uh course um before we um finished this course uh does anyone uh whether online or onsite have any question sorry I've been a bit too long um so if you develop a new molecule or or if you find a promising molecule that can be used as a drug do you take out the patent on that or is this done by your customers or so so we help our customers um integrating those tools or using them we are not our self developers or of diagnostic tools or drugs so we are mainly service providers MH so if if there's intellectual property get generated it's taken by your customers or yeah yeah okay W um well um yes H thank you very mucho just by cuity the the tool that you show us the video uh from GSK is it only internal used by GSK or something that everybody can the video is publicly available the tool is only for their internal use as I said it's it's very exciting I think there is the road is still long before you will have a major tool which you can trust and move on and in now you have been involved in the creation of the the process or we contributed to its development and we still collaborate to some use case but we are we have no I have no Shar in there it's purely operated by Healthcare professional representative and that is something we we fought for in the design of that platform if it would have been led by governmental agencies the healthare practitioners will feel would feel um constantly observed and monitored if it would have been by the Private Industry people would quite legitimately a fear that their data would be used for anything maybe without control so it's just the right balance here it's used managed by Healthcare professional but also with the intention of making private sector benefit from it but in a in a control it's really meet the requirement from the the data governments actor it's kind of data organization excellent thank you well um if you have any further question obviously you can contact Mr tiut he he share his contact on on his presentation um well in the meantime thank you uh both Mr Hut and Mr Alo for coming here at blueo obviously we um um we also thank Mr Briganti for um giving the lecture um from uh abroad um so for the next class the next class will be online and obviously stay tuned for more thank you very much for coming