Spotlight Talk: Drug Discovery using Chemputation by Lee Cronin

uh so yeah I'm just going to introduce myself I'm yakob uh I a uh PhD student at Imperial College London and I've been part of nucleate since uh last uh July and yeah I've co-founded the a b initiative um which um this talk is a part of uh and uh I would just like to I I will send the link uh in the in the in the in the chat uh where we include all the amazing things that been running uh in January this year we've run an AI for drug AI for protein design uh sessions and then this month we're running AI for drug Discovery sessions uh in March and April there's going to be uh more sessions coming uh on cr1 Automation and then on uh data management and regulation compliance and then another biomedical Ethics in AI uh session what I mean by sessions uh We've run two virtual panels uh we've got amazing amazing people coming in from all over the world uh then we've got also overview talks uh and then we've also created uh resource Pages uh where you can see you know the most recent models or data sets used Within These spaces uh and and go from there it's I think it's a very nice like intro uh that either PhD students or people just trying to get into the field might use well uh also uh we've got a couple uh imperson events uh coming in uh quite soon for example there's one in Los Angeles so uh will send the link tree uh in a bit in into the chat so if you're in Los Angeles you might you might stop by but then there's also events uh coming in Toronto and Montreal and Munich uh in March and April and there's probably probably more coming uh after that as well so make sure to to follow us on on Twitter and Linkedin uh so you can keep keep in touch with us uh but that's all from me uh thank you firstly uh very much for attending Lee uh and very much for for the attende to to come in as well as fully uh to to to to be eager you know to give us a nice talk about his uh company or right person is going to be about his company cifi um and yeah Lee I'll just I'll spare you you know this me trying to introduce you maybe you can introduce yourself uh in in a few minutes and can then we can start off uh can you share the screen you should be allowed to do that right yeah yeah I will be a to do that yep awesome that one one second oops one second share screen okay hopefully you can see that awesome yeah we can all right after to you leave okay great so it's great to be here so what I'm going to do is um I'm so my name is Lee Honan I'm the Regis professor of chemistry at the University of Glasgow and I'm also the CEO of chemi and chifi is uh digitizing chemistry and enabling um the uh I suppose the production of molecules and design of molecules in a a completely new way um what I'm going to do today is kind of give you a lot about the background of the technical stuff that gave rise to it so a lot of this is kind of how it what was being done in in Cronin lab a lot of what's happening in KFI is um obviously um as we're you know we came out of stealth a little bit last year but a lot of the the main underlying kind of operationalization that we're doing is remains obviously uh a kind of part of the secret source of the company but don't be disappointed because what I'm going to actually explain to you is far more important is about the underlying uh ability to digitize chemistry and what that means and there's a lot of interest in this area because getting getting chemistry digitized is the first step for making AI work in chemistry I'm I'm not a really big fan of current AI right now as it understand because I think there's a lot of people misunderstanding what AI does but hopefully I'll explain you why I'm super enthusiastic about using embodied Ai and methods for exploring chemical space and today we're going to focus on how digital chemistry and uh will help change drug Discovery in the real world and so that's kind of what I'm going to talk about so I'm going to talk about and it's could be quite geeky high level but I'm going to go through the slides quite quickly and and hopefully when we get to the questions there'll be a lot of questions in terms of uh understanding this technology because I think it's really important that people do understand this isn't just a new uh kind of incremental change in one area it's an incremental change in a lot of different areas which give chemistry Universal reach so so what I've tried to do is enable chemistry to become Universal so I'm going to explain what that means and what computation in the computer is and comp the computer was invented by me in in my research group in about 2012 the concept of computation kind of emerged over the last few years and I think um it's important to understand uh the what what computation actually is for chemistry it's not just a fancy word for a machine it's actually uh a significance to chemistry is of equal significance to computation the word computation to mathematics which is a pretty profound change I'm going also talk a little bit about reaction Discovery and how to explore chemical space and also drug Discovery and uh and we'll actually Focus just on drug discovery not materials and complex systems really um so the let's let's start at the beginning in digital chemistry um and the important thing about understanding the digitalization of chemistry is that chemistry is kind of got a lot of abstractions um in terms of you know the concept of molecule and a bond and things um it's important to understand um how chemists do that every day but how the connection with um digitization has kind of not been made yet and I'm going to try and see explain how we've been making in Glasgow in the digital chemistry Center and also using it in chemy now if you start at the hardware matware you've got you start off at the bottom of the computer you have devices and transistors and chemistry you have atoms and bonds and then you can go up in in computer science Gates and registers and chemistry we and solvent in in uh again in computer science or electrical engineering you have microarchitecture and in chemistry of reaction Hardware now you can see parallels between this what we do in the lab this may seem really boring and Abstract but it's absolutely critical if we are going to be able to do drug Discovery using AI because we need to be able to make molecules for real robotically programmatically flawlessly um uh in the same way that we're able to run code on microprocessors with error correction as you go up from the microarchitecture of instruction set architecture and chemistry of synthetic conditions machine code in chemistry process control in Assembly Language reaction procedure programming language retrosynthetic analysis algorithm chemical programming and all the way to molecular design and application and look at the parallels here you're able to go from matter Weare to software and um and make a a connection between the molecular design and the ability to make that molecule programmatically and so chemy is kind of the full stack in digital chemistry it is a we're able to go from you know the concept of the molecule to making the molecule but again this talk is really focused on um uh non-confidential um stuff that I've been doing in Academia which really underpins a lot of doing chemy and you can put a lot of the other things together um few years ago we started off trying to think about how could you actually implement this right how can you make sure that you can come up with a a literally a computer to make molecules what would you need so from a very practical point of view we made a reader um which seems trivial now in the work the day of chat GPT but actually chat GPT doesn't solve this problem in fact you would never use chat GPT to do synthesis despite the fact many people are kind of you know using it to run robots that's totally the wrong thing to do if you want formal reasoning to work so we basically built a programming language called kidl and kidl is the python of chemistry and we and we made this kidl code is to kind of just take the literature the chemical literature and develop a machine independent um Hardware sorry pro programs to run those those uh programs on a on a on a computer uh on a computer and we have a virtual machine and a way to compile the graph to make that work to do this you would the reason this is was super hard and never happened before in chemistry is that chemists didn't really understand the need for for universality and um it's actually very important that we have universality in chemistry um we already have it conceptually but there is a cheering thing called a cheering machine which explains mathematically how computers today are basically um able to kind of emulate each other it be a digital if you like for each other and that's what we wanted to to do in chemistry It's like because Universal chemical processes will be provably complete what does this mean for the lay person it means reliable it means that you can make complex molecules using Code it means that you can move code around and the recipe um you you don't need to worry about because everything is in uh encoded precisely with error correction so um for fun I made a in the lab using some of the hardware we have in the Cronin lab you can see this really complicated looking graph on the left hand side which has heads and a camera but I can show you um uh this fun thing here where we made a little touring machine just moving reagents backwards and forwards and you can see this is the input tape and this is the output tape where we basically took chemical saw the chemical with the camera and put it below and all what we show what can show here is possible using image recognition moving solvent around using pumps and valving to basically emulate a touring machine and this is kind of hilarious but actually very profound because it shows if we can have a if we're able to move chemicals from one place to another and see that they're moved and then do operation on them and use a sensor we can start to see how we can program uh at the abstract level uh chemistry and this is the a for a computer scientist it's trivial um from a informatics point of view but from a chemistry point of view it's trivial from a chemistry point of view but both needed to be done um in code and to be programmed so why is making computable molecules why is computation important for molecules well if you can think about chemical space this video is showing here chemical space is very big the number of molecules potential drug candidates and chemical space even for for the modest set of elements is essentially infinite so how do we address this issue how do we get access to these molecules because you have a set of molecules in this space which are physically Allowed by the laws of chemistry in physics and then you have molecules that are synthetically accessible by the synthetic chemist in a fumehood but what we're trying to do is make sure that the a large fraction of the molecules that can be made by a chemist by hand can be made using computation that is those molecules are computable that is that the the co code and reagents can be used to make those molecules ules rather than a chemist doing it by hand now how do you do this well you do it by automation but the current Automation in in chemistry does does a mere fraction of what is possible so the examples of automation include peptide synthesis nucleic acid um um Olo nucleic acid um synthesis and some HTE for say cross coupling chemistry so it's really very limited so it means the complex molecules that people make cannot be done automatically um but computation tends aims to solve this Pro this gr this problem it's a grand problem so what we started to do in the Cronin lab before we we went to kind of you know um kind of develop this further we came up with this standard for software and Hardware which would literally allow us to add liquids add solids add gases and to basically um use these modules to do chemical reactions and to explore new chemistry and to make new molecules and then once you've made those molecules then test those molecules so you have this idea of design uh in in silico in code using kidl and then make it using the robot and then test it and this is where computation for drug Discovery really begins to take shape the dream I had was this back in 2012 was to basically build a robot that would you know um take code and produce uh in this case it's a 3D printer 3D printing an architecture and then the 3D printer doing the actual chemistry um and this was kind of fascinating that we were you know even imagining this integration and then go fast forward to 2014 where we used a a rip wrap printer um to actually make um a test tube a 3D printed test tube and in that test tube we added in the reagents to make ibuprofen and so you can see here this 3D printer that was kind of a kit that you bought and put together we 3D printed some pumps and then we connected the pumps to com ports and we together with the the STL code the code used to to print the test tube we modified it so we could then use the same STL code to do chemistry so think about that the same Co type of approach used to 3D print objects on a 3D printer we kind of took and used the build this kind of programming language that would be kind of incorporated into existing robotics Jeep Robotics and that was kind of fun to do that that was when I realized you know between 2012 and 2014 that the computer concept was going to work but the need for any Ai and for any uh robotics you need an abstraction and in chemistry the abstraction that to implement to encode is the con we just have to do four things in this case have a way of of doing a chemical reaction a workup isolation and purification those four steps put abstractly con constitute um the the process of doing computation and so the kind of programming code we built was able to kind of we started to translate the literature into kidl and this allowed us to teach the robot to do more and more chemistry and so and we could develop the chemistry of the robots and give it more modularity that you can see here there's a picture of look quite complex looking robot actually from Cronin lab where we're able to do um uh kind of maab boronate chemistry cost coupling chemistry um make some diazines so for labeling and also to make um peptides in the same robot so kind of uh example of how you might go about you know using uh the robot to do computation is to basically come up with a route and turn the route into a code you know using a search engine engine and then basically take that code and add it to the robot and this is what this video just showed you here it showed you the process of right the beginning submitting a smile then it was just running the retrosynthesis on that smile and that smile then is drawn into these molecules and then basically we pick the the robot knows what what um reactions it can already do so it automatically devises a roote from reactions it's done before so this for the first time turns the process of retrosynthesis into a a robotic process where the the chemist doesn't need to do any work they just input the molecule and say please make the molecule and then that code is in sent to the robot as you can see here and the robot goes through the four steps I told you reaction and work up you're just seeing here really sped up Ultra sped up and then followed by work up um you then do isolation that you can see here evaporation in the rotary evaporator and then after the the the evaporation to produce crude product the final step is the purification by crystallization in this case and you can see end to end we've just put a code into the robot and the robot just makes the molecule using that code and and the drug is in in this case the API this one I think is the final step and Viagra symphasis actually um is put in the jacketed filter at the end so you bet produce the product that way so kind of That's great so we've kind of shown how we can um make this U make a molecules using this programming language I've shown you how you can run the robot and we can you know we can now start to really show how kidl is really going to be the standard for uh chemical description language and as we added in more logic it's become a programming language fully T and complete and now there's hundreds of thousands if not millions of kidl available and we've done we use it to do projecting group chemistry heter aculation CC B formation multic component reactions generate libraries and do multi step synthesis you can see from a point of view of just using the robot to for sheare chemistry techn technical expertise the robot is really the system is really taken uh accelerating quite dramatically and I can't um overemphasize the fact that the cultural barriers we face in synthetic chemistry are quite large that a lot of people don't believe that this is even possible because chemistry is so hard so manual so dangerous and hopefully in this talk not only am I going to get to explaining how we do AI get clean data for AI but also dispel some of these this ideas that the chemistry cannot be encoded um these are all the kind of reactions that that the that we the Cronin lab published in the 20 uh 22 science paper um where we were just able to show that you could just take a wide variety of different reactions encode them make the kidal files and then run them on the the robots and then it was pretty amazing that the chemistry worked okay um there were lots of kidal we digitally validated there are now hundreds of thousands possible we've done over 50,000 chemical steps 880,000 unit operations you know thousands of hours of Total runtime so that's kind of awesome that you've got the statistics people would say oh but can you do column chromatography and hlc and all of this of the answer is of course yes I mean there's a lot of automatic purification systems out there we just add them onto the onto the system um that we call in the chronin lab the computer but uh in in the in the chemy it's called something different um and where we have this kind of kmu that controls the symphysis you have a chromatography module you run the pep do sample injection return the product so you really have this end to endend kind of idea for molecule make the molecule purify the molecule um and so this really at the end of this section I've really want to kind of just finish this section by saying look we really can program chemical synthesis but what I'm going to do now is explain some other modules and then get into the drug Discovery part because one of the things that's critical in chemistry is you cannot do AI in chemistry without having data so now ai is I think a bad term I would call I would like to call it machine learning or kind of data uh exp exploring chemical space with proper chemical data um because people misunderstand what AI is and it's kind of infuriating to say the least um we can also extend lots of different modules into the system we can do photochemistry microwaves electrochemistry there's an example here of doing microwave synthesis you have to be very careful because you can see you might have seen on the left hand side here the bottom there is actually this is where the microwave is we have an actual microwave on the coax into a that goes into a round bottom flask where we have to be super careful to make sure that's in a faraday cage because you don't want to MIP way the chemist um so this is very much a proof of principle and this is nowhere near any Deployable technology because it's too dangerous but we can we can do very effective chemistry um as a function of the microwave chemistry we can do also highly reactive chemistry in in in these systems so um and this is super important this is one of the mo major things that we we are going to be able to do from to impact um industry from Cronin lab be at chemy and other Industries just to make sure that using computation using computers we can actually um uh um uh um do chemistry automatically you would not let a chemist do manually because just too dangerous um and so we can do chemistry like click chemistry aside chemistry nitrations do organolithium chemistry ozonolysis all this chemistry is super important for drug Discovery and making complex molecules but you wouldn't want to make too much of some of these molecules some of them are hazardous you might want to find other roots and what we basically the way we've enabled this is we built um a load of uh sensor systems that connect with the chemical programming language to check that the the the chemistry is going as proceeded so what you can see here is um let me just get the video on so I'm going to play this video here on the left and you look in the top left hand side you can see the reaction flask changes color and we can monitor this change in color and uh as we're add adding The Bu and uh making sure that the reagent which would is pyrophoric it would spontaneously catch fire in air uh does reaction and we're able to move the uh the reaction and quench it and the way we do that is we monitor the addition of the LDA make sure the the valve pump is moving look at the valve reactor the reactor and the quencher and if all those things are moving correctly we know the reaction is failed has succeeded and not failed and the event that any of these things don't go to plan we fail to safety and add it to the quench as quickly as possible so there's no danger of fire and that's really important to be a to do this type of chemistry might seem mundane but really critical um we've also kind of Taken robot the robots and Blown Blown them up on purpose using uh um explosives and under control conditions this is a video here just showing that what actually can happen um just to make sure it's safe you can see it's this is actually using quite a lot of plastic explosives and it's contained um you'd never have this amount of uh reactivity in a robot um but we just wanted to make sure that we designed the systems so that they could explode with no um and with with no failure of the contain the the secondary containment within the fum Hood so we're really becoming very excited about the scope of the chemistry that can be done in the reactivity control again critical for data collection critical for active learning so to kind of get the data that we need to discover new reactions particularly important for drug Discovery we need to be a to collect real-time data and have all sorts of sensors for this a color sensor pH sensor liquid sensor uh conductivity have integrated NMR hlc Ramen and a good oldfashioned very basic um web camera you can get so much from videoing the outside of the in videoing the whole F hood and uh and we have kind of you know world leaders kind of been able to use this type of let me sorry go back using um computer vision with chemistry with the programming language to basically set up the reaction execute it optimize and then check everything is safe and that's what this slide is showing that we can do this have this uh have this Dynamic execution you know you could basically tell the robot to um add reagent to color is orange or add reagent to bubbling stops things like this a chemist would would do normally by eye and by hand but you can explicitly encode this and so these are examples of sensor data here we can monitor exop firms we can monitor um discoloration um and we can look at U RGB and look at kind of the opacity as well so there's lots of things that we can do um to to get realtime feedback and this is when we we'd use another type of kind of self-driving if you like so in the same way that you have uh a kind of autopilot in a Tesla well a lot of the chemistry that we can do we can train the robot and then have an autopilot semi- autopilot so the the robot can learn by doing new chemistry on the Fly and and encode it explicitly um I won't say too much about because I I know I'm I want to finish in the next 15 minutes or so to our time for questions but we can also do things like repurpose the computer to do peptide symphysis this is solid phase peptide symphysis and then that gives us access because we use the same programming language to get access to all the other types of chemistry so you know you know you're not then restricted now to just doing peptide symphysis you can do all the other chemistry that's available to organic chemists under automation at the same time when you're you're doing the um the peptide elongation you just have to make sure that everything is compatible and if it's not compatible then you can change the you know cleave off the peptide move it to another flask do further chemistry um with that and move it backwards and forwards and that's really exciting that we've um and also the cost of these systems now is very very low compared to uh what it cost now to BU to get some of these commercial um multi channel synthesizers so the the peptides also use kidl there's a general kidl blueprint for making peptides um which is compatible with the uh organic symphysis kidl which is beautiful because it means we can express the entire synthesis of a peptide including nonoral non orthogonal um chemistry um that you might you might not be able to do and a peptide synthesizer explicitly and uh and that's pretty awesome uh yeah so we've also done Organo catalysis in the robot again this is very detailed so I'm kind of apologized for the complex of these slides but the basic idea is to make chyro molecules you can do organometallic chemistry um deprotect and then reprotect and do this in cycles and make very highly complex chyro molecules um and um and and also recycle the materials so you can you can you can make it very cheap um and this allows us then to imagine using the kidl files and the process chemistry as a kind of blueprint like a like an um and also like a li library in your in your program in the same way in Python you will call different libraries to have different functions available you can do the same thing in kidl call a specific function that will do a specific set of chemistry and then Returns the molecule back to the robot in a certain way ready for for further elaboration so um so what I'm going to do now is I'm now going to move into how we've been using uh a different approach to drug discovery AI using proper causal graphs um but what I want to do is summarize this by saying look um it's possible now to build robots that can do all of chemistry in principle they are Universal they right now they can do probably about in about 90% of all the chemistry in Med chem space that's what chifi is exploiting um and um although there might be some specific unit operations may be outside uh the remit of the hardware in principle it's possible to develop a hardware module to to do that right it's just a matter of cost but the fact it's all using the same software environment the same programming language the same kind of uh digital twin means it's very extendable and this is really important we have developed the first Universal extendable programming language that is general purpose and we'll do all chemistry Material Science formulation and this is something that is quite extraordinary I think um now one of the things I wanted to do in in looking using a new approach to evolutionary approach to a type of AI which I call evolutionary uh AI this is not um to drug Discovery is I wanted to work out how we could search chemical space uh using uh a new approach rather than doing fragment-based drug Discovery and so one of the things I developed in the last few years is a thing called assembly Theory which it actually allows us to tell how selected a molecule is and on this slide here I kind of give an intro to assembly Theory but I don't really want you to worry about that I just want you to know that there is a Technique we can use to basically start to design molecules using uh um assembly algorithms so that's the point I want to make um and to give you an example here is we we have been thinking about how you might take say the opiate family tree um and the opioids um to kind of understand how we might make new molecules in the library that may have better um um capabilities less addictive better at pain relief less poisonous if we can understand how we can you know um make hit the the the various receptors in a way that you can reduce addiction and reduce toxic make molecules that are less toxic this could be transformative I mean Fen is a terrible molecule right so it's like so one of the things we've been doing for fun is looking at the evolution of um of molecules darwinian Evolution this is not not genetic algorithms actual darwinian Evolution and apply this kind of causal Theory to it and what we've been able to show is that by sampling chemical space we're a and we are able to kind of you look at what building blocks are conserved and to make new type of structures this is kind of similar you see this in protein sequence protein Crystal structures all the time but what we're able to do is make a chem informatics process where you can actually look at the how evolved a molecule is by the biological system and then build new molecules in silico that are going that that are promising drug candidates for hitting a particular Target and the way you do this is is I call these like alien molecules where you're able to take them take the natural product space take the mole UL that are known already and then uh kind of erase their history and apply new selection pressure to them and come up with weird and wonderful molecules that are just completely different to what you would imagine using traditional fragment based methods um in this example here we basically took the coconut database of the natural products built a joint assembly space of all the natural products and then went back in time and then re reused the basic units to make new molecules reassembled so what assembly Theory literally does it says hey let's just take all the assembled natural products that are interesting for for a type of biological process disassemble them a bit mix them together reassemble them and look at how different they are and then come up with synthetic roots to make them and see if they are val valuable potential nutron candidates and so what we've able to show is that assembly Theory Maps chemical space and and this is a a totally new way to um to map the evolution of natural products and therefore to explore how we can develop new drugs based on non-natural products that are U that could be the next generation of natural products built by biological evolution indeed or by thinking about how we might do New chemistry that biology doesn't have available to it so we are able to map The evolutionary trajectories of the molecules and then disc c new chemistry all the time um and this is just an example here of how we can take take well-known molecules with high assembly index the complexity of the molecules given by What's called the mcro assembly index and then we can erase the complexity and then build new molecules and this is example here of the kind of new molecules we can get as we start off on the left hand side with well-known molecules that we know exist and then erase their contingency um and rebuild them and make these weird and wonderful molecules that of just completely new uh um to to chemistry so what I'm going to do now in the in the final kind of um uh 10 minutes or so uh eight minutes or so is kind of Crose a loop on this I've talked a bit about Robotics and Ai and how um you know a new technology has been built in glason digital chemistry lab obviously this technology is a is kind of inspiration for digitization of chemistry and and uh companies like chemi have used that inspir using that inspiration to indeed uh build endtoend processes for digit chemical digitalization but I really want to talk about the price here like what what is the most important thing that I think the the the digitization of chemistry will give drug Discovery in the next few years well this is the ability to basically design and make molecules faster than we can right now with a human in the loop and it doesn't remove the human from the creativity quite the opposite the human the chemist is now center stage in terms of the creative process but they're no longer limited by manual unit operations and so what you want to be able to do is come up with a new kind of design CR new specification for molecule and if you can design that molecule in silicone um one of the things you want to be able to do is understand how easy it is to make that molecule not just using a um uh a synesis a synthesizability metric but actually outputting the code the actual code to make the molecule you know in chemistry the proof of the pudding is in the making right having a synthesizability index is not good enough um because these are just benchmarks developed by computer scientists with arbitrary data sets and those data sets are don't have the accuracy um that we need and it's very easy to generate surrogate data that doesn't reflect what we see on the laboratory floor as it were so one of the things we've been building is a chemical search search engines for chemical space so having a hybrid human AI that will use basian uh um interpretation of chemical reactivity for high throughput reaction Discovery so this is something that is very important for the first time in um in chemistry we are automating Discovery all if you think about how reactions are discovered they are discovered by accident nobody not you to wake up in the morning go you know what I'm discovering a reaction today some people might have an inkling of transformation was possible there obviously Nobel Prize has been given to genius chemists that have you know really thought about this and got gone and try to make these discoveries but they normally take tens of years and um it's a it's based upon Serendipity here we want to use basian operations to to do to make this work but the problem is if you didn't have a kind of programming language and a way to State the the process chemistry encode how do you encode for Discovery right it's really important and that's why the computation Paradigm allows you to build on that for reaction Discovery and to do this practically you have a robot that has a reagent module reactor module and Analysis module we do the analysis with NMR MPC and hlc and you can see this kind of quite simple setup right but the flexibility that we have with this compared to say an XY uh uh robot that would basically just do simple unit operations is quite phenomenal um so one of the other things we've been doing Beyond this is also using and this is going to be P this is coming being published next week I think where we've made an electron density uh uh uh model system that where we can take a host and it will generate generate electron density within that host uh using uh well- defined machine learning methods and deep neural networks and we've been using this to invent new Mo invent new electron densities that we then go synthesize in the lab so the idea is you have a chemical space we trained a model on to work out how to make electron densities we would then feed that model a Target host um it would then generate electron density to bind that that host and then output um the the the molecule to be made the smile we would then make the smile in the C computer and then test it and um and the workflow is pretty uh um in depth here where we use a a kind of have a theory database we we been basically turn the electron we combine that with the the electron density with the UCP um and then train the model to to sample the the chemical space and then um and then at the end we can then use that model to generate Smiles that maximize uh um the electrostatic potential minimize hysteric hindrance so we can basically come up with binders so you can see how we could do protein Lian design here um and so we were able to go from densities to targets for symphysis and we had a kind of little um uh GPT model for this and there were a couple of examples here where we took these hosts this uh this CB Host this plaum cluster host and then we generated these new guests using the electron density system and then either made them and bought them and show that they binded so um and that was really amazing because you could imagine now taking a protein where you what you think is a target then showing it to this system generating electron density where you want it to be in the protein and then generate new lians and you can also play around with a confirmation right you can do kind of um similar Pro steps that kind of people were using with um Alpha fold and so on but but this is a new way of generating molecules using electron densities it's computationally reasonable expensive but it's a lot of fun but now you know closing us down to the final kind of points is we've now got this robotically enabled drug Discovery chemist where we're able to kind of generate new molecules in silico uh kind of come up with a workflow for them to bind to uh see how how well they interact with our Target um we can use electron density methods to generate new new new lians as well put them through our retr symphysis engine and literally output the kidl code so the robot is like a search engine for chemical space and it can test biochemical biological molecular pharmacological problems depending how you express them and make the actual molecules and those molecules can then be tested um um either in phenotypic screens we binding assays and then eventually in more complex systems maybe organoid models and then go going all the way into the clinic and that's kind of the dream that we've kind of trying to to generate here whereby we have a we have this discoverable chemistry is increased dramatically using Computing because we can establish design for the molecule we can turn that design now into a code and make it more General we execute that chemical code um and we can and the outputs of molecules and the functions are cutable kidal codes um we're also having some fun and this is this is not in chifi this is in Cronin lab um building portable robots that we can put around the world um and just to show how simple the chemistry is becoming now sorry how simple the engineering is it's really important the engineering becomes simpler and cheaper for doing Universal chemistry it's very much um an important point that if you make the robots more and more complicated and more and more costly they're just they're not going to work the way you want the programming language isn't going to be Universal um the the the actual the benefits will be decreasing and this is really important to understand that I think still um in digital chemistry lots of people are throwing robotics and code at things which is non-uniform non nontranslatable and people spending too much money on robots and not enough money on data uniform data and and making sure that we can uh collaborate using that data um so kind of philosophical thoughts before I finish you know why why program chemistry why does chemistry need it well it needs it for reliability and operability collaboration remove ambiguity lower cost increase safety open up Discovery molecular customization but most importantly most importantly if we want to use AI um for drug Discovery is we need to be a to publish executable chemical code and collaborate and this is one of the things that I'm really excited about doing in in my research lab also beyond my research lab across collaborations in Academia and Industry and also through chemi is the ability to turn code to molecules and molecules to code for publication of executable chemical code and ultimately chifi you know core stack is literally about turning um digital designs into real molecules um and that's kind of um I hope U being um uh interesting to know this is my the Cronin lab um um that have been working on this we've been funded from DARPA Google um breakthrough initiatives Bill amender Gates Schmid futures um NIH um and and of course um if you want to know about chifi you just have to go to chifi DOI www.m.ssaze.com so hopefully um this has been a kind of little insight into digital chemistry and what KFI is doing and I'm happy to take questions awesome thank you very much that was very very insightful and I've heard you a couple of times on Lex Freedman as well as heard you talk at the London event and for the first time actually have saw some like slides and actual actual signs normally it's like uh just just verbal right this has been very insightful so anyone who has questions please post them post them in the chat or in the Q&A session I just opened up the chat I've just opened up chat and I can see there quite a lot of questions we can you're the boss so we can just go down so I just wanted to you know use the privilege of of of of being here and actually ask you the first question myself and maybe that that can uh to give some time to others to to to to you know think of something but as I so I wonder like with this kyl probably like I'm not sure how old that programming language or how all of this research like how old is it but it is but uh I wonder uh how does this compare to uh the stuff that for example Andrew White has been working on with chem where you try to sort of have this generaliz General like llm algorithm that picks out the tools to do the scientific discovery for you whereas in your case it seems that it's much more um let's say maybe I guess rule based uh Bas and optimization based so how do you how do you compare these two things and do you see a a point where you would also start using some some of these agents uh to I don't know to as I say automate scientific discovery or what is your take on that um so I I have a lot of time for what Andrew White is doing I think he an exceptionally talented individual and I think he's now um at future house which is about using AI to automate science um I must declare that I don't I don't I'm I'm a My Philosophy is that um AI can't be used in Discovery because AI is inductive and you need to be gener rate explanations having said that having said that the stuff that Andrew and is doing and chem cro and all these are really great for generating taking out the tedious parts and making new suggestions and finding what's already out there so I I don't know whether it's going to um be be you know generative in a real Discovery sense um but I there's no question those tools are going to be extremely powerful and I'm excited to use their use them and work with them I think that people have generated hype associated with them and that hype is doing the wrong thing and and I know that there's a great young Professor in I think it's in uh where is he I think it's in uh in in um Cari melon I think but Gabe Gomez he is using llms to play open trons and that's cool but there's no way you should ever be doing that without supervision and I think to give Gabe is is not doing that without supervision why well the same way that generative AI they just lie they're probabilistic and so the synthesis chemist you need to basically make sure you're doing non the chemistry is compatible with your hardware and also that you're doing things safely and there is a formal method to trace it now formal methods of computer science are really important but and I don't need to go over them here but I think to kind of answer your question very quickly those techniques using generative Ai and the agents are great for giving you information to help you aggregate stuff to come up with new experiments they should not be used to run robots they might help you um in some way and we are using those tools but having a formal programming language means it's formable can be error corrected and checked and will make sense to people so I see there there's they go together wouldn't it be great to use those agents um to help come up with new ideas and suggestions and then you g develop new code and you run that but we mustn't allow people to connect the agent to the robot directly because it won't end well but and I'm not an AI Doomer it's just that the the just the robot will probably catch fire and chemists don't want their robots on fire and that's kind of it but I think there's a lot of scope and there's some very good people out there doing stuff you know I didn't mention that I'm working with d Dar funed project with Alan aspu Jason High Marty Burke and B and Bartowski doing fantastic work there uh there's great work going on at MIT uh in Cambridge and UK everywhere right in looking at this so there's a big rush on but the programming language is something is quite unique to what we're developing I've been working a lot with Jason Hine people in Berlin um ALS also kind of worked with uh some pharmaceutical companies and I think it's really important that people start to consider using the programming language going forward because it is the way to basically make sure the chemistry has greater reach all right all right that that's a very extensive and comprehensive answer I think you touch upon uh good good points uh maybe to tie it I have a really good question that somewhat ties to what you said about you know using llms for maybe interacting with the literature or or creating hypothesis so does your uh does kidl or the pipelines that you're using currently use any extraction of of information from literature to to to quote the question specifically you see much value in literature data when you when there's little standardization in reporting such differences between groups so um let's be nice to the literature the literature 200 years of literature all these chemists work in good faith there is a absolute bucket load of really interesting stuff out there I actually have funded by Schmid Futures in my kind of nonprofit haton I've been basically making a kidl uh database that connects the o to kidl and if you go to www.s syy n-ext tx.com you can log on there and just basically download um lots of procedures that you can get into standardized language and into into kidl and I think the the the literature and the new database is extremely powerful as starting points and what I think we can do is not only use them as starting points but we can then test them and then upgrade them and then validate them and make a GitHub for chemistry now chifi might may or may not be doing that all right but I I encourage Academia to start doing that and I'm trying to figure out how to do that for the good of all um chemistry academics wouldn't it be awesome if if you could go to kemit and download a reaction code and then make it better in your lab reversion it and put it back and say oh use this Catalyst of this conditions and then people oh cool so I don't I think we keep B beating up on the literature literature is great it's just not digital and if because it's not digital it's difficult to error correct and it's costly and there have been no incentives to do that but it's a great place to start and and if it wasn't for the literature and all and the patents we wouldn't have any drugs we wouldn't have any materials we have anything so I would I'm you know I like to to kind of say yeah people lie in the literature of course they lie in anything but it's a great place to start and we will make it better and we need to have incentives to make it better so yeah so Kido is the way to do that but we are building tools to do that both in Cronin lab and I'd love to work with some of the other groups around the world to to develop standards and nice thing about kidl it's true and complete so we don't just have to make it K we do it in any standard they're interconvertible what is the standard that works well I think kid my guess is that kidl is the base layer of chemistry and we should use it but you know I designed it it's going to be flawed in some way right oh this idea of kemit I think it's it's very interesting and hopefully at some point we'll also have like a get for biology which I think is much more difficult yeah I think so I think biology is because biology has been claiming to be programmable for decades and you've got DNA and so on and sure there some parts programmable but chemistry is much more programmable it follows that chemistry is programmable so is biology what is the ultimate programming language of chemistry it's called the periodic table and so you know it's kind of get matching the you've got to find the level that you can match to robotics together and drop the cost increase the speed and make sure it scales and there you know and there's the sad thing about you know not sad thing it's capitalism optimizes you to do three things cut the cost increase speed go at scale do that and you'll be fine and that's one of the things that chfi may or may not be doing right for chemistry uh and a question that might relate to the controversial paper or the paper how they say it the controversial paper of assembly Theory uh is a question from SRA uh how do you make sure that the molecules coming up in assembly Theory are also synthetically accessible oh that's simple what you do is you put them through a retrosynthesis engine and you make sure that you you have code available for them um and and assembly theory if you if you use assembly Theory based upon a molecules you know have already been accessed right then the Transformations on the graph are not going to be that tricky so you can then restrict them to that but yeah the assembly Theory paper wasn't controversial it's just that people just it was you know it's kind of funny I me controversial because everyone read it and went wow now I wasn't expecting that but yeah we can talk that's another podcast or another not another podcast another another Zoom discussion not for today uh and so in terms of uh the retrosynthesis like algorithms that you use uh firstly like my question is like how how good are those at the moment uh and another question from the audience is that do they use machine learning at the moment because I know that for example IBM was using some was language models for retro synthesis so what do what do you use use and what is I mean there there's lots of people out there I mean how can I put this politely um sure I mean IBM actually published a little um video explaining how ARA was terrible and their system was was brilliant right but what they failed to say is that their system their code they put out they didn't make any all of them right and so I think the problem is that people kind of misunderstand the there is a need for for a retr symphysis engine well let me put another way retr symphysis is a process human beings do to take a Target molecule cut them up and then identify all the right Transformations and do it what we're doing is a bit different right we've uh using wouldn't it be great if you could break a molecule into Parts identify the code for the Transformations put the code together make sure the process code was going to work on the robot and you had the reagent to plug them in and you didn't have to do any manual checking that's what we're doing that's the dream that's what makes it relevant using all the other gimmicks the llms are just gimmicks because they don't help you do the one critical thing is make sure the code runs in the robot right and that you don't have to redesign the robot what I've developed with the abstractions ensures the code runs in the robot it doesn't divide by zero and that's where I think a lot of people kind of when you've got people working computer science and chemistry the computer scientists kind of do the pet chemistry and the chemists kind of do the pet computer science the trick is to combine the benefits of those disciplines together and I've been very lucky both in my laboratory and also in chemi to combine those together and that's why we're kind of able to make massive leaps and so it will happen but that's a very long way of saying if you don't validate your code right in real world it has no value and it's so that valid that benchmarking you know running a script to Benchmark it in some way in some kind of digital twin not good enough um because people make false assumptions that aren't tested or testable so yeah so I I think that the IBM have got great kind of um you know dat playing with data but they are they they're on the conceptual side um and what we decided to do was actually because I'm a synthetic chemist is actually build the robots that can really do the chemistry and synthetic chemists are a hard Bunch right they really they Tak they take a lot of convincing and so quite rightly so right so yeah it's important to close that Gap and close it quick all right uh yeah so I think uh we've had the we've H the timeline uh so we can close it up there uh or yeah awesome uh thank you very much Lee again for coming and making Mak the time uh I know that you probably have a very busy schedule and thank you very much everyone uh else in the audience for attending and make sure to stay Ted for the a btic initiative uh as we're going to have an announcement of the next panel on sio and automation coming up in about I think one or two weeks so yeah thank you very much see you later see you take care bye bye right

Channel: Nucleate

Views: 571

Keywords:

Id: 6GgGHVAyclc

Length: 57min 42sec (3462 seconds)

Published: Tue Mar 05 2024