New Capabilities in Computational Protein Design Webinar

so thanks to the um 298 and counting um people out there who are logging on um at this time greetings to everybody in whatever time zone you might be in i'm amy keating i'm president of the protein society and i'd like to welcome you to our webinar on new capabilities in computational protein design that was put together by chris ball and while we wait for others to log in i'm going to make a few announcements about the society so for those of you who don't know this the protein society is an international organization we have members from um 30 countries we run an annual meeting that i'll tell you about more in a minute and we publish the journal protein science to which i hope you'll all consider submitting your work um oops advance here uh we're in a period of transition for the journal as john curion is preparing to take the helm from brian matthews who's been providing strong leadership as editor-in-chief for the past 15 years and we're really excited for the future of the journal even though i have trouble imagining it without brian um leading it one of brian's innovations is the introduction of our special issues on tools in protein science and i wanted to bring to your attention the 2021 issue which um was released in january and as you can see from a few of the contributions that i highlighted here this includes articles about modeling and tools and databases and other resources of particular interest to this community at this webinar today so i wanted you to all be aware of those things because most in-person protein science has been on hold for the past year we've been busy at the society running a series of webinars organized by our members these have been a great success um i'm showing you some of the topics and participants here and i'm super proud of the science that's been covered in the amazing lists of speakers our fourth webinar was particularly special this was organized by bill clemens who's the protein society diversity equity and inclusion committee chair and this featured a panel of protein society stars who have expertise in this area and talked about issues of diversity equity inclusion broadly and in their own lives and their talks are recorded as are many of the other webinar talks so i encourage you to um look at for those on our website if you missed them when they were live i would also like to encourage everybody to visit our website diversifyproteinscience.org this includes a database of protein scientists from underrepresented groups and if you are a woman or a person of color or a member of another underrepresented group it would be terrific if you could create a listing here this will help meeting organizers find you and access your talent and expertise and if you're organizing an event please check here for excellent protein scientists that you can include so this july will hold our 35th annual symposium and we recently made the difficult decision to offer this as an entirely virtual online event so we're pretty sad that we won't be able to meet in person but nevertheless we have a phenomenal program lined up that was assembled by program planning committee chair jeannie hardy and her team that you can see here and in addition to more than 50 great talks there will be live interactive posters and also lots of networking and community events so i encourage you to look for details about that meeting on our website to register um starting on february 15th at extremely affordable rates and to consider submitting an abstract to give a talk or a poster at that virtual meeting let's see as usual we have a few open opportunities for protein society members if you would like to organize and host a webinar like the one we're having today you can submit a proposal online and also as we look forward to our 2022 meeting which will be held in san francisco we will soon open a call for member contributed sessions so if you have an idea about a topic and speakers that you would like to have represented please look out for the call in march and submit your idea then okay so with that i will hand things over to member and organizer chris ball chris is known to many of us in boston as the organizer of the boston protein design and modeling club which has been a great addition to the scientific landscape here i'd like to send a shout out to the members of the um bpdmc who are here for today's webinar and with that i will turn things over to chris awesome thank you very much amy thank you all for coming um i'll keep my opening comments brief because i'm really excited for the to hear from our speakers today um so to kind of set the landscape just a little bit um computational protein design has undergone something of a renaissance in the past few years as many scientific fields i think are undergoing right now um and you know for a long time you know the dream of the field has been to be able to figure out how to control proteins how to control their structure how to control their function and to be able to build this from first principles for a long time understanding protein structure and function has taken something of a of a top-down approach where we look at extant proteins and we make mutations and we fiddle with them in the lab and we try to understand how the bits of structure and function go together by breaking intact machines or modifying intact machines but now with protein design we can really start to build function from the bottom up from first principles and to be able to really understand how these things work and there are many aspects of controlling proteins structure and function some is controlling their shape others is controlling their behavior like binding or catalysis sorry um function like binding and catalysis and others is behavior like dynamics and solution behavior it's solubility it's thermostability kind of the intersection of all of these different parameters lies the complete control of proteins you can make a protein do kind of whatever you want behave under the conditions that you want it to be active when and where um and this really i think has the potential to revolutionize society and improve human health and well-being in many different areas from um therapeutics to diagnostics to climate change manufacturing proteins can do lots of stuff so it's it's a really exciting time to be a protein scientist and in particular to be a protein engineer using computational tools and so with that i will turn things over to our first speaker of the day i'm really excited to introduce uh professor roberto chica um so so professor chica um got his phd from the university of montreal and then did postdoctoral studies with steve mayo and um at cal tech and now he's a faculty at the university of ottawa and um roberto has received many uh prestigious fellowships and awards and in particular he's now at the helm of uh preaching engineering design and selection pets which is a really great journal on our field and and his leadership is already uh you know sort of shaking things up over there and he's really moving this journal forward so i'm also really excited uh to have him uh uh leading off our session so uh roberto please please take it away great thank you chris for the kind introduction and invitation it's a pleasure to uh participate in this uh session with not only excellent scientists but also friends so uh let me just share my screen here so we can get this going okay okay so thanks again for the opportunity chris and today what i want to tell you about is some of our work in computational enzyme design and i want to start by drawing your attention to the images at the top of the slide here which highlights uh what we're trying to contribute to this field so in the past researchers have used a um a single structure as a template to design as you see here on the left but i think that we should start moving away from designing enzymes as a single structure and start thinking of them more as a complex landscape of conformational states that's what we see on the right here because a lot of enzyme activities depend on conformational changes that occur during the reaction and today what i want to tell you about is our work in developing ensemble-based computational enzyme design methods where instead of using a single template we're going to use these ensembles to approximate the conformational flexibility of the native folded state okay so what is computational enzyme design well it's the creation of an enzyme from scratch by building an active site on a protein scaffold devoid of the target catalytic activity and methods to do this were pioneered in the groups of steve mayo at caltech as well as david baker at the university of washington and typically to perform this you you need a protein scaffold as well as a theozyme which is a transition state for a reaction of interest in a catalytically competent arrangement with amino acid side chains and in the first step of this process you try to fit the theozyme onto the protein scaffold we call this theosine placement and once you have your theozyme inserted onto the scaffold then we perform an active side repacking where we optimize sequences around the transition state to identify stabilizing interactions that will further increase activity now this process has been used successfully for many different reactions most notably by the group of david baker and even though this has worked there are still challenges that need to be overcome so the first challenge is that all of the novo design enzymes have had low catalytic efficiency so what i'm showing here is the catholic efficiency of design enzymes here and you can see that this is orders of magnitude lower than that of natural enzymes and typically if we want to improve this we can use direct evolution and get to these values here which are starting to approach those of natural enzymes another issue is that there's still a lot of prediction in accuracy so what i'm showing here is an overlay of a design model in pink as well overlaid on top of its crystal structure in cyan so you can see the transition state analog here doesn't adopt the same pose as the design transition state so there's still inaccuracies that need to be overcome and we need to do this to enable robust design of highly active de novo enzymes so as i mentioned before you we can design the novo enzymes and we can improve them using direct evolution and perhaps the most successful uh example of the evolution of the novo design enzyme has been that of the hg series of camp eliminases which were designed in the group of steve mayo and then evolved by the group of don hilbert at eth zurich so here what i'm showing you is the camp elimination reaction on the left you have the substrate in the middle you have the transition state and you can see that there are two catalytic residues an aspartate that acts as a base to the protonated substrate and then a h-bond donor that stabilizes negative charge up on the oxygen at the transition state and here we have the product so as i mentioned we can use direct revolution to increase the activity of the novo design enzyme by several orders of magnitude so starting from the ag3 computational design which had a fairly low catalytic efficiency here it took 17 rounds of directed evolution to increase the catalytic efficiency by about three orders of magnitude and bringing it up to the level of natural enzymes and this required 17 mutations and i've color-coded the mutations here by these spheres in green are those that occurred at positions that were optimized in the original computational design and in pink are those that are found further away from the active sites and typically when we design enzymes we focus on optimizing the active sites so we don't consider these distal positions because they would they would be very difficult to identify a pre-orie so these mutations of course must be doing something to increase activity so it begs the question can we even design highly active the neuroenzymes by focusing exclusively on the active site well to start addressing this question we decided to generate mutants of these evolved de novo enzymes that contain only mutations that were found by evolution at positions within or close to the active site so those positions that we would have normally optimize anyway in our calculations so we call these variants core variants as you can see here i'm keeping only mutations that are within a certain distance cut off from the bound ligand in this case eight angstroms and what we're interested in knowing is our d score variants active and if so are they more active than the original designs comparable uh inactivity to the evolved enzymes we'll see okay so we did this for four uh camp eliminations that were designed and then evolved so the two on the left were designed by the group of david baker and evolved by the group of den tophik where the two on the right were designed by the group of steve mayo and then evolved by the group of don hilbert and again i'm using this color code of pink for mutations that are for far from the active site and green are mutations that are within or close to the active site so if we make the the mutant here on the left where we keep only those green mutations found by evolution and we measure the activity and we compare that to the evolved variant you can see that we have a decrease in activity but the activity is still 160 fold higher than that of the original de novo design and it's about one and a half fold less than the evolved variant and this trend is observed for all of these that means that the evolved variant is always more active of course it contains more mutations but the core variants are at least tenfold and up to a thousand fold more active than the original design so what this is telling us is that we should be able by design to find these sequences and therefore get orders imagine more activity than what we've been able to do before and out of these mutants the one that we found most interesting was this hg3 variant because you can see that the k category km of the core variant is about a hundred thousand per molar per second and now this is really approaching the catalytic efficiency of natural enzymes so that was really interesting to us so when we when we saw this we asked ourselves is the structure of that core variant that we call ag4 similar to that of the highly evolved ag 3.17 from which it is derived so to answer this we teamed up with my dear friend and colleague james frazier at ucsf to solve the structure of these enzymes and here i'm showing you an overlay of ag4 which is the core variant and age 2.17 which is the evolved variant and you can see in the active site here we have you know excellent agreement near identical structures in the active site so when we saw this obviously the next question became well can we design this can we design the structure of a g4 using the same procedure that yielded hg3 which was the original computational design okay so we ran a positive control calculation where we took the crystal structure of hg4 with bound transition state analog called 6nt here in the middle and we optimize the rotomers for the ag4 sequence using the same protocol as the one that produced hg3 and what you can see is that we get very good agreement between the design model in pink and the crystal structure in white except for this one methionine residue which is closer to the surface so it can adopt it has more space to adopt a different watermelon nevertheless there's good agreement and the energy was very favorable so the next question we asked ourselves well what if we designed the ag4 sequence using the exact same template that was used to design hg3 so that's what we did we took the one gore crystal structure which is a xylonase and we designed the ag-4 sequence on it and now we observe that not only is the binding pose of the transition state different from that of the crystal structure there's a lot of these rotamers in the active site that no longer adopt a similar confirmation as the one seen in the crystal structure and the energy is much more unfavorable and we also tested other backbones we tested the agg3 crystal structure with bound transition state analog and you can see that we get similar but less severe effects so that means that the input template is having a big effect on the prediction accuracy because everything else is identical so we asked ourselves well can we improve upon this by using an ensemble instead of a single template so to generate the ensemble we use this approach called ensemble refinement of crystallographic data which works like this so we take a crystal structure of a protein of interest and we perform an electrodynamic simulation that is constrained by the electronic density so if we zoom in here on this loop what we see is that atoms that are in areas of high density don't move as much as those that are in areas of low density so from the simulation we can extract an ensemble and we think that this ensemble is a good representation of the experimental data therefore of the actual confirmation ensemble of the the protein so using this ensemble we went ahead and we designed the ag4 sequence again on the left i'm showing you the exact same figure i showed you earlier that's our positive design calculation you've seen this before if i use an ensemble that was generated from the original template that was used to design ag3 you can see that we get similar issues to when we use the single crystal structure although the energy is a little bit more favorable however when we use the structure of hg3 now we're starting to see something that's much more similar to the crystal structure of ag4 and the energy now is getting closer to that of our positive design calculation so what this suggested to us is that enzyme design with a backbone ensemble derived from a low activity enzyme could obviate the need for direct evolution okay so what we're proposing is basically can we change the design procedure where we start we would start off in the same way as we've always done we use a template a theozyme we run our design calculation we create a low activity with no enzyme this has been done successfully many times but then we crystallize this low activity enzyme and we generate our ensemble and presumably within this ensemble there could be an alternate conformation that is more um adapted to catalyze efficiently the same reaction so then by performing a second round design we could further increase the activity and get a higher activity than elbow enzyme so we are in the process of testing this we've reached the point of our second design step so we're currently doing this so hopefully next time we meet i'll be able to report some some interesting data all right so to conclude i remind you that there are still challenges that need to be overcome to make computational enzyme design robust for the creation of highly active binocular enzyme but i hope that i've convinced you that we should be able to design highly active lenovo enzymes because ag4 again contains mutations only at sites that were already optimized in the original calculation so we should be able to find that and uh here i'm showing you that we can also use an ensemble to improve the accuracy of our predictions by you know changing slightly the um the confirmation of the backbone okay so where are we going with this well obviously we want to keep using this approach of using an ensemble in computational enzyme design to create the novo enzymes for other reactions of interest but where we really want to take this is to start designing enzymes in the context of a large conformational landscape where we have multiple conformational states that may need to be sampled during catalysis for it to be efficient so for example we could think about you know shifting the equilibrium between these states to favor a more active state and therefore increase activity or maybe we need to sample these two states during catalysis in a multi-step mechanism so these are all ideas that we are currently exploring so to finish i just like to acknowledge the people who did the work so everything i showed you was performed by three group members so aaron right here was a former postdoctoral fellow who initiated this project and conceived it with me he was assisted by woods was right here she performed a lot of the ensemble refinement and design calculations and a lot of the kinetics were performed by niajesh who's right here at the bottom so thank you to all of them for their hard work also i'd like to give a shout out to my collaborators james frazier and a former postdoctoral fellow in his group mike thompson who recently started his independent career at uc merced so they helped us a lot with the the crystallography and we're still collaborating with them and i'd like to thank all of our sponsors for money of course and you for your attention thank you awesome uh thank you for a fantastic seminar we've got a few questions lined up so um start there was one asked in the chat so i'll just remind everybody uh please ask your questions in the q a rather than in the chat but we'll read this first one from the chat um so in the core mutants the kempo elimination sorry this question is from paul shanda in the core mutants for the chem eliminases you remove mutations to which amino acids did you mutate the distal sites alanine no no no sorry uh i didn't mention this we reverted it back to the amino acid that would that was at that position in the original design which presumably came from the wild type template used as scaffold for the original design awesome so we've got quite a few questions in the q a so we'll try to try to get through these so um see roberto will be the easiest would you like to just uh oh if you want me i can read them and answer them that might be the most the fastest way here how about chris i will help you why don't you sort of scan pick one ask roberto and roberto if you can um stay on and answer like later the ones that we don't get to live that would be really appreciated by participants no worries and thank you everybody for the questions so let's start at the top then um the mutations that are part of the engineered modifications uh these mutations are they part of the engineered modifications or are they happening in vivo so these mutations were um so the ones the the green balls that i showed you are mutations that were found by directed evolution and these occurred in the active site or close to the active site so we basically inserted them on the original de novo design enzyme have a question from shorya have you tried to skip the crystallization step and generate an ensemble from normal mode analysis or something similar does the ensemble resemble the ensemble you get from md with constraints right excellent question so of course over the years we've developed computational methods to generate ensembles that wouldn't necessitate you know this ensemble refining procedure with the crystal structure so there's many methods that can be used to do that and in this particular study we did use um unconstrained molecular dynamics we still use the starting crystal structure but not the the density to guide the molecular dynamics and uh we did improve the the designs compared to using a single crystal structure but it wasn't as good as using the ensemble refinement procedure so i still think that ensemble refinement is probably going to give us really good ensembles for this approach but it doesn't mean that it's the only way to get access to the correct confirmation of the backbone to design more more accurately high activity enzymes let me answer a question from us a student fourth year graduate student odessa asks um during the ensemble generation were there surprising results that came from the density-constrained md simulations like did you observe movements outside of floppy loop regions um that's a good question so of course most of the movement was in uh in floppy uh loop regions but interestingly depending on the the starting structure and its density the loop movement wasn't identical even though the um you know the space groups were the same so it seems that it's not not every part of the protein moves the same and depending on how it does move that affects the prediction accuracy and it's very hard to know beforehand you know how to which which one is going to be the one that works the best should we let bruno ask his question i i guess i yeah if you if you if you want to go this way let's let's try so roberto you know this is really beautiful like completely like extremely hard problem such a fundamental problem in science really great to see that you're working on this i guess my question for you is the following so one of it is actually in agreement with uh one of these questions that you have also here in the chat is you know it was very clear that you had very high differences in energy in these different poses which often times when we start scoring ligands and proteins we don't really see this level of differences in energies right so like there's not really this clear signal so what kind of energy function did you use here that's point number one and then i guess point number two is uh in terms of these structural changes are these like very minor structural changes or do you actually can sort of like spot them by eye right so is this sort of like some kind of invisible state that you're sampling or that you need to sample or it's really sort of like a discrete state that it's far away from a template where you need to be right excellent question so for the energy function we're using the phoenix energy function developed in a group of speed mayo and basically it's a potential energy function that has um you know the usual linear jones electrostatic h-bonding solvation terms uh so it has those four energy terms and um that energy is basically the sum of all the pairwise interactions so that's why it gets to be so low uh you know but that's typically what we observe when we use this energy function in design uh so we're more interested obviously in difference in energy between different sequences and not absolute values of energies which are not physical energies i mean they're not there's no entropy contribution so it's not a delta g value and then for the the ensembles so the ensembles we typically characterize them by measuring the pair the average pairwise rmsd for the backbone between the different members of the ensemble we call this diversity and what we've observed is that diversity like a sweet spot it can't be too low you can't be like you know 0.05 angstrom rmsd because then the structures are very similar we like to be in the you know 0.1 2.5 range that seems to work best but we can see differences and depending on the on the flexibility that we obtain from ensemble refinement some parts of the backbone are much more flexible than that as you saw from that animation some loops you know are are moving more than those few angstroms thanks a lot maybe there are a few questions um including one that i had along the theme of designing allosteric mutations and is there a reason to constrain your redesign to the active site if your ensemble provides the opportunity to maybe design distal mutations that might stabilize productive states and let me just read paul chanda related asked what is your perspective for designing allosteric properties do you go in this direction and what is more challenging there is compared to focusing just on the active site evidence right so i think ultimately we'd like to be able to design distal mutations as well at this point i don't have a clear answer on what would be the best way to identify those sites and then i design a mutation that's going to impact what happens in the active site so i can't answer that we are exploring the these ideas but we definitely want to start considering that the good news though is that my view is that if we can get at least the active site mutations to be really good and get a you know 10 100 or 1000 fold more active enzymes than what we were able to do previously that level of activity should be sufficient for a lot of biochemic applications it may not be you know as high as natural enzymes but it may be sufficient for many applications at least that's the hope but we're definitely interested in identifying distal sites but that's a different problem altogether and i don't have a clear answer to tell you how to do that yet chris how are we on time um i'm thinking maybe uh yeah this is excellent we have a lot of really awesome discussion but i think maybe we should um continue with our next speaker um make sure we stay on track uh thank you uh again uh roberto for a really excellent seminar and um you've got 26 questions in the q a if you have uh uh some time you could uh go through and you know and answer so you know folks if you've asked questions in the q a then um even if we don't get to answer it live then um the speakers can can answer them uh by text after um all right thank you everybody awesome so i'm really excited for our next speaker uh professor tanya cortemi um so um professor cortemi uh she got her uh uh phd at um the embl in heidelberg and she did a postdoc there as well and then joined david baker's group in seattle washington for um for a postdoc now she's a faculty at ucsf again very highly decorated uh many many awards uh major contributions to the to the field really um uh i'm super excited so um uh yeah tanya please please take it away well thank you very much chris and um really for organizing the symposium i feel really privileged to to speak among my fabulous colleagues and friends here and also i think we already have a really terrific discussion so i'm hoping for more of that and i think you might see some of the themes that you already saw so beautifully illustrated by roberto also um coming coming up in in the kinds of problems that i will discuss um let's see okay so so we're obviously fundamentally interested in the problem um how to build new biological functions entirely from scratch um and there's a number of different motivations ours are we really like to probe the design principles of biological functions we'd like to be able to use engineered systems to control biological processes to to really understand and be able to dissect them and obviously there are practical implications as well um in particular what we like to do um like like many of us here i think we're interested in designing proteins from uh first and physical principles uh using computational predictions in our case um in the program rosetta that we've co-developed for for many years um and i just want to you know start by acknowledging sort of the two global challenges that that we're all facing uh one is that the space of possible sequence structure combinations for proteins you can imagine it's obviously really enormous so you need to carefully think about how to sample that space um and also how to represent that space what is the conformational flexibility that you allow as we just heard from roberto also and then secondly we need some objective function because ultimately computational protein design is an optimization problem and we're only as good as our objective function really is and we need to make simplifications in order to allow for many fast computations but we still like to base our energy functions really on the physical chemistry of of the problem so that so that we're hopefully also advancing on those methods so despite these considerable challenges of course the field has made considerable advances that are really driven by fundamental work over three decades or more and in particular there's a lot of progress in design applications i would say in the last 15 years and we now in particular have many examples where small and in some ways idealized structures can be designed really with amazingly high success rates and many of these designs will follow the following principle where you have a a topology blueprint so you have a topology of the number and relative ordering of helices and sheets and then from those topologies you assemble idealized structures through taking fragments helices sheets and loops often from their protein data bank and then these idealized structures can be functionalized and in particular a wealth of functions has been built from helical bundles making really strong protein binders and even now getting into new therapeutic candidates so so really exciting developments still i would argue that these designs or what we've achieved so far often deals with largely static structures static components and the design of function and in particular dynamical functions is typically much more difficult in many of the functions that we would like to design out of reach roberto illustrated the really hard problem of enzyme design um other problems might be sensing dynamically sensing molecules that we currently cannot sense proteins that switch their conformations when they control signaling and many really of the composite functions and and beautiful molecular machines that nature has invented so so here i will argue that some of the key challenges are that um first of all that these functions in contrast to the idealized structures that i described above require precise control over often not ideal geometries in functional sites and then moreover the structures that we can currently design are often extremely stable but for function we often need to switch between multiple uh roughly iso energetic states and that really is one of their challenges in the field so so with these challenges in mind i will describe three very short stories of recent progress that we've made with the difficult challenge of precise control over geometries the first is to build um new geometries or build the geometries of functional sites the second is to precisely position functional groups within these sites and here we're building on methods for sampling protein conformations borrowed from the field of robotics which we are now adapting increasingly to protein design and the third is to systematically tune the geometries the shapes of entire proteins into which new functional sites could be placed with hopefully better accuracy so here's the first application we wanted to design proteins that can detect new signals in our case these signals are small molecules for which we didn't have a sensor we wanted these proteins to work in cells where they would not only detect the signal but also convert signal detection directly into a cellular response or a change in cellular behavior and ideally we wanted to engineer modular systems where we could choose different potential cellular output responses and so fantastic former graduate student dan mendel came up with one idea of how to do this which is utilizing the principle of small molecule induced heterodymerization that nature obviously has used but we wanted to build a system like this synthetically and so the key idea here is that two protein sensor components preferentially interact in the presence of the small molecule signal now you can link the two sensor components with two halves of a split reporter that when the small molecule signal is present the sensor components dimerize and upon dimerization assemble a system that reconstitutes a functional reporter and in the sense the system is modular where of course the detection of the signal has to be specific to the signal but the sensor can then be linked to many potential different output reporter systems so that leaves us with a difficult problem however how to build the geometry of a binding site for the small molecule from scratch into a protein protein interface um as a first target where i'm going to illustrate the strategy we take this molecule here fibrosil pyrophosphate and we mainly picked it for its interest because it's a key metabolic intermediate in engineering pathways that can be used to engineer value-added compounds such as precursors for therapeutics or fuels in bacteria yeast so in principle a sensor for fpp could be useful biotechnology logically this is also a challenging target it has 11 rotatable bonds so dan developed the following design approach and again some of the themes will resonate with what roberto described before we first defined the binding site geometry in this case by the relative orientations of side chain functional groups binding to the ligand in optimal physical chemical geometries and we used not just one but we tried a number of different geometries here the second step then is to build these geometries into protein protein interfaces de novo so now building them into sites where before we didn't have a binding site of a small molecule that we can borrow from um here we were completely agnostic which system we wanted to use instead we built these binding sites into more than a thousand protein protein interfaces from the protein database and then for each of those hits we now diversify it the area around the binding site using ensemble-based approaches that you've also heard about to generate a vast number of confirmations for each of our designs and then we design sequences that stabilize that modeled binding site and then using computational quality metrics in rosetta we selected the top models on three different scaffolds that are highlighted here so in total we had three potential protein-protein interaction interfaces from this entire computational screen and we made nine designs and so then a really great postdoc ming huang developed a strategy to test the functionality of these designs in e coli and so he coupled the function of the sensor to a growth based signal in bacteria i'm going to show you data for the nine designs for our three interfaces first you can see the yellow interface here did not show any sensor signal so in a sense that's a negative control then for the second interface here you see some small signal but then for the third interface we were really encouraged to see that two of the designs on this interface showed a very robust sensor signal any coli straight out of computational design this is without any experimental optimization here and then of course for a sensor to be useful the signal has to be proportional to the signal that you're trying to detect um our signal fpp does not get into cells when we add it externally so instead what we did also in this experiment here shown on the left what ming did is he expressed a pathway of five enzymes so that he could add the precursor mervalinate that then would be converted by these five enzymes to fpp and he could see a proportional signal of the sensor with increasing concentrations of external methylene and to to highlight or to summarize how we got to these designs our computational designs had a large number of mutations and then moreover they were computationally screened from a thousand possible protein protein interfaces so this is certainly not something that we could have done by experimental screening the computation was really essential to get us to these design sensor constructs so then of course to uh demonstrate that we are right for the right reasons the best way to do that generally is to solve the crystal structure of the actual design in our case the ternary complex um this is in collaboration with mike thompson and jamie fraser you've heard from about both of them already and roberto's work so mike was able and this was actually rather challenging mike was able to solve the ternary structure of the complex showing the de novo design binding site here for the small molecule and encouragingly even at the level of detail of all of the design side chains we really have a quite reasonable agreement between our original design model shown here in gray and mike's crystal structure in the colors although not perfect and i want to come back to you to that problem also um so then finally um a terrific postdoc in the lab adam glasgow wanted to see whether our total our complete design strategy was actually modular as advertised and she also wanted to get more quantitative data on the sensitivity of the sensor to the actual ligand fpp so she um coupled the sensors to two different split reporters either a split luminescent system or a dimerization dependent fluorescence reporter and now she expressed the system in an in vitro transcription translation assay where we don't have the cell membrane so now we can add the ligand fpp directly and test the sensitivity to fpp and adam could show that for our two best designs here particularly for design two now the sensitivity to the ligand again out of straight out of computation is in the nanomolar range and the same was observed also with the dimerization dependent fluorescence reporter uh really demonstrating um that we can link the signal to different modular outputs so in conclusion for this part we are excited that we seem to be able to build functional site geometries lenovo by computational design and we can do this at atomic accuracy as confirmed by crystallography and these systems generate synthetic small molecule-induced heterodimers that can be used to program cells to sense and respond to new signals however what you could already see from if you look very carefully at the crystal structure um the binding site geometries when we put them into existing uh proteins given the backbone of these existing proteins are often not optimal so we screened a thousand protein protein interfaces we only found three and even for those it really wasn't optimal so that prompts the question to us at least can we control these binding site geometries to precisely place side chains um much more much more accurately with different methods and so what we thought here is that we can use methods that we in our lab had developed for a number of years which are borrowed from the field of robotics so very simply in robotics if you imagine a robot iron the problem that robotics really solves is to precisely position and move the robotic hand by basically setting the degrees of freedom in the robot arm and so this lends itself i think very nicely to putting the same mathematics in the same problem to proteins where now our problem is to precisely position side chain functional groups by utilizing what we know about the physical chemistry about the degrees of freedom in the protein backbone and we developed many years ago some methods to sample these possible in conformations that are mechanically accessible to proteins and so now the idea was to apply those to design to precisely position functional groups and this is work um developed by uh kale and cody our two graduate students in the lab so again this starts with the defining the geometry that we want to optimize for so the precise positioning in this case of a side chain functional group relative to a ligand and then we utilize these robotics inspired mechanical sampling methods to build backbones with restraints so that they are successfully optimized to support the desired geometry that we want to optimize for and then of course in the final step we designed the environment around it to stabilize the new confirmation and kale and cody picked a very simple but i would argue and actually in fact rather difficult problem to test whether this design strategy had any chance of working and the the simple idea here is to replace the precise positioning of a carboxylic acid group of carboxylate of an aspartate with a glutamate which meant that if you succeeded in precisely positioning the carboxylate the design had to be accurate on the order of the length scale of a carbon-carbon bond so a very challenging problem and using this methodology and again collaborating with mike and jamie um cody was able to solve the crystal structure of this reshaped region here in the design and you could show by comparison of the design model and the crystal structure that globally the backbone indeed adopted the desired structure and i'm also showing you here a comparison of the design model with the original wild type structure so you can appreciate to some extent on what the extent of remodeling in this active site here really was and so the rmsd um here is larger than three angstroms and the accuracy here is a little bit over one angstrom rnsd and then of course what was really on the key point here was to show can we actually precisely position the carboxylate um and you can see here the overlay that the positioning of the new glutamate with the rearrangement of the backbone is nearly perfect so it seems we are starting to have methods we would can think about positioning functional groups using this robotics inspired computational design um yet moreover the positioning although it is not perfect is sufficiently accurate to now indeed distinguish aspartic acid from glutamic acid here so what we did here is we measured this is an enzyme we measured the enzymatic activity of this enzyme it's much worse than wild type so we clearly have not understood the enzyme design problem but the enzyme is active with now the glutamate as the catalytic residue but if we now change the glutamate to the original wild-type aspartate now we essentially kill enzymatic activity showing that positioning seems to be sufficiently accurate to distinguish between these two residues um so i told you that the problem that we're still having is that the binding side geometries are often not optimal but we're starting to have methods to reshape binding sites using these robotics-based methods um but finally we wanted to address the problem right now what we've done is we've done everything in the context of existing proteins with the limitations of their backbone um and so an obvious um next step would be can we build custom shaped proteins entirely lenovo into then into which we can then place these functional sites such that these lenovo proteins are really optimized for the geometries that we want and so this is a problem that um a really phenomenal phd student changed took on and the idea is borrowed from natural proteins where natural proteins diversify their folds again i'm showing here in topology by essentially having variations in the position orientation and size of secondary structure elements to create distinct geometry variations within a certain fold family that they're then optimal for function so what xinxi developed is a computational method that systematically samples these units to create a vast number of potentially distinct geometries lenovo so in essence his method creates artificial shape families of proteins with hopefully really precisely tunable geometries so to test this idea took two different de novo designed protein topologies and diversified them computationally creating thousands of new geometries he then took a selection of these and tested them experimentally and showed that 38 of the proteins were folded by biophysical methods with purified proteins and many of those had predicted geometries that were not seen in nature he then determined several structures of these the lenovo design proteins i'm showing you three here and he could not only show that these structures are clearly distinct from each other even though they all have the same full topology but they also match the design model with atomic accuracy as you can see here again from the overlay and then finally zhinxia went on to quantify the potential structure space accessible to these proteins so what you see here on the left are projections of these two designed helical elements the centers and orientations of those helices onto the underlying sheet in this beige color and you can see very sparse sampling very sparse coverage in naturally occurring proteins in the pdb um however you can see that the space of accessible geometries by xingius method greatly surpasses that scene in nature and you can see the examples in yellow and green that jinji is sampled in this space for experimental testing so in summary why i'm excited about the possibilities for new capabilities in computational design is we're starting to think about not only creating new functional sites entirely from scratch but we're starting to have methods to finely tune the geometries and the precise active site residue positioning in these functional sites and moreover we're starting to be able to design artificial proteins with light shaped varieties to match these new binding sites and going forward as an outlook of course now that we have the ability to tune the shapes of artificially designed proteins now we're really interested in exploiting that capability um to build switchable proteins that that can buy a variety of signals can really precisely traverse landscapes switching between different conformational states and so i'd like to end of course with thanking the phenomenal people who've really been driving this work i've told you about dan ming and adams collaboration on the modular sensors jinjia developed all the ideas underlying the project of gen generating these shape families cody and kale were really driving the methods to apply robotics um to precise active site positioning um and a number of fantastic collaborators who helped us with structural prediction um and the galley could see us who introduced us to the robotics-based mathematics and i thank you for your attention fantastic uh yeah tanya thank you so much um we've got some great questions um in the q a and i'll just one more quick reminder everyone please um uh use the q a for your questions um so we'll start off maybe with a question from colin smith um he asks um how sensitive was the dynamic range of the biosensor biosensor in vitro to the concentration of the sensor oh hi colin um great to at least um talk to you remotely here um yeah so um these uh small molecule-induced heterodimers are very sensitive to concentrations right because um of course what we're looking at as the sensitivity of the sensor is the difference in dimerization um with and without the small molecule and so if the sensory components are at high concentrations uh then of course they will dimerize they will they'll be above the kde on in the absence of the small molecule signal so they will dimerize and give a give a false positive signal without the small molecule actually being present so those need to be tuned and that's one of the one of the issues with these types of assembly-based systems and so the reason or the the way to improve upon that is of course to maximize the difference in affinity of the sensor protein components in the presence and absence of the small molecule catherine goodman asks a variation on a question i wanted to ask so i'm gonna i'm gonna squeeze it in here um she writes great talk tanya i'd love to hear more about the sampling space in the last example it looked like the tested constructs kind of fell in the middle of the larger smears that are possible how do those overlap with natural space and is zinji considering testing some of the constructs that are more at the borders of the smears yes hi catherine um excellent questions um and so i guess the answer is yes to all of those things so um yes so we tested um in the examples that i showed you i'm probably um the the most conservative um samples out of that space uh still though we did a quantification showing that the um if you bin um that space and you um count uh how many of the bins of the helix orientations are covered by naturally occurring proteins and then how that compares to the tested designs in that particular um story we pick bins that are not covered by naturally occurring proteins um but as you say we're still sort of in the middle of the larger sphere and what you're saying is exactly what we're doing now is to basically maximize the sampling uh within that space for experimental testing to really see how much we can push the methods um got a question from um eva hyde uh this this one is a really hard question and one i think all of us grapple with all the time so i'll pick this one out and it uh forward it to you here um you did a huge number of calculations to choose your proteins can you see why some of the ones you chose did not work yeah yeah so so obviously the ones that we chose so in this in the particular example um that i showed right we only picked you know three different protein protein interfaces we obviously chose them because we thought they're gonna work um and it's always really hard to um you know once things do not work um to do a proper um proper investigation of what all the possible reasons are and so for the cases where we've done that to some extent um the reasons where either that the proteins aren't folded or they're just not binding a small molecule is i think what we're thinking but that's sort of very general and then really sort of to get at the the deeper question here which is you know how how can you figure out um what what the what the actual problems are and the actual parameters that can help you guide find more successful computational protein designs um i guess the the best answers to those questions we have from large scale experiments um where really a large number of designs are being tested and gabe brocklin and and others have done have done things like that and really taken all of the computational quality metrics um and compared them between successful designs and non-successful designs and trying to to tease apart on which quality metrics are most informative for getting for getting successful designs okay maybe i'll ask a question that came in early from che yang how did you usually evaluate as criteria that the generated backbone is suitable as a template to scaffold the binding site usually i know people pick it up by local matching of rmsd just wondering if there's any other way yeah so how to pick backbones that accommodated your design binding section yeah i mean it's basically what what you're saying it's basically how closely can we approximate the ideal geometry that we want uh given the background that we have and giving given the model side chains in that pathfind environment by some sort of rmsd type metric by yeah an armistice type metric or um you know you can look at you know precise distances and angles which are probably better than that chris do you want to pick one more or i'm not sure how we are on time but um do we have this one's maybe is pretty quick so i think we got one more um eckerd um mosner asked um uh in the aspartate to glutamate example so i think your robotic example here how big is the change in the backbone how many mutations did you engineer yeah so in the uh remodeled region the change in the backbone from the wild type is on the order of three angstroms uh we made a couple of different designs i only showed you one and the number of mutations um was between 12 and 18. excellent um so yes thank you again tony you have some more questions in the chat um if you mind answering those by text and um yeah i'm very excited now to introduce our next speaker professor bruno correa um so professor crea did his um his postdoctoral work with with bill chief at uh scripps and was really one of the pioneers of immunogen design and now as a you know faculty at um uh in the sun and i'm really excited uh bruno please uh please take it away oh you're still on mute chris the first thing i'm going to do is to thank you for having me here the second thing is to ask you to please call me bruno and not professor correa because he gives me a headache man so all in all you know it's really true pleasure to be here with this so distinguished group of scientists and also friends and you know the only regret that i have in us not being in the same room is that i can absolutely predict what was going to happen after the talks which would be we would have gone to a bar and we would drink and laugh until very late and so i really hope that this can happen like any time soon but so to this i will tell you um a few of the things that we've been doing here in lausanne and again the scope is is about the same of what robertos and tanya which are of course you know references in doing this type of things that i'm going to tell you and here the idea is how can we actually expand this this landscape of functional proteins by computational design and so as tanya mentioned you know that the problem is is just so it's of a magnitude that is so large that it's actually hard for the human brain to comprehend and so if you start sort of like thinking about what the theoretical space of possible combinations of amino acids when you start putting proteins of 100 amino acids together you know you're very quickly like in in a larger scale than the the number of atoms that actually are predicted to exist in the universe and so clearly you can think about the protein spaces this very large uh uh space that nature by by actually so like you know this was an observation that comes now from from genomics experiments where nature really only sampled sort of like small subsets of this space right and i think you know the question that that really here begs to be asked is what is uh uh present on this on this on this on this like gray space or this dark space where we actually maybe there's a lot of like great proteins to be discovered there and that's that's really truly what i think that that we all believe that is uh that that there's there to find out so just this sort of like you know to put you a picture on how uh in general protein design works so a lot of us uh and a lot of the community that does protein design uses modeling programs which sample backbone and and sidechain confirmations and ultimately you have a an objective function or a scoring function which tries to to optimize that search for you and ultimately what happens in the computer somehow resembles very briefly what i'm just showing you here in this movie eventually you get a couple of sequences out or a lot of sequence outside that you will then will have to um experimentally characterized so but speaking about the design of function right so i think if we would like to characterize how the design of function is is done using computational methods you could probably you know perhaps this is not the the fully inclusive categorization but but there's likely two types of approaches right you can either take what we refer to as sort of like a top-down protein design approach where you use uh the natural protein repertoire that you see here on the left and then you know you search for for for backbone similarities and you try to graft in some type of functional motifs and what what you get out of this process is basically a natural protein which has been functionally repurposed what i will tell you about today or part of what i'm going to tell you about today is about the exercise of doing a bit the opposite right which is starting from a bottom-up approach where you eventually have the the only the structural building blocks and you start assembling proteins around dysfunctional motifs uh and ultimately what comes out of here is a topologically assembled uh assembled process where um you uh will obtain um a functional protein so uh what i will tell you about today uh has a very well defined design objective where the idea is to design the novo proteins which are functionalized with structural motifs and and binding sites and so and these binding sites are basically recognized by um by by antibodies the application that i will tell you about here today is related to the design of vaccine immunogens i will not tell you all the biology that we have done in order to figure out how these immunogens actually work but uh but i will i will give you some highlights of how we actually assess the function of this design proteins but mostly i'm going to spend time telling you how the the the functional approach actually works so we have the rosetta topo builder approach which was what we developed here in the lab and and um essentially what what happened here is that we have this um [Music] computational design approach where we start with a very rough definition of which protein we would like to design so in this case a 2d topology definition where you can you can define which which types of secondary structure elements you're going to use you also know what your motif looks like which is shown here and then you sample through the different connectivities that you can um that you can obtain through this um different secondary structure elements and ultimately you turn you'll translate this into a three-dimensional uh motif what you get out of this process is essentially um simple topologies with uh a binding site embedded on them and then they have potentially different types of structural configurations so a better vision of how this looks i'll show you here in this movie is that so what i'm showing you here is basically a viral protein which is recognized by a monoclonal molenta body and you see our motif here in blue and basically we extract this motif and then we start assembling um different secondary structures around uh around the the motif we extract distance constraints and once we have these constraints we start simulating the folding process and after having many of these decoys designed we actually optimized for the sequence and then that's when uh we start testing which uh testing experimentally uh this sequences so to give you an overview of the types of proteins that we tried to design so you see that we have a number of different alpha beta topologies where uh uh this particular functional motif that is recognized by a monoclonal body has been embedded uh we also have fully alpha topologies with with a complex binding motif which is um which is composed by a helix and a loop so this these motifs are a bit more challenging because you have to not only have a segment of the protein a segment of the protein right but you actually have to have the three-dimensional orientation of the motif and so let me just connect this laser here and and also we designed a protein which also has then two different motifs which is actually able to be to be bound by two different antibodies and so um the way we did this was was basically by doing the computational design first and then select the best designs building libraries based on these designs performing a high-throughput screening on yeast display to select the both the best binders and also the proteins that were folded using a protease resistance essay and then a readout of next generation sequence ultimately we identify which sequences look the most promising in terms of folding and binding and then we make some of these proteins and here i'm showing you a a summary of the data where you have different kind of topologies which we then see if they are folded many of them have well folded uh cd spectra they also have in many instances high thermal stabilities and then more importantly they do bind their monoclonal antibodies as you see here measured by surface plasma resonance as mentioned before the ultimate proof that the design is correct is when you solve an experimental structure so here i'm showing you the comparison between the computational design models and the experimental structures which uh it's always in the range of the one point uh eight angstroms when we start thinking about root mean square deviations okay so now an application why of where we use this so ultimately one of the the the the global aims of our lab is to design vaccine antigens and so uh what we did here we we we took on this this virus which is uh the the respiratory situal virus and um and and we extracted a couple of epitope motifs which uh are shown in the in the in the viral protein and so then we use the tuple builder we designed a number of different antigens and we made a synthetic cocktail that then we used to immunize a number of uh animal systems and then we studied the the immune responses so i won't tell you much details about this but i will just show you a few key results where here in these experiments in mice we had three groups one of them that only has seen the computational designs the other one that has seen the viral protein as a positive control and the other one where we're doing a natural logos prime boost by priming the immune system with the viral protein and then boosting with them uh with the computationally designed scaffolds so what you see here the first the first thing that we found out is that in fact the the computational design scaffolds alone could actually elicit neutralizing antibodies basically if this neutralizing antibodies if you have them in your body will be protected against infection that's what what this result means and then what you see here is that in this heterologous prime boost scheme that we've used in one of our immunizations we actually improve the antibody quality and basically what this this this means is that we have a ratio of antibodies uh that uh that neutralize the virus that is superior to that one that is that is triggered by the viral protein and so this for us was an interesting finding then also as a last point when we tested the computationally designed um vaccine cocktail in non-human primates we actually obtained uh responders out of six animals which uh comparing to other experiments that we have done in the past was the highest level that we that we could have ever ever seen with um with computationally designed uh antigens so just a couple of conclusions and an outlook before i go to the second part of my talk uh ultimately uh i will highlight that we were able to design a number of different topologies that were functionalized with um epitopes or binding motifs and i think you know one point that i still want to to highlight this is not necessarily about the achievements but rather about um what we don't understand right and i think one thing that we don't understand here is related to a point that uh that also tania has touched in in our in our presentation which is you know what are these features that allow us to design um to design successful proteins right what what are the qualities of these backbones that we need to have in order to sample the best sequence possible and i think even although we screened a lot of a lot of designs and we stringed a lot of sequences and yet from the analysis that we have done these features are still not completely clear so so i wanted to still highlight that this is still you know the problem is far from being solved so as the second part of my talk i will switch gears now towards a different type of um a different type of approach and basically what we have here is about this idea of on how the question that we had was can we actually identify surface patterns that reveal functional features of proteins and i like to make this this this this analogy in between identifying patterns in people's faces and patterns in protein structures and so um and so we started thinking about this problem and essentially what this our approach this problem is to develop this this uh this software which we call massive which stands for molecular surface interaction fingerprints and basically the way that this um that this software works is by taking a protein surface segment this protein surface in patches and then once we have these patches of protein surface which by the way they don't really know what is the amino acid identity but what we do is we actually map into these patches the geometric features of the surface and then the chemical features of the surface together with the polar coordinate system just so that each point knows exactly one relative to the other and then we put this into a machine learning framework so that we can try to optimize the descriptors of these surfaces for different tasks and so the tasks that we've used this this algorithm here for was for pocket classification uh interface site prediction and basically the idea here is given a protein structure can we actually identify which site on that protein structure you are more likely to interact with other proteins and then another exercise which resembles that one of protein protein docking which is the prediction of protein complexes okay so just a very brief overview about the results that we were having here and so in terms of the site prediction we could see that massive could predict with much higher accuracy than other softwares to do the same type of task uh the sites on which other proteins are able to to engage even though you don't know what the other protein looks like okay and then in terms of uh the benchmark for uh for for docking we compare it massive against a number of other different docking programs and essentially the result that came out of there or the most noticeable result is essentially this one where um we were able to do the same type of computations but in much less time and when i mean in much less time i mean in like four orders of magnitude uh difference so here you see the results for zvog z rank so in terms of performance we're on par a little bit below but in terms of time uh we're in fact much faster and the reason why we're this this is happening is because we're doing a lot of the computation in the vector space rather than on the three-dimensional space and that's where the speed up comes from okay so this is how it started but i want to bring you back to where um to where i left you in terms of designing function right so a lot of our approaches that we are having are are based on functional motifs that we already know that that can make the particular binding interaction or that particular function so the idea here was could we actually use this massive approach in order to come up with new motifs that could for instance mediate novel protein-protein interactions and so the problem here is this one of designing new proteins that engage another target protein in the defined site with a predictable structure the way we went about this was okay so we have a protein target in this case we used pdl one it's not very important uh what the protein is right now but we defined a site that we wanted to target we extracted a surface fingerprint from this site and then we did a search on a database of 100 million fragments so you know now because of this this speed up that we are able to achieve we can actually do this in in a very wide scope and ultimately what these searches produce is this um this seeds that are bound into a protein and so then once we have the seeds bound into a protein we transfer them into um into globular proteins and then we can use this for high throughput screening okay so here we tried a number of different protein motifs to to to bind to this target and we um we transplanted them to different um two different protein scaffolds and then we did experimental evaluation so many of them was as expected not very high affinity but some of them after being uh improved and evolved using yeast display so these results are not what you are you actually get out of the computer um they they could in fact bind to this target uh with relatively high affinity and so and again there's always this need to solve a crystal structure so eventually we we picked up one on one of these um one of these designs we tried to solve a crystal structure in collaboration with with george gau and ultimately what you see here is an overlay in between the actual crystal structure and the model that we generated in the computer and also of the seed that massive actually outputted and you see that you know we're we're pretty much at uh spot on in the prediction that the computer had so for us this was of course a very interesting result that we're very excited to to push forward but it's it's still just just the beginning for us okay so the last point that i want to make is you know because we are actually searching for motifs that are present in the protein structural database you might wonder if we are actually finding new motifs or if we're just recovering motifs that already could mediate this interaction and basically what we can see is that we can certainly recover the motifs that mediate the native interactions but here i show you the native interaction that is known in between pdl1 and pd1 and with the seed model that that we that we actually calculated and you see that essentially we don't use the same type of secondary structure we don't use the same type of side chains to engage the surface and so this for us was a very exciting realization in terms of uh showing that in fact the program is able to come up with different uh binding solutions to to engage the same site okay so with this i will just close and uh we'll close with with a couple of uh ideas in terms of uh what what i've described in the second part of the of the of the talk and so ultimately it means with that with these vector fingerprints we are actually able to find functional signatures from protein structures and we can identify a small molecule and protein binding sites we also are able to do very fast docking simulations and we can generate protein binders out of the computational stage at the micro molar level and then they can be improved using in vitro evolution and so when two of these design binders are actually in i only showed you one but we actually have a different one which is also in agreement uh with a with a crystal structure okay so now the most important slide and i have been an incredible uh incredibly fortunate uh person in having been had the opportunity to to work with the the group of people that i've been working in the past five to six years and it has really been an exciting journey for us here and um with this i would also like to thank the funding and i will be happy to take any questions if you might have them thank you bruno that's a really exciting seminar i think um thanks i i'll take the the privilege of uh the organizer here and ask the first question before we dive into the to the q and a uh questions here so i mean massif is really exciting um i mean that it's going to open up a lot of doors i'm just wondering um uh how sensitive is is it to um changes in rotamers on on surface residues yeah so what we have found is that for now it's still more sensitive than what we would like it to be so small changes in um in these surfaces in fact will throw the signal off in a way so we have seen that for instance sometimes you have the crystal structure of the same structures okay and they have um you know they have very little differences in terms of rotamers and you know it's clear that the signals there get a bit confused so i think this is something that we still need to work on it's for instance to get to the level of you know could we actually predict accurately you know the effect of point mutants right which honestly right now i think we cannot do yields okay that's that's basically my most honest opinion actually bruno related to that can you just speak a little bit more to what the fingerprints are like what resolution and types of features they're capturing so yes i didn't show you any of the data that we uh or any of the the the types of experiments that we try to do in order to understand that so it really depends which type of task so for instance we have noticed that whenever we're trying to predict complementary uh interfaces that for instance geometry plays a big role so complementarity you know this has been known right we're we're in a way just sort of like you know trying to resuscitate what was in the books um but we also have noticed that uh the electrostatics also plays a role so we have done this ablation studies where you you sort of like kill one of the features or you just you play games with the data and see how the signal degrades and ultimately it means that um depending on the task one or one or the other will dominate but generally when they work together they work the best now in terms of what these fingerprints means you know so again these vectors come out of the the network and essentially you know they you just don't know what they mean that's basically it so the only the best thing that we could do for now is basically making these ablation tests there's so many more things i could ask um but let's um go to the q a chris i've got i got one um for the um the first half um so catherine goodman also says that thanks for the great talk bruno for the first half are you able to handle discontinuous epitopes using this method oh so that's that's a good question yes so um one of the examples that i've shown you is actually this continuous epitope so you know again i'm going to be honest with you right you know this is a discontinuous epitope that we work during five years so i think every time you work on a problem five years you know you've got acquainted with that problem so much that you know eventually i i don't think if it's fair that i can just tell you yeah this thing is just gonna solve any like discontinuous uh like epitope i don't think this is true you know i think in this case you know we have a helix and then there's there was a loop on the side and the helix was easy enough to to embed in the protein the loop we had to do a little bit more work so i think you know what this data shows is that there's hope and i think we could probably do it more consistently in the future but um but it's not it's not an easy one okay venita sood is asking whether massif can be used to improve docking as well as design which i think you hinted it can yeah so so let's see so i think you know everything comes kind of like with with a price okay so i think you know right now i think what what we're gaining with massive is this ability of making very large scale experiments uh you know so we can we can dock many things against many things of course that you know what i didn't tell you is the the weakest point right and the weakest point is we also know that mass is terrible docking unbound structures okay which by the way it's not just massive a lot of programs are terrible but docking and bond and bond and bond structure but we're also particularly terrible so we're sort of like excited about thinking if there is one way where we can tackle this problem of unbound docking even if in in proteins which the backbone doesn't move too much but right now we're we're not there yet be time for just one more question um david de sancho's asking um the range of binding affinities is rather broad um what have you learned from the refinement steps that you took to get down to the nanomolar binding range yeah okay so that's you know that's you know that's like the 1 million dollar question because we all want to do this from from the beginning right we don't want to really be doing any optimization so i can tell you what what we have learned so what we have learned can be summarized very briefly and you know so massive gave us a really good um hydrophobic core to ground the interaction but when we started designing the outer ring of the interface we have designed it pretty poorly and um and then what got optimized ultimately in our in vitro evolution experiments was the outer ring and what we saw was basically you know networks of hydrogen bonds you know a lot like you know like what has been described as tanya's talked about you know very earlier work that she has done in protein-protein interactions on how hard it is to to actually find um the proper hydrogen bonds and electrostatic interactions so i think we're missing a lot of that still so this this low affinity comes initially from a a good interaction is a bit sort of like analogous to this to this enzyme design aspect right you know you can get sort of like to some point but then when it's time to sort of like refine the details uh and particularly refining the details for interactions which are more challenging for for the scoring functions that we are currently using that's when things break are breaking down fantastic bruno thanks so much again for yeah really exciting seminar thank you alrighty um it's been a really just super fun session um and uh uh you know to to sort of bring us home here i'm really excited to introduce professor beer to hawker um so professor hawker she did her graduate work at the university of cologne did post-doctoral studies with hama helenga at duke and then started her academic career at university of tubing and now she's moved to buy a ruth recently and uh oh yeah really excited um she's really uh i think one of the first people to take design in a major way and use it to help us understand protein fold evolution and it's really i've learned so much from from her over the years and i'm really excited um please take it away thank you so much chris for the nice introduction and let me just start this up here i hope you can see right screen okay thanks a lot for having me here it's a pleasure and it's been so much fun already listening to all of you others um talking about this really great um yeah new ex developments in the field and so well there's many of these things that i'm interested in as well and um but i as chris said often take this view that i i actually i always say i'm so impressed what nature does and this is really also i guess what we all compare ourselves to right so this is just a picture of from david goodson who's using crystallographic data to um to show structures and in the context of the cell so in the background you see all those contract all those proteins on the surface of cells um or that are interacting with other proteins or um involved in the gel emotion or um i should actually use a pointer um and uh and then you have here the polypeptide chains coming off of ribosomes and folding and interacting with dna so all this diversity and even though we've already heard um many of you have said maybe we are not actually sampling or nature has maybe not sampled all protein space it's amazing of what it can do and how precisely it can do this and so i guess um there's always these questions um how did this come about how did nature sample all these different possibilities and can we get close to this and can we actually also do this and so i guess this is always a challenge this is always what we compare ourselves to but i want to argue here it's not only a challenge i think it's also a big big resource that we can use in building proteins and um yeah that's i think the message that i want to bring across but let me step back a moment and uh kind of show you the view that we've assembled in one of those reviews where we um how i see what you can do with protein engineering and design approaches because it's all about coming getting methods together and actually using things together so it's not only anymore we do direct evolution or design it actually has to be working together but there are different ways we approach this by either starting with natural proteins and um and reusing them in order to build new functional sites into them binding sites enzyme science whatsoever and then also the whole idea can we build proteins from scratch so that's i think the two sides of uh of the spectrum and i actually want to tell you a little bit about the middle part about actually using parts of this but just briefly we are also trying on both sides so and i think this is just one of the things i'm very excited about we were able to use a repressor protein and change the binding site so it actually now binds a very related compound that is a plant uh hormone and now um we combine this protein which is a dimer and actually you can attach fluorescent protein make a frag sensor out of it so you take domains you take existing proteins you change binding sites and then you can make a sensor and this is actually something that now by tinkering with this by engineering um changing the binding pocket of a protein we now can make senses that our colleagues in plant science can use to follow a small molecule so i think we can go a long way also this really just using these using this really um great proteins that nature has can provide us and just sort of tinkering with them on the other end we are also very excited in trying to learn about how some of those protein falls are um can be designed and we've been dabbling in de novo tim barrel design which is one of my favorite folds and we worked together with alejandro velasco and david baker in building actually a tim barrel from scratch and we've then tried to actually learn how we can diversify um make a whole family of different stable structures uh stable timbers and actually also extending them in order to make them more towards something functionalized but i have the feeling we are still a long way even though i've seen you guys doing so much great progress here i still feel doing de novo is a long way so why not use what nature is providing us as well and that's why i'd like to focus on this middle part because this is also i guess sort of my special view over the years that i think that nature can provide us with very interesting pieces and actually not taking a full domain by taking fragments and combining them and then diversifying on that level and actually taking interesting pieces that might already bring something with it that we don't have to design in over them all right why do i think this is something one can do because nature i think has done this all the time this is the way how nature has been making all these diverse proteins and so we've already heard that if you take a random sequence it will you know it will most likely not fall so there are many many sequences that will not do well so as soon as something works we all stick to it right so we like to use use it again and that's what nature is doing if there is a domain that makes a nice function turn over something then you can reuse it again and use these models and combine them with other modules to tweak the function and that's basically what nature has been doing it's using domains to end assemble them with other accessory domains in order to tweak the function so domain i think everybody agrees is an independently uh is is um a building block but what we have seen in the recent years is also that domains themselves can actually be cut into smaller pieces because domains are already pretty large so they can't really have evolved in oval so um actually there seem to be smaller fragments that have led to the rise of domains and i would like to show you that here in a range of different protein domains on the left we have my favorite beta alpha 8 or tim barrel and we have evidence that this one evolved from half a barrel or maybe even a quarter barrel but this is one that has a clear two-fold symmetry and you can kind of see this beta alpha four element here that's um in there twice there's also the beta tree falls where you have the threefold symmetry for example or then beta propellers this one is a wd-40 with uh seven blades and each one of those blades is very similar so the better propellers is something where it has been shown that they evolve through duplication infusion or um like um beta hairpins like in membrane barrels so this is something that has been tried with um designing our proteins by actually just taking these fragments and um putting them together so i myself have done this together with dana we've taken her half barrel duplicated infused it optimized it and this is actually crystal structure of a duplicated half barrel and so this is a way forward how you can take one piece and by um duplicating and fusing making an uh a similar protein structure and also for the bitter tree foil um rheumatol and um labor they have actually used a similar approach similar approaches to build these structures but this is only one way you can do it and i think what might be even more interesting is starting to recombine things because then you can take different parts with these pieces and so recombination can also be done and we have done this within the tin barrel by taking one half of a of um a protein from history by synthesis pathway and then from another one that is just um catalyzing the successive step in in the pathway so those two tin barrels they look very similar but they only have about 25 percent sequence identity so they're pretty far away in that respect so taking the n terminal half of one c turn half of the other recombining them we actually end up with a really nice protein and by then evolving it further you can actually um there are only a few mutations important uh necessary to actually make this a protein that is just as active as a as a natural enzyme so these are ways you can make proteins and obviously we think by recreating these steps um i think we can nicely show that likely nature has done it this way but now what about taking actually pieces from different folds from recombining between different folds and that's something that we have tried and um actually it started as a summer project with tanmaibalat and where we actually just based on this high structural similarity of these py this is a response regulator protein that part here in green it looks very very similar to the part here in this in gray and this timber protein is f and just based on the similarity we decided let's recombine see what happens whether we can actually make a nice protein out of it and it worked really nicely and you could solve the crystal structure you can see it here that it's a nice barrel like shape interestingly there was also the c terminal end that intercalated between the first and second beta strand of the qy fragment so we are not really getting exactly what we we expected but we could actually learn then from the structure and start to optimize it and we optimized the interface actually working together with ian smiler using rosetta design and we introduced five targeted mutations and with that we got an even more um well-behaved protein from which we couldn't solve the crystal structure but the nmr structure this was together with miracles at the mpi intriguing and that's shown here so basically what i can show you here is also that you can take fragments from different proteins recombine them and then actually um with a few mutation at the interface um have a very uh compact a nice structure and it's not only that the structure is nice it's actually that we carried over a functionality and that is a phosphate binding site here in the his f from the his f protein it actually has two phosphate binding sites one we lose the other one we keep and so this is an anchor point that you can use and we actually used it in this um optimized tim barrel and it already binds a small molecule um from an analogous reaction so we can't use the one from his f because we don't have the second phosphate binding site but there is an analogous reaction tryptophan biofences and by just tweaking it a little further with two mutations we actually hit an um a binding affinity that is like this these well-type enzymes so what i want to show you with this is then you can recombine tweak it a little optimize the interface and you end up with something that could be that is competitive to a natural world so we feel that with these approaches we are following um evolutionary events or mimicking possible evolutionary events you might actually end up with something that you didn't want exactly something we call a hopeful monster that um actually could even be an interesting other fold we still need to follow up on that this nine stranded barrel because that has not been observed in nature um but then we actually converged back here towards the normal tim barrel and in the end we end up with something that is functional well-behaved and actually something that you can't really differentiate between flowerdogs from normal timbers and actually you can't really see that this comes from a fluvadoxin protein so this triggered all of this triggered this question what is actually affect with these uh what is actually the case with these two folds are they maybe even truly evolutionary related we only did this based on structural similarities and so we started searching and uh analyzing all those sequences and we did a larger scale approach because both folds the beta alpha eight barrel and the flower dachshund like fold there are super full there's so many different um copies of them known and when you go to for example a stop database where you find on the fold level you find these two folds you find many super families the tim has 33 um flavadox and has 15 super families and so we started comparing all of those or rather jose did this he actually built hidden markov model profiles and then um compared all against all and that's what i'd like to show you in these next slides just trying to put all of that data into one one graph so on the left we have all the 10 barrels on the right all the fluvadoxin proteins and i'm going to draw a line whenever we have a bidirectional hit with a significant key value in hh search and so when you draw those lines within the folds then you see already very different views on those two folds for the twin barrels you have many many connections so it looks as if there's a monophyletic origin of this protein fold while on the flavoxon side you actually have very few so either they have diverged beyond recognition or they actually have arisen multiple times now what was much more interesting was that there's plenty of hits that we get between two superfamilies from floridoxen and all of those phosphate binding timbers so this number one here are the um qy like proteins actually the fragment that we used for our chimera genesis and then we also have the vitamin b uh b12 binding um fluorodoximes and these seem to be really related to all the phosphate binding barrels now we went further and we looked for sequences that are ambiguous in their sequence um uh signature where we found similarities towards both fluvadoxin and timbrels and we actually found a full family that's sort of in between those those two super families and solve the structure of one of them to see whether it actually how it actually looks like and it turns out it looks very much like a flower production protein but it has a few features of a half barrel because it's much more round it has more curvature here and it actually also had ends up with a with an extra beta strand well what we also noticed was when we started looking into the sequence similarities where all those similarities are most um most obvious they're not all over the full fragment or full protein they are actually within a fragment and in this case it was mostly in this fragment these alpha beta alpha beta fragment around 45 amino acids in length and these seem to be actually fragment sizes where you most um very often find these similarities but we also find larger and we also find smaller ones all right so this kind of led to the next question if we can do this now on um on those two folds and if this are actually fragments that we also can use nicely for chironogenesis what about all of those other folds that are out there what do they actually compare and can we maybe identify interesting fragments for more chimera genesis um designs and so we actually set out to build a fragment database and we did these hfm profile generation of all the ones in the pdb uh or in scop actually where we know what the um whether it is associated with a with a certain fold so stop um this was stop 95 and then um we compare all against all match those hits and put them in our database which we call fuzzle for fold puzzle because we want to use them to build new proteins we also connect them to the data that when we do structural superpositions so that once you look at a hit you can also see whether they're structurally really looking the same and so we have many many hits and so now there are many pieces that one can play with so to save if you want to use it you can actually use this on our fossil website you would get you can search for um your favorite protein for example you can put in your pdb or a sequence up here and then when you look at a at a certain um and a combination when you have hips for example i like here between is this is again a twin barrel versus a fluvroxane protein you actually have these graphs so we have made network graphs so that you can actually um look at the so the nodes or domains and the linkages are then showing the similarities and you can click on those and look at the structural similarities and explore them so if you have a favorite protein please feel free to check it out we wanted to check out what um i wanted to look at what is in fuzzle actually um or what the connections are in there so we made a big network or rather noilia did and so this is um the well not the full data set it was filtered because otherwise it was too much but we actually have here now all the different nodes each dot is one domain and uh the connections are the languages and they're colored by their um scop fold and so you can see that many of them have some connections but there's this major component so there's a couple of them that are really tightly connected and that's shown here on the right hand side and you have the stop identifiers up here so c1 are our timbers and c23 are flavidox and proteins so again the players that we already saw before and there are many others that others have been picking out so um interestingly you actually have the c1 and c23 twice that was very interesting to us because we get the hits um because you can have fragments multiple fragments in one protein that have a hip so you can have the same protein in these um in these um networks multiple times and so here you actually have one half of the tim barrel linking to the flowerdoxin and here's the other half of the timber linking to the flowerdoxin showing that most likely timbers evolved through duplication of a flavidox and light fragment or something like this i mean we can't go back and check all right so um this is the surface of the or this is a fault fuzzle full puzzle database um if you want to access it please use this um this website um so what have we done we actually i started out telling you that we've been building uh camaras we actually have not only built them between timbers and flower toxins we've also been trying between uh pvp like proteins for example we have evidence that the pvp's most likely also evolve from the flavored oxide but in a different manner like the tin barrels there's also the humd like proteins so this is a playground where you can now pick fragments that nature apparently has been um reusing them and so in some way they are very there are some popular fragments and those um seem to be interesting also for us to use for some of those camaros we have also looked into the functionalities especially for the hemd because they bind vitamin b12 and so that's um a vitamin b12 derivative and so this is something that we're also trying to explore so let me just sort of wrap up by saying what have we learned so far and what is still missing so where do we need to go one thing especially towards functionality it's ligands so ligand information is really something that we want to have in there we actually have been building it in there now and it's supposed to be released very soon so we hope that in a so very soon fossil 2.0 you should be able to also then look at the interactions to the ligands um in fragments and then i guess the other thing is um making these cameras automated in an automated fashion because we've just been doing this by hand a lot a lot of tweaking and i think it would be really good to do this in a larger scale so this is something that we still need to do but we have our first tool that um noelia phillis has actually been developing where you can automatically now fetch um fetch targets from the puzzle database and you then end up with their networks you can pick one of those for example in the manuscript that is out there we compare the rosman and the p-loop proteins and then you can build chimeras where you just um combine these fragments with each other and then we score them you can screw them by energy and analyze the structures at least some of the basic things so far like hydrogen bond networks salt bridges and something that we feel seems to be important hydrophobic clusters because we feel that if those are broken that is often not good for folding and so um this is something that we're looking into at the moment we don't really know yet which of these pieces are most important i think this is what we need to figure out over time but this tool is out there and we encourage you to use it you can actually get it on github and um please feel free to use it so let me wrap up um well what did i show you we think that proteins evolved not only from the domains but actually from domain size fragments fuzzle is a resource where you can actually explore those um and kind of look at them we look at them as a natural building block that we might use actually also for designing proteins and we already showed that some new proteins can be built from these fragments and that the fragment properties that we carry over can be an advantage so maybe you have a faster way towards certain functionalities and we started um now uh providing automated tools for this camera design and evaluation so i see this as a complementary approach kind of um combining um with either de novo fragments that you might have or actually things that we already learned from uh designing um or or changing engineering natural proteins and so this fragment based design is just a way to use uh what nature is giving us as a resource and trying to build new proteins new variations of what we have so with that i'd like to actually thank my great team that's um the current team here on the right-hand side the people in um bold are the people working on camera genesis and then also many alumni and many people who helped us along the way and great collaborators in uh tubing as well as pyroid and thank you for your attention i hope i didn't talk for too long no it's fantastic uh virtua thank you so much um we have a lot of excellent questions and we'll start off with one from a graduate student at uc davis uh huang asks um for the recombine uh fold protein do they perform different mechanisms or the same mechanism but lower activity like it's sort of asking about the activity of the fragments that you're pushing you're combining together well i assume this is probably for some of the first fragments so and um we haven't really so for the for this chimera out of a um out of the tim barrel and the flavidox and protein we're not looking at enzyme activity here we've been looking at binding so far so we've been using this binding site the phosphate binding site and then improved on the binding we haven't really looked at turnover but for the recombination of two halves from two timbres we actually we looked at a um that was a that was an activity from uh listed by census pathway so this is an enzyme and there we actually measured we measured activity i'm not i mean it depends on what you actually want to look for because the flavadoxin protein it doesn't have a typical enzyme function it actually has a it's something that gets phosphorylated so that's a very different function to look at so i guess it always depends on on on what you combine and what function you're looking at i hope that covers it i don't see the questions right now um but i see tanya didn't either chris you can see the questions yep okay that's strange we're looking maybe you guys can prove it here okay well tanya can ask a question see super neat bird has always always been um so let's see uh verner braun asked um sometimes putting a piece of another virus into a viral envelope protein works but other times the piece gets pushed out already at the dna plasmid level um by e coli or we do not get virus at the rna level do you have methods to predict how this could be better integrated so i think it's questions about how viral proteins are combining and recombining oh wow that's uh that might actually be quite special also to vira to the var environment i mean we've been not really looking at at specific um now specifically at viruses i guess we've been mostly looking into whether those proteins will then whether when you recombine them whether they have the biophysical properties that you look for you know structural integrity whether they still have a certain shape so it might be that those can be recombined but just the uh the virus doesn't want them i mean if that is pushed out maybe i don't know amy is reacting there oh sorry no i just i'm technically so so um so the thing what what we usually look at is then um when we take two pieces we model them together we actually use the same um kind of analysis like i guess of all of us here are using just trying to to look for um structural integrity whether you have surfaces on all the all the residues at the interface are interacting um whether there's any holes in there whether these pieces are fitting i guess that's that's the way we would predict it um but that's nothing specific to viruses back to flavadoxins colin smith wants to know if the flavonoxins are a common ancestor do you have any idea where they evolved from that's a very good question um we don't i mean we we can't necessarily even say that they are the ancestral ones to the tim barrels so many people have tried to uh understand or predict which ones are the most um ancient protein fulls and tim barrels as well as flavored oxides are considered two out of the five most ancient ones um but just the way i think about it is uh you know the larger things get the more complicated so in that respect i would always assume something in terms of the flavadoxin would be ancestral there's many many rosman folds that are also very similar but then again when we look at those they're not all they don't all share show these relatedness so it might rather be that it's a smaller fragment maybe a quarter actually a beta alpha beta element those have been seen or observed as described as super secondary structure elements because once you have such a smaller element you actually can have an hydrophobic core and so that could be a starting point for something to fold by themselves and that might actually just be the starting point to all those different protein folds so i i i would hesitate to say a one-fold is the insist of the other i think they have a common thing that was before and that was either something around 40 amino acids or maybe a little longer and people have speculated where those came from and there's a lot of hypothesis going around whether they evolved in the context of rna or maybe in the context of helping other molecules fold or whether it was actually driven by metals so there's lots of literature that we can discuss this for a long time i heard messner wants to know if the repeat modules um from tim barrels and other structures match with genomic exons um no not not necessarily no they don't um but then i not not as far as i've looked but i haven't really looked in that detail um so the repeat is also something that you don't find in all timbers um it's not that obvious anymore so you actually have to look for it um it's not a high sequence similarity so within this um 10 barrels from histidine biosynthesis it's also only around 25 sequence identity um and uh and that's already fairly high and then when you um when you look at other timbers it seems to it seems to start to die diverge but the de novo tim barrel that we actually designed that was a four-fold symmetric one and so um i think it has to do also with the shape that you the timber needs to adjust to so there are some of those that are for photometric that are very round and then the ones that have the two-fold symmetry they're more oval-shaped and i think that has to do with something how the protein then also might adapt to find the right binding pockets because all those um or enzyme um enormatic sites because all of those timbers are enzymes and i think they they adapt properly to those shapes and then we won't be able to see all of those symmetries so well anymore in natural proteins excellent so we have many more questions um but i think the we're at 2 p.m so i want to be respectful of people's times um so so um first is we've got um some spillover time so anyone who wants to stay and discuss and get questions asked please stick around um but uh i would like to really just thank all of our speakers one more time for uh for really excellent presentations today it's really fun to see all the cutting edge stuff that's going on in our field um i especially like to thank uh the protein society amy keating for really making all this happen and uh also reluca kadar who organized everything behind the scenes today um yeah so thank you reluca so much um uh she she's been managing the questions behind the scenes and keeping us on track and so uh yes just really excellent um and um and thank you all of you for coming out today um it's really great that we can still get together in some in some way uh during these uh during these challenging times so everyone for your you know the audience for your enthusiasm and questions and um uh yeah thanks again um and so um with that yeah let's uh let's maybe continue with some of these awesome questions um we've got four outstanding questions in the q a um from many of our speakers tanya did you want to ask britta something because maybe your question didn't make it into the q a i do yeah um so so bert is super neat um so what we are seeing with our um sort of shape designs is you know they're based on rosetta they're not using sort of um motifs i guess um but we find tertiary motifs in the designed proteins in the end right and so you know we compare that with some of the things that you know gavor gregoria is doing um and so so that prompts the question for your elements that are sort of being reused um to what extent can you now link the fragment-based approach to the the physics-based approach right so to what extent do you find sort of common um physics-based interaction motifs in your database that then maybe are the strongest predictors of being able to mix and match things hmm what do you mean by a physics-based motif tiniest before yeah what do i mean by this so so what i mean is so maybe there are particular um stable arrangements of side chains that you know are then um sort of uh micro course where you basically get um stable interactions between the different fragments or between different elements and so that also reminds me of you know things that that bill degrado has recently done right with the the vandemere based design you know so maybe you want to have sort of not just a sidechain functional group but you want to look look at the context you know what side chains are next to it in sequence but also next to it in in three-dimensional space to form these these defined motifs that are just reoccurring in many structures yeah and that's a that's a very good point and it's very very interesting would be to look into how for example the terms of org are comparing to some of those or or maybe just looking into what what um what terms pop up again in those fragments we haven't really done that but it's it might actually be worthwhile to see whether there is something special about this um the thing is that with some of those fragments that we have um you you might actually then have longer loops you have longer parts there that you where it's not the same but where you can actually use those fragments also as a as the the point where you can merge these things so it's not only about so with these fragments it's not only about what is the most stable thing but maybe this is also a way to to then use it as linkers so that's that's one of the other things that i haven't really been talking about but um yeah we haven't we haven't really compared it like that i would say it's i think things are coming up right now since we started doing also folding studies on some of those chimeras and trying to figure out um whether they behave the same way because we've seen this with the nova design proteins that they have sometimes different folding pathways you know and with the um with with these chimeras um what we were hoping to to do is also that or maybe to find with these fragments something that can be a pathway to what's folding better and so that's something we're trying to address but and one thing we've really been seeing is that um these hydrophobic clusters seem to be important that if you break them or if you extend them that that makes them better but this is just superficial so we still need to dive into that more deeply i guess yeah and sort of a follow-up question sorry if i um but this is such an interesting you know space to to do analysis in i think that you know of course um a lot lots of excitement about deep learning right and the question um what what are deep learners learning and so i i wonder um and i'm i'm i'm a fat bruno is um having fun i don't know whether he was having fun about what i just said or something else but um so so but i'm i'm a fan of deep learners that that actually learn something and um and so so i'm wondering whether there's something interesting in in your space that would lend itself to the architecture of a deep learning model that can actually tease out some more of the interesting relationships that you see this is a very undefined question but but there might be something interesting there to start thinking about it yeah so you're suggesting that one should just try to learn on those fragments well but maybe you want to learn on the fragment and the space that you stand with the relationships between proteins yeah you know what the difficulty is with some of those fragments there are the definition of them is also i mean we always talk about i always talk about these fragments but they're not that greatly defined because it's sometimes hard to say where it starts and where it ends and the way we look at some of those fragments once we have a hit we actually see that it extends you know so you go from one protein that where you find a similarity to the next and you see a small fragment and then you look to the next one and it's not using exactly the same but it uses it shifted and so you can actually walk through one protein and hit with different areas different other proteins and so i think that might actually be a problem in defining what kind of space we are actually looking at because i wonder whether we in the end not end up with almost the same space general space you know what i mean when do we or maybe we have to focus on just the general comp on the major component the pieces that seem to be really highly connected and get rid of all that other stuff out there and then focus on that that might be the the parts to look at and then we probably end up with almost all those beta alpha proteins that seem to be really really highly connected no but very good points to think about not entirely sure yet how how one would do that and i would love to know what uh machine learning is really design uh learning on absolutely can i just rephrase to see if i understand like tanya are your two questions that are related in that you're wondering if these conserve these these recurring motifs share some fundamental properties and you refer to that as maybe they're you could recognize them through their physical energy scores as being favorable but maybe it's not so simple and instead we could use advanced machine learning methods to try to understand what the properties are that make them have this special role in defining the protein universe that we've got yeah yeah well so i mean you know what what the physical scores between you know residues that that's what we all use right and and and we all know where this breaks down and you know in terms of higher order interactions that we just don't capture very well um and all the pathologies that you know i think each different design scoring function has you know slightly different pathologies but i think they all suffer from um not taking into account some of the higher order interactions that we just we just can't compute properly um and so so i'm wondering whether um there is something in in in the patterns of higher order interactions that a um that that an advanced deep learning model could capture and maybe the using this this relationship between proteins that that you have bird on is is a useful way to to span the space of relationships between proteins and maybe you know i hereby to what you're saying which is you sort of get these these you can't just define one pattern and then look for it everywhere but it sort of gets added on somewhere and subtracted on somewhere else um and i think that might be sort of an advanced pattern recognition um um idea that that that could potentially be captured but maybe you should stop talking i really haven't thought this through to the level um more discussion i think this is great actually maybe i'll throw just like a general discussion topic out for all the speakers this is kind of touched upon many areas as machine learning becomes more and more important for our field right we run the risk of sort of deep learning but shallow understanding right um and so how can we what are your thoughts on how we can best leverage machine learning to help us understand what's happening to not just be able to predict it so i think i share the same nightmares that's just basically all that i have to say right you know ultimately in the end sort of like the fact that you have some type of functional protocol uh it ends up being sort of like not very rewarding if if you don't really know what you're capturing so that's the bad side of things but i guess sort of like picking up on on on what tanya was saying right which which i agree that is sort of like a complex problem but i guess the interesting aspects of of the more machine learning techniques is that they perhaps will be able to take us out of this space of the the two body interaction right which is basically the space that we've been locked in our scoring functions right it's always like pairwise right and and sort of like we are pairwise pairwise pairwise everything but perhaps here you know we may not we might not be able to understand exactly what they're capturing but at least we know that they might go beyond the pairwise interactions which i think there could there should be some scheme which would allow us to do this okay so so this that this part i find exciting how much you take out of it i i find it scary bruno can you maybe relate it to you said you did these ablation tests to try to look at the influence of electrostatics versus geometry and but also you suggested that like geometry is not really a feature that features some sort of obscure high order thing maybe can you just comment on like how when you're analyzing the models with ablation how are you trying to figure out yes exactly so i i think i can i think i think what we've seen sort of like you know i think we actually have seen sort of like what what intuition tells us okay which was for instance if you have like a very um hydrophobic interface then typically geometry captures it really well right because it's mostly about complementarity so you see that but when you actually start having a lot of polar interactions then this is no longer the case this models kind of break down a little bit more um and so so that's that's what we have seen there is a dependence from the constitution of the interface itself to the factors you you're taking in now i think the key here is that you know what is more important right you know so like well there's one matter more than the other and i think this actually varies from interface to interface but i think ultimately you know what what the approach that we have sort of like explores is basically this this this concept of complementarity but not only geometric complementarity but also chemical complementarity which you know in a way it's kind of written in the books right so like we all know that these things to interact they have to be complementary but um this to say that we understand that they both contribute to better performances uh and depending on the constitution so we know that for instance this model if the if the if the interfaces are unusually polar it doesn't do well so that's that's very clear to us now you know we can argue well is it because of like salvation or it's because you know one thought that we have is also because you know our the training sets are highly biased towards um hydrophobic interactions so sort of like you know the patterns that it picks up is more on the hydrophobic side than the taller sides so that's basically what i have to say let me contribute charlie brooks's comment because it's stuck in the q a charlie weighs in with i would suggest that we're just at the beginning of understanding how and where we can use machine learning in applications to protein design and related aspects i think that deep learning does have the opportunity to capture the higher order terms that tanya is suggesting this is certainly the case in prediction of contact in proteins i'm sorry if other people have things in the q and a that are someone asked about alpha fold 2 to all speakers would alpha fold 2 be a suitable tool to more quickly evaluate designed folds maybe the de novo people versus the designing from natural fragments people well i i'm still surprised we have a job after alpha fall 2. you know i thought that we were all redundant now i think we could use alcohol too then we might know yeah so but okay i think i think related to your question maybe i can just say sort of like a very short comment you know i i i i am not sure how much it relies on alignment information yet i think i don't think we know apparently from what what the experiments on casp was saying was that you know even for proteins which you didn't have very deep alignments it it could already do very well that's that's what we saw so i think it's fair to say that you know perhaps it will work well for design proteins i think it's fair but we i guess we'll have to see most likely the next cusp is gonna test this by adding some proteins that are not folded right so i think at the moment in casp you have only folded proteins that you test so the next test will be can it also predict the ones that are not folded and that's the problem we often have with the designed ones right i take the conversation briefly back to structurally conserved tertiary motifs because i want it which we were talking about a minute ago because tanya i wanted to ask in your new topologies that you said don't look native like in their arrangements well this is what i wanted to probe in what ways they looked non-native-like and if you go to um smaller units of of structure maybe these compact motifs that govord's been defining are those can your um designed proteins be described with those do you know um yeah so i think i should first define what i mean by non-native right so we have a what the analysis that we've done um is is very specific where we basically so we're moving uh helices around um in that initial test and so we can define the geometry space of helices by the centers of the orientations of helices we can bin that space and then we say do we are our design proteins do they occupy bins that we don't see in the pdb for that topology and that's an easy test because we don't have that many proteins in the pdb so so um so so but the space is large and you know you saw the smear right so so we're sampling many different um um he likes uh centers and orientations there so that's the definition of there's something new here um the so then the second question you know how does that relate to um to gevort's term analysis yes so we do see sort of the smaller pieces of you know how a loop would interact with the helix or how a helix um interacts on a sheet we do see those represented when we look for what's there in the works database of the terms and we found that very interesting because we didn't put those in are we getting them out but we haven't done sort of this systematic analysis how common is this and you know what is the you know how large is that signal over what one would expect sort of by random chance kind of thing that's very cool yeah so if i can follow up with tanya a little bit so tell me just one question about about those designs do do you also think that sort of like the stability of the initial protein allows you more space to sample like more aggressively other motifs that you don't find i i guess i think i think it's clear what i mean right in the sense in your designs right you know did you see sort of like you know stability kind of like i read the paper but i forgot now so the designs actually got more like stable less stable uh yeah so that's probably on you know supplementary page 53 or something so yeah so so the proteins we characterize are also very stable um so so we're again sort of driving those into into the deep minima so that you know we basically you know you have to cook them to unfold them um but so but what you're getting at is you know functional science are you know they are um intrinsically they mostly making proteins less stable right because they're there in the protein to be complemented by whatever the protein does it binds or they might have to you know have the dynamics that that we're all talking about might be important um and so they are not these deep minimize so basically what we're hoping is that maybe these geometries um we can design very i see that i have someone else listening to me bruno from the best well maybe she wants to do a rotation in the lab um so so the the shapes that we have right now are also very stable this might just be you know how conservatively we selected them but we're hoping that we can then if we make those very stable we can afford destabilization because of dynamics or other functional states thanks uh shuguang wants to know how we're gonna deal how you're gonna deal with intrinsically disordered proteins that bind fold upon binding there are many naturally unfolded parts domains they fold upon encountering other complementary parts domains how to understand and design them i mean we do know you can design their bound states um and often get them to do what you want so the simple answer not to be facetious is like ignore the unbound intrinsically disordered state which that certainly works sometimes i think roberto should handle this one the way we design proteins would have to change for the intrinsically disordered ones because we're always scoring to favor folded proteins so the energy function is designed for that i don't know how we can handle something that's unfolded bruno can i ask you a question about mo topo builder like when you're trying to can do bottom up construction around a pre-chosen motif um can you just elaborate on how you're choosing what you should use to complement it yeah so i think that's you know that's essentially it's sort of like it's the first problem that one faces right basically we tried to kind of keep the topology small and which we try to build topologies that support the motif right where you actually embed with uh with backbone contacts so of course that what we've done it's a very small sampling of the space so oftentimes we do have to kind of like you know take some ideas from other proteins that are around we we do sort of like very rough matches in terms of the the backbone so like to see which kind of topologies this can fit and then so like we go from there there's of course you know the other you know one of one of the functionalities that we built in is is based on you know so like this very old work from willie taylor uh which you know many many years ago proposed this this idea of like a periodic table of folds which you could describe them with strings and so actually something that we built in is sort of like you know a string based method so that you can kind of just do different strings and this will tell you which topologies they represent in the 3d space of course some of those look proper some of those don't look proper um so that's still sort of like you know one important aspect to solve is to come up with the initial template like it wasn't you sort of showed a slide where you had your maybe helix turn strand motif and then you showed additional secondary structure elements popping up to complement it but it wasn't clear to me if you were choosing them one at a time and sort of building outward or whether you were looking for a topological framework into which the motif could fit yeah so no so i think there you know that movie basically it what it does is you know it's just assembling already a predefined topology that that we've sampled in this in this initial stage just with these 2d diagrams or the string based diagrams um so there's not really much of a smart thing going on it's just that we we predefined which topologies should should should be built we try to sort of like pick things that already have local contacts with the motif they still need to be compatible with with the binding partner right so that's that's an important limitation in terms of what the architecture of the fold can be um but that's that's about it okay and so the search for things that can complement the motif is done in 2d before moving into uh-huh there is actually a question about enzyme mechanism for roberto that we didn't get a chance to ask during the main um but i think it's really exciting i wanted to ask um from uh govind and subramian um how do you comprehend the non-core mutations as relevant to enzyme activity and how do you assess the contribution of these yeah so there's a few distal mutations that we think we understand what they could be doing uh some of them are found for example at the entry of the channel the substrate entry channel so you know you make the channel wider it facilitates substrate entry there are some mutations like that others we think are shifting the confirmation ensemble towards the more active state the one that stabilizes the transition state and and obviously you can have also distal mutations that allosterically will impact the confirmation of residues within the active site and possibly help to pre-organize them so this is all i mean i i can't confirm with each individual distal site is doing in these enzymes but that's what we are hypothesizing could i follow up with roberto a bit on the on the enzymes i guess you know so in some of your of your mutations you know i i guess they were sort of like surface residues that were very far from uh you know from the active site you know and of course you know i think the second shell you know intuitively we get it right why it's important but when when you're sort of like out there in that protein space or that protein structure that is so far away from the active site what do you make out of these mutations yeah i mean that's a challenge it's very challenging some some of these mutations that were found by direct evolution probably were found not because they directly impact the you know catalysis maybe they just help to stabilize the protein so that it can accept these other mutations that are important for catalysis so it's going to definitely be very challenging to identify distal sites that will impact catalysis positively so this is going to be probably the the next challenge i think you know as our data shows we can make better de novo enzymes than previously possible by focusing on the active site by making it more complementary to the transition state but eventually if we want to get to really high activities we need to start thinking about how to install distal mutations that are going to affect you know either the confirmation landscape of the enzyme or allosterically impact the active site or maybe help to draw the to pull the substrate towards the uh the active site in the correct orientation etc and so roberto i'll ask you another one about this like last paper from dorothy sorry that i'm like kind of like monopolizing the whole thing but this this last paper from dorothy where they actually studied this this designed enzyme you know like i think it was the first time right we're so like this very detailed study with with dawn where where you actually can kind of like see how the enzyme breeds and so like what what were the lessons that did did you took sort of like any big lessons from there or there was anything surprising for you there yeah well um those enzymes were actually the same ones that we were studying so this happened at the same time and when we saw the paper it was super interesting because there was obviously good agreement with what we saw but they pushed it further and one thing that was really interesting is that they solved a high temperature structure i think it was 70 degrees celsius of one of these a higher activity evolved um temple eliminases and they saw this inactive state this confirmation of the backbone that we didn't observe in our crystal structures and when i saw that you know i asked myself how can i even predict this really different confirmation of the backbone starting from the you know this a structure that doesn't have any density for it and i think to me that is the big challenge is how do i know how mutations and conditions are going to impact the confirmation of the backbone because that's going to impact what i'm designing in the active sites in any case the the take-home message for me from that paper was really there's this equilibrium between an inactive and active confirmation and as you evolve the enzyme you start shifting the equilibrium towards the active state and that contributes to increasing activity but without knowing what this competing state is how do we design against it that's the challenging part roberto it seemed like your approach could potentially get at well if i don't understand the details necessarily but if you're doing ensemble-based design and you're identifying ensemble members that do better positioning of the active site can't you then also target mutations at distal sites to stabilize that ensemble member directly or are you not selecting out like a specific member of the ensemble but rather designing on the whole ensemble right now that's a very good point we want to do that we want to let's say we have an ensemble of 100 states and one of them is really good for stabilizing the transition state much better than the others what are the positions the distal positions that would favor that state over the others can we predict that so that i have a few ideas of how to do that but i haven't really tested it so i can't answer but you could evaluate the the the effect on the stability of each of the states of distal mutations the question is are we encompassing the correct states in the ensemble to do this that's going to be challenging and you know i think a lot of those considerations also be very mechanism dependent right because it really depends on you know what what what steps and what specific potential dynamics or confirmations are you actually limited by in catalysis and one thing i didn't mention is that here we're working with a single step reaction with a single transition state but a lot of enzyme reactions have multiple transition states multiple steps and we want to be able to exchange between them and it's not the same backbone conformation that's going to stabilize each of these individual transition states in the same way so we can't over stabilize one and then not be able to switch to the other ones that are also important for catalysis so now we started designing these multi-step reactions and still using an ensemble the idea is that maybe one member of the ensemble is better for stabilizing transition state one and then another member is better for stabilizing transition state two can we you know in doing so get better enzymes for multi-step reactions than just using a single backbone i am really sad but i have a 2 30 meeting i have to go to so i'm gonna have to log out but i i really miss seeing you all i wish we could have this conversation in the bar as bruno says but san francisco 2022 if not before okay um i'm gonna bow out i encourage you guys to entertain me questions and chat as long as you wish but thank really fantastic talks so exciting bye for now thanks amy thanks amy i am you see you around here in 22. indeed yes all right can i ask a question to my rosetta colleagues as you know i don't use rosetta so i'm naive about this but i find it really amazing how it seems to be performing really well designing the novel structures and i'm trying to understand whether it's really that easy or it still is a lot of work i mean what you show is really impressive so is it becoming plug and play or is it a lot of involvement a lot of trial and error to get those really nice structures um maybe maybe i take this one um unless bruno if you want to go for it i see you see you uh i was gonna i was gonna ask tanya because she has done this for a lot longer than i did so and maybe just breakfast my question uh i remember um someone told me that the design of small proteins is basically quote-unquote assault problems for rosetta it we can get highly accurate small proteins i you know i tend to be very careful with the phrasing of a solved problem i don't i don't know what that even means you know do we really understand it no can we do it sometimes if we pick the right problem yes is it topology is it a loop is it certain small changes you know i guess it's also a question what what is solved in that respect yeah yeah but yeah the answer your question i would say maybe the the easy part of de novo design is the sequence design the hard part is the backbone design getting the protein main chain in a confirmation that can hold a sequence is the really hard part once that part's done putting the sequence on it is fairly much easier only because um [Music] convergent evolution is a thing and there are many solutions that you can find for that one protein uh confirmation you don't have to find one and actually um nobu koga and um uh had a paper i think la just last year in um i think pnis where they they took some of the de novo proteins and they basically just mutated the chord of valine and the things still folded up so it's just like this polyvalene core and their conclusion from that is it's like yeah you get the backbone right the sequence is easy it just you got to get the backbone right right okay well and i think you know there's some really interesting so if you if you look at you know the no boost papers but then also anastasia in the beta barrel design that's some really interesting considerations right so there so there are certain things that you have to get right certain topological considerations also refer to your tim barrel right there's you know there's certain topological considerations that you really have to get right which i think speaks to chris's point with you know you really have to get the backbone right otherwise you're lost um and then you know i'm you know i i think that those considerations of you know what what the energy gap is i think still make a lot of sense to me right if you drive something into a deep minimum you can be quite wrong with your energy function you'll still find something that looks like what you designed um as long as you don't you know commit a horrible error just like putting you know a charge in the core of the protein which rosetta will happily do if you don't prevent it from doing it um you know so so i think it's those features like you have to learn something about the topological rules of protein space to to make your backbone and then you follow i think rather simple heuristic rules on how to put the sequence onto it um and then you hope that the energy gap is large enough so that you can actually con tolerate a considerable error and you don't have to deal with the the complex geometries of hydrogen bonds in many cases and and bruno i guess you you you also um you know raised that in the discussion right that you know for for getting hydrogen bonds right not only does the geometry really have to be quite precise uh which you quite often can't do because you just don't have the backbone in the right place to put an actually occurring amino acid in the right place to actually satisfy all hydrogen bonding geometries um but then also to understand the energetics of polar networks is is is not something that we or uh perhaps anybody else can can can do really on predictively and so if you know if you just generate a structure you avoid all these complexities in a sense yeah and does it happen sometimes that you you want to design a specific structure but you end up getting something very different but it's still folded does that happen because it does me like a good way to learn how to you know improve yeah so that's a good point so you know we have not seen that happening a lot we have not had a lot of wrong structures we might have had one wrong structure um although you know it had like missing density and whatever and from the missing density we could sort of like piece it out that it was not quite right but i think you know i would like also to pick up a little bit on what um what what tanya was saying you know which i i think i think she's right on this idea that you know there's a few determinants on the topology that you have to have right i think there is one thing though right is that right now i feel like we've been doing this you know in a survival mode which is this idea that okay look you know let's just build this thing as simple as possible and tanya's work sort of like you know in a way kind of like the parts from that by trying to sort of like building kind of like irregularities into this very regular structure but if you feel right you know these the structures are just so perfect in every sense right you know the loops are so short you know the packing is so optimal so i think every time that you go a little bit out of that space right and you start to like putting a longer loop or so like making a kinked helix or you know so like you know these irregularities which we typically call function i guess um you know things things just go just go kind of balloony right you suddenly lose the ability to discriminate what's right from what's wrong exactly like what you are living in your in your enzymes and the other point that i will just make is you know i think our experiments with making for our experience with making high-throughput experiments and and and really sort of like trying to um to test many sequences at the same time are are very humbling to say the least okay because basically when we start looking at the sequences which basically made the pool of the winners and start to try to discriminate those from the losers um you know i guess we don't we don't really have many good ways of discriminating them yet which which is highly concerning at least you know it concerns me right now um which in a way it means that you know we're somehow roughly on that space but we don't have good discrimination metrics to understand what the goods from the bad sequences are at least when we when we put them on the three-dimensional space so we had an example where we tried to insert a helix into this de novo tim barrel and it actually wasn't a normal helix but it made it 310 helix and it made a totally different interaction but crystallized fine i guess you need to be lucky to get those i think most cases it's so tough to actually get structures of these things anyways and so who knows how often we're actually capturing them so that's just one thing i wanted to add to i fully agree otherwise bruno i will tell one thing you know i i was kind of surprised with the differences in energy that you were capturing in between your different ligand poses this to me sounded like you know something that you know for many scoring functions would basically look kind of the same right you know it would be like one or two units above or below and i think you know there you said like you really had like these gaps which are like 40 units you know maybe this is sort of like you know some kind of like psychological aspect but i don't think it's a psychological aspect because one was right and the other one was wrong so this to me is interesting okay you know and i don't think i've ever seen i don't think i've seen this very often this kind of this level of discrimination between two structures that are kind of so similar you know what i mean yeah i guess i'm not surprised because small changes in confirmation can affect the energy greatly at least in the energy function that we're using um so so that wasn't so surprising and and the way i i took it was you know the energy of the positive control is what is as close as we can get to what would be like a perfect structure given the that backbone and when i saw the other one that was 40 kcals per mole higher for me it just meant okay this backbone's not that great for this particular sequence i might ask one more question from the q a this is for everybody it's from patrick allen the great talk everyone i wanted to ask an open question for any of the wonderful presenters today for someone with experience in the top-down approach and some molecular dynamics what advice would you give to someone to prepare for entering your particular lab for a phd or protein design in general i've seen the rosetta software in several presentations we're learning that software be something you recommend thank you having a community is super important so maybe as somebody who's not from originally from there but who's been you know everybody's been so welcoming i can definitely say that rosetta community is a community where you can learn and where you can engage so i think that's a tough part if you try to do things by yourself and there are not that many people to engage that's hard so i would definitely encourage you to to reach out and use that i think that's one of the really strong or really big strengths thanks to all the rosetta guys i guess one one other thing as a general consideration maybe isn't really specific to protein design but but really rather general which is pick a problem right pick a problem that you're interested in um and then you know engage with the people working in that space to find sort of the good overlap between it's an exciting unsolved problem but it's the next problem that could be solved by someone who makes advances and i think that requires really discussions within the community to really figure out where are people at what are people thinking about what are the challenges and what what is on the horizon yeah i guess if maybe i can say something also perhaps from the experience that i have of having students in the lab which basically know nothing about protein design no nothing about rosetta some of them might not even know anything about protein structure i i do think that they're sort of like two sort of like survival you know tools which is you know to do to computational protein design in whatever software you want and i think you know one one aspect is you know to to to know protein structure kind of like inside out to be very comfortable with sort of like those principles and then the second one is sort of like you know having like uh some basic uh coding skills that you can get on your own right you know sort of like you know just sort of like you know parsing files maybe taking a book of tutorials running a few things you know working on a problem like like uh like tony saying you know put yourself your own problem if you need to sort of like i don't know count amino acids in 10 000 proteins how would you do it if you cannot do it manually okay and now like you know try try to figure that one out if you're not used to it and i think this generally sort of like breaks the rock breaks breaks a lot of the the the energy barrier to entry to entry the field because then you kind of know the language already right you know you understand that we're talking people are talking about structures you know these are so like the toolkit that you need to survive in the midst of like so much data and i think these two for me at least seemed like a a good starting point yes i would add that um you know any tool can be used or unlearned the important thing is to to understand what they can be used for their limitations and their their benefits and like tanya said to think in terms of what is the question what are you trying to do and what is the best tool to help you answer this question whether it be md rosetta or something else and you can work on any project with any software you're not limited by one so even though it'd be great for you to learn these tools before you join a lab maybe when you join the lab you realize that there's other tools that may be better suited for what you want to do so don't limit yourself on the tools basically awesome thanks everybody with that um i think i need to to sign off so thanks again roberto tania bruno and birthday i really enjoyed getting to hang out with all of you today um and bruno especially i think it's pretty late for where you are and tania it's a bit early for you so everyone who made adjustments and beerta as well i think it's also quite late for you so everyone announcements for time zones thanks so much um thanks to everyone in the audience yeah it was a lot of fun to get to see you guys and i really hope that we get to meet together pretty soon thank you so much chris and and everybody and great discussions and great questions a little bit too early in san francisco to have that drink bruno but i will i will have one tonight in your honor [Music] yeah thanks so much everybody good fun thanks

Channel: The Protein Society

Views: 994

Rating: 5 out of 5

Keywords:

Id: xrPu1WQ8Q7A

Length: 164min 16sec (9856 seconds)

Published: Tue Feb 09 2021