 Okay, so I think we're going to start. So welcome all to the botoxycomputational biology seminar series. Today we have the pleasure of with a senior research scientist at the molecular modeling group here at the C2 Institute of Bioinformatics. She was trained as a physical chemist at Ludwig Maximilian University in Munich, Germany where she obtained in 2004 a PhD in computational chemistry at the Swiss Federal Institute of Technology of Zurich in Switzerland. Then from 2005 to 2006, she was a postdoctoral research assistant at the Enrico Family Research Center in Rome, Italy. And then from 2006 to 2011, Ute did a second postdoc at the Ludwig Institute for Cancer Research and the Swiss Institute of Bioinformatics in Mosin. And since 2011, she is a senior research assistant here at the CIBA, where she continues her work on in-cylical drug design for cancer immunotherapy. The Ute areas of expertise among others are computer-aided drug design, docking, homology modeling, and classical molecular dynamics. Today, Ute will share with us a work on computational drug design for the therapeutic targets in cancer. So Ute, thanks for accepting our invitation, and the floor is yours. Thank you very much. Thank you all for being here. It's a nice weather and it's the upcoming holiday. So my talk will be a bit divided in two. So I will first talk about one specific target in cancer that we've been working on for many years now. And the second part will be about some methodological developments that we are doing. And so the first part is about endotamine 2,3-dioxygenase, or IDO1, which is an important target in cancer immunotherapy. IDO1 is an enzyme that catalyzes the rate-limiting step in the panoramic pathway of a tryptophan degradation, and this pathway has been recognized as one of the central, being of central importance in cancer because it's used by tumors to evade the immune system. It has been shown that the depletion of tryptophan and the accumulation of tryptophan catabolites induces T-cell energy and apoptosis. Therefore, IDO1 inhibitors are in development for cancer immunotherapy, and importantly, these inhibitors are not thought to be used as a single-agent therapy, but in combination with other strategies such as immune checkpoint inhibitors, for example. And how do we go about very roughly to design inhibitors for IDO1? So the first step is done on the computer. It's done in silico and mainly by myself and Vincent Zouet in the group. We then go with our ideas to the chemists we collaborate with who are located at EPFL and if a compound is synthesizable, they will try to do so. They come back to us with a synthesized molecule and at the moment it's myself who does the enzymatic testing. Then successful compounds will in the following be tested in cellular tests up in the parlance by Nasli Dilek and Melinda Irving. Again, successful compounds can be moved into mouse models that are currently being developed by Nasli and of course the final goal of the whole story is to go into clinical trials at the sheet. I would just really very, very briefly show one approach that we use. So one approach we like is fracking-based drug design on discovery. Just one example, so here we docked small fragments into the ideoactive site. This is the structure of ideo which is available. Here we docked a small benzene ring, a small trisor ring into the ideoactive site. Our algorithm detects that those two are in a good position to be linked to each other. We do the linking, we dock again the resulting molecule and if you obtain good results for this, we go to experimental testing. So we did it for this compound and it actually proved to be active in the micromolar range on ideo1. In the next step, we try of course to optimize these molecules again using our docking strategies and so on. And for this specific scaffold, so meanwhile synthesized about 130 compounds for us and roughly half of them were active at least in the micromolar range and this might not sound so great but if you compare it to a brute force approach, so we had some collaborators in Belgium that said okay, we'll just buy everything that has the scaffold and that is commercially available. They did so they got 200 triosols and they actually found only two to be active. So it is quite a good way to design new compounds. Of course we looked again at the confirmation of these derivatives that we synthesized in a few or part of the structure of typical relationship you can really understand from the classical docking approach that we use. So here we have our parent compound. We see that in this position there's not much space in the active site when you put a substitute in there. The activity drops while we see that in this position, the meta position, we have a bit more space. We have a small sub-pocket. If you put a chloride ion there, it's relatively by a factor of 8. And also if you do incharmo hydrogen bonds within the ligand, we can obtain better compounds. But this, in case of IDO, it's only part of the story because yes, you do have the protein here, but you also have this very important in-cofractor active site. And there things become a bit more complicated because in order to calculate really binding energies or binding strength between a small molecule and this in-cofractor, you need quantum mechanical methods. And so we turned to this approach here. And there's just a small slide to show that really the binding strength is strongly modulated by the identity of this compound. So here we see that really the deprotonated compound binding to the oxidized form of the protein gives the best binding energy. And so our calculations, these and more calculations, we did show that electronic effects strongly modulate ligand binding strength. So at this point, you could say, okay, these are only calculations. Can you observe this also experimentally? And the answer is yes. And so for this, I'll show you this slide. This is a four-final imidazole or PIM. It's a classical IEO1 inhibitor, which has been co-crystallized in the enzyme in 2006. So we know that this is binding to PIM and how it is binding to IEO. And here in red, I marked the nitrogen atom that actually lines to the iron in the heme core. And over the years, we and many other people have tried basically every possible biohistorical replacement of this ring here. So all these compounds are very similar. They do have exactly the same shape. They all do have this nitrogen in the right position to bind to the iron. But when you go to experimental testing, you see that they're actually almost all of them are completely inactive. And the only active compounds are actually the imidazole and then the one-final imidazole. So yes, also experimentally, you could show that these subtle electronic effects are very important in a lot of good IEO1. All these calculations that I showed you, so the classical docking together with the quantum mechanical calculations, they allowed us to develop a good QSAR model. So a quantitative structure, a specific relationship that explains the observed activities of the different compounds. We only need two terms, one based on the classical docking, one based on the quantum model, two to obtain a very good. And when you look at this graph, you see that there's one very good component here in the lower left corner. And this component we called MMG0358, it is the best component we've developed so far. It shows a good shape complementary to the active site. It forms a hydrogen bond with a serine residue and it has this chloride ion in the position that it showed before. And when we tested this in the enzymatic assay, we got good nanomolar activities. It's also very active in cellular assays for most IEO. And interestingly, it also shows a very high selectivity for IEO over TEO, which is a related enzyme, and it doesn't show any seriatricity. So these were very nice results that we obtained. And in the following our collaborators also tested this in a legal model of the mice. And here, I just want to remark that in this case, the single agent therapy is the only thing that they have in the mice. And the room for the first slide, the principle idea, when you're talking about nation therapy with other strategies, but already in the single agent therapy, they could observe a significant amount of the mice as compared to the control room. And so, all right, this was the first part. In the second part, I will try to show you a bit of different strategy to come up with initial scaffolds over the lead component. Because I mean, during all this work, we noticed that really difficult step is to find the initial active scaffold that then you can try to optimize to increase different properties of the compound. So we had the idea to kind of put a step zero to our pipeline to initiate the whole thing by actually doing it through the screening. And this was made possible by the presence of the Biomelecular Screening Facility which at the time had basically two compound collections. One is the Prestric Chemical Library, which is composed of 1,200 FDA-approved drives. So screening those is interesting on the one hand side in the perspective of drug repurposing. So it's quite fashionable at the moment to do this, to test known drugs for other indications. But it's also very interesting because about all of these compounds has a lot of data. So bioavailability, safety, but also a lot of other data is there. And on the other hand side, we used the Maybridge Hitfinder Collection, which is a discovery-oriented collection of 40,000 small molecules. So it's not a huge high profit screening, but it could be done in a reasonable amount of the instrumentation that is attended at PFL. And these compounds are supposed to have drug-like properties and a high diversity. And this is just an example from one of the 384-welled ways that we screened during this campaign. And you see everything looks really nice. We have our negative control on the left-hand side, the first 32 wells. These are the library compounds. And then on the right-hand side, you have our positive control, that is a molecule that we developed previously. And you see that we have a very nice signal-to-noise ratio and good hits there. And so on the y-axis here, you have the high profit screening score that basically goes from zero, which is the average for the negative control to one, which is the average of the positive control. And when we used this approach, we got a nice hit rate, both for the Christopher chemical library and for the Maybridge Hitfinder Collection. Everything is very nice. And the problems actually started when we went to the experiment to follow up for these hits. So that's an example of one of the molecules from the Maybridge Hitfinder library that we obtained. And things look very good when we first did standard assay. Also, again, the lab. So we get a nice dose response curve. And things look good. And we became suspicious of this type of compounds a bit later. And we decided to actually add some nucleophile to our assay solution to see what happens. And what you observed was really a very strong shift of the acidity. The compound basically became inactive because now instead of currently covalently modifying idea one in order to inhibit it, it was reacting with the nucleophile. So it was a strong indication that this is actually a discursive reactive compound in a specific way. And this is a second example. So the molecule looks very similar. It just basically changed the position of two halogen substitutions on it. So you might think, okay, this might do the same thing, but no. It actually does something else. So here you see already a suspiciously steep dose response curve. And actually it's known that these steep curves often are a product of aggregation. And in order to test this, we added some detergent to our solution. As it has been suggested by Shashi and co-workers a few years back. And again, we still have strong shift in IC50 and it's probably just an aggregator. So going through the list of hits that we obtained from the discovery-oriented library, we noted that mainly almost all the compounds that we that were detected as hits were promiscuous compounds, either because of chemical reactivity, redox cycling, iron, chelation, aggregation, interference, we essay read out because there were dyes. So this came as kind of a shock and it also took a bit of time to discover all this. Then we discovered that actually we're of course not the first one to have these type of problems. And actually over the last years it's filters published to detect exactly this type of problems. And the most prominent one is the Spain's essay, the Pan Essay Interference Compound filter, sorry, filter, not essay. Which is used very widely and I think mainly because it's very nice to see. And actually I really personally prefer these Lillimet Chem Rules which are a Lillin company and which are also smart filters to detect all kinds of problematic compounds. And then there are others and I will not mention all of them, but this one is another nice one for IDO because actually it has, so I think some of the problems you see are actually that compounds will react with cysteine residues that are in the IDO site in that way inhibit IDO. So it's nice to detect these kind of problems. And in order to assess these compounds and also compounds that are in the literature supposedly as IDO-1 inhibitors we made this radar, radar scheme for compound classification. So on the upper part you have the three filters that I mentioned before and we give a score of one if the compound passes the filter, score of zero if it fails the filter. And on the lower half you have experimental data for the compounds. So enzymatic ligand efficiency and cellular inter-efficiency which is basically the binding layer which gives you a measure of how good an efficient. And when we look at this data for three of our top hits from the Maybridge Hitfinder collection we see so from the screening we get a very good score so all these are basically as good or better than our control compound. When we look at these classifications we see at least the two on the left are really bad so they fail all of the three filters. And here this is mainly because of the genome scaffold that is hidden inside and also this actually has a bit more hidden but it's also a genome-like structure. With this compound we lost quite a bit of time because we thought it wasn't so bad and we actually see it fails only one out of three filters and it looks quite efficient and in the end we concluded that also this is actually a promiscuous inhibitor and it's a micro, it does micro to some parts of the protein. So these are bad examples and this answers the question of why promiscuous inhibitors, A, they exist but why do we have to release so many among our, among the hits that we observed and a partial answer to this is probably that the enzymatic assay solution that we need in order to have an active compound is quite, to have an active enzyme is quite complex so IDO is a redox enzyme so we need reducing agents in the solution to have active IDO. We need catalysis for H2O2 the optimal solution that can inactivate IDO and then of course we look at the enzymatic reaction so we need oxygen tryptophane in solution we will form this product during the enzymatic reaction so there are really many things in solution but we add a small molecule to it we can observe the two reactions that we hope to observe but there are many other side reactions that are in place and this is why we're actually very happy about the new developments a new machine that was recently bought by the Olympic center for cancer research at the Unil it's a Biapour SPR device which is so sensitive that it allows you to detect the binding of a small molecule to a macro molecule to a protein and in this way during these type of experiments we can greatly reduce the complexity of the assay solution because we don't need to observe any more of the enzymatic reaction so we don't need any of these compounds in the solution we just need to reduce IDO during the assay and another nice thing is that this really allows you to measure a quantity of the Kd which is directly related to the binding free energy which directly relates to quantities that we can calculate with our computational approaches so this really is a nice device for us to use and we're looking forward to doing those experiments and just to finish a bit on a later note so we did not observe only promiscuous inhibition in our head throughput screening we also found specific compounds so these antifungal imidazoles they all 15 of them in the press rig library they were all detected to be IDO inhibitors and we believe that those are really specific IDO inhibitors which have a good score on the screening and they have no other word detected and with this I've just come to a very last slide about IDO it's a bit to put our work into relation with what's currently available so this compound is in phase 2 clinical trials now it's the most advanced compound developed by Insight a compound similar to this one it's exactly this one but something similar to this one clinical trials developed by Nulink genetics here you see our compound which really has a very good profile in the properties that we looked at then you have a classical IDO inhibitor it's a tryptopane analog TU which has also very good properties but it's not very efficient and finally recently these type of compounds have also been crystallized with IDO-1 and they are very interesting so much about IDO-1 for the moment I will go to the second part of the talk which should be a bit shorter it's about some methodological developments that we are doing that are mainly being done by a press hub in the group and these so we're trying to show you before that we use methods to try to do drug design and Fressat is trying to really use or put it in a regular frame or to use quantum calculations coupled to classical calculations directly on the fly docking and what I mean by this I will explain in the next few slides so here again some very basic notions about how we do a ligand protein docking so we need a 3D structure of our target it can be a straight structure it can also be a large model then we have different methods to design the ligands we can use databases we can use a fragment based approach or anything you can think of then we dock this compound into the target binding site we try to find the optimal position within the receptor do geometry optimization to obtain the energy the estimated energy of this from this taking also salvation into account we can estimate the KD of this complex and with this information we go back we design new ligands for the cycle many many times until we find a molecule that has good properties and the challenges in this type of approach to calculate KDs is that you cannot treat chemical reactions so whenever a bond is formed or broken you have a problem because you have to define in the beginning where are your bonds so here we have an example of a EGFR boundary which is actually making a chemical bond a covalent bond and another problem that is difficult to treat in this approach is polarization so when you have a highly polarizable ligand it's charge distribution will change in response to the environment that it's in and this you don't actually discuss and yeah so probably not all familiar with what is the classical approach or how does it work so in this approach for each molecule be it a small molecule be it protein, DNA whatever you want you define the atoms of your molecule so each atom is assigned a radius a mass partial charge and 3D coordinates and then you define the topology and so you say this atom is found to this atom and the bond length is one point whatever angstroms and basically you define bonds as small springs and then you do the same for angles and so on and so forth and when you do a geometry optimization of this compound as an output you get new coordinates which correspond to the minimal the most optimal structure of this protein and you get an energy which leads you which is associated with these new coordinates on the other hand you can do the same using quantum approaches which is a more physics based approach and here you actually don't know molecules anymore all you know is on the one hand are atomic nuclei so each nucleus has a mass it has an integer positive charge which is given by the number of protons that are within a certain chemical element and it has again a 3D coordinate and then the second ingredient of the system are the electrons which are negatively charged which hold the system together and which also a quantum approach and again you can do a geometry optimization and again as an output you get new coordinates and new energy but in this case you also get additional information so you can get a new topology and a new charge as you do so you get additional information on this approach and of course it also has a drawback so that's a computer time you can do the more high level viewer quantum approximation and it's very unpractical to treat a full biological system with a quantum method and that's why people have a particle circular approach so the bulk of your biological system with a classical method what I've shown before and then you just choose a subsystem basically the active side of the enzyme wherever chemistry is happening you choose that part of the system and treat it like that by the QM method and of course the trick of the trade or the real science is how you treat the interface between those two systems and I don't know if you might remember but this is exactly the type of approach that 2 years ago the Nobel Prize was given for so Martin calculus, Michael Leavitt and Ariel Warshall they were attributed the Nobel Prize for developing this type QMM approach alright I see the time alright so we have a few results from Prasat it's really just a few slides we applied this QMM or scoring function approach first to zinc metal proteins because they are very diverse they are only present in all major enzyme classes and this zinc atom in the active side actually induces a strong polarization which is a problem for many classical doping algorithms and it's also very suitable because there are many high quality structures available in Phoebe's so for benchmarking set it's very convenient and last but not least many of these zinc metal proteins play a role in cancer development so they are interesting as a target too before showing you the results I would just very briefly visually show you what we call a success in docking and what we call a failure so our reference is always an x-ray structure here on the left side we have a small molecule binding to a zinc in an active side of course if our docking pose it's here this purple this purple pose in the middle if the overlap with the original with the x-ray structure is good so the RMSD is below 2 angstrom then it's become a success and if this measure the RMSD is above 2 angstrom so here you see an example for this where the ligand is completely flipped around and not in the good position then we call this a docking failure and when we look at this observable for a test set of 230 zinc metal proteins treated either with a classical scoring function or a cumulative scoring function so that we get a significant improvement using a QMM approach which is due to the better description of our coordination and charge distribution that is induced by the zinc atom in the active side and here that's just one example from this benchmark set again you have the x-ray structure in orange so this is the correct solution this is the solution that we get with our QMM docking approach so you see a nearly perfect overlap between the codes generated by the docking and the x-ray structure it's a very small RMSD and these two codes are classical docking codes that are freely available and you see that AutoDock gets the ligand confirmation completely wrong so the two rings are exchanged the whole ligand is flipped and the RMSD is very big and this AutoDock vina also gives you a good result so the RMSD it's a success but actually if you look at the details here on the zinc binding functionality you see that actually the ligand binds from the compound this is of course very important so just one last example for the QMM docking where it works very nicely is also this TNF alpha converting enzyme case which is what we call the Adam 17 and it is known for the system it's known that binding of the ligand influences the acid dissociation constant of an active cyclic amide and just by chance we have 15 complexes of this enzyme in our large benchmark set and when we looked at the result for this sub for these 15 complexes we were very surprised to see that actually there was a big gap between classical and the results classical didn't get any of the causes correctly when QMM was perfect perfect success and this prompted us to look closer into what's happening here so here you have an animation hydrocemic acid ligand binding to this case and you see the active site here so now we have a protein around you see the active site the zinc ion and the hydrocemic acid ligand binding to the active site and what's actually happening it was very quick in a second is that the proton jumps from the ligand to this active site and then this it shows a very good binding so once again there's the proton sitting on the ligand and then it jumps to the coboxylate of the group acid so these are exactly the type of topological changes that you can see in the QMM approach classical approach and for this the end I would just like to give a little outlook on what else you can do for approaches and it's more toy system than it is in relation but of course when you treat an enzyme you'd also like to know what's exactly happening obviously exactly taking a place and this can also give you of course what is important or what should you say in your particular place so we looked a bit in more detail into this dioxygenation reaction of tryptophan to end your knee so here you have tryptophan bound in the ideal one active site it makes nice hydrogen bonding interactions with this arginine so this kind of force tryptophan in the good position and also in this interaction here with this coboxylate bound to the hemi you have the dioxygen which first attacks this carbonate so you saw one oxygen is transferred to this carbon atom which in the following was from an epoxide and later on moved to this carbon atom so this gives way to the second oxygen atom to attack again the same carbon atom and then the system relaxes and this is basically almost in formula of panel release or the reaction product that just needs to detach from the hemen to create a reaction product with this system so with this I would like to come to an end to acknowledge the work of all collaborators for these projects so as I mentioned for the QMM development it's the work of Prasad Chaska the chemistry is done by Somi Somi Redi Majigabu in under supervision at EPFL and everything that pertains to modeling the stunning collaboration and I would like to thank for quite a few computer time the BSF at EPFL for screening support and for both support health and the and the National Science Foundation who produced this project