 Good morning everybody, I'm Francesco Colizzi from IRB Barcelona, the Rothschild lab, working in the Rothschild lab. Today we will see in this short talk how all that you have learned so far and what you will learn tomorrow on free energy calculation can be combined to face scientific problems a different kind. Although I'm the one presenting now this work has been mostly done by von Westermayer and Adam Hospital. Right, so there are three keywords in the title, AGFR as chemical transformation and building block, and we'll see what they means. Hopefully, you know already a little bit what they means. AGFR is the Pidermal Growth Factor. It's a receptor. It's a serial receptor and the misregulation of this receptor creates programming in the cell and this problem in the cells may need to cancer mostly. There are drugs that have been developed to tackle this. There are mostly ATP competitive inhibitors that bind to the kinase domain of the Epidermal Growth Factor receptor, but resistance can be easily developed during the development of cancer. So new drugs are always needed to tackle new mutation, basically. The receptor is also over-expressive many times. 60% of patients with metastatic no small, so longer habit are over-expressive and mutation may cause loss of therapeutic efficacy. So that's why it's critical to understand and to be able to predict whether or not the mutation. For instance, a new mutation will create resistance or not. So we have collected a little bit of experimental data that are shown here in the table with binding affinity and there are mutations that are a little bit spread everywhere around the active site, but also far away from the active site and we will see how this can affect or can be problematic sometimes for the simulation that we will discuss here today. For chemical transformation, how do we want to predict the effect of a mutation? We do it by calculating the free energy difference of the transformation from one amino acid into another and we do this with a non-physical, but they call it alchemical transformation. In this alchemical transformation, we have a tonion of one system. For instance, here we have a valine that is coupled with the shifting parameter lambda and when lambda is zero, the system resembles to a valine. When lambda is one, then the system is coupled with the serine. And what we can do is to shift the system from one side to the other and we can do this in several times, in several ways. The approach I will be looking at today here, it's an equilibrium approach, so which means that we are doing this transition, this transformation very fast in the range of 50 picoseconds or 100 picoseconds. I will do this exploiting Gromax thermodynamic integration feature and PMX to generate the topologies and also to analyze the data afterwards. Here in this slide on the left, you have a little plot of what we are doing. The system goes from A to state B coupled to lambda, it goes from zero to one, we pull it basically from one side to the other and then we pull it back. Then there is a lot of statistical mechanics that have been developed, starting from the Jartzinski equation and also the Krupp's situation theorem for the bidirectional approach and basically that help us to discount all the dissipated work because we are out of equilibrium and to recover the free energy difference of the process that we are simulating. Graphically speaking, this equation can be interpreted like the delta G shown in this equation is basically the intersection between the two Gaussian distribution of the work for the forward and reverse transition. You will see a lot of this probably tomorrow in the free energy calculation lessons. All right, so then all we will be doing, we are using several tools, different tools and we want to combine them in an efficient way. What we are doing this by combining building blocks, putting these tools into building blocks and then assembling building blocks in the way we like the most. This is very useful because then we can give these building blocks to a workflow manager that can parallelize for instance on a supercomputer. So building blocks makes everything more easily. You can take the feature from different tools and put them together. We have seen this already, Adam explained it very well. One interesting feature also, building blocks, is that they can combine it with a workflow manager such as PyCom's and then scale it on a run on a supercomputer. For instance, here we are using Marin Ostrom to do this kind of calculation. So you can get a terrific parallelization of this simulation. Also, I would like to add that a chemical transformation because we are doing this each realization independently from the others. This is an intrinsic added value because it is intrinsically parallel. You can run a lot of simulation at the same time and then gather together the results. So the scalability is very high. So all these, once you put everything together, workflows, building blocks, knowledge of what you are doing, then you can apply everything to real-sense problems. Also, let's go back to our real-sense problem, real-case scenario. We want to predict the resistance of a mutation to a drug. So here we have on the left AGFR, the enzyme, the kinase domain. When we add the drug, the drug binds to the TPSY binding side with the nanomolar affinity and this gives inhibition of the enzyme with positive therapeutic effects. But what happens if we have a mutation, passions that have developed a mutation? We have the same drugs and the question that we would like to answer or we are, we will for sure wondering is whether or not the drug is still effective in treating the passion. And this is what we are going to do here. Now in the next step, how do we do that? By calculating the free energy difference with of the ligand or the drug for the enzyme and the drug for the mutated enzyme. How do we do that? We do it by building a thermodynamic cycle. Here the vertical branches of this thermodynamic cycle basically shows us the HAPL wild type structure and here if we add the drug, we can calculate the delta G of binding of the drug and on the parallel side, here on the right, always on the vertical branches of this thermodynamic cycle, we have the mutation of our enzyme, we add the drug and then we can calculate the delta G of adding the drug to the mutated enzyme. But the problem here is that adding a drug can be quite disruptive alchemical transformation. So what we could do and to have every few more under control is to add as little as possible perturbation to our system and this is achieved by exploiting the thermodynamic cycle and using, calculating the delta G, one for instance, to move from the HAPL wild type to the mutated enzyme and then to do the same transition transformation with the OLO enzyme with the drug bound to the active site and calculating delta G4. And what this thermodynamic site tells us is that basically we can get the free energy difference in the delta G of binding between the mutated and the wild type enzyme in this way. All right, some preliminary results. We're still running simulation, honestly, on this and I will not go into the details. So we have assembled 23 clinical mutations. We are testing three drug inhibitors. This is the table of the results and overall what we are observing is that we're correctly cascifying mutation with the 90 percent accuracy, which is a good and 10 percent that we are missing are what we are mostly related to cases where the delta G of binding is small and is within the basically is within the root mean square error of our calculation that we have assessed to be around 1 kcal per mole, which is also what all the people doing this kind of calculation have observed. So the future application what we are looking at is to use the tools that we have in our hands, the workflow that we have assembled by combining together different building blocks to perform a systematic drug resistance profiling of a GF4. So we will mutate all around the active site recipes and we will calculate, we could calculate the delta G of binding and this way we could see whether or not a mutation is affecting the binding of a known drug. And to do this we may want to use a 10-fold affinity loss threshold criterion basically if the delta G of binding is larger than 1.36 kcal per mole, then we may estimate we may estimate the demutation generating resistance. If the delta G is lower than 1.36 kcal per mole, then we may assume that the binding is still effective. There is no 10-fold loss of affinity. And the drug may still be effective in binding the enzyme, treating the pathology. All right, when we do this, this kind of approaches may have a terrific impact on the clinical, even on the clinical treatment. Let's say that we have a new mutation, nobody knows what it's doing. You can ask your computational friend to do the simulation for you. And basically in less than 24 hours, you know whether or not the mutation is likely to generate a resistance or or not. And this may guide clinician basically in defining the therapeutic pathway. Of course, like with all the captures of the case, considering these are prediction and also for the caveats that I will show you that are soon here just at the bottom of this page. So this kind of study may also help in prioritizing candidate compounds for clinical trials and ultimately help in expanding the domain of application of personalized medicine. The caveats are several and we should very well know them. So mutation can generate a conformational change, okay, after conformational equilibrium of your protein. For instance, in kinase there are two states at least, active and inactive states that are in equilibrium and mutation may alter this equilibrium, shifting the population, the protein towards a conformation that we are not taking into account in our calculation. Also because we're doing the simulation so fast out of equilibrium there is likely in a poor sampling of a ion redistribution it may also affect the calculation and the generate artifacts in the free energy calculation that we're doing. And finally we're looking only or we're not looking only what we may attempt and we may be tempted to look only at the free energy of binding of the drug but for instance in this case of AGFR the mutation may also alter and modulate the affinity of ATP which is the natural substrate of this enzyme. There are no mutations such as the cavekeeper one that alters the affinity to ATP and it is this that indirectly affect the or not the binding of drugs so you need to be careful also with this and also here the substrate affinity can be modulated but this is like because the substrate is usually a protein this becomes a little bit too complicated to model and we are not handling that for now so far but these cavities should be like kept in mind because like the number seems to be correct, but there might be exception to this correctness so wrapping up what we have seen in this short talk is that like a bunch of computational tools varietal computational tools, although we have gone into the details of them, but where user were combined together into a new workflow that was then scale a scale it on a supercomputer to have a a super efficient parallelization basically yeah on thousands of core you can run the simulation simultaneously on thousands of core and in a couple of hours get the maybe in a couple of hours get the results for one mutation even less than a couple of hours and and yeah and do really impactful research because you can really give the answer in real time to the question that is it might be very critical for instance for a new mutation that appear in a passion we have seen that the HFF mutation have been correctly classified in about 90 percent of the cases it was a promising this is a good promising results that put us in a position an optimistic position to look at the next step the outlook like systematic drug resistance profiling that may enable real-time clinical and clinical support guiding basically clinicians into the most effective therapeutic path we still need to be aware of the limit of this method and if we handle the limits then we can better understand how much farther we can push our prediction this is the end of this short talk these are the people the many people probably not all of them I'm sure there are other people that are not listed here. That is the main contributor to this project and I would like to thank you for your attention and let's see if we can have a question-answer session live session