 The tools we introduced last lecture, in particular docking, have been remarkably successful in aiding modern drag design, not because they're accurate, but because they're fast. However, as computers get faster and faster, this leads to another challenge. Computers can get so fast that we can't really use the computational power for docking, there simply aren't that many molecules we can just test. At the same time, we would like to use the computers to do something more accurately. And that brings us back in part to the molecular simulations. Remember, when we spoke about molecular simulations, we brought up free energy calculation. And in theory, the molecular simulation can help us to determine how much is the binding energy in a particular pose. This is going to be significantly more expensive than docking. But again, hopefully that problem is going to be solved by faster and faster computers. There are a couple of challenges with that. First, we're going to need a protein, right? In most cases, we can hope that we might have a crystal structure, or in particular, we might have a docked pose. So again, this might be a protein where we have docked the molecule in that particular binding site. And now we want to check, how good is the binding site? I accept that the docking pose is just a rough guess, but can I use the computers to get a more accurate answer? And in most cases, we can simply by applying that free energy cycle that you remember. Remember what we had? So we had the protein with the ligand bound. And then we had the protein and the ligand separately. And we had the protein with the ligand turned into a dummy. And finally, we had the protein and the dummy separately. And then we needed to get this part for the binding up here. We calculated these two legs instead. That is turning the ligand into a dummy, either when it's bound in the protein or when it's out of water. Today, this might take a day or so on a reasonably fast workstation. But the point is, a supercomputer can have tens of thousands of processors, meaning that we can run through not a billion compounds, but possibly a few hundred thousand or so. And in particular now with COVID, that is actually done in very large scale. For instance, in folding at home, four docked poses of small compounds bound to the spike protein or the protease, for instance. The challenge here is that while it's better than docking, it's still usually not as good as experimental values. At best case, I would say that you can get to an accuracy of two kilo calories per mole or so. But that's sorry, one kilo calorie per mole or maybe half a kilo calorie two kilo joules per mole if you're really lucky. It turns out that there are two reasons for this. First, there are some fundamental shortcomings in force fields and everything. There might be polarization involved, changes in charts that the code simply can reproduce. And then there are of course also the motion in the protein. The docking certainly doesn't include that. We get a bit of it in the simulation, but it's not perfect. The other challenge here, if we're doing this turning a very large ligand into a dummy, this is going to be still going to be a relatively complicated process, even though it's easier than this one on the top. So it will be noisy. And that in particular means that it's very difficult to get the exact absolute binding energy. If I'm literally taking my ligand and moving it from water into the protein, there is going to be fairly large. However, when I've done this for one ligand, it's much easier to calculate a series. That is, if I replace say one carbon here with a hydrogen or something, that's just a small change from one ligand to the next ligand. And that type of series, they're much easier to calculate. And in this case, we get a fairly nice correlation. The correlation is not always that good. This is another example for an ion channel. Actually, the correlation here is nice, but do you see the difference between the y and x scales? The correlation is beautiful, but there is an offset of several kcal per mole between them. And that is exactly due to that. It's difficult to get the absolute binding free energy right from a simulation, but it's somewhat easier at least to get the relative binding energy of a series right. These are very usable results and that will help us to rank at least a handful of compounds. And as computers get faster and faster, who knows, in a few years, we might be able to screen through a million compounds in a day or so. More slow and the development of, say, GPUs here is certainly helping us.