 The next example is from a colleague of mine in New York, John Shadera, who's using a concept called Markov State Models. I won't have time to take you through the gory details of them, but on a conceptual level, the idea is that if in a simulation, you can cluster conformations into, say, an open state and a closed state of a protein, then we are able to identify more conformations like that and gradually get a better and better definition of these clusters. At that point, you can also start to say, if I am in the open state, what is the probability of going over to the closed states and vice versa and calculate the rates between these two states? If you continue that and adding more data, you can have a more and more fine grade definition of your states. We can capture intermediate conformations, maybe a third states that opening in another way or maybe a states that binding something. John has been doing this, in this case, for a protein called metal transferase. Exactly what metal transferase does is not super important here, but this is a case of metal transferate without the ligand bound, and here we have metal transferate with a small ligand bound. And I'm gonna show you what happens in these simulations. What you have in the very middle of the pane here is the schematic Markov state models that shows what conceptual states it's visiting, and then you show the very average trajectory of the protein on the top and bottom. Here we go. So do you see that once you're in a state, you frequently tend to stay in that state for a while, and then you might be jumping over to another part of this state space, and then the top one jumped to the right again, and it will continue like that. The other part is that these conceptual states are quite different from the top and the bottom one, simply because the protein behaves in different ways. Binding this small ligand changes the complete dynamics of the molecule and the states in which it's gonna be stable. Now, we can capture an average states from an x-ray experiment perhaps, but there is no way any other experimental method can obtain as much information about these states and the motions here as it can from the simulation. John has been using folding at home for this project too, and in this case, they've been able to collect something like two milliseconds of data for the protein, which is amazing. With the caveat though, that is not one simulation, but many simulations, and that's why we need these sampling concepts such as Markov state models to stitch all the data together and being able to model things on longer time scales. Pretty cool.