 So that was the first step of docking, generating lots of conformations and at that point generate anything You don't have to know whether it's good or bad. Just generate it The second phase though is scoring. Now I need to say given all those Conformations which ones are good and which ones are bad There is more than one way to skin this cat, too There are three common ones. We could do something that is force field based Remember the force field in the molecular dynamics simulation? This is pretty much exactly the same thing. We try to calculate all the interactions between atoms This might be accurate We have to do it a bit more sloppily here because I don't have the water So I have to represent the water Indirectly would say with the surface I'm exposing to water and whether my atoms are hydrophobic or hydrophilic This can be a best costly, but it's frequently pretty accurate So at least for the atoms closest between the receptor and the ligand we normally do that We can add empirical terms though for instance hydrogen bonds Do you see the difference here in the molecular dynamics simulations? We relied on detailed balance sampling according to the Boltzmann distribution Then it was very picky that we had a proper potential and calculated forces and generated new Confirmations according to the Boltzmann distribution here. I couldn't kill us If we know that it's it's usually good to have hydrogen bonds So if you have many hydrogen bonds awesome Let's just count the number of hydrogen bonds and if the molecule confirmation I'm testing here has lots of hydrogen bonds Add that to column 2 if column 2 has a high value. You should keep that molecule the third option is what we called Knowledge-based or statistical scoring You actually know this this is pretty much Boltzmann inversion So if we look at the protein data bank and all the proteins and we know that if I have a ligand bound It's very common that the distance between say the aromatic ring and an oxygen is three angstroms If we have enough statistics like that, you can calculate the probability for that, right? And if we have the probability we can turn that into potential Delta G equals minus RT ln the quotient of the probabilities So knowledge-based scoring really just means that look at statistics of known structures known complexes And when I see patterns that fit that I should score them high We could decide to use kilo calories per mole for our scores, but really I can use anything I want I could introduce the Lindahl score that is a number that should be high All we care about here is that I can I find a way to separate things that are likely to be good binders from things that are likely to be bad binders Do you notice one thing here? I'm only predicting binding. That's really the only thing we can do with talking So here I need to rely and hope for that high affinity is going to be the same as high Efficacy, which is usually the case, but it's not the guarantee