yunhuige
commented
4 years ago This is my hypothesis based on my memory: Both figures (Fig10,11) are correct in the paper. In this case, we have 25 states clustered using all data from Hongbin's TRAM estimator. However, for each ensemble, we may not have all 25 states sampled due to the temperature/bias. For example, 300K 0.5kJ we only have 7 states sampled. So the prior goes into BICePs calculation (in Fig10) is the population of these 7 states. So are the rest of the models in Fig 10. After determining the best model for each temperature (yellow star in Fig 10), now we need to compare these best models and find the best one among them. To do that, we "fake" potential energies (1000000000.0 kT) of those missing states for each model to make sure they all have 25 states. Then these priors go into the BICePs calculation and get the BICePs score in Fig 11. So as I said, both figures are correct. The difference between the score is due to the different priors used in calculation. This may not be the best way of comparing these models but it is what we did in practice. I can work with Rob to confirm my hypothesis.
If you can do this; then it will serve as our HDX example in the code.