Can we find protein conformational ensembles in the data we already have?
Summer RosettaCon 2025
Figure 1: Feynman-Kac steering for ensemble sampling
Instead of single states, we resample promising ensembles at intermediate steps according to density fit.
Figure 2: Steering appears struggles to infer conformational changes for proteins with conformations not represented in the training set
Perhaps because the conformational change is too far off the data manifold, the steered model fails to recover the conformations present in the density used to steer sampling. RMSD shown is calculated as the average between the ensemble and the target conformations.