i wouldn't be surprised if free energy estimations are much less accurate for random sequences and de novo designed proteins than for natural proteins, as pLMs are learning coevolutionary statistics as discussed here. curious what you think! www.pnas.org/doi/10.1073/...
Yes, I think I agree. There are two problems: 1) the approach they use isn’t very accurate even for natural proteins and 2) any model will behave very differently for sequences close to the training set than for those far away, so their result is expected just from how the model is trained