I'm pretty skeptical of the methods underlying those simulations given that most are fitting their parameters to either fully folded structures or molecules much smaller than any protein. Protein folding itself is difficult to experimentally measure, and you usually don't end up with a lot of data to base your fitted model on.

But then again, having never run an MD simulation before, I'm probably not the right guy to ask on that front...

Personally, having done completely computational analysis of other peoples data sets during undergrad, I'm much happier putting myself in a place where I'm on the data-generating side of things. In the case of protein folding / binding, there's still plenty of space for experimentalists to apply high-throughput techniques to screen new / larger libraries of proteins for various properties.

Going for the gold of de novo prediction of structures is just so hard to get right when you're basing it on, say, the structures in the PDB, where quality control is often low and the conditions between structures varies so much.