Hi Christian, I've seen this behaviour as well. I'm not an expert but I rationalized the situation as the TLS parameters for an incomplete model getting overfit as the missing/wrong parts of the model pull on the TLS parameters. REFMAC doesn't alternate between TLS and coordinate refinement, so you can get stuck in a local minimum that moving atoms and refining individual B-factors won't get you out of. As more of the model is built correctly, the drift of this refinement gets less and less.
I've found that refining comparatively few TLS cycles (I usually do 3) before the restrained refinement tends to avoid this problem most of the time - don't let the TLS refinement converge. Sometimes I'll let the coordinate step converge, then run another TLS refinement to bootstrap things along, but usually the stats only marginally improve at that point. In cases where I've made substantial rebuilds, it's sometimes been necessary to turn off TLS and refine without it like you do, then re-introduce the same TLS groups, setting the B-factors to fixed values. Not really ideal, but it gets there. Hopefully others have some more theoretically-sound advice. A good thing to look at doing, especially if TLS is behaving weirdly, is to check with the PARVATI server (http://skuld.bmsc.washington.edu/parvati/). If your TLS group boundaries set off flags, you have some more work to do on the refinement. Shane Caldwell McGill University On Thu, Jun 18, 2015 at 4:30 PM, Christian Roth < christian.r...@bbz.uni-leipzig.de> wrote: > Hi all, > > I refine a structure with TLS and restrained refinement in Refmac. After > the run finished, I fixed a few outliers manually in Coot, saved that file > and used that with the original TLS file from TLSMD for the next round of > refinement. As soon as the restrained refinement starts the R-factors are > going up and stabilise than at a higher values compared to the previous run. > If I switch of the TLS refinement and just do restrained refinement with > the same input files the R-values start initially at a higher values and > than decrease till they stabilise. Unfortunately the final values are about > 1.5 % higher, compared to the run with TLS before manual rebuilding. So TLS > seems to me beneficial and I would like to use it, but somehow a strange > combination of wrong files?, wrong parametrization? or whatever prevent me > of doing that. > Has anyone an idea what might be the cause of that strange behaviour. > > Thank you very much in advance. > > Cheers > > Christian >
