You need much higher resolution for anisotropic B-factor refinement. You could try TLS refinement though. However, while that is likely to help you improving the R-values, it probably won't cause major changes in map quality. If it's the anisoptropy that gives you the problems, which is not unlikely, you might get substantial improvements if you subject your data to anisoptropic truncation and scaling (https://services.mbi.ucla.edu/anisoscale/). Another thing is that it may help trying different refinement programs, as they apparently handle bulk solvent differently. I personally like Phenix and Buster. Finally, I've had a case where B-factor sharpening helped in improving map quality for a structure refined in Buster using very anisotropic data (sharpening can for example be done directly in Coot).

@Esben many thanks for the hints! I'll use the server you provided, though.

I have 2 models for MR available. One is the one that I mentioned above (WT and mutant are in the same space group, C21), and the other model has  ~98% identity and is solved in different space group (P212121) but to better resolution 1.7A with almost perfect fit into density...Does it really matter which one I'll use? As far as I know during MR the starting model is inserted into P1 group anyway, so maybe if the solution looks siginifcantly better, it is more reasonable to use this one?

Sure, you can also use the other model! However, the choice of starting model should not make a big difference if the poor map quality is due to the data being severely anisotropic. So if you anyway get a markedly better map with the other model, it may simply be that it's conformationally less different to your target protein than the first model you used. However if structural errors in your original model is what gave you the problems with poor map quality, it'd be weird that your map should worsen further as you refine the model to better R-values. Btw, your Rfree is very high for 2.4 Å data. Of course you are still early in the refinement, but be aware that if you get stuck at a very high Rfree, it could be that you have deeper problems than anisotropy. Twinning should then be considered (you can do an easy twinning check using Phenix Xtriage), and so should the choice of space group.

@Esben

I run again MR after anisotropic truncation and got terrific improvement in the map quality. Thanks a lot! Not to mention that after first refinement, Rwork/free have lower values than after several rounds in the previous attempt, but the overall structure looks much more better! Once again many many thanks :)

Cool! I am happy to hear it!

You can't compare the r-factors before and after truncation since you get rid of a lot of poor reflections. The proof is in the electron density maps - and when you say that those get better then it's very clearly the correct approach.