TLS Refinement

[kmdalton]

Hi @HatPdotS , I'm working on integrating TorchRef as the primary refinement method in abismal, but I'm hitting a wall where I can't get R-values to square with PHENIX refine. I believe I've narrowed down the likely cause to be TLS refinement. Isotropic ADPs are worse than PHENIX. Anisotropic ADPs overfit badly but produce a much lower Rwork. I need a way to model a bit of anisotropy without overparametrizing the model. TLS would seem a good first step. I wonder if you have a plan to implement TLS and/or the Extensible-Component Hierarchical TLS (ECHT) B-factor model from Nick Pearce (https://www.nature.com/articles/s41467-021-25814-x). If it's not a near-term objective, would you welcome a PR on the topic? I'm very keen to have this feature.

[HatPdotS]

Sure, that sounds great. What kind of distance to Phenix do you observe? There were some very recent changes to the scaling that should now put it on par with Phenix in the reported rfactor. Also, the anisotropic adps had some restraints not working at some point. For the TLS, I was thinking of defining nodes in space that project a TLS field, and for each atom the TLS would then be the distance-weighted combination of the nodes' values. This would have the advantage of both the splitting as well as the TLS being freely refinable but this is just an idea at this point, and a working implementation would be great.

[kmdalton]

I clawed some performance back by tuning the ADP target weights. I'm still assessing, but it's looking like the performance for anisotropic ADPs is on par now. I'll keep digging on my isotropic example, but I think its R-values are only slightly higher than PHENIX (<0.01 difference for R and Rfree). I think this is acceptable and probably within the uncertainty in R anyway.

I like the spirit of your TLS field idea, but I do worry it will make the TLS tensors less interpretable in terms of domain motions. It's likely okay from the perspective of getting the best maps and fitting the data.