2024-01-11 Protein FF meeting note

@Chapin Cavender

Open

Chemical shift predictors

• (Slide 8-10)

  • JW – Ways to quantify agreement between histograms?

  • CC – It’s normalized by the observed shift in our references - So even though the shiftx results might have a chemical shift difference of 2.5 ppm between coil and helix (compared to 3.0 from expt), we’ll use 2.5 as a denominator to normalize all shiftx results.

• DM – Even AMBER ff14SB isn’t helical enough?

  • CC – With TIP3P it’s more helical. With OPC and TIP3P-FB it’s less helical than the reference.

  • DM – If I were very pessimistic I’d recommend starting with ff14sb parameters.

    • CC – I did a hybrid of that in the original “specific” FF, where I took all the protein-specific torsions, and that performed about the same (no helicity)

  • YM – Details of calculations? Using folded structure from beginning or starting from linear?

    • CC – Start from linear

    • YM – Would this hold a helix if it started from there?

    • CC – I haven’t tried that. I tried it for ff14SB and earlier versions of null and specific, and ff14SB holds it, while null and specific change to beta hairpin.

    • YM – Do you know why?

    • CC – My hypothesis is that we don’t model the basins well enough.

    • MG – We’re finding that we’re getting the QM surface right, but the ramachandran populations wrong. So we get a better fit to the QM, but ff14SB has a greater favorability for alpha vs. beta. We’fe thought about going into the PDB, looking at prevalence of backbone confs, and using that to guide the conformations in our training set. Though that would inherently bias the training

    • CC – It’s possible that the pairwise energy fitting will help with this.

    • MG – I wonder if we could weight the pairs based on their energies to ensure that we prioritize getting low-energy regions right.

    • CC – Right, I think PB has some code to do this.

    •  

    •  

@Chapin Cavender

QM parameter fits

• JW – ELF10 assignment failures - New?

  • CC – Just 20 or so in a dataset of thousands.

  • JW – Please send me a few reproducing cases if you have time!