You are here

Can I compare binding affinity of different ligands using Rosetta?

5 posts / 0 new
Last post
Can I compare binding affinity of different ligands using Rosetta?


I have 50 experimentally tested ligands and I know which one is a good inhibitor of GSK-3 and which one is a bad inhibitor of GSK-3. I did ligand docking for all of them separately using the protein-ligand docking protocol learned from the 2020 virtual Rosetta workshop with 10000 nstruct; I sorted the outputs by interface_delta for each ligand and I picked the best interface_delta score for each ligand and then I sorted the 50 ligands by their best interface_delta score. Should I see any correlation between the experimental results and the Rosetta docking results. In other words, good inhibitors have lower interface_delta scores. I was wondering if it is OK to compare different ligands using Rosetta ligand docking. Also, should I look at interface_delta or something else like total score? 



Post Situation: 
Sun, 2020-12-06 11:30

Hi Athena,

It's been a long time since I've docked ligands with Rosetta. I also did not attend the virtual workshop. But here are a few things to consider:

  • The score function matters. Take a look at this recent manuscript put out by the Meiler lab. I'm not sure which strategy was presented at the virtual workshop but you might re-score your models with the RosettaLigand score function. In direct answer to your question, take a look at Figure 2B of the linked manuscript: RosettaLigand performs well in Spearman rank correlation tests.
  • The pNear metric has been shown to correlate with binding energy. Have you calculated this metric for each ligand in your set?
  • Have you performed any clustering to ensure the decoys chosen for the interface_delta score are well populated? If the conformation is only sparsely populated it may be unlikely to actually form. Try correlating the average-of-top-5 interface delta of the largest clusters to the wet-lab results.
  • In protein:protein interface design the binding density is sometimes utilized to track best models. (Binding density is binding energy / buried surface area.) For small molecules the parameter is ligand efficieny, which is binding energy divided by the number of atoms.
  • On occasion, a model with good binding energy can actually have a poor total score. It can sometimes be useful to cull the lowest 5-10% of models by total score before sorting by binding energy.


Hope that helps,


Mon, 2020-12-07 07:30

Hi David,

Thank you for your detailed answer to my question. That was quite helpful. How can I find out what score function has been presented in the workshop? Honestly, I don't know how I can use different score functions. Where can I find more about all score functions available? Which part of the XML file should I change? I appreciate your help.




Wed, 2021-01-06 19:16

The protocol presented in the 2020 Virtual Workshop should be the currently recommended one, including the proper score function. (Shannon, the first author on the paper linked, was the one who presented the ligand docking tutorial in the Virtual Workshop.)

The scorefunction itself is controlled by a combination of the SCOREFXN section of the docking xml as well as various flags in the options file. In particular, to use the recommended ligand docking scorefunction from the paper above, you'll want to use the -restore_pre_talaris_2013_behavior flag in the options file, along with the ligand weights and ligand_soft_rep weights in the SCOREFXN section. (Again, this should be the values in the provided example.) -- If you did want to change the scorefunction, consult the SupMat in the Smith & Meiler paper for how the various versions differed, and adjust the SCOREFXN section and options files appropriately.

Tue, 2021-01-12 12:40

I might add a reminder to double-check that the OUTPUT section is present in the docking xml as this is the mechanism for having the xml-defined scorefunction reported to the score file.

Tue, 2021-01-12 13:24