# How to model antibody-antigen complexes with partial experimental data?
**URL:** https://ask.bioexcel.eu/t/how-to-model-antibody-antigen-complexes-with-partial-experimental-data/6164
**Category:** HADDOCK
**Created:** [June 5, 2026, 11:16am UTC](https://ask.bioexcel.eu/t/how-to-model-antibody-antigen-complexes-with-partial-experimental-data/6164 "2026-06-05T11:16:25Z")
**Posts on this page:** 4
**Page:** 1
### Author:  [@mberksoz](https://ask.bioexcel.eu/u/mberksoz)
#### Post date: [June 5, 2026, 11:16am UTC](https://ask.bioexcel.eu/t/how-to-model-antibody-antigen-complexes-with-partial-experimental-data/6164/1 "2026-06-05T11:16:25Z")
Dear Haddock team,
I have been using a local version of Haddock3 to dock antibody clones that we have partial epitope for. Clone X is shown to share an overlapping epitope with a reference antibody on SPR epitope binning experiments, i.e they block each others binding to the same antigen. Clone Y is shown not to block the reference antibody’s binding, so they potentially share non-overlapping epitopes. Reference antibody has a crystal structure in complex with the antigen, so we know its exact epitope. Inputs for candidate clones are AF3-Rank1 models.
I have done docking runs for the two candidate clones by defining all the CDRs as active and epitope we know from the reference antibody as passive restraints. Since I don’t know the exact epitope for these clones but I know it must be binding somewhere near, I defined all residues on the antigen within 4A of the antibody in the crystal structure as passive restraints.
My aim was to see whether haddock can differentiate the true binding clone vs the non-binding clone, i.e give a more favorable score for the clone that we know binds somewhere near that patch. I’m sharing the results for both clones which I find hard to interpret; it seems the non-binding clone gives similar scores to the binding one. Obviously the most favorable result belongs to reference antibody when I used the true epitope/paratope information.
What kind of docking protocol do you recommend in these kind of scenarios where we have partial epitope information? Or how can i interpret these results ? Please let me know if I need to clarify the situation further. Thanks a lot !
---
### Author:  [@VGPReys](https://ask.bioexcel.eu/u/VGPReys)
#### Post date: [June 5, 2026, 12:35pm UTC](https://ask.bioexcel.eu/t/how-to-model-antibody-antigen-complexes-with-partial-experimental-data/6164/2 "2026-06-05T12:35:20Z")
Dear Melike,
Thanks for your interest in using HADDOCK3 for your research.
The HADDOCK scoring function is used as a ranking method within the same docking run, and usually cannot be used to compare docking runs.
It is also NOT a solution for binding affinity prediction.
Because restraints energy is related to the size of your system, they can have an impact on the resulting HADDOCK score when comparing two docking runs.
For binding affinity prediction, you could try to use the `[prodigyprotein]` module in HADDOCK3.
> [@mberksoz](#):
>
> Since I don’t know the exact epitope for these clones but I know it must be binding somewhere near, I defined all residues on the antigen within 4A of the antibody in the crystal structure as passive restraints.
In my opinion, 4A is a little short and you could increase it a bit.
> [@mberksoz](#):
>
> What kind of docking protocol do you recommend in these kind of scenarios where we have partial epitope information?
The one you are setting up is good.
But I would pay a particular attention at the CDR conformations, which can have a huge impact on the docking runs, please see [https://academic.oup.com/bioinformaticsadvances/article/5/1/vbaf161/8185414](https://academic.oup.com/bioinformaticsadvances/article/5/1/vbaf161/8185414) and [https://research-portal.uu.nl/ws/files/242909600/btae583.pdf](https://research-portal.uu.nl/ws/files/242909600/btae583.pdf).
You can get inspired by the protocols that are published there.
Best regards,
Victor
---
### Author:  [@VGPReys](https://ask.bioexcel.eu/u/VGPReys)
#### Post date: [June 5, 2026, 12:45pm UTC](https://ask.bioexcel.eu/t/how-to-model-antibody-antigen-complexes-with-partial-experimental-data/6164/3 "2026-06-05T12:45:47Z")
Additionally, you could use scoring function that are not using restraints, such as the newly published DeepRank-AB ([DeepRank-Ab: a scoring function for antibody-antigen complexes based on geometric deep learning | Communications Biology](https://www.nature.com/articles/s42003-026-10408-4)), that could predict the DockQ of your complex.
Because with HADDOCK you enforce the binding site, DeepRank-AB could maybe discriminate between binder and non-binder.
---
### Author:  [@mberksoz](https://ask.bioexcel.eu/u/mberksoz)
#### Post date: [June 5, 2026, 2:04pm UTC](https://ask.bioexcel.eu/t/how-to-model-antibody-antigen-complexes-with-partial-experimental-data/6164/4 "2026-06-05T14:04:43Z")
Thanks a lot Victor! I’ll try the DeepRank-Ab.