epitope-specific binder design

[lizhanwei]

Hi,

Thank you for releasing ESM and the binder design workflow.

I have a question regarding epitope-specific binder design.

Suppose I have a target protein and I would like to design binders against a predefined epitope (for example, residues 125–140 on the target surface), rather than allowing the model to bind anywhere on the protein.

From the documentation and examples, it seems that binder design is performed against the full target structure, and I could not find a way to explicitly specify a desired binding epitope.

Is there a recommended approach for epitope-constrained binder design using ESM?

For example:

• Can a specific residue range or surface patch be provided as a constraint during design?
• Is cropping the target structure around the epitope the recommended strategy?
• Are there plans to support explicit epitope conditioning or interface constraints in future releases?

Any guidance or best practices would be greatly appreciated.

Thank you!

[robert-verkuil]

Hi @lizhanwei, apologies for the delayed reply. Here are answers to your questions:

Can a specific residue range or surface patch be provided as a constraint during design?
Are there plans to support explicit epitope conditioning or interface constraints in future releases?

Not currently in the code we've shared. We've internally experimented with losses similar to contact loss (but with additional masking / settings tweaks) to achieve this in silico, and hope to share more details and a code option in the coming few weeks, here!

To roughly sketch the approach, we've tried simply masking contact_loss to only be calculated between target epitope residues and the full binder, on these lines. And to either replace inter_contact loss with this new loss, or add the new loss to the overall loss mixture, in either case with coefficient 0.5, following the other contact losses.

Edit: Expect to hear more from us and see a code update in the coming few weeks.

Is cropping the target structure around the epitope the recommended strategy?

Generally we crop to a small and globular subregion of the target which ESMFold2 assigns high confidence, as a rule of thumb. Using cropping as a tool to encourage a specific epitope may work, but we haven't tried encouragement via this mechanism.

[robert-verkuil]

@lizhanwei and other viewers - a quick correction: The loss we've tried internally is actually a bit different from just the masked inter-contact loss described in my post above. More details to come.