Skip to content

Interfaces with strong hallucination-stage i_pTM consistently fail to survive MPNN resequencing on a hard, non-aromatic epitope: expected behavior or tuning issue? #375

Description

@NicoloGennari

Hello,
thank you for developing and sharing BindCraft. I have encountered some issues in my campaign and I would like to ask for your advice.

Environment: Google Colab
Target protein model: AF3 prediction (PTM: 0.56)

Parameters:

  • default_4stage_multimer_hardtarget.json (under default AF2 prediction protocol the most confident trajectories were the ones where the target itself deformed most, hardtarget solved this problem)

  • default_filters.json

  • template_protocol: Default (we would like to use the binder to stabilize the target in a closed conformation).

Target description: protein of 236 residues formed by two domains (roughly 120 + 100 aa) joined by a short flexible linker, cross-validated against XL-MS crosslinks. AF3 alone gives inconsistent inter-domain orientation across seeds, the target model was selected based on conformation and hotspots (see below). The protein contains no Trp, no Met, all 5 Tyr are buried and the vast majority of exposed residues are charged or polar.

Goal: a bidentate binder that bridges both domains across their ~24 Å interdomain gap.

Hotspots: selected via SASA + Cα distance-to-opposite-domain + pLDDT filtering, split across both domains. It is difficult to select hotspots with the suggested target residues.

Problem observed: every hallucination trajectory that reports high pLDDT/i_pTM/i_pAE at the trajectory stage fails after MPNN resequencing. failure_csv.csv shows i_pTM/i_pAE as the overwhelming dominant rejection category (e.g., 180/180 and 344/354 in two successive batches), with pLDDT/pTM playing a much smaller role, and 0 MPNN designs ever reaching Rosetta-stage scoring. This holds even for trajectories with i_pTM up to 0.85 at the hallucination stage.

Questions: before committing to a full campaign of 1000 to 10000 trajectoriesIs, we ask for an opinion on these considerations:

  • Is the pattern "confident hallucination but MPNN-fragile interface" a known/expected failure mode for low-signal (non-aromatic, charged-dominated) epitopes, or does it suggest something is misconfigured on our end?
  • Would raising weights_con_inter relative to weights_iptm/weights_pae_inter be a sensible way to push toward more MPNN-robust interfaces?
  • Would switching from MPNNsol to MPNN help in any way, since the target exposes negatively charged residues?

I thank you in advance.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions