STRC Research

STRC AF3 Interface Triage Protocol

4 min read

STRC AF3 Interface Triage Protocol

Use this before any AF3, AF-M, Boltz, docking, peptide-interface, antibody, or binder-screening run that could affect a STRC hypothesis. The goal is to stop one-off confidence scores from becoming evidence without controls.

This protocol is adapted from Sang 2026 AF3-TurboAb and the STRC method notes. It is a starting checklist, not a universal validation claim: antibody-antigen thresholds do not automatically transfer to STRC/TMEM145, WH2-actin, homodimers, or small molecules.

When to use

  • New AF3/AF-M/Boltz protein-protein or peptide-interface screen.
  • Any screen using ipTM, pTM, PAE, pLDDT, DockQ, contact counts, or interface RMSD as evidence.
  • Any candidate-selection step that might move a hypothesis tier or next step.
  • Any library-scale design run where preprocessing, MSA reuse, or target-feature caching matters.

Protocol

  1. Define the target class. State whether the run is antibody-antigen, protein-protein, peptide-protein, homodimer, glycoprotein, or protein-ligand. If it is not antibody-antigen, do not inherit Sang 2026’s ipTM >= 0.8 threshold as a proof threshold.

  2. Lock the biological input. Record exact sequence, construct boundaries, isoform, mutation, glycosylation state, oligomeric state, cofactors/ligands, and target species. If a number or biological constant matters, resolve it via sources/lit/ before writing scripts.

  3. Create controls before running candidates. Include plausible positives when available, known negatives, decoy targets, scrambled/unrelated binders, and cross-pair permutations. A screen without decoys can rank candidates but cannot make a specificity claim.

  4. Check leakage and overlap. For benchmark-like claims, remove or label pretraining-overlapping structures/sequences. For antibody or peptide libraries, check sequence similarity to known PDB binders where relevant.

  5. Set seed policy up front. Use at least 3 independent seeds by default. High and low confidence cases may be stable; intermediate scores need deeper sampling before interpretation.

  6. Record the reusable feature plan. If many candidates share one target, precompute or reuse target-side MSA/template/features where the tool permits it. Log tool version, run date, hardware, seeds, and feature source.

  7. Score more than one metric. Capture ipTM, pTM, interface PAE, local pLDDT, contact counts, interface residue lists, and any reference-based metric such as DockQ/RMSD when ground truth exists.

  8. Calibrate locally. Interpret confidence by separation from the decoy panel, not by an imported literature threshold. A threshold from antibody-antigen work is a hypothesis for STRC, not evidence.

  9. Cluster contacts, not just sequences. Group models by epitope/contact footprint, pose, and hotspot residues. Look for pose convergence and contact-patch stability across seeds.

  10. Flag induced-fit risk. If a target domain is flexible, glycosylated, disordered, or known to move, treat epitope identity as more reliable than full-pose RMSD. Require MD or experimental follow-up for conformational claims.

  11. Separate enrichment from affinity. ipTM can enrich true binders in validated regimes. It is not a Kd predictor unless a local Kd calibration exists for that target class.

  12. Close with STRC bookkeeping. If the run becomes a phase proof, add ## Ranking delta, prepend the hypothesis log, update hub frontmatter if rank/next step changes, and update STRC Computational Scripts Inventory for new or modified scripts.

Minimum evidence language

Use these phrases to keep claims honest:

  • Allowed: “AF3 confidence enriches candidates relative to decoys in this target-specific screen.”
  • Allowed: “The model identifies a recurring contact footprint across seeds.”
  • Allowed: “The result is a structural triage signal requiring orthogonal validation.”
  • Not allowed: “ipTM implies nanomolar binding.”
  • Not allowed: “An antibody-antigen ipTM >= 0.8 cutoff proves a STRC peptide/protein interaction.”
  • Not allowed: “A single AF3 model establishes mechanism.”

Source anchors

Connections

Graph View

Backlinks