STRC Research

h26 Gate Pilot — PeptAI Discipline Import 2026-04-25

4 min read

h26 Gate Pilot — PeptAI Discipline Import 2026-04-25

First STRC adoption of the per-stage gate discipline pattern observed in PeptAI’s 9-gate pipeline and discussed in the broader Bankless / DeSci framing. The kernel imported is declare each phase’s pass/fail threshold up front, then let the calculation decide, not the looser “use more LLMs” reading of agentic AI. See the upstream Brain note for the conceptual basis: obsidian://open?vault=Brain&file=notes/Nine-Gate%20Discipline%20for%20Computational%20Drug%20Discovery.

Why h26 first

Of all live hypotheses, h26 has the cleanest fit:

  • The objective is structural (engineered disulfide stabilises a weak homodimer interface), so AF3 + Boltz + MD map directly onto PeptAI’s G1–G5.
  • Phase 1 has already failed once (R-R repulsion at ARM 1579-1581), Phase 1c has produced a re-clustered candidate set, and Phase 1d is teed up. There is enough computational scaffolding to populate gate states with real verdicts, not aspirations.
  • It is downstream of h03 (drop-in to Ultra-Mini AAV payload), so G9 has a concrete “compatibility” definition — not a vague “ship it” placeholder.

h09 (peptide hydrogel) is a partial fit but self-assembly geometry doesn’t map onto receptor-binding gates cleanly; deferred. h01 (small-molecule pharmacochaperone) needs a different gate set entirely (FEP ΔΔG_fold, not interface ΔΔG_dimer); deferred.

What was imported

Six gates carry over with adapted thresholds:

  • G1 structure quality — AF3 ipTM ≥ 0.50 AND pTM ≥ 0.60. Tracks Phase 1d’s existing criterion verbatim.
  • G2 interface preservation — homodimer contacts ≥ WT count, closest Cb-Cb < 8 Å. Phase 1 had this implicitly; making it numeric.
  • G3 disulfide geometry — Cb-Cb 4.5–7.5 Å, χ_ss within native cystine range. Phase 1c PASS already on analytical grounds.
  • G4 interface energy — ΔΔG_dimer ≤ −2 kcal/mol vs WT monomer (PRODIGY or MM-GBSA). Conservative engineered-stabiliser bar.
  • G5 MD stability — 100 ns explicit-solvent MD, RMSD < 4 Å, disulfide intact ≥ 90% of frames. Sized to M5 Max wallclock.
  • G9 drop-in compatibility — combined-construct AF3 against h03 Ultra-Mini × TMEM145 retains the GOLD-pocket binding contacts. Real downstream check, not ceremony.

What was dropped

Three gates from PeptAI do NOT carry over:

  • G0 ChEMBL agonist baseline — h26 has no agonist set. Equivalent ground truth would be “WT homodimer interface CIF” + “known-destabilising mutations” (Phase 1’s failed quartet). That’s already encoded in G2 (≥ WT count) and the negative control (A1078W) in Phase 1d. Calling it G0 separately adds bookkeeping without information.
  • G6–G8 ADMET stack (PROSPERousPlus / OpenSol / ToxinPred3) — calibrated for blood-borne peptide drugs. h26 is an intracellular protein-domain edit translated from an AAV payload; the relevant pharmacokinetics are AAV biodistribution + cochlear delivery, both of which inherit from h03’s existing PK/PD model. The peptide-drug ADMET tools would produce noise, not signal.

Recording the drops explicitly so future agents don’t reintroduce them by analogy.

Operational impact

  • Hub frontmatter (hypotheses/h26-engineered-homodimer/index.md) now carries a gates: block with the 6 adapted gates + history strings.
  • Hub body has a ## Gates section reproducing the table.
  • No tier / mech / deliv / misha_fit / next_step change — h26 stays B-tier with next_step: Phase 1d AF3. The pilot is a discipline overlay, not a research finding. Hence no Ranking delta propagation, no row in STRC Computational Scripts Inventory (no script written or modified).
  • Rollout decision: keep the gate block on h26 only until Phase 1d closes. If the gates produce useful decision pressure (i.e., the threshold-up-front discipline catches a borderline call that would otherwise be eyeballed), expand to h09 next, then h01 with its own gate set.

Connections

Graph View

Backlinks