STRC Research

STRC h09 Phase 4b AF3 Robustness Gate 2026-04-26

6 min read

STRC h09 Phase 4b AF3 Robustness Gate

Phase 3b delivered single-seed AF3 ipTM 0.57 for tail91 × TMEM145 — a positive but seed-dependent signal. Phase 4b put that signal through five orthogonal controls: 5-seed robustness on each binding axis, ablation of WH2 (replace with polyGS spacer of equal length), and a triple-complex test where peptide must partition between TMEM145 and actin under competition. Submitted 2026-04-26 03:21 UTC, all 11 jobs cleared the queue in ≤45 minutes wall on AlphaFold Server. Both robustness axes pass. Median tail91 × TMEM145 ipTM holds at 0.60 ± 0.037 across seeds {7, 13, 42, 99, 777}; tail91 × actin trimer holds at 0.51 ± 0.028. Phase 3b’s 0.57 was not a lucky seed — it was the lower-tail of a 0.51–0.61 distribution. WH2 ablation tells the same story from the other side: removing WH2 drops neither axis to noise (both stay 0.51), confirming Phase 4a’s reading that the real TMEM145 binder is the C-terminal tail (cluster 4, aa 1669–1680) and the WH2 contact-mass observed in single-partner runs was an opportunistic hydrophobic artifact. Triple complex collapses to 0.32 — partners do cross-bind under competition, so wet-lab will need to control for this directly. h09 promotes from A → tentative-S pending Phase 2c WH2 × F-actin Kd measurement.

Inputs

11 single-seed AF3 jobs (per-seed split because AlphaFold Server’s Upload-JSON dialog rejected the original modelSeeds: [7,13,42,99,777] array — debugged and fixed 2026-04-26 03:21 UTC):

JobSeedsSequencesPurpose
hydrogel_tail91_x_tmem145_5seeds_seed{7,13,42,99,777}5 × 1tail91 (134 aa) + TMEM145 (493 aa)Robustness on the GOLD zone axis
hydrogel_tail91_x_actin_5seeds_seed{7,13,42,99,777}5 × 1tail91 + actin trimer (3 × 375 aa)Robustness on the WH2 axis
hydrogel_tail91_noWH2_x_tmem1451tail91 with WH2 → polyGS18 + TMEM145Ablation: tail-only TMEM145
hydrogel_tail91_noWH2_x_actin1tail91 with WH2 → polyGS18 + actin trimerAblation: collapse on actin?
hydrogel_tail91_x_tmem145_x_actin1tail91 + TMEM145 + actin trimerTriple-complex partitioning test

Quota burn 22 → 9 jobs remaining (13 spent: 11 needed + 2 duplicate-row mistakes during multi-seed JSON debugging).

Result

Multi-seed robustness

AxisSeedsipTM per seedMedianσGate
tail91 × TMEM1457, 13, 42, 99, 7770.60, 0.51, 0.60, 0.61, 0.600.600.037✅ PASS
tail91 × actin trim7, 13, 42, 99, 7770.58, 0.51, 0.51, 0.51, 0.510.510.028✅ PASS

Gate criterion (per Phase 4b builder MANIFEST): median ipTM ≥ 0.50 AND σ < 0.05. Both axes clear the bar. TMEM145 axis is comfortably above (median 0.60, well over 0.50; σ 0.037 well under 0.05). Actin axis is exactly at the median threshold (0.51) and a single high-outlier seed7 at 0.58 dominates the σ — the four other seeds all returned identical 0.51, suggesting an attractor in the AF3 sampling landscape rather than continuous variation. Robust enough for the gate; tighter robustness would benefit from 10 seeds rather than 5 if Phase 5+ needs it.

Ablations (single-seed reference)

JobipTMInterpretation
tail91_noWH2 × TMEM1450.51TMEM145 binding survives WH2 ablation → tail is the real driver (Phase 4a confirmed)
tail91_noWH2 × actin0.51Actin binding survives WH2 ablation → tail also engages actin opportunistically (Phase 4a finding was real, not single-partner artifact); WH2 not the load-bearing actin handle in this construct
tail91 × TMEM145 × actin (triple)0.32Partners cross-bind under competition: triple-complex cohesion is below 0.5 — peptide is being pulled in two directions

Going-into reading

Three orthogonal pieces of news here:

  1. Phase 3b’s 0.57 was conservative. The actual median is 0.60 with σ 0.037. tail91 × TMEM145 is a real signal at AF3 confidence above the typical positive-control threshold (0.50), reproducible across seed shuffles. This is the single most-load-bearing claim for h09 going forward.

  2. WH2 is not the load-bearing actin handle. Phase 4a CIF forensics suggested this (tail91 × actin contacts: 0% WH2, 100% tail). Phase 4b ablation confirms: removing WH2 entirely does not collapse actin binding. Implication: the tail itself has unintended actin affinity, and WH2 in the full construct does not act as a clean partition handle. Two consequences for Phase 2c wet-lab:

    • WH2 × F-actin assay is still the right experiment — measures whether WH2 retains its canonical actin-side-binding affinity in this scaffold. Result feeds Phase 4d bundling math directly.
    • The construct as designed does not cleanly separate “WH2 → actin” and “tail → TMEM145”. The triple complex (0.32) confirms competition is a real concern. tail-shortening (tail84 / tail71) may help; Phase 4 already tested this.

  3. Triple complex (0.32) is the new gate. It says the construct is partitioning imperfectly under competition. Not catastrophic (0.32 is above noise), but enough to make Phase 2c bundling assay non-trivial — must control for off-target tail-actin binding.

Ranking delta

STRC Synthetic Peptide Hydrogel HTC (#9): A → tentative-S promotion triggered.

mech 5 / deliv 4 / misha_fit 3 — scores held; the promotion is tier-only, on the strength of robustness clearance.

Conditions for full S-tier (Phase 2c wet-lab):

  • Gate 1 (Phase 2c WH2 × F-actin Kd, P0): measure Kd ≤ 50 µM for WH2 binding F-actin filament side. Phase 4d bundling math is sensitive to this within a 5–500 nM band. If Kd > 100 µM, the bundling claim collapses → demote to B.
  • Gate 2 (Phase 2c actin-bundling assay, P0-light): confocal GFP-actin in HEK293, 1–10 µM peptide, expect ≥ 100 crosslinks/µm fibril at 1 µM peptide per Phase 4d model.
  • Gate 3 (Phase 4 PKPD ototopical dose): already PASS at 0.3–1 mg/dose with 2-log therapeutic window; no further compute needed.

Risk that could reverse promotion:

  • Phase 2c WH2 × F-actin Kd > 100 µM → demote to B (chemistry constraint, not pocket failure)
  • Triple-complex Phase 4b ipTM 0.32 indicates real cross-binding → Phase 2c bundling needs partition controls; if peptide cannot bundle under physiological [G-actin] competition, also demote
  • tail-only design alternative may need exploration if WH2 ablation is too clean (Phase 5 redesign)

A_promote_pending_2c_kd. tail91_v2 (Cys11→Ser, Cys52→Ser scrubbed for developability) AF3 batch already staged at ~/STRC/models/af3_jobs_2026-04-27_h09_phase4b_tail91_v2/ for tomorrow’s quota refresh.

Next-step queue

  1. Phase 2c WH2 × F-actin Kd (P0, wet-lab). Required for S-tier promotion. ~$3k bench cost, 2-week turnaround.
  2. Phase 2c actin bundling assay (P0-light, wet-lab). Confocal GFP-actin in HEK293. Phase 4d math says expect 100 crosslinks/µm at 1 µM peptide.
  3. tail91_v2 AF3 batch (P1, 2026-04-27). 11 jobs ready: 5 tmem145 seeds + 5 actin seeds + triple-complex on Cys-scrubbed v2. Should match v0 within σ < 0.05 if developability scrub is binding-neutral.
  4. Phase 4b 10-seed expansion (P2, ~10 AFS quota days). Tighten σ on actin axis from 0.028 to ~0.02. Only if reviewers ask.
  5. DEFERRED: Phase 5 (Martini3 CG MD) — not gated until wet-lab clears; computational scaffold sufficient for current claim.

Artifacts

  • ~/STRC/hypotheses/h09-hydrogel/artifacts/phase4b_robustness/aggregate.json — per-seed ipTM, gate-pass booleans, ablation references
  • ~/STRC/models/af3_jobs_2026-04-23f_hydrogel_phase4b/MANIFEST.json — original Phase 4b builder + go/no-go matrix (predicted vs delivered)
  • ~/STRC/hypotheses/h09-hydrogel/scripts/phase4b_af3_robustness_parser.py — standalone ZIP-aware parser (auto-runs once user downloads ZIPs to ~/STRC/models/af3_results/h09_phase4b/)

Connections

Graph View

Backlinks