STRC h09 Phase 4p Boltz WH2 Orthogonal Controls

Phase 4p PASS-supportive: Boltz-2 WH2 control panel keeps h09 in the positive-reference band and below no negative control.

Problem

The reviewer pack needed one orthogonal structure-model sanity check for the Phase 4m claim. Phase 4m used AlphaFold Server WH2-family calibration; Phase 4p tests whether a separate local model family, Boltz-2, preserves basic WH2 rank separation when matched positives, h09, and sequence-destroying negatives are run under one protocol.

Method

  • Builder: hypotheses/h09-hydrogel/scripts/phase4p_boltz_wh2_control_builder.py
  • Parser: hypotheses/h09-hydrogel/scripts/phase4p_boltz_wh2_control_parser.py
  • Input YAMLs: models/boltz_jobs_2026-04-27_h09_wh2_controls/
  • Boltz output: models/boltz_runs/h09_phase4p_wh2_controls/
  • Artifact: hypotheses/h09-hydrogel/artifacts/phase4p_boltz_wh2_controls_2026-04-27.json
  • Command: ~/.venvs/boltz/bin/boltz predict ... --use_msa_server --model boltz2 --accelerator gpu --devices 1 --sampling_steps 50 --diffusion_samples 1 --output_format mmcif
  • Device: M5 Max MPS; 9/9 examples completed, zero failures. One SVD operator fell back to CPU inside PyTorch/MPS.

Results

ConstructRoleBoltz ipTMConfidenceInterpretation
MIM WH2 x G-actinpositive0.9250.928high positive
Tbeta4 x G-actinweak reference0.9140.926high in Boltz despite weak ITC Kd
WASP WH2 x G-actinpositive0.8730.918positive
h09 WH2 x G-actinquery0.7060.861inside positive-reference band
WAVE2 WH2 x G-actinpositive0.6990.881positive but not top-ranked
scrambled h09 WH2 x G-actinnegative0.6330.867below h09, but close
alanine h09 WH2 x G-actinnegative0.5880.855below h09
WIP WH2 x G-actinpositive0.3400.781false-low positive
poly-GS spacer x G-actinnegative0.3190.774low negative

Verdict

  • Outcome: PASS-supportive.
  • Supportive part: h09 remains in the positive-reference band and all three negative controls score lower than h09 by ipTM.
  • Caution: Boltz does not reproduce the known Chereau ITC rank order. WIP is false-low; Tbeta4 is false-high; scrambled h09 is only modestly below h09. Treat this as an orthogonal sanity check, not affinity evidence.
  • Reviewer use: The panel answers “does an independent model immediately invert h09 below negatives?” with no. It does not strengthen the absolute Kd claim beyond Phase 4m.

Ranking delta

No ranking change. h09 remains S with mech 5, deliv 4, misha_fit 3. Phase 4p is reviewer-pack support only; gate-closing evidence remains STRC h09 Phase 4m Kd Mono Soft Floor Result 2026-04-27, and STRC h09 Phase 4k Fibril Actin Geometry 2026-04-27 remains the geometry caution.

Connections