STRC Research

STRC h09 Parameter Provenance Audit 2026-04-23

8 min read

STRC h09 Parameter Provenance Audit 2026-04-23

Systematic verification of every numerical constant in models/hydrogel_phase*.py against primary literature. Triggered by Egor’s question “have you actually verified these numbers?” — answer turned out to be: partly, and the gaps are load-bearing.

Method

4 parallel Sonnet agents (Domain 1 Cochlear PK/PD · Domain 2 Actin/WH2 · Domain 3 Biomechanics · Domain 4 RADA16/SAP), each tasked with: find target papers → download (open or Anna’s Archive) → MinerU parse (with VLM backend fallback for math-heavy PDFs) → extract numbers with page refs → write paper notes. Results consolidated into ~/STRC/literature-params/ (5 topic files + hub).

Papers added: 20 new paper notes in papers/ (total now 49).

Temp reports: /tmp/strc-lit-params-domain{1,2,3,4}-*.md (preserved until user approves deletion).

Findings — three severity levels

⛔ Critical (hypothesis-level)

1. 118 aa tail on RADA16 exceeds characterized parameter space by ~10×. Empirical limit for appended sequences preserving RADA16 β-sheet assembly is ~12 residues (2021-frontiers-rada16-clinical-review). Our tail91 construct appends 118 aa. No published study has ever tested a >50 aa globular domain on RADA16 — behavior unknown and almost certainly disrupts self-assembly at standard 1:1 composition. The PEPTIDE_TAIL91 at 100% composition probably does not form fibrils. Only viable path: 9:1 molar blend (plain RADA16 : RADA16-tail91) — established in literature for hybrid scaffolds. Requires explicit Phase 4e re-design. See rada16-geometry.

2. WH2 × F-actin side-binding Kd = 5 μM has NO primary literature backing. Zero papers measure WH2 × F-actin side-binding affinity. Closest analog Tβ4 × F-actin = 5–10 mM (1,000–2,000× weaker than model’s 5 μM claim). WH2 binding site is structurally disfavored in the filament (buried by longitudinal actin-actin contacts, per 2026-04-23-chereau-wh2-actin-pnas + 2026-04-23-dominguez-2016-wh2-nucleation-review). The model’s 5 μM is speculative optimism. h09 does not fail trivially — avidity from multi-WH2 on RADA16 scaffold could compensate — but the single-molecule premise is unsupported. Phase 2c wet-lab actin-bundling assay is the only resolution. See actin-kinetics.

3. STRC × TMEM145 Kd has NO experimental measurement in literature. Neither 2026-04-17-derstroff-tmem145-ohc-stereocilia (Neuron 2026) nor the parallel NatComms 2025 TMEM145 paper (Reimann lab) reports SPR, BLI, or ITC. Only CoIP + nanoSPD (qualitative). Verpy 2011 has no binding data. The 10 nM claimed in Phase 4h competition math is fabricated. The 100 nM claim in Phase 4e is an AF3 ipTM → Kd heuristic, not a measurement. Every dose-occupancy calculation in Phase 4e/4h inherits this. See strc-tmem145-interactions.

🔴 Wrong values (parameter-level)

ConstantClaimedLiterature-backedError
K_RWM (Phase 4e)0.02/h cites “Salt 2011”Citation phantom. Scaled TMPA permeability (1.9e-8 cm/s, 2001-salt-ma-quantification-rwm-permeability) → 14 kDa: 0.002–0.004/h5–10× too high
K_PROTEOLYSIS (Phase 4e)1.4/h (t½ 30 min)Assembled RADA16 in vivo t½: 7–14 days → k ≈ 0.002–0.004/h (2021-frontiers-rada16-clinical-review)350–700× too fast, wrong regime — free monomer vs assembled gel conflated
PERILYMPH_VOL_UL (Phase 4e)70 μLHuman ST alone: 34.2 μL (2021-dhanasingh-scala-tympani-volume); total ST+SV: 93 μL (2016-ekdale-human-cochlear-hydrodynamics)Unsourced interpolation, affects all concentrations ±2×
K_CLEAR_MIDDLE_EAR (Phase 4e)0.35/h (2h)Gentamicin 75 min (k=0.55); dex-phosphate 40 min (k=1.0) (2018-salt-plontke-pharmacokinetic-principles-inner-ear)2–3× too slow for free solution; depot mechanism conflated
WT_BUNDLE_STIFFNESS_PN_PER_NM (Phase 1)1400 pN/nm5.12 pN/nm mouse OHC apical (dulon-2019-htc-bundle-mechanics); 2.5–8.6 pN/nm rat 1–4 kHz (2019-tobin-elife-tiplink-stiffness-gradients)150–560× wrong. Also: dead code (gate3_bundle_stiffness() never references it)
K_HTC_PN_PER_NM (Phase 1)7.5 cites “Tobin 2019”Wrong citation — Tobin 2019 measures K_GS + K_SP, not HTC. Real: 0.16 pN/nm per link (Dulon derived); 20 pN/nm aggregate (Kozlov 2011, bullfrog model)Citation fabricated, value between real bounds but undocumented
STEREOCILIA_INTER_FILAMENT_NM (Phase 4d)12.07.9–9.7 nm native OHC (plastin-1/fimbrin, Krey 2016 FFT-EM); 12 nm is espin-specificWrong crosslinker class

🟡 Defensible but comment wrong

ConstantValueIssueFix
RADA16_MONOMERS_PER_NM_FIBRIL4.35/nmComment cites “Zhang MIT” — Zhang 1993 does not tabulate thisCorrect derivation: Cormier 2012 WAXD 0.47 nm × Paravastu 2014 bilayer → 4.26/nm. Model value 2% off, within uncertainty. Update comment.
fibril_length_nm1000 nmUpper-range, not meanDocument as: mean 615 nm (2005-yokoi-kinoshita-zhang-rada16-reassembly), range 200 nm – few μm
WH2_KD_GACTIN_M100 nMComment cites “Chereau 2005, Husson 2010” for 100 nMReal range 50 nM – 3 μM by construct; 100 nM is WAVE-favorable end. Update comment to range.

New parameter discovered (Krey 2019)

  • Local [F-actin] in stereocilia ≈ 3–5 mM (derived from 200,000 actin/stereocilium, 200 nm × 3 μm volume). Value is 3 orders of magnitude above any plausible WH2 Kd → F-actin concentration is not limiting; affinity is.

Effect on Phase 4 verdict

Prior verdict (STRC Hydrogel Phase 4 Computational Campaign): “Integrated verdict GREEN.”

Revised verdict: Phase 4 computation is valid in structure but hollow in parameterization. All 10 gates need re-evaluation against literature-backed inputs. The most consequential shifts:

  • Phase 4e dose window: likely 5–10× higher doses needed than the 0.13–1.32 mg window (K_RWM correction); offset by scaffold not clearing at K_PROTEOLYSIS = 1.4/h (gel dissolution is 7–14 days, not minutes).
  • Phase 4d bundling: still works via avidity argument, but single-molecule premise is aspirational; wet-lab validation mandatory.
  • Phase 4h competition: math is sound; the Kd baseline (10 nM WT STRC × TMEM145) is fabricated, so Misha displacement numbers are speculative.
  • Phase 1 assembly: 118 aa tail likely disrupts RADA16 β-sheet — requires blend strategy re-design, not yet modeled.

Ranking delta

STRC Synthetic Peptide Hydrogel HTC (#9): A-tier HELD, but with explicit axis recalibration.

AxisBeforeAfterRationale
Mechanism54118 aa tail exceeds empirical RADA16 modification limit by ~10×. Blend strategy is the credible path but not yet computationally or experimentally validated.
Delivery43K_RWM off 5–10×, PERILYMPH_VOL off ±2×, K_PROTEOLYSIS wrong regime. Dose window recalculation pending; feasibility preserved but margins tighter.
Misha-fit33Unchanged — but note that the competition math (Phase 4h) inherits the fabricated STRC × TMEM145 Kd.

min(4,3,3) = 3 → A-tier. No tier change. Not promoted to S-tier until: (a) blend-strategy assembly validated computationally (new Phase 4e_v2 with 9:1 blend, fibril persistence test); (b) wet-lab Phase 2c actin-bundling confirms WH2 × F-actin avidity; (c) real SPR/BLI on STRC × TMEM145 obtained (from Holt/Oliver/Reimann or generated in-house).

Promotion gate to S (was: Phase 4b triple-complex result). Revised: Phase 4b triple-complex PLUS Phase 4e_v2 blend-strategy fibril-persistence model PLUS Phase 2c bundling assay. Triple gate, not single.

Scripts affected (need update before re-run):

  • hydrogel_phase4d_factin_bundling_model.py — update WH2_KD comments, STEREOCILIA_INTER_FILAMENT_NM to 9 nm, add warning on WH2_KD_FACTIN_M being unvalidated
  • hydrogel_phase4e_cochlear_pkpd.py — K_RWM lit-backed range 0.002–0.004/h, PERILYMPH_VOL split into ST-only 34 μL vs total 93 μL, K_PROTEOLYSIS → k_gel_dissolution 0.003/h for assembled, k_free_monomer_proteolysis separate
  • hydrogel_phase1_self_assembly.py — K_HTC comment correction (Dulon/Kozlov not Tobin), WT_BUNDLE_STIFFNESS corrected or removed (dead code)
  • hydrogel_phase4h_endogenous_strc_competition.py — flag STRC_NORMAL_OHC_M and Kd baseline as estimated

What changes in practice

  1. Before any new computational proof in STRC research, a parameter-provenance table is produced first — every constant points to a paper note that points to a parsed PDF. This is now encoded in ~/Brain/AGENTS.md §0c + ~/.claude/CLAUDE.md + memory as feedback_literature_first.md. Triple-redundant.

  2. Phase 4 of h09 gets a v2 pass with:

    • Blend-strategy scaffold (9:1 plain : tail91) — requires modeling
    • Lit-backed PK/PD parameters
    • Explicit flags on unmeasured STRC-TMEM145 Kd with sensitivity scan
    • WH2 × F-actin single-molecule Kd treated as hyperparameter over the Tβ4-to-ABD range (5 mM to 10 μM)

  3. No more standalone “estimates”. Every numerical claim has a paper-note citation or a flag.

Post-audit re-run 2026-04-23 (task #7 closed)

hydrogel_phase4d_factin_bundling_model.py, hydrogel_phase4e_cochlear_pkpd.py, hydrogel_phase4h_endogenous_strc_competition.py updated in place with lit-backed parameters. Scripts run clean.

Phase 4d param swaps:

  • STEREOCILIA_INTER_FILAMENT_NM 12.0 → 9.0 (Krey 2016 J Cell Biol plastin-1/fimbrin; 12 nm is espin-specific — wrong paralog for OHC)
  • WH2_KD_GACTIN_M 200 nM retained (within Chereau 2005 + Dominguez 2016 WAVE-like range)
  • WH2_KD_FACTIN_M 5 μM retained but flagged LOAD-BEARING unmeasured (no primary F-actin side-binding Kd; Tβ4 analog is 5-10 mM, 1000-2000× weaker)
  • RADA16 and F-actin geometry citations clarified (Yokoi 2005, Oda & Namba 2009)

Phase 4e param swaps:

  • PERILYMPH_VOL_UL 70 → 34 (Dhanasingh 2021 ST-accessible, μCT n=30)
  • K_RWM 0.02 → 0.003 (Stokes-Einstein from Salt & Ma 2001 TMPA; prior “Salt 2011” cite was PHANTOM, 6-7× too fast)
  • K_CLEAR_MIDDLE_EAR 0.35 → 0.7 (Salt & Plontke 2018 gentamicin/dex-phos mid)
  • K_PERILYMPH_CLEAR 0.35 → 0.18 (Salt & Hartsock 2015 J ARO FITC-dextran t½ 230 min)
  • K_PROTEOLYSIS 1.4 → 0.05 (t½ 30 min → 14 h; serpin-rich perilymph, still ⚠ unmeasured)
  • KD_TMEM145_M retained 100 nM but flagged as PLACEHOLDER (AF3 ipTM, not Kd)

Phase 4h param flags:

  • WT_STRC_KD_TMEM145_M 10 nM ⚠ PLACEHOLDER (no SPR/BLI measurement)
  • STRC_NORMAL_OHC_M 1 μM ⚠ UNSOURCED (no paper quantifies STRC/OHC; Krey 2015 cite was phantom)
  • TMEM145_TOTAL_OHC_M 5 μM ⚠ UNSOURCED
  • E1659A scenario scan preserved — is parametric output, not point claim.

Re-run outputs (Phase 4e with lit-backed params):

  • Dose for 1 μM perilymph peak: 0.19 mg (was lower with phantom fast proteolysis)
  • Dose for 10 μM (upper therapeutic): 1.95 mg
  • Dose for 100 μM (toxic): 19.5 mg
  • Therapeutic window: 2.0 log units (stable vs pre-audit run)
  • Multi-dose 0.6 mg/day 7-day: 36% time in 1-10 μM window, mean occupancy 0.73 (anchored on placeholder Kd 100 nM)

Verdict after re-run: therapeutic window claim (1-10 μM, 2 log units) survives the lit-backed parameter swap. But: (a) blend-strategy Phase 4e_v2 still not modeled (task #8 pending); (b) WH2×F-actin Kd single-molecule premise still unsupported (wet-lab Phase 2c required); (c) absolute occupancies parametric on unmeasured STRC×TMEM145 Kd.

Connections

Graph View

Backlinks