Step-by-step methodology to reproduce these results
No genetics degree. No lab access. No budget. Just a laptop, a phone, and an AI agent (OpenClaw + Claude Opus 4.6) that can actually do things: download files, search databases, parse data, build sites. My job was asking questions. Good ones. Here's exactly what I asked and what came back.
Michael's WES report from Hong Kong Children's Hospital (Lab No: 23C7500174, December 2022) listed two STRC variants. One was labeled "Pathogenic" (a whole gene deletion from his father, confirmed by MLPA). The other was labeled "Variant of Uncertain Significance" (a single letter change from his mother, confirmed by Sanger sequencing): NM_153700.2:c.4976A>C p.(Glu1659Ala). I needed to know: is this second variant actually harmful?
I asked Claude to look up the STRC protein. It searched UniProt and found the ID: Q7RTU9. Claude then pointed me to AlphaFold, which has the predicted 3D structure. The confidence score (pLDDT) at position 1659 was 95.69 out of 100, meaning the structure prediction at this spot is very reliable.
AlphaMissense is a tool by Google DeepMind that predicts whether a protein mutation is harmful. Claude downloaded the AlphaMissense predictions file for stereocilin and searched for "E1659A" (E = Glutamic acid, the original amino acid; A = Alanine, Michael's variant).
The result: 0.9016 out of 1.0 (Likely Pathogenic). Anything above 0.564 is considered likely harmful. I then checked all 19 other possible changes at position 1659. Every single one scored above 0.846. This means position 1659 is structurally critical: any change there breaks the protein.
| protein_variant | am_pathogenicity | am_class |
| E1659A | 0.9016 | LPath |
| E1659D | 0.9483 | LPath |
| E1659G | 0.9191 | LPath |
| ... all 19 substitutions: LPath (0.846-0.999) | ||
If a position is important for the protein, it should be the same amino acid across different species. Claude pulled stereocilin sequences from 9 mammals on UniProt (human, mouse, rat, cow, monkey, pig, dog, bat, bear) and searched for the motif around position 1659 in each.
Result: 100% conserved. All 9 species have Glutamic acid (E) at this position. The surrounding 13-residue motif (PEIFTEIGTIAAG) is identical across ~80 million years of evolution. This is PP1 Supporting evidence under ACMG criteria.
Normally, geneticists use SIFT, PolyPhen-2, and CADD to check variants. Claude tried all three through the Ensembl VEP API. They all returned nothing for this variant.
The reason: STRC has a nearly identical "twin" gene next to it on chromosome 15 (a pseudogene called STRCP1) that confuses sequence-alignment-based tools. This is why AlphaMissense is uniquely important for STRC: it works from the protein's 3D structure, not from the DNA sequence, so the pseudogene doesn't affect it.
ACMG/AMP guidelines (Richards et al., 2015) are the standard framework geneticists use to classify variants. Each piece of evidence gets a code and strength level. I learned the rules and applied them:
2 Moderate + 2 Supporting = Likely Pathogenic. Per ACMG combining rules (Table 5), this meets the threshold for Likely Pathogenic classification.
I compiled all evidence into a formal letter addressed to the Chemical Pathology Laboratory at Hong Kong Children's Hospital, requesting a reclassification review of the variant from VUS to Likely Pathogenic. I attached the AlphaMissense data, conservation analysis, and ACMG criteria breakdown. I also built this website so the evidence is transparent, reproducible, and accessible to anyone reviewing the case.
If the hospital accepts the reclassification, Michael's molecular diagnosis will be confirmed: biallelic pathogenic STRC (DFNB16). This is a prerequisite for future gene therapy clinical trials. Dual-AAV gene therapy has already restored hearing in STRC-deficient mice (Iranfar et al., January 2026). Human trials are expected within 2-3 years. Michael will be 7-8 years old.
Reclassification is the immediate goal. But once you start asking questions, you can't stop. Can we make the gene smaller? Fix just one letter? What if we test it computationally before anyone spends a dollar on a lab? These aren't genius insights. They're obvious questions. The difference is having an AI agent that can actually go look for the answers.
Instead of replacing the whole gene, what if we could fix just the one wrong letter? Claude downloaded the genomic sequence around Michael's variant from Ensembl and checked whether gene editing tools could target it.
Base editing (CBE/ABE): cannot fix this variant (C>A transversion is outside their range). Prime editing: feasible. Claude found a suitable PAM site just 4 base pairs from the mutation. A prime editor could theoretically correct the single base change, though this approach has not yet been tested in inner ear cells.
Current gene therapy for STRC requires two viruses (dual-AAV) because the gene is too long for one. Two viruses means lower efficiency: both must enter the same cell. Claude analyzed the AlphaFold structure and identified that the first ~600 amino acids have very low structural confidence (pLDDT below 50), suggesting they may not form a stable structure and might be dispensable.
If those regions are removed, the remaining "mini-stereocilin" (1328 aa, 3984 bp) fits in a single AAV vector. This is a computational hypothesis. It needs lab testing. But the precedent exists: micro-dystrophin (removing non-essential parts of dystrophin) is now in Phase 3 clinical trials for muscular dystrophy.
To test the mini-STRC idea further, We submitted a job to AlphaFold 3 Server to predict the 3D structure of stereocilin bound to its interaction partner TMEM145 (a protein recently discovered to be essential for stereocilin's function, Nature Communications 2025).
First results received (Job 1). ipTM = 0.47, pTM = 0.48. Low confidence in direct binding. PAE matrix analysis shows best cross-chain contacts at N-terminal residues 174-185 (but still poor at 8.6 A).
I then submitted 5 more jobs to systematically test the mini-STRC hypothesis:
| # | Experiment | Status | Tests |
|---|---|---|---|
| 1 | Full STRC + TMEM145 | Done (ipTM 0.47) | Baseline interaction |
| 2 | Mini-STRC + TMEM145 | Done (ipTM 0.43) | N-terminal dispensable (0.43 vs 0.47 baseline) |
| 3 | STRC E1659A mutant (solo) | Done (pTM 0.64) | No structural damage. Fold intact. E1659A affects function (charge), not structure |
| 4 | STRC wildtype (solo) | Done (pTM 0.63) | Baseline: 16% disordered (N-term drags it down) |
| 5 | Mini-STRC solo | Done (pTM 0.81) | YES! Mini-STRC folds excellently (7% disordered) |
| 6 | N-terminal solo (1-615) | Done (pTM 0.27) | CONFIRMED: 38% disordered, pTM 0.27 |
| 7 | mini-STRC + Piezo2 CED | Submitted | Does mini-STRC interact with mechanosensitive channel? |
| 8 | NFATC1 + Calcineurin A/B | Done (ipTM 0.73) | VALIDATED: CnA-CnB ipTM 0.91, NFAT-CnA ipTM 0.72. Cascade confirmed |
I emailed the leading researchers working on STRC gene therapy at institutions in the US, France, and China. I shared the reclassification evidence, the mini-STRC hypothesis, and a link to this website.
I received encouraging responses confirming the computational approach is sound and that the analysis has been shared with research teams working on STRC gene therapy.
On day three, we asked a question that changed the direction of the research: instead of a constitutive promoter that's always on, what if we used a promoter that responds to sound? Hair cells already convert sound to calcium signals. That's a built-in sensor we can hijack.
Hair cells have mechanotransduction (MET) channels that open when sound deflects their stereocilia. Ca²⁺ flows in, activating the calcineurin-NFAT signaling pathway. This pathway is well-characterized in the sonogenetics literature (Wu et al., Nature Communications 2023), where researchers used it to achieve 62-fold gene induction with zero background leakage.
We designed a construct: 6xNFAT promoter + mini-STRC. The 6xNFAT promoter is a synthetic element with 6 copies of the NFAT response element, creating a cooperative digital switch that requires sustained calcium signaling to activate. Combined with mini-STRC (from Day 2), the total construct is 4,401 bp, fitting within the 4,700 bp AAV packaging limit with 299 bp to spare.
To go beyond speculation, we built an ODE (ordinary differential equation) model of the complete signaling cascade: sound level → MET channel open probability → Ca²⁺ influx → calcineurin activation → NFAT nuclear translocation → transcription → translation → protein accumulation. Every parameter comes from peer-reviewed literature.
Result: with a realistic hearing aid schedule (16h on, 8h off), the model predicts therapeutic stereocilin levels (>15,000 molecules per OHC) in just 13 hours. In silence, the system produces only 6.8% of the target (promoter effectively OFF). The full Python code is available on GitHub for anyone to reproduce or extend.
To test the mechanosensitive hypothesis structurally, we submitted two new AlphaFold 3 jobs. Job 7: mini-STRC + Piezo2 CED (does stereocilin physically contact the mechanosensitive channel?). Job 8: NFATC1 + Calcineurin A/B (positive control, validates the cascade model).
Results pending. If Job 7 shows high ipTM (>0.6), it would mean stereocilin directly interacts with Piezo2, implying a feedback loop: broken STRC affects mechanosensation itself, not just structural connections.
Gene therapy currently requires surgical injection into the cochlea. What if ultrasound could push the treatment through the round window membrane non-invasively? I found papers showing that microbubbles + ultrasound create temporary pores in the RWM (Shih et al. 2019: 5.2x permeability increase, full recovery in 24 hours, zero hearing damage).
I built an ODE model of the full delivery chain: ultrasound → pore formation → nanoparticle diffusion → cell uptake → protein production. Honest result: baseline parameters fall short (0.39% per session, 258 sessions needed). Optimized LNPs with ionizable lipids: 78% per session, 2 sessions total. The bottleneck is endosomal escape, not the ultrasound. The most realistic path: one-time AAV surgery, plus non-invasive LNP top-ups years later if expression fades.
OpenClaw is open source (free). Claude Opus 4.6 is available via Anthropic API (pay per use). All scientific databases are free and open access. AlphaFold 3 Server requires a Google account.