Methods & Claim Governance
Methods v0.1 · computational-only; wet-lab validation not yet contracted · GitHub
Claim Governance — the 3-LLM Gate
No claim goes external without: meta-analysis across ≥2 independent datasets with I² measurement, orthogonal structural check (Chai-1 cross-validation if Boltz-2 iPTM ≥ 0.5), 3-LLM consensus gate (Claude + Gemini + GPT-4o), and visible claim-status badge. When a claim fails re-derivation it is RETRACTED in-place with a banner; original content is preserved for transparency.
See the Corrections Log for every retraction.
Translational Score (v0.1)
A weighted composite of eight dimensions:
- Disease relevance — causally linked to SMA MN loss, or downstream?
- Reproducibility — direction consistent across ≥2 datasets, I² < 75%
- Tractability — druggable pocket, antibody epitope, ASO target
- Selectivity feasibility — paralog similarity check
- CNS / tissue delivery — BBB, intrathecal, muscle uptake
- Safety / ADMET — hERG, DILI, CYP, AMES
- Combo compatibility — pairs with nusinersen / risdiplam / onasemnogene
- Validation depth — published, modifier-proven, patient-derived, or none
Limits: judgment-weighted aggregate, not a p-value. Per-finding FDR across tracked targets is NOT yet applied. v0.2 will add permutation-null FDR and explicit weightings published as JSON.
Phase 4 — Mutation Validation Campaign (Queued)
A computational-only platform must prove it can recapitulate known disease-causing mutations, or its predictions cannot be trusted. This campaign runs AF3 and Boltz-2 on clinically-characterized NMJ disease mutations to test whether the pipeline’s ΔiPTM matches published clinical phenotypes.
Gate: Pipeline must reproduce the direction of the clinical effect on at least 3/4 mutations before any novel-binder claim is published. Results will be posted as a dedicated news article when the campaign completes.
Reproducibility
- Pinned container images (Boltz-2 PyTorch 2.11+cu128 sm_120, AF3 server, NIM saturator)
- Meta-analyses: DerSimonian-Laird random-effects with reported I², pooled log2FC, 95% CI
- All structural predictions report iPTM + pTM + per-chain pLDDT
- Raw compute outputs rsync’d to Dropbox Tier-1 with SHA-256 checksums (≥ 90 days)
Wet-Lab Status
None of the computational claims on this platform have yet been experimentally validated.Every campaign page carries an “Experimental validation required” banner. Wet-lab collaboration is actively sought but has not yet been contracted.
Established Biology vs. Our Hypotheses
Established (textbook): SMN1 deficiency causes SMA. SMN2 copy number modifies severity. Nusinersen, risdiplam, and onasemnogene abeparvovec are approved. Motor neuron degeneration, NMJ dysfunction, and muscle atrophy are the core pathology. None of these constitutes a cure.
Internally supported: p38-MAPK14 is the current lead axis under evidence review. Other active axes include glycolysis, Kv2.1, HDAC2, NRF2-KEAP1, and NMJ signaling through MuSK, AGRN, and LRP4. All claims remain subject to active re-derivation.
Our hypotheses (not validated): p38-MAPK14 modulation as an adjunct therapeutic layer. NMJ structural rescue via AGRN/LRP4/MuSK. TDP-43 cross-connections to the SMN complex. Metabolic and redox rescue as combination strategies.
License: AGPL-3.0. Repository: github.com/Bryzant-Labs/sma-research. API: API Guide.