Four engines. One pipeline.
Each module is independent (run alone or stack), but designed to share data: fitness scores feed phase-3 prediction; BBB scores filter the fitness output for CNS programs; predictive-levels frames where each module sits in the evidence hierarchy.
AAV Capsid Fitness Predictor
CapsidFitnessCNN (256K params) trained on 2,014 Bryant variants. Ranks engineered and natural capsids by tissue-specific fitness.
Contrastive AUC 0.793 · accuracy 0.761 · ESM-2 640-dim 02BBB Transcytosis Model
Receptor-biology-grounded brain penetration scoring. Replaces ADME-style heuristics with first-principles transcytosis: LRP1, transferrin, GLUT1, integrins.
CNS 0.641 vs non-CNS 0.244 · PHP.eB 0.710 03Predictive Levels L1–L8
Eight-tier framework from physics (L1: pocket detection, surface curvature) through evolution (L4: variant fitness) to outcomes (L8: clinical durability).
4 core + 3 data modules · 80/80 tests passing 04Phase 3 Outcome Predictor
Eight feature categories (efficacy signal, expression durability, safety, dose selection, trial design, capsid biology, immune response, regulatory).
AUC 0.82+ · gene therapy specific15 capsids modeled.
Natural and engineered. BBB transcytosis score, dominant receptor, tissue preference, and seroprevalence (corrected per Boutin 2010 where the literature was wrong).
| Capsid | Class | BBB score | Primary receptor | Tissue tropism | Seroprev. |
|---|---|---|---|---|---|
| AAV9 | natural | 0.647 | Galactose / AAVR2 | broad (CNS · heart · muscle) | ~47% |
| PHP.eB | engineered (AAV9-derived) | 0.710 | LY6A | CNS (murine) | — |
| AAV-Spark100 | engineered (AAV8-lineage) | 0.412 | LamR | liver (Beqvez) | ~38% |
| AAVrh74 | natural (rhesus) | 0.483 | HSPG · LamR | muscle · CNS | ~18% |
| Anc80L65 | ancestral reconstructed | 0.534 | HSPG | liver · retina | low |
| AAV5 | natural | 0.218 | PDGFR · sialic acid | airway · CNS | 3.2% (Boutin 2010) |
| 4D-R100 | engineered (745aa) | 0.427 | retinal-targeted | retina | — |
| 4D-C102 | engineered (739aa) | 0.292 | cardiac-targeted | cardiomyocytes | — |
| … 7 more (AAV1, 2, 3B, 6, 7, 8, DJ — see full inventory under briefing) | |||||
Real data, not synthetic.
The platform replaces mock features at every level with peer-reviewed published data — and where the literature was wrong, we corrected it (AAV5 seroprevalence 30–40% → 3.2%; Beqvez capsid mis-assigned as AAVrh74 → corrected to AAV-Spark100/AAV8 lineage).
30 perturbation experiments
From 8 published papers. Used to calibrate Level 2 scaffolding (receptor → tissue → fitness coupling).
7 Kd records
4 SPR + 3 BLI measurements for capsid-receptor binding affinity. Used to anchor the binding-affinity module.
5 PDB structures
With binding residue maps. Used for surface curvature, pocket detection, and contact-map generation.
U.S. Provisional 63/986,270.
Filed 2026-02-19. Covers the predictive levels framework, the BBB transcytosis model, the surfaceome calibration approach, and the contrastive-learning pipeline for capsid fitness.
Modeling a capsid program?
Send us a candidate list (sequences, target tissue, or program name). We come back within 5 business days with a ranked report against the 15-capsid baseline.
Contact Us →