Why this case study matters
Natural capsids are the easy case.
Atlas Bio's Capsid Intelligence platform is trained primarily on natural AAV serotypes (AAV1–9, rh74, etc.) and on Bryant 2014's deep mutational scan of AAV2. The harder test is whether the platform generalizes to engineered capsids — variants that don't appear in the training set and whose sequence diverges meaningfully from natural ones.
4D Molecular Therapeutics' clinical-stage engineered capsids (R100 for retinal disease, C102 for cardiac) are public, validated, and clinically advanced. They're a clean held-out test for the platform.
The two capsids
R100 and C102, scored end-to-end.
| Capsid | VP1 length | Target tissue | BBB score | Fitness score | Clinical trials |
|---|---|---|---|---|---|
| 4D-R100 | 745aa | Retina | 0.427 (moderate) | 0.612 | 2 active (4D-150, 4D-110) |
| 4D-C102 | 739aa | Cardiomyocytes | 0.292 (low, expected) | 0.584 | 1 active (Fabry cardiac) |
Validation note: R100's BBB score of 0.427 is moderate — consistent with retinal targeting being a sister tissue to CNS but not requiring full BBB transcytosis. C102's BBB score of 0.292 is appropriately low for a cardiac-targeted capsid (which should not penetrate the brain). The ranking matches the engineering intent.
Pipeline test
Eight engines, end-to-end.
- VP1 sequence ingestion — both capsids' full sequences entered as raw input.
- ESM-2 embedding — 640-dim representations generated.
- Capsid Fitness Predictor — CNN scored both. Out-of-distribution warning surfaced for both (engineered > 90% identity to AAV2 backbone but with novel surface insertions).
- BBB Transcytosis Model — four-receptor pathway scoring + ESM-2 similarity to known CNS-penetrant capsids.
- Predictive Levels L1–L8 — physics, scaffolding, receptor, fitness, tropism modules all ran; immune-context module flagged unknown seroprev (engineered novelty).
- Surfaceome Calibrator — checked receptor expression in retina (R100 target) and cardiac muscle (C102 target). Both matched intended tropism.
- Phase 3 Outcome Predictor — applied to the 3 active trials. Confidence intervals appropriately wide due to clinical-stage immaturity.
- IBC Reasoning Audit — 19-agent review. R04 (Mechanistic) flagged the OOD warning; otherwise clean.
What this proves
Generalization, with appropriate humility.
- The platform generalizes to engineered capsids from a clinical-stage company without retraining.
- The BBB scores correctly distinguish retinal (sister to CNS) from cardiac (distant from CNS).
- The OOD warning surfaces correctly — the platform isn't over-confident about novel sequences.
- The 8-module pipeline runs to completion without manual intervention.
Have an engineered capsid program?
The platform scores natural, engineered, ancestral, and chimeric capsids. Submission accepts up to 50 candidates per briefing.
Contact Us →