The AAV gene-therapy field has eight named challenges. Atlas Bio's platform is built to address all of them.
Gene-therapy delivery hinges on the capsid — the protein shell that gets the payload to the right tissue, past the right barriers, with the right manufacturing yield, in the right species. The hard problems in this field are widely known. This page lists them, and what Atlas Bio is solving against. Methodology and weights are NDA-gated; the problems are not.
Eight named problems in AAV capsid intelligence.
Each card states the challenge, the cost when it goes wrong, and the kind of capability the field needs to solve it. The platform’s specific implementations against each challenge are gated under MNDA — the problems are not.
The mouse → NHP → human translation gap
A capsid that works brilliantly in mice often fails completely in non-human primates and in humans. The most-cited example uses a receptor that doesn’t exist outside mouse strains. Every major AAV program has been burned by this once.
NHP cohort cost dominates pre-clinical AAV budgets
A program testing 20 candidate capsids at NHP scale spends $4–$48M on animal cohorts alone. Most candidates fail. The marginal cost of running one more capsid through NHP is what determines how many shots on goal a program gets.
In-sample headlines vs out-of-distribution generalization
It is common to see AAV-prediction models report 80–95% accuracy on held-out splits, then fail under leave-one-serotype-out cross-validation by an order of magnitude. The headline number and the generalization number are often confused in industry decks. Programs commit resources against the wrong number.
Engineered capsids don’t generalize from natural-serotype training
Models trained primarily on the natural serotypes (AAV1–9, rh74, etc.) and on deep mutational scans of one or two parent capsids often fail on engineered variants. The engineered capsids carry insertions and point mutations the training set never saw, and the model’s feature space collapses them onto the wrong cluster.
Receptor identity is often ambiguous or contested
For several clinically-important capsids the receptor itself is debated — one paper says galactose, another says a GPI-anchored protein, a third says co-receptors. Models that assume a single receptor label per capsid encode that ambiguity as false confidence.
Public training data is severely class-imbalanced
The publicly available AAV-binding datasets are dominated by HSPG-binding (AAV2-family) and galactose-binding (AAV9-family) variants. Sialic-acid, laminin-receptor, integrin, transferrin-receptor, and mouse-only LY6A classes are tiny by comparison. Models trained without class-imbalance handling default-predict the majority class on novel capsids.
Mouse-specific receptor dependencies hide as “BBB crossing”
The most widely-used CNS-tropic AAV variants in academic literature rely on a receptor that is mouse-specific and either absent or differently-expressed in primates. Programs that scale from mouse results without flagging this dependency are systematically over-confident about their CNS reach.
Pre-NHP ranking has no clean industry standard
Every gene-therapy sponsor has to decide which capsids to take into NHP studies. The decision compounds through the lifecycle — cohort choices made now determine which clinical programs exist in three years. There is no widely-accepted standard for how to make that decision under audit.
How the AAV stack is structured.
The AAV platform applies Atlas Bio’s broader agentic methodology to gene-therapy delivery. Predictions are tagged L1 / L2 / L3 by evidence level; every output passes through a multi-node reasoning audit before release; cross-domain validation requires a single parameter set to clear multiple test cases. Tools below are the surfaces a partner team would actually use.
Capsid Intelligence
The end-to-end AAV stack: capsid fitness scoring, blood-brain-barrier transcytosis prediction, surfaceome calibration, phase-3 outcome forecasting, and the published-data ingestion layer.
Capsid Fitness Predictor
Ranks engineered and natural capsids by tissue-specific fitness, trained on publicly-available deep mutational scan data. Output is calibrated and audit-traceable per candidate.
BBB Transcytosis Model
Receptor-grounded brain-penetration scoring. Replaces ADME-style heuristics with a model that names contributing factors and surfaces species-specificity warnings when relevant.
Predictive Levels L1–L8
The evidence-graded prediction framework for AAV: physics-grounded pocket detection (L1), evolutionary variant fitness (L2), through to clinical durability (L8). Every prediction carries its level.
Phase 3 Outcome Predictor
Gene-therapy-specific phase-3 success forecast from phase-1/2 public data. Calibrated against a public-data cohort with explicit limitations documented.
4DMT Capsid Integration
End-to-end pipeline test on engineered clinical-stage capsids (R100 retinal, C102 cardiac). Demonstrates the platform on engineered, not just natural, variants — the harder case.
Want the AAV technical pack?
Briefings cover the platform topology, the published-data ingestion layer, the validation pipeline, and how the engine handles each of the eight named challenges above. Useful for capsid engineers, gene-therapy sponsors, and translational reviewers.
Contact Us →