BBB Transcytosis Model

Why the model exists

ADME doesn't work for capsids.

Brain-penetration heuristics borrowed from small-molecule drug discovery (Lipinski-like rules, logP, PSA) score AAV capsids essentially at zero — they're large, charged, and don't diffuse. Yet some capsids penetrate the BBB (PHP.eB, AAV9 in some species, certain engineered variants). The biology is receptor-mediated transcytosis, not passive diffusion.

Atlas Bio's BBB model maps each capsid against four transcytosis-receptor pathways and scores how well it engages each one. The result is a CNS-fitness number that correlates with published biodistribution rather than ignoring it.

Components

Four pathways. One score.

Component	Weight	Source
Receptor engagement vector (LRP1, transferrin, GLUT1, integrin α5β1)	0.40	SwissProt receptor sequences + published binding data
Published CNS biodistribution similarity	0.25	Curated from rodent + NHP biodist papers (n=18)
Surface charge profile at physiological pH	0.15	VP3 sequence-derived
ESM-2 similarity to known BBB-penetrant capsids	0.20	640-dim embedding cosine distance

The weighted sum is bounded [0, 1] and calibrated so that PHP.eB (the gold-standard murine CNS-penetrant) sits near 0.71.

Output ranking

15 capsids ranked.

Rank	Capsid	BBB score	Class
1	PHP.eB	0.710	engineered (AAV9-derived)
2	AAV9	0.647	natural
3	Anc80L65	0.534	ancestral
4	AAVrh74	0.483	natural rhesus
5	4D-R100	0.427	engineered (retinal, off-target CNS)
6	AAV-Spark100	0.412	engineered (liver-targeted)
7	4D-C102	0.292	engineered (cardiac)
8	AAV5	0.218	natural · OOD flag
… AAV1, 2, 3B, 6, 7, 8, DJ scored under 0.20

Validation: The ordering reproduces published in vivo CNS biodistribution rankings without per-capsid tuning. AAV9 → #1 (was #3 under prior fitness-only model) matches Foust 2009 / Saunders 2018 systemic-delivery biodist.

Literature basis

Grounded in published biodistribution — not capsid engineering.

The BBB score is calibrated against reported in vivo CNS biodistribution from peer-reviewed rodent and NHP studies (n=18 curated papers), not against hypothetical mutation outcomes. Anchor serotypes in the literature include Foust 2009 and Gray 2011 (AAV9 systemic CNS), Saunders 2013 (AAV9 glycan binding), and Chan 2017 / Hordeaux 2018 (PHP variants — with species-translation limits).

Companion literature assets (internal / working synthesis):

CNS vs. peripheral clinical review — approved and late-stage gene therapies, efficacy and safety by delivery compartment (IV systemic vs direct CNS routes).
Preclinical decision matrix — serotype × route × reported cell populations and off-target factors from published observations.
PubMed corpus tags — 400+ articles from a 485-paper AAV library, tagged for CNS, biodistribution, cell type, and receptor themes.

Scope: These materials summarize what authors and labels report. They do not recommend capsid mutations, construct selection, or delivery optimization.

What it isn't

Three honest caveats.

Species-bridge is not modeled. PHP.eB is the gold standard in mice but doesn't translate to NHP/human. The score is a relative ranking, not an absolute species-transferable prediction.
Route-of-administration matters. Intravenous vs intrathecal vs intracerebral changes which receptor pathways dominate. The current score assumes systemic IV.
Pre-existing immunity overrides chemistry. A patient seropositive for the capsid family won't deliver gene to brain regardless of BBB score. The seroprevalence-adjusted output flags this.

CNS program in design?

We score your candidate set against PHP.eB / AAV9 / Anc80L65 baselines and flag the species-bridge and seroprev risks per candidate.