The canonical failure, in numbers
Hordeaux and colleagues (Molecular Therapy 2018) tested the AAV9 variant PHP.B, which had been selected in vivo for central nervous system delivery, outside the model in which it was selected. In C57BL/6J mice, brain transduction was approximately four logarithmic orders higher than AAV9. In BALB/cJ mice - a different strain of the same species - transduction was no different from AAV9. In rhesus macaques, central nervous system transduction was very low, comparable to AAV9, despite robust transduction of liver and muscle.
The study also reported a serious safety event: one macaque dosed at 7.5 x 10^13 genome copies per kilogram developed thrombocytopenia and haemorrhage and was euthanised on day 5.
The headline is not that one capsid disappointed. It is that a strain change inside a single species was enough to erase the effect, which means the original result was never a property of the capsid alone.
The mechanism, and why it generalises as a warning
Two groups identified the responsible receptor in the same year. Hordeaux and colleagues (Molecular Therapy 2019) showed that the GPI-linked protein LY6A drives AAV-PHP.B transport across the blood-brain barrier, identifying for the first time a ligand for this GPI-anchored protein. Huang and colleagues (PLoS ONE 2019) independently reported LY6A as a cellular receptor for AAV-PHP.B capsids and demonstrated that this mode of binding and transduction can occur independently of other known AAV receptors.
Batista and colleagues (Human Gene Therapy 2020) then closed the strain question, concluding that the expression level of Ly6a in the blood-brain barrier is likely the determining factor for the differential efficacy of AAV-PHP.B between mouse strains.
LY6A is a murine protein, and the enhanced transport did not appear in macaques. The general lesson is mechanistic, not anecdotal: a selection experiment optimises against whatever receptor is available in the host it runs in, and an in vivo screen cannot tell you whether that receptor exists in your target species. Only the receptor biology can.
Even the same capsid changes cell type across species
Species differences are not limited to engineered variants. Gray and colleagues (Molecular Therapy 2011) characterised intravascular AAV9 delivery in adult mice and juvenile non-human primates. In mice, AAV9 transduced roughly twice as many neurons as astrocytes across the central nervous system at doses from 1.25 x 10^12 to 8 x 10^13 vector genomes per kilogram. In non-human primates at a comparable middle dose of 9 to 9.5 x 10^12, the authors observed reduced peripheral-organ and brain transduction relative to mice, together with a clear shift toward mostly glial transduction.
They also found that low levels of pre-existing neutralizing antibodies mostly occluded both central-nervous-system and peripheral transduction, and concluded that high peripheral tropism, limited neuronal transduction in primates and pre-existing antibodies all represent significant barriers to human translation of intravascular AAV9 delivery.
So the quantity that changes across species is not only how much vector reaches the tissue. It is which cells it reaches - which can invalidate a therapeutic hypothesis even when the biodistribution number looks acceptable.
What this means for a ranking model
A model trained on rodent biodistribution learns rodent receptor biology. That is useful for ordering candidates within the same species and route, and it is extrapolation the moment either changes. A defensible capsid ranking therefore carries species metadata as a first-class part of the output, not a footnote.
- Ask which species, strain, route, dose and assay produced every label in the training set
- Require an explicit flag when a candidate's predicted mechanism depends on a receptor that is species-restricted or whose ortholog is uncharacterised in the target species
- Treat a cross-species claim as an extrapolation with its own uncertainty, reported separately from the within-species ranking
- Check whether the receptor is expressed in the target species at the barrier of interest, not merely present in the genome
- Hold route constant when comparing: intravenous, intrathecal and intraparenchymal delivery recruit different pathways
- Treat an unexplained species gap as an unidentified mechanism, not as noise
A sequence that respects the trap
The workable order is: rank in silico to decide where to spend, confirm that the predicted mechanism's receptor is present and expressed in the target species, then design the primate study to falsify the ranking rather than to confirm it. The cohort is expensive precisely because it is the only step that resolves the species question.
Stated negatively, which is the way programmes actually get hurt: no amount of murine evidence, however large the effect, substitutes for a primate study when the mechanism is unknown. A four-log result in one inbred strain is the strongest possible demonstration that a capsid can work, and almost no evidence that it will.
Atlas Bio's AAV platform is built for pre-NHP ranking, translation-trap detection and audit-traceable evidence aggregation, attaches a species-specificity flag to every central-nervous-system tropism prediction, and states on its public pages that it does not predict non-human-primate outcomes and does not substitute for in vivo validation.
Does a strong mouse result mean anything at all?
Yes - it establishes that the capsid can cross the barrier and express payload, which is a real result. What it does not establish is that the mechanism exists in primates. The published case where that gap was investigated found a mouse protein doing the work.
Can a model detect a species-restricted mechanism before the primate study?
It can flag the risk when the predicted receptor is known to be species-restricted or has no characterised ortholog in the target species. It cannot detect an unknown mechanism, which is why an unexplained effect size should raise suspicion rather than confidence.
Is the problem specific to central nervous system capsids?
It is most visible there because the barrier makes receptor dependence stark, but the same logic applies anywhere tropism is receptor-driven. Published work comparing mice and primates for the same serotype found differences in both the amount of transduction and the cell types reached.
- Hordeaux J et al. The Neurotropic Properties of AAV-PHP.B Are Limited to C57BL/6J Mice. Molecular Therapy 2018 (PMC5911151)
- Hordeaux J et al. The GPI-Linked Protein LY6A Drives AAV-PHP.B Transport across the Blood-Brain Barrier. Molecular Therapy 2019 (PMC6520463)
- Huang Q et al. Delivering genes across the blood-brain barrier: LY6A, a novel cellular receptor for AAV-PHP.B capsids. PLoS ONE 2019 (PMC6855452)
- Batista AR et al. Ly6a Differential Expression in Blood-Brain Barrier Is Responsible for Strain Specific Central Nervous System Transduction Profile of AAV-PHP.B. Human Gene Therapy 2020 (PMID 31696742)
- Gray SJ et al. Preclinical differences of intravascular AAV9 delivery to neurons and glia: a comparative study of adult mice and nonhuman primates. Molecular Therapy 2011 (PMC3129805)
Talk to Atlas Bio.
Sponsors, researchers and clinicians: tell us the decision you are trying to make and we will show what the platform can and cannot answer.
Contact Us →