AAV Immunogenicity and Pre-Existing Neutralizing Antibodies

Short answer

What can models predict about AAV immunogenicity and pre-existing neutralizing antibodies? Models can rank capsids by expected cross-reactivity within a serotype family and estimate the share of a population likely to be excluded by screening. They cannot predict an individual patient's titre, the titre at which transduction fails, or the cellular response after dosing. The limiting factor is not modelling capacity; it is that the assays themselves are not standardised.

Two immune problems, not one

Pre-existing humoral immunity and post-dose cellular immunity are different failure modes with different timing. Ertl (Cellular Immunology 2019) frames both: AAV-neutralizing antibodies reduce transduction rates, while CD8+ T cells directed to AAV capsid antigens cause rejection of AAV-transduced cells. The first keeps the vector from arriving. The second removes the cells it reached.

Mingozzi and High (Blood 2013) set out why this is structurally hard: because AAV vectors are administered directly to the patient, the likelihood of a host immune response is high, and pre-existing or recall responses to the wild-type virus from which the vector is engineered, or to the transgene product itself, can interfere with therapeutic efficacy if not identified and managed optimally. They also state the constraint that still holds today - a comprehensive understanding of the determinants of immunogenicity of AAV vectors, and of potential associated toxicities, is still lacking.

A field without a comprehensive account of its determinants is a field where predictive models can rank and flag, but cannot be expected to forecast individual outcomes.

How common is pre-existing immunity? It depends on the cohort

The reference dataset most capsid work still anchors on is Boutin and colleagues (Human Gene Therapy 2010), who measured total IgG and neutralizing factors in healthy donors. Total IgG prevalence was highest for AAV1 (67%) and AAV2 (72%), with AAV5 at 40%, AAV6 at 46%, AAV8 at 38% and AAV9 at 47%. Neutralizing-factor prevalence was highest for AAV2 (59%) and AAV1 (50.5%), and lowest for AAV8 (19%) and AAV5 (3.2%).

Those figures are not universal constants. Qin and colleagues (Human Gene Therapy 2024), screening healthy participants across multiple centres in China alongside patients with Duchenne and Becker muscular dystrophy, reported far higher adult neutralizing-antibody positivity - 97.4% for AAV2 and 86.6% for AAV9 in adult controls - against 66.9% and 32.4% in the patient group, with AAV9 positivity rising with age and tending to plateau after 40.

The decision-relevant reading is not which study is right. It is that seroprevalence varies by serotype, geography, age and assay, so a seropositivity figure quoted without its cohort and its assay is not a number you can plan an eligibility screen around.

The assay is the bottleneck

Weber (Frontiers in Immunology 2021) states the problem directly: establishing protocols to determine therapeutically relevant titres of pre-existing anti-AAV antibodies, and approaches to deplete those antibodies, is more urgent than ever, and a central challenge is measuring those antibodies by methods that are most predictive of their influence on the therapeutic efficacy of AAV gene transfer.

That is the sentence to read twice. Binding assays and cell-based neutralisation assays measure different things, laboratories use different cell lines, multiplicities of infection and cut-points, and a titre of 1:10 in one laboratory is not a titre of 1:10 in another. Until that is harmonised, any model trained on pooled published titres is learning a mixture of assays as much as a mixture of patients.

The animal literature shows how little antibody is needed to matter. Gray and colleagues (Molecular Therapy 2011) reported that in juvenile non-human primates receiving intravascular AAV9, the presence of low levels of pre-existing neutralizing antibodies mostly occluded both central-nervous-system and peripheral transduction.

What a model can usefully do today

Scoped tightly, computational work on AAV immunogenicity earns its place. Scoped loosely, it produces numbers that read like patient-level risk and are not.

Can do: rank candidate capsids by expected cross-reactivity with the serotypes a population is most likely to have seen
Can do: convert published seroprevalence into a seropositivity-adjusted eligible pool, with the cohort and assay named and an interval attached
Can do: flag that a capsid's family carries high neutralising prevalence before a programme commits to it
Can do: surface the eligibility consequence - a seropositive patient will not be dosed effectively regardless of how good the capsid's tropism score is
Cannot do: predict an individual patient's titre, or the threshold at which that patient's titre blocks transduction
Cannot do: predict capsid-specific T-cell responses, their timing, or whether steroid prophylaxis will contain them
Cannot do: transfer a seroprevalence estimate from one country, age band or assay to another without re-anchoring

How to use the number you do have

The practical use of an immunogenicity estimate is portfolio arithmetic, not patient selection. If one capsid family excludes half of an indication's population on screening and another excludes a fifth, that changes recruitment timelines, site counts and trial cost, and it changes them before any patient is screened.

The patient-level question stays where it belongs: in a validated assay, run on that patient, close to the dosing decision. A model that respects that boundary is useful. A model that crosses it is making a clinical claim it cannot support.

How Atlas Bio does this

Atlas Bio places immune context at Level 6 of its published L1-L8 capsid framework - seroprevalence corrected against the published literature, with neutralizing-antibody escape prediction listed as partial rather than finished - and its brain-penetration scoring states outright that pre-existing immunity overrides chemistry: a seropositive patient will not achieve delivery regardless of the score.

Common questions

Can a capsid be engineered to evade pre-existing antibodies?

Evasion is an active research area, and Weber's review surveys both antibody-depletion approaches and capsid-side strategies. What has not been demonstrated is a model that reliably predicts in advance how much evasion a given set of surface substitutions will buy in humans.

Does a low seroprevalence serotype solve the problem?

It improves eligibility, not immunity. Boutin's data put neutralizing-factor prevalence for AAV5 far below AAV2, but patients still mount responses after dosing, and capsid-specific T-cell responses are a separate mechanism that low pre-existing prevalence does not address.

Why does seroprevalence differ so much between published cohorts?

Age, geography, the assay format and the cut-point all move it, and published studies differ on all four. That is why a seroprevalence figure is only interpretable with its cohort and method attached, and why pooling across studies produces a number with no clean referent.

Sources

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.